Autonomous Weapons, Human Sovereignty, International Governance and the Strategic Competition for Artificial Intelligence

Scope: A global strategic assessment of the five artificial-intelligence governance propositions articulated by Pope Leo XIV on 30 October 2026, their interaction with national-security priorities and international law, and their potential implications for military, industrial and institutional developments through 2030.

Executive Summary — Bottom Line Up Front

Principal assessment — High confidence: The five propositions establish a coherent ethical framework but do not themselves create international legal obligations. Their strategic weight lies in their ability to influence future treaty language, domestic regulatory standards, defence-procurement rules and institutional accountability.

The central conflict — High confidence: The requirement that humans retain effective authority over decisions affecting life and liberty is compatible with existing obligations to comply with international humanitarian law. A comprehensive prohibition on lethal autonomous weapons, however, goes beyond currently established international restrictions and conflicts with the military value attributed to autonomous capabilities.

The negotiating reality — High confidence: The United Nations General Assembly’s December 2025 resolution urged progress toward an instrument on lethal autonomous weapons. The Convention on Certain Conventional Weapons process remained directed toward negotiating elements of an instrument without prejudging whether it would be legally binding. mandates.

Institutional asymmetry — High confidence: International scientific assessment and political dialogue have developed faster than international inspection, enforcement and remedies mechanisms. Scientific coordination cannot substitute for compulsory jurisdiction or national implementation.

Corporate exposure — Medium-high confidence: Stronger human-accountability requirements could increase spending on system evaluation, traceability, operational safeguards and legal documentation. Their eventual distribution across developers, deployers, manufacturers and public authorities will depend on jurisdiction-specific rules.

Competitive asymmetry — High confidence: States and firms controlling advanced computing infrastructure, semiconductor supply chains, critical software and proprietary model capabilities have greater practical influence over access conditions than institutions whose authority is principally advisory.

Distributional concern — Medium confidence: Common-good commitments will have limited distributive effect without financing, accessible computing infrastructure, technical capacity-building and mechanisms that allow less-resourced countries to participate in governance rather than merely adopt externally defined standards.

2030 outlook — Medium confidence: Selective institutionalisation of transparency, lifecycle assurance and human-control requirements is a plausible pathway. A comprehensive, globally verifiable prohibition on lethal autonomous weapons would require substantially greater agreement on definitions, military exemptions, inspection rights and enforcement than the current record establishes.

Decision consequence: Governments should distinguish three different objectives: improving accountability under existing law, negotiating additional restrictions on autonomous weapons, and building international institutions capable of monitoring compliance. These objectives have different legal authorities, implementation costs and diplomatic prospects.

Decisive uncertainty: Whether the principal military and technological powers will accept restrictions that reduce unilateral operational flexibility, expose sensitive capabilities to external scrutiny or limit their ability to exploit advances in autonomous targeting.

Artificial Intelligence in Critical Infrastructure SOCs: Force Multiplier – Attack Surface and Operational Resilience Through 2030

Papal AI Ethics and State Power: Who Will Enforce the Limits of Artificial Intelligence?

The decisive question raised by Pope Leo XIV is not whether artificial intelligence should remain under human authority, but whether governments will accept enforceable restrictions when autonomy delivers military advantages, industrial power and geopolitical leverage. His five propositions of 30 October 2026 reject machine personhood, prohibit the delegation of decisions affecting life and liberty, demand equitable technological benefits and call for effective international institutions. Yet the Convention on Certain Conventional Weapons has not produced a binding prohibition on lethal autonomous weapons. The European Union regulates high-risk civilian AI while excluding relevant military and national-security activities. Meanwhile, governments are investing billions in computing infrastructure and competing for control over advanced semiconductors. The contradiction is immediate: ethical agreement is expanding precisely as the economic and security incentives to preserve technological discretion become stronger.

Fifty Governments Want Negotiations. None Has Yet Secured a Global Ban.

The diplomatic record is more substantial than the conventional opposition between ethical ambition and governmental indifference suggests. On 4 September 2026, the Convention on Certain Conventional Weapons Group of Governmental Experts adopted by consensus a report containing elements for a possible international instrument on lethal autonomous weapons. Its provisions address target identification, selection and engagement without human intervention, responsible chains of command, operational restrictions, testing and accountability. On 22 October 2026, a coalition of 50 states, including Italy, France and Germany, welcomed those elements and called for a negotiating mandate. The International Committee of the Red Cross separately reported that more than 70 states supported beginning negotiations. These figures establish political momentum, not agreement on a binding prohibition. The distinction will be tested at the CCW’s Seventh Review Conference in Geneva on 16–20 November 2026. CCW Final Report, CCW/GGE.1/2026/3 — United Nations Office for Disarmament Affairs — September 2026.

The disagreement concerns the precise point at which machine execution becomes an unacceptable delegation of human judgment. The United States Department of Defense’s Directive 3000.09, reissued on 25 January 2023, regulates autonomous and semi-autonomous weapons through requirements concerning appropriate human judgment, testing and oversight rather than prohibiting every autonomous function. The United Kingdom accepts context-dependent human involvement, including parameters established before operation. France rejects systems falling within its nationally defined category of fully autonomous lethal weapons. Germany distinguishes constrained rapid defensive responses from autonomous operations in more complex environments. All can endorse human responsibility while maintaining different operational doctrines. A treaty that leaves these distinctions unresolved risks either restricting capabilities governments regard as indispensable or defining prohibited autonomy so narrowly that existing programmes remain largely unaffected.

International humanitarian law already requires distinction, proportionality and feasible precautions in attack. Article 36 of Additional Protocol I also requires its High Contracting Parties to review new weapons for compatibility with applicable international law. But an obligation to conduct a legal review is not equivalent to a prohibition on autonomous engagement. Nor does successful technical testing demonstrate that a system can lawfully operate in every environment. The central treaty question is whether certain autonomous functions must be prohibited independently of national assessments of military necessity. Additional Protocol I, Article 36 — International Committee of the Red Cross — June 1977.

The Military Advantage Lies Precisely Where Human Control Becomes Hardest

Autonomous engagement becomes militarily attractive when communication is disrupted, threats arrive rapidly or several platforms must operate simultaneously. A human commander may define an operating area, identify authorised target categories and approve activation, while the system subsequently selects and engages objects without a separate command for every action. That arrangement is fundamentally different from allowing software to determine independently whether an individual should be killed. Yet the legal distinction becomes difficult when target characteristics, operating conditions and the consequences of engagement change after activation.

The September 2026 CCW consensus report recognises direct and indirect human control, while addressing operational limits on geography, duration, target categories and scale. This reflects the operational problem rather than resolving it. A human authorisation recorded before activation may be legally significant, but it cannot automatically validate every subsequent engagement. Conversely, requiring a separate human command for every defensive response could impose limitations that some governments consider incompatible with the speed of the threat. The central requirement is therefore not the ceremonial presence of an operator but demonstrable authority over the weapon’s permitted behaviour.

Verification compounds the disagreement. Autonomous functionality may depend on software configuration, sensor inputs and operating parameters rather than on easily identifiable physical characteristics. A system inspected in one configuration may subsequently be updated. A successful test in a controlled environment does not prove compliance under different conditions. Inspectors might require access to source code, operational records or classified targeting information that governments and defence contractors are reluctant to disclose. Without credible arrangements for examining those functions, even a formally binding prohibition could leave compliance dependent heavily on national declarations.

The Real Enforcement Deficit Is Institutional, Not Philosophical

On 26 August 2025, the United Nations General Assembly established the Independent International Scientific Panel on Artificial Intelligence, comprising 40 members, and the Global Dialogue on AI Governance. The Dialogue’s inaugural meeting in Geneva on 6–7 July 2026 brought delegations from 163 countries. Its next meeting is scheduled for 3–4 May 2027 in New York. These institutions broaden scientific assessment and political participation, particularly for governments lacking extensive domestic expertise. They do not possess general authority to inspect military systems, compel proprietary disclosures or impose sanctions. Participation by 163 countries therefore measures diplomatic reach, not regulatory jurisdiction. UN General Assembly Resolution A/RES/79/325 — United Nations — August 2025.

The European Union offers a different model because Regulation (EU) 2024/1689, the Artificial Intelligence Act, imposes enforceable obligations within its legal scope. Yet the AI Omnibus, effective 27 July 2026, revised the timetable: relevant Annex III high-risk requirements now apply from 2 December 2027, with corresponding Annex I product requirements following on 2 August 2028. The amendments reflect the collision between regulatory ambition and implementation capacity. Documentation, testing and human oversight require technical expertise and investment; smaller suppliers cannot necessarily absorb those costs as easily as established technology groups. The regulation also excludes relevant military, defence and national-security activities, leaving the most consequential autonomous-weapons questions to separate legal and diplomatic processes. AI Omnibus Enters into Force — European Commission — July 2026.

The Council of Europe’s Framework Convention on Artificial Intelligence and Human Rights, Democracy and the Rule of Law, opened for signature on 5 September 2024, supplies a further treaty pathway. The European Union deposited its instrument of approval on 15 May 2026. But treaty commitments, domestic enforcement and international inspection remain distinct. The Convention’s entry-into-force conditions and national implementation requirements cannot be replaced by political endorsement. The institutional deficit is therefore specific: the international system can formulate principles and negotiate obligations more readily than it can establish compulsory access to the evidence needed to investigate violations.

Computing Power Has Become an Instrument of Foreign Policy

The economic contest surrounding AI is governed by infrastructure that ethical declarations cannot redistribute. The International Energy Agency estimated global data-centre electricity consumption at approximately 485 terawatt-hours in 2025 and projected approximately 950 TWh annually by 2030 under its updated central scenario. The difference, approximately 465 TWh, is additional projected annual electricity consumption, not generating capacity already commissioned. These figures cover data centres generally, not exclusively AI. They nevertheless expose the physical constraint on technological expansion: computing ambitions require electricity, grid connections, specialised equipment and finance. Key Questions on Energy and AI — International Energy Agency — 2026.

The industrial consequences are equally direct. In January 2026, the US Bureau of Industry and Security revised licensing treatment for specified advanced semiconductors exported to China, including the NVIDIA H200 and AMD MI325X, allowing case-by-case consideration subject to security conditions. Control over advanced chips therefore translates into influence over foreign computing programmes, investment decisions and supply relationships. Access to AI services does not confer control over the infrastructure or intellectual property required to develop them independently. A government can endorse equitable technological benefits while restricting strategically sensitive inputs; the resulting distribution of capability is determined by licensing, capital and productive capacity, not by the universality of the ethical language.

The divide begins below frontier computing. The International Telecommunication Union estimated that 2.2 billion people remained offline in 2025. Internet use reached 94% in high-income countries, against 23% in low-income countries. These figures measure connectivity, not advanced AI capability, but they demonstrate why access cannot be reduced to the commercial availability of models. Countries lacking reliable networks, technical expertise and computing resources cannot exercise the same practical control over AI development as those possessing the infrastructure. Common-good commitments acquire economic meaning only when financing and capacity-building produce operational capabilities rather than additional declarations. Measuring Digital Development: Facts and Figures 2025 — International Telecommunication Union — November 2025.

Europe Can Regulate AI More Quickly Than It Can Build Independence

Europe’s national programmes demonstrate the distance between regulatory authority and productive capacity. Italy established differentiated responsibilities for the National Cybersecurity Agency, ACN, and the Agency for Digital Italy, AgID, through its 2025 AI legislation, alongside a €1 billion programme supporting startups and SMEs in AI and related technologies. The IT4LIA AI Factory secured an infrastructure procurement contract in April 2026, strengthening the computing ecosystem associated with CINECA and Bologna. The programme creates capacity for industrial research and development, but a procurement contract is not proof of completed deployment or measurable productivity growth. IT4LIA AI Factory Procurement — EuroHPC Joint Undertaking — April 2026.

France announced €109 billion in prospective AI-related investment in February 2025, with particular emphasis on computing infrastructure and the attraction of data-centre projects. Germany’s Hightech Agenda Deutschland pursues a different industrial transmission mechanism: by 2030, it targets AI-based activity representing 10% of German economic output, a fourfold increase in AI startups relative to 2025 and AI adoption by more than 50% of manufacturing SMEs by 2029. The French figure concerns prospective investment, while Germany’s figures are policy targets rather than achieved output or independently estimated GDP gains. Their difference matters because large computing installations and industrial deployment generate economic value through distinct channels.

Outside the EU regulatory framework, the United Kingdom’s 2025 Compute Roadmap envisages up to £2 billion for public computing and expansion of its AI Research Resource from 21 AI ExaFLOPS in 2025 to 420 by 2030, using the government’s stated performance measure. At EU level, the Commission reports 19 AI Factories and 13 associated antennas being established, while the Chips Act is associated with expectations of more than €43 billion in public investment and over €100 billion in broader policy-driven public and private semiconductor investment through 2030. These envelopes must not be added together as completed expenditure. They reveal a common constraint: European governments are attempting to enlarge their technological capacity while remaining exposed to international semiconductor supply chains, infrastructure-delivery schedules and competition for capital. UK Compute Roadmap — UK Department for Science, Innovation and Technology — July 2025; European Chips Act — European Commission.

The Next Twenty-Four Months Will Separate Governance from Declaration

Between November 2026 and November 2028, the institutional test will become measurable. The CCW Review Conference must determine whether the September consensus elements become the basis for substantive negotiations. The UN Global Dialogue will reconvene in May 2027, while the EU’s revised high-risk AI requirements take effect across relevant categories in December 2027 and August 2028. None of these milestones automatically establishes effective compliance. The decisive evidence will be binding language, national implementing measures, technical assurance procedures, actual regulatory enforcement and credible access to information when violations are alleged.

The financial consequences will fall unevenly. Defence ministries that cannot establish reliable human-control and testing procedures will face procurement uncertainty, operational risk and potentially expensive modifications. AI suppliers will bear documentation, evaluation and contractual compliance costs, with smaller firms less able to spread fixed expenditure across large product portfolios. European governments that fail to convert computing commitments into functioning facilities will continue paying for technological dependence through external infrastructure, software and specialised components. Countries lacking independent assessment capacity will remain dependent on suppliers not only for technology but also for the evidence needed to evaluate it.

The choice now confronting governments is therefore concrete. The five propositions of 30 October 2026 can remain a standard against which conduct is publicly judged, become the foundation for national assurance obligations, or contribute to binding restrictions on particular autonomous-weapons functions. Each outcome carries different costs and grants different degrees of discretion. Inaction will not preserve the existing balance: military and corporate programmes will continue under their applicable legal regimes, while compliance burdens, infrastructure investment and technological dependencies evolve. The governments and firms that finance the infrastructure will shape the available capabilities; those that neglect implementation will inherit the costs of dependence, weak oversight and legal uncertainty.

AI Infrastructure Capital Cycle: Credit Discipline – Cash-Flow Constraints and the Economics of Artificial Intelligence Through 2030

---

Navigational Index

PILLAR I — Ethical Authority, Legal Responsibility and Military Power

• Chapter 1 — The Five Propositions: Normative Authority and Operational Translation
• Chapter 2 — Strategic Competition, Autonomous Weapons and International Humanitarian Law
• Chapter 3 — Human Responsibility, Machine Personhood and the Allocation of Liability

PILLAR II — Institutional Capacity, Industrial Competition and Distributional Power

• Chapter 4 — International AI Governance: Institutional Mandates, Oversight and Enforcement
• Chapter 5 — Computing Infrastructure, Export Controls, Corporate Incentives and Equitable Access
• Chapter 6 — European and Transatlantic Implications: Italy, France, Germany, the United Kingdom and the European Union

PILLAR III — Strategic Trajectories, Verification and Decisions Through 2030

• Chapter 7 — Three Governance Scenarios: Normative Influence, Partial Institutionalisation and Restrictive Arms Control
• Chapter 8 — Verification Problems, Strategic Indicators, Policy Options and Final Net Assessment

---

The New Architecture of Autonomous Warfare: Military Deception – Artificial Intelligence and the Global Struggle for Battlefield Superiority

Master Abstract

The fundamental distinction: moral authority does not automatically become legal authority

The intervention of Pope Leo XIV on 30 October 2026 introduces a structured conception of technological governance based on five principles:

• the reversibility of technological choices,
• the instrumental rather than personal status of machines,
• the preservation of human authority over consequential decisions,
• the equitable distribution of technological benefits ,
• the creation of institutions capable of enforcing shared norms.

These propositions are neither legally equivalent nor equally difficult to operationalise. The assertion that technological development is not predetermined supports precautionary governance, authorisation requirements and the possibility of refusing particular applications. The rejection of machine personhood preserves the distinction between AI systems and human or corporate legal subjects. The prohibition on delegating decisions affecting life, freedom and lethal force introduces a more demanding operational constraint. The principles of common benefit and effective institutional oversight require affirmative political and economic measures whose scope cannot be established by ethical declaration alone.

The primary Vatican text expressly advocates concrete progress toward banning lethal autonomous weapons at the November 2026 Geneva discussions. It also distinguishes system performance from the ability to foresee and control harmful consequences. These are relevant distinctions because military effectiveness, legal compliance and political legitimacy are not interchangeable measures of technological acceptability. Address to the Participants in “The Future of AI” — Holy See — Oct 2026

The accompanying encyclical Magnifica Humanitas, issued on 15 May 2026, develops a broader argument about the human person, technological concentration, governance and the moral limits of autonomous decision-making. Its propositions constitute the underlying doctrinal framework, not a source of binding international law for sovereign states. Encyclical Letter Magnifica Humanitas — Holy See — May 2026

The military contradiction: human control versus strategic advantage

The central strategic difficulty arises from the relationship between meaningful human authority and the military incentives to automate identification, selection and engagement.

Autonomy can offer operational advantages when communications are degraded, reaction times are compressed or large numbers of targets must be processed rapidly. Those advantages are not sufficient to establish the lawfulness of a particular system or attack. They nevertheless create powerful incentives to retain operational freedom.

The CCW negotiations demonstrate the importance of definition. The Chair’s April 2026 summary records discussion of a functional characterisation involving identification, selection and engagement without human intervention in those tasks. Importantly, pre-setting target parameters would not necessarily remove a system from the proposed category. This distinction matters because a broad definition could cover existing and emerging military capabilities that states regard as essential.

---

Key Components Explained

• The CCW negotiations: This refers to the discussions surrounding the United Nations Convention on Certain Conventional Weapons (CCW). This treaty aims to ban or restrict weapons that cause unnecessary suffering or affect civilians indiscriminately (such as landmines, booby-traps, and laser-blinding weapons). Recent CCW debates have focused heavily on Lethal Autonomous Weapons Systems (LAWS), often called “killer robots.”
• Demonstrate the importance of: This means these specific meetings serve as a clear, real-world example of why something matters.
• Definition: This refers to how a specific term or concept is legally described and understood by all participating countries.

Why “Definition” Matters in this Context

In the CCW negotiations, progress often stalls because nations cannot agree on what words actually mean. For example:

• What is “autonomous”? Some countries argue a weapon is autonomous if it selects a target on its own. Others argue it is only autonomous if it operates entirely without human oversight.
• What is “meaningful human control”? Countries must define how much human involvement is required before a weapon is no longer considered an “unregulated robot.”

In short: If nations cannot agree on the definition of a weapon, they cannot write a law to regulate it. Therefore, the phrase highlights that precise language is the foundation of international law.

---

The same record establishes that the mandate concerned elements of a possible instrument, without predetermining its legal character. The negotiation therefore concerns both substantive limits and the institutional form through which those limits might become enforceable. Chair’s Summary, CCW/GGE.1/2026/WP.2 — United Nations Office for Disarmament Affairs — Apr 2026

The legal and ethical positions must remain distinct. Existing international humanitarian law applies to autonomous weapons, but the International Committee of the Red Cross does not interpret that fact as eliminating the need for additional prohibitions and restrictions. Its proposed framework distinguishes weapons whose effects cannot be adequately predicted, systems designed to target humans directly, and other autonomous weapons that would remain subject to limitations.

This is narrower than an unconditional prohibition of all autonomous weapons. Whether the wider ethical prohibition can be translated into an acceptable treaty will depend on the exact scope of prohibited conduct and the exceptions, if any, that participating states are prepared to recognise. Frequently Asked Questions: Autonomous Weapon Systems — International Committee of the Red Cross — Oct 2026

The institutional contradiction: coordination exists, enforcement remains dispersed

The governance problem is not the complete absence of international institutions. It is the mismatch between their current mandates and the authority required for binding oversight.

The United Nations has established a scientific assessment mechanism and an inclusive intergovernmental dialogue on AI governance. These arrangements can improve evidence-sharing, participation and the identification of common risks. They do not, by themselves, confer authority to inspect national military systems, compel disclosure of proprietary models, invalidate domestic authorisations or award compensation to persons harmed by AI.

At the regional level, binding regulatory approaches introduce enforceable obligations within their legal scope. However, military, defence and national-security activities remain subject to distinct legal regimes. Consequently, regulatory progress in commercial AI cannot automatically be interpreted as progress toward international controls on autonomous weapons.

The emerging architecture is therefore characterised by increasing normative coordination but uneven coercive authority. Its practical effectiveness will depend on how international standards are incorporated into national law, procurement contracts, technical certification, judicial remedies and military command systems.

The distributive contradiction: common benefits within unequal industrial structures

The demand for equitable access to AI benefits encounters an industrial structure characterised by uneven access to advanced chips, large-scale computing facilities, skilled personnel, investment capital and proprietary technological ecosystems.

The relevant issue is not whether unrestricted technology transfer is universally desirable. States may have legitimate security reasons to limit access to sensitive capabilities. Rather, the institutional question is how to differentiate justified restrictions from arrangements that entrench dependency without materially reducing risk.

Common-good principles could acquire operational content through public-interest computing infrastructure, accessible scientific research capacity, concessional financing, interoperability obligations, transparent access criteria and technical assistance to developing economies. Their impact would depend on financing, enforceable allocation rules and implementation capacity.

A framework that imposes sophisticated regulatory obligations on countries lacking the infrastructure to comply could unintentionally reinforce the concentration it seeks to correct. Equitable governance therefore requires both safety standards and the means to meet them.

Strategic assessment

Through 2030, the five propositions can influence governance by changing the legitimacy of particular practices, informing domestic legislation, shaping multilateral negotiations and establishing benchmarks against which governments and corporations can be publicly scrutinised.

Their effectiveness will remain conditional on three forms of translation: from ethical principle into precise legal obligation, from legal obligation into technically verifiable compliance, and from nominal compliance into meaningful responsibility when harm occurs.

The governing distinction is between agreement on desirable human oversight and agreement to surrender operational discretion. The former has considerable support in international instruments and diplomatic statements. The latter involves security, commercial and sovereignty costs that have not been resolved through normative convergence.

Key Evidence and Regulatory Baseline

The following records establish the principal documentary baseline. They distinguish moral propositions, diplomatic commitments, negotiating texts and enforceable legal regimes.

Indicator	Value / status	Reference date	Definition / scope	Issuer and exact source
Five ethical propositions	Published in full	30 Oct 2026	AI development, personhood, non-delegation, common benefit, international oversight	Address on the Future of AI — Holy See — Oct 2026
CCW negotiations	Instrument elements under consideration; legal nature not predetermined	Mar–Apr 2026	Autonomous weapons, human judgment, operational restrictions and accountability	CCW/GGE.1/2026/WP.2 — UNODA — Apr 2026
UN resolution A/RES/80/57	4 operative paragraphs	2025	Multilateral cooperation and progress toward an instrument	Lethal Autonomous Weapons Systems — UN General Assembly — 2025
UN scientific governance	International scientific panel established	26 Aug 2025	Independent scientific assessments; no general inspection or enforcement mandate	Independent International Scientific Panel on AI — United Nations
UN policy coordination	Global Dialogue established; inaugural meeting held 6–7 Jul 2026	2025–2026	International dialogue and policy exchange	Global Dialogue on AI Governance — United Nations
European AI regulation	Regulation (EU) 2024/1689; consolidated version dated 27 Jul 2026	2024–2026	Risk-based AI obligations; military, defence and national-security exclusions within the applicable scope	Artificial Intelligence Act — European Union — Jul 2026
ICRC autonomous-weapons position	Prohibition of specified categories; restrictions on other autonomous weapons	Oct 2026	Unpredictable systems and systems targeting persons directly	Autonomous Weapon Systems FAQ — ICRC — Oct 2026

Evidence-status qualification: The Vatican address, CCW Chair’s summary and UN mandate registry were accessible for direct examination. Some UN and EU pages restricted full-text retrieval; their status entries are correspondingly limited to the official information accessible during verification. The CCW Chair’s summary is an account issued under the Chair’s authority, not an agreed treaty text.

The Five Propositions: Preliminary Operational Matrix

Proposition	Operational requirement	Compatibility with current practice	Institutional prerequisite	Principal obstacle
1. Technology remains a human choice	Prior evaluation, purpose limitation, withdrawal authority and lifecycle review	Partial; existing risk controls coexist with strong deployment incentives	Capable regulators, independent testing and effective withdrawal powers	Speed of innovation and strategic competition
2. Machines are tools, not persons	Assign legal obligations to natural and legal persons; maintain identifiable responsibility	Broadly compatible with existing legal structures	Clear rules on developer, deployer, operator and command responsibility	Multiple actors, opaque systems and cross-border harm
3. Life and liberty must not be delegated	Effective human oversight, operational limits and restrictions on autonomous lethal engagement	Contested, particularly regarding degrees of autonomy and the timing of human intervention	Agreed definitions, weapons reviews, compliance measures and accountability	Military necessity claims, technical verification and divergent concepts of control
4. AI must serve the common good	Equitable access, public-interest capacity, accessible benefits and distributional assessment	Uneven; investment and advanced computing are concentrated	Funding, infrastructure access, technical capacity-building and transparent allocation rules	Export restrictions, proprietary rights and industrial concentration
5. Effective international institutions are necessary	Shared standards, independent scrutiny, accountability and meaningful remedies	Partial; coordination mechanisms exist but enforcement is fragmented	Treaty authority, domestic implementation, technical capacity and legal remedies	Sovereignty, jurisdictional limits and resistance to external inspection

Analytical qualification: The first proposition also has a restrictive dimension: the duty to reject applications when their consequences cannot be adequately controlled. The second establishes a categorical ethical position but does not itself enact a legal prohibition on machine personhood. The third imposes the strongest operational constraint, although effective human oversight is not necessarily identical to real-time human approval of every individual engagement.

The CCW discussions illustrate that distinction. Delegations have examined human judgment across the weapon lifecycle, including development, deployment and operation. The Chair’s summary records disagreement over whether the proposed concept of context-appropriate human judgment and control reflects existing law or would introduce an additional standard.

Chair’s Summary, paragraphs 35–57 — UNODA — Apr 2026

Competing Strategic Pathways Through 2030

These scenarios are alternative governance configurations, not mutually exclusive predictions about every jurisdiction. Their relative standing is qualitative. No probabilities are assigned because the public record does not provide a defensible forecasting base rate.

Scenario	Supporting evidence and mechanism	Evidence against or principal constraint	Indicators	Current analytical standing
A — Normative influence without binding constraint	Ethical declarations, political commitments and continued military AI development	Growth of national regulatory requirements and pressure for more specific arms-control restrictions	Nonbinding declarations, divergent procurement rules, continued autonomy investment	Plausible; low institutional burden
B — Partial institutionalisation	Existing UN mechanisms, national regulation and CCW attention to human-control requirements	Disagreement over technical definitions, confidentiality and verification	Agreed human-control terminology, testing requirements, national weapons reviews, reporting standards	Plausible; significant compatibility with existing institutional structures
C — Restrictive autonomous-weapons regime	Humanitarian proposals, multilateral diplomatic engagement and potential treaty pathways	Military competition, treaty-scope disputes, verification challenges and nonparticipation	Negotiating mandate, binding prohibitions, inspection provisions, ratifications and credible national enforcement	Conditional; requires substantial additional political agreement

Scenario implications

Scenario A — Normative influence only

Military deployment trajectories remain driven predominantly by national defence requirements and procurement programmes. Corporate compliance expenditure is determined mainly by existing jurisdictional obligations. Unequal access to advanced AI capabilities persists unless separate industrial and development policies address it. States retain considerable freedom to pursue national military applications, subject to applicable international law.

Scenario B — Partial institutionalisation

Military AI programmes increasingly incorporate testing, lifecycle governance, documentation, human-control requirements and traceable authorisation. Compliance costs rise, particularly for suppliers lacking established assurance processes. Common technical requirements may improve interoperability but could also disadvantage smaller organisations and lower-capacity states. Unilateral action remains possible within the bounds of domestic law and agreed international obligations.

Scenario C — Restrictive arms-control regime

New restrictions would affect the development and procurement of covered autonomous capabilities, with implementation timing dependent on treaty scope and transitional provisions. Military organisations would need to distinguish prohibited functions from permitted automation. Verification costs would increase, and states could face obligations to disclose sensitive information. Developing countries could benefit from reduced proliferation risks but might remain excluded from sophisticated defensive technologies and assurance infrastructure.

The decisive transition from Scenario A to B is the adoption of operationally specific requirements. The transition from B to C requires agreement on prohibited capabilities, legally binding obligations, inspection or verification arrangements, national implementing legislation and responses to noncompliance.

Principal Evidence Gaps and Watch Indicators

Strategic uncertainty	Required evidence or observable indicator	Significance for the 2030 assessment
Will major military powers accept binding restrictions?	Formal state positions, agreed negotiating mandates, treaty signatures and ratifications	Determines whether prohibition can move beyond declarations
Can autonomous engagement be verified?	Technical inspection protocols, validated testing methods and disclosure obligations	Determines whether compliance can be demonstrated independently
What constitutes sufficient human control?	Agreed definitions, national weapons-review procedures and operational doctrine	Determines whether controls have substantive or merely procedural effects
Will AI developers retain identifiable accountability?	Statutory liability rules, court decisions, audit obligations and public procurement contracts	Determines access to effective remedies
Can international institutions exercise independent authority?	Legal mandates, budgets, inspection powers, compulsory reporting and enforcement mechanisms	Distinguishes coordination from supervision
Will developing countries receive material benefits?	Public-compute allocations, financing commitments, equipment access and implementation outcomes	Tests whether equity commitments change the distribution of technological capacity
Will regulatory fragmentation intensify?	Incompatible national requirements, divergent military standards and cross-border recognition arrangements	Measures market and alliance-coordination costs

Decision-relevant monitoring priorities

The most consequential diplomatic observation will be the outcome of the 16–20 November 2026 Seventh Review Conference of the Convention on Certain Conventional Weapons. The ICRC identifies that meeting as an opportunity to move from discussion toward negotiation of new international rules. Its outcome must be assessed through adopted decisions and negotiating mandates rather than anticipated agreement. International Committee of the Red Cross

The second priority is the emergence of technically testable human-control requirements. General endorsement of accountability is insufficient unless national authorities can identify which person authorised deployment, which operational constraints were imposed, how the system behaved and what mechanisms existed to suspend or terminate its operation.

The third priority concerns international institutional capacity. The practical indicators are not the number of declarations or meetings alone, but whether institutions acquire competent oversight personnel, dependable financing, independent access to relevant evidence and recognised procedures for addressing violations.

The fourth priority is distributive. Commitments to accessibility should be evaluated against documented access to computing infrastructure, technical knowledge and economic opportunities, with particular attention to whether safeguards are financially and technically achievable in lower-capacity jurisdictions.

Opening Net Assessment

The available record supports a distinction between the political influence of ethical standards and their direct enforceability.

High confidence: The five propositions provide a clear basis for evaluating the legitimacy and accountability of AI development. They do not independently bind governments, companies or military organisations.

High confidence: Existing international law already imposes obligations on states and individuals in the conduct of hostilities. The unresolved question is which additional restrictions, if any, should specifically govern autonomous identification, target selection and engagement.

Medium-high confidence: The most operationally significant developments through 2030 will concern the quality of human oversight, the traceability of authority and the ability to verify system behaviour within its intended environment.

Medium confidence: International standard-setting and national implementation can advance without comprehensive agreement on autonomous-weapons prohibition. Such progress would reduce some governance gaps but would not eliminate military competition or resolve the distribution of technological power.

Low confidence: The public record permits no reliable determination that major military powers will accept a comprehensive prohibition supported by intrusive international verification before 2030.

The strategic question is therefore not simply whether governments endorse human dignity or responsible AI. It is whether those commitments are translated into restrictions that remain effective precisely when operational advantage, commercial revenue or national-security discretion would favour a different course of action.

The significance of the papal intervention lies in making that institutional contradiction explicit. Its practical consequences will be determined by treaty negotiations, national implementing measures, military doctrine, corporate liability arrangements and the distribution of technological capacity—not by moral authority alone.

PILLAR I — Ethical Authority, Legal Responsibility and Military Power

Chapter 1 — The Five Propositions: Normative Authority and Operational Translation

Principal judgment — High confidence: The five propositions articulated by Pope Leo XIV on 30 October 2026 constitute a coherent normative framework for governing artificial intelligence, but their strategic importance depends on whether they can be translated into legally defined obligations, technically enforceable restrictions and identifiable institutional responsibilities. The distinction is fundamental: an ethical rule can establish that a practice is unacceptable without determining which authority possesses jurisdiction to prohibit it, which technical characteristics make a system noncompliant, which persons bear responsibility for its operation or which remedies become available when violations occur. The most consequential proposition concerns decisions affecting life, liberty and lethal force because it challenges the degree of discretion that public authorities may delegate to automated systems. The rejection of machine personhood reinforces the allocation of responsibility to human and corporate actors. The remaining propositions address the authority to reject technological applications, the distribution of their benefits and the institutional arrangements required to convert shared principles into operational governance.

Normative authority and the boundaries of institutional competence

The address delivered at the Vatican on 30 October 2026, during the meeting organised jointly by the Pontifical Academy for Life and the Future of Life Institute, identifies five propositions that operate at different levels of political and legal authority. Their collective significance does not derive from the creation of an immediately applicable international regime. It derives from the articulation of a hierarchy in which technological capability remains subordinate to human agency, institutional responsibility and the protection of fundamental interests. The address explicitly rejects technological determinism, denies that machines should be treated as persons, requires effective human oversight of consequential decisions, demands that innovation serve the common good and identifies effective international institutions as an indispensable condition for translating ethical principles into practice. The discourse also connects these propositions to the broader doctrinal framework of the encyclical Magnifica Humanitas, particularly its treatment of human responsibility, technological concentration and the legitimacy of delegating consequential authority to artificial systems.

The legal consequences, however, cannot be derived from moral terminology alone. The Pope exercises moral and religious authority, and the Holy See possesses a distinct position in international diplomacy, but the address is not a treaty, a Security Council decision, a domestic statute or an instrument conferring compulsory jurisdiction over governments and corporations. Its propositions may influence national legislation, diplomatic negotiations, procurement rules, technical standards and institutional expectations. They do not automatically amend international humanitarian law, impose obligations on states that have not accepted additional commitments or create private rights of action against commercial AI developers. Consequently, the analytical question is whether these principles can be attached to existing legal duties or incorporated into new ones without losing the substantive limits that give them meaning.

The speech is particularly important because it does not confine its analysis to the reliability of artificial systems. It differentiates the ability of a system to perform a specified task from the legitimacy of allowing that system to exercise consequential decision-making authority. A highly reliable classification system can still be inappropriate for determining whether an individual should be deprived of liberty or subjected to lethal force. Conversely, the mere presence of a human supervisor does not establish meaningful control if that person lacks adequate information, sufficient time, practical intervention authority or an effective capacity to reject the system’s output. This introduces an institutional test that cannot be satisfied through engineering performance metrics alone: governance must establish who is entitled to decide, under which legal conditions, with what information, and through which mechanisms that decision may subsequently be scrutinised.

The central documentary basis is the Address to the Participants in “The Future of AI” — Holy See — October 2026, delivered on 30 October 2026. Its five considerations are sufficiently explicit to permit a differentiated operational assessment, provided their ethical implications are not presented as existing universal legal obligations.

The first proposition: technological development remains reversible and politically governable

The assertion that the future of artificial intelligence is not predetermined carries a stronger implication than the general observation that governments can regulate technology. It rejects the proposition that technical possibility or international competition is sufficient justification for deployment. If technological development remains subject to deliberate human choice, responsible authorities must retain the ability to discontinue, modify or prohibit applications even after significant financial investment, operational integration or commercial adoption. This makes reversibility a substantive governance requirement rather than a discretionary preference. A technology programme that can be authorised but cannot realistically be suspended, withdrawn or replaced may satisfy an initial approval process while failing to preserve meaningful political control over its future consequences.

Operational translation requires distinguishing decisions about research, acquisition, market access, deployment and continued use. Research approval does not necessarily justify commercial release; commercial release does not establish suitability for critical infrastructure; military procurement does not establish that a weapon is lawful in every operational environment; and demonstrated performance in testing does not remove the need to reassess a system when its data, software, operating conditions or capabilities materially change. An effective authorisation framework must therefore specify the intended purpose, identify foreseeable harmful uses, establish the conditions under which permission becomes invalid and allocate responsibility for monitoring the system after deployment. Such controls are especially important for models that undergo updates, depend on external components or operate in environments that differ substantially from their validation conditions.

This proposition also has an industrial dimension. Advanced AI development involves substantial investments in computing infrastructure, software development, specialised personnel and long-term customer relationships. Those investments can create organisational pressure to proceed even when new evidence calls the original authorisation into question. The relevant economic mechanism is not necessarily deliberate misconduct. It is the accumulation of sunk costs, contractual dependencies, reputational exposure and institutional commitments that make cancellation increasingly expensive. A governance framework designed around reversibility must therefore make the power to stop deployment credible before the moment of crisis, rather than relying on voluntary withdrawal after harm becomes evident.

For national-security authorities, the requirement introduces a direct question about procurement architecture. Systems designed to operate within military command networks, intelligence-processing infrastructure or integrated defence platforms may become difficult to remove without disrupting other capabilities. A credible withdrawal mechanism would require configuration control, alternative operating procedures, documented technical dependencies, defined suspension authority and prior assessment of the consequences of disabling the system. The requirement does not imply that all high-risk systems must be prohibited. It implies that continued deployment must remain a reviewable decision, supported by evidence proportionate to the consequences of failure.

The most relevant existing legal analogy is the obligation in Article 36 of Additional Protocol I to determine, during the study, development, acquisition or adoption of a new weapon, whether its employment would be prohibited in some or all circumstances by applicable international law. That obligation is directed to High Contracting Parties and does not independently establish a universal ban on autonomous weapons. Nevertheless, it demonstrates that legal review can be attached to the development process before battlefield deployment, rather than treating technical completion as sufficient authorisation. Additional Protocol I, Article 36: New Weapons — International Committee of the Red Cross, Treaty Database — June 1977.

The second proposition: machines remain instruments, not legal substitutes for human responsibility

The rejection of machine personhood performs two separate functions. At the philosophical level, it refuses to equate computational capability with the moral standing of human persons. At the institutional level, it rejects the possibility that attributing autonomous status to a system could relieve developers, deployers, commanders or public authorities of accountability. These functions should not be confused. The existence of legal personality in domestic law does not always depend on biological personhood: companies and other legal entities may possess rights, obligations and procedural capacities despite not being natural persons. The more precise institutional question is whether assigning a form of legal status to AI would improve the allocation of obligations or instead create a new intermediary through which existing actors could evade responsibility.

For governance purposes, the proposition favours a structure in which every consequential system remains connected to identifiable natural persons and legally constituted organisations. Developers control design choices, data selection and testing procedures. Providers establish contractual conditions and disclose known limitations. Deployers determine operational environments and authorise particular uses. Operators monitor performance and take decisions within their assigned competence. Public authorities establish legal requirements, exercise enforcement powers and determine whether the system may be used in sensitive functions. The distribution of responsibilities need not be identical across jurisdictions, but legal obligations must attach to actors capable of complying with them and responding to claims for harm.

A practical complication arises when no single organisation controls the full technology stack. A high-impact system may incorporate third-party models, licensed datasets, externally supplied software components, cloud services and a separately developed user interface. It may subsequently be modified by an integrator and deployed by an organisation that lacks direct access to model development records. An accountability regime that identifies only the final operator may overlook defects originating in design or integration. Conversely, assigning every adverse outcome to the original developer may ignore operational misuse, inadequate training or the deliberate removal of safeguards. The proposition therefore supports human and organisational accountability but does not eliminate the legal need to establish causation, the applicable duty, breach and the allocation of responsibility among several participants.

The speech’s concern with personhood should not be represented as an enacted international prohibition on granting any conceivable legal status to AI systems. Its stronger and more defensible operational content is the requirement that an automated system never become a jurisdictional or procedural barrier between an injured person and the actors legally responsible for the relevant conduct. This distinction is especially important for corporate structures that might seek to exploit uncertainty about autonomous behaviour. If a company can describe an outcome as an independent machine decision while retaining control over development, deployment or commercial benefit, the allocation of responsibility risks becoming disconnected from the distribution of effective power.

These issues are addressed more fully in Chapter 3, where the analysis separates the normative rejection of machine personhood from civil liability, state responsibility, individual criminal responsibility and military command obligations. The underlying ethical position is established by the Address to the Participants in “The Future of AI” — Holy See — October 2026.

The third proposition: non-delegable decisions require substantive rather than ceremonial human authority

The third proposition introduces the most demanding constraint because it links the legitimacy of automated decisions to the preservation of effective, informed and responsible human oversight. It encompasses decisions affecting life, liberty and the use of force, while explicitly calling for progress toward banning lethal autonomous weapons. The proposition consequently addresses both civilian administrative decisions and military engagement. Its central concern is not simply whether a human appears somewhere within a process. It is whether that human retains the practical authority and sufficient understanding necessary to make the decision legally and ethically attributable to human judgment.

In administrative contexts, a distinction must be maintained between systems that assist human decision-makers and systems whose outputs determine outcomes in practice. A border-security database may produce a match that triggers further inspection. An automated assessment may identify an individual for additional review. Neither function necessarily amounts to an unlawful delegation of final authority. The governance problem emerges when a generated classification effectively determines detention, exclusion, investigation or another significant consequence without an independent assessment by an authorised official. Even when a formal human approval exists, the process may be substantively automated if the reviewing official is required to accept the recommendation, receives inadequate supporting evidence or must process a volume of cases inconsistent with meaningful evaluation.

For military applications, the distinction becomes more difficult because systems may operate after activation without requiring a new human decision for each engagement. A commander might select a limited operating area, define authorised target categories and impose time restrictions, while the system subsequently identifies and engages objects meeting predefined conditions. Whether that arrangement provides sufficient human judgment depends on the intended target, the predictability of the operating environment, the degree of uncertainty, available intervention mechanisms and the law applicable to the attack. It cannot be resolved merely by identifying the location of a human within a technical control loop.

The April 2026 CCW Chair’s summary demonstrates that this question is central to the negotiations. Delegations considered forms of human involvement across the weapon lifecycle and discussed whether continuous direct control was necessary. Some supported context-appropriate human judgment as a means of ensuring compliance with international humanitarian law. Others questioned whether the terminology represented an established legal rule or would create additional obligations. The resulting disagreement is substantive because different definitions could permit very different combinations of human authorisation and autonomous execution. Chair’s Summary, CCW/GGE.1/2026/WP.2, paragraphs 52–57 — United Nations Office for Disarmament Affairs — April 2026.

The papal proposition therefore contains an internal distinction relevant to future negotiations. Effective human oversight of decisions affecting life and liberty is a general institutional requirement. The call to ban lethal autonomous weapons is a more specific policy objective. The latter requires agreement on the prohibited class of systems. An absolute prohibition extending to all autonomous selection and application of force would have substantially different military consequences from restrictions targeting unpredictable systems, autonomous attacks against persons or systems deployed outside sufficiently controlled environments. Ethical consistency does not remove the need to specify which technical functions and operational circumstances trigger prohibition.

The fourth proposition: the common good requires more than universal declarations of technological benefit

The requirement that AI serve the common good introduces a distributive dimension into a field often governed through safety, security and economic competitiveness. Its significance lies in distinguishing the social availability of technological benefits from the simple expansion of AI markets. Greater commercial adoption can coexist with concentrated ownership of essential infrastructure, restricted access to advanced capabilities and unequal capacity to influence the rules under which technologies operate. An institutionally meaningful interpretation of the common good must therefore consider who possesses the productive resources required to develop AI, who controls access to those resources, who receives economic returns and who bears the risks associated with deployment.

A distributional obligation cannot be evaluated exclusively by counting the number of products accessible in a particular market. Access to AI services differs from the capacity to develop, evaluate, adapt and govern AI systems independently. A country may have extensive access to commercial applications while remaining dependent on foreign providers for computing infrastructure, model development, cloud services, cybersecurity and technical standards. Such dependence can create strategic vulnerabilities even where end users receive immediate productivity benefits. The common-good proposition consequently implies a distinction between consumption access, technological capability and governance participation.

In operational terms, public authorities could interpret the principle through funding for public-interest computing infrastructure, research access for universities and smaller institutions, technical assistance to developing countries, transparent eligibility conditions for subsidised services and procurement policies that discourage avoidable dependency. These are possible institutional translations, not obligations created by the speech itself. Their relative merits depend on implementation design. Public computing resources allocated without governance safeguards could subsidise commercial activity without materially improving scientific independence. Mandatory openness could undermine legitimate security restrictions or intellectual-property protections. Technology transfer without training, maintenance arrangements and sustainable financing could produce nominal access without durable operational capacity.

The economic tension is particularly important because the development of frontier AI capabilities depends on inputs that are not equally distributed. Advanced semiconductor manufacturing, specialised computing infrastructure, electricity availability, engineering skills and access to investment capital shape which actors can develop and deploy high-capability systems. Governments may seek to expand the distribution of benefits while simultaneously restricting access to particular technologies for national-security reasons. The common-good proposition does not resolve that conflict through a universal presumption in favour of unrestricted transfer. Instead, it requires an institutional distinction between restrictions justified by identifiable security risks and restrictions whose primary effect is the preservation of economic or geopolitical advantage.

This principle is therefore directional in legal form but potentially significant in economic consequence. Unlike the non-delegation proposition, it does not establish a single prohibited act. It calls for positive institutional arrangements whose effectiveness depends on resources, eligibility criteria, financing, enforceable commitments and measurable participation. Its broader industrial and international implications belong to Pillar II; the concern here is the transformation of a distributive ethical demand into obligations sufficiently precise to be implemented and reviewed.

The fifth proposition: international governance must possess operational powers, not merely deliberative legitimacy

The fifth proposition identifies institutional effectiveness as a necessary condition for implementing the preceding principles. It shifts attention from the content of ethical declarations to the legal competence, technical capability and independence of the organisations expected to supervise AI development. This is a distinct proposition because institutions may endorse a norm without possessing authority to monitor compliance, compel disclosure, investigate misconduct or provide remedies. A scientific panel may evaluate technological developments but lack inspection powers. A diplomatic forum may encourage convergence but lack authority to bind participating states. A national regulator may impose detailed obligations within its territory while being unable to investigate conduct occurring wholly outside its jurisdiction.

For international oversight to acquire more than advisory significance, governments would need to determine the legal basis of the institution’s powers, the activities subject to supervision, the standards of evidence required for enforcement, the treatment of confidential information and the procedures available to challenge institutional decisions. These are not secondary administrative questions. They determine whether enforcement is possible without making the regime unacceptable to the states and companies it seeks to regulate. An authority unable to inspect relevant systems may depend entirely on declarations supplied by regulated entities. An authority permitted unrestricted access to military systems may be rejected by governments concerned about operational secrecy. An institution whose decisions cannot be challenged may generate legitimacy problems of its own.

The most consequential design distinction is between standard-setting, conformity assessment, independent investigation and compulsory enforcement. These functions require different powers and resources. International technical standards may be incorporated into procurement contracts or domestic legislation without creating a supranational inspectorate. Independent assessment may be conducted through accredited institutions operating under national authority. Compulsory international inspection generally requires more extensive legal commitments, including rules governing access, confidentiality, findings and consequences for obstruction. Remedies for injured persons raise additional questions about jurisdiction, standing, applicable law, evidence, compensation and enforcement of judgments.

The papal proposition is thus institutionally demanding precisely because it requires the creation of arrangements capable of functioning when compliance is inconvenient. Its implementation cannot be measured by the number of international declarations expressing support for responsible AI. More appropriate measures include whether an institution can demand relevant records, verify compliance independently, identify responsible authorities, require corrective action and ensure that affected persons have access to meaningful review. The address makes this shift from principle to institutional capacity explicit in its concluding consideration. Address to the Participants in “The Future of AI” — Holy See — October 2026.

Comparative architecture: from ethical principle to enforceable obligation

Ethical proposition	Legal character at the assessment date	Necessary operational translation	Responsible institutional actors	Evidence of meaningful implementation
Human choice over technological development	Ethical principle, with partial analogues in existing regulatory and weapons-review duties	Authorisation, purpose limitation, change control, suspension and withdrawal	Legislators, regulators, acquisition authorities and system owners	Documented review decisions, withdrawal powers exercised where required, recorded changes of approved use
Rejection of machine personhood	Normative position broadly aligned with existing attribution of legal duties to people and organisations	Identify human and corporate duty-bearers throughout the system lifecycle	Legislators, courts, developers, deployers and enforcement bodies	Traceable allocation of duties, access to defendants and effective adjudication
Non-delegation of decisions affecting life and liberty	Ethical constraint; specific legal requirements vary by jurisdiction and context	Human decision authority, informed oversight, operational limits, challenge and intervention	Public authorities, military commanders, regulators and operators	Demonstrable ability to review, reject, suspend or reverse consequential system outputs
Common-good orientation	Directional distributive principle	Public-interest investment, access mechanisms, capacity-building and accountability	Governments, development institutions, research organisations and infrastructure providers	Verified access, sustained capability development and transparent benefit allocation
Effective international institutions	Institutional reform objective	Shared standards, independent verification, national implementation and remedies	States, treaty bodies, regulators, courts and accredited technical institutions	Legal mandates, inspection or audit capacity, actionable findings and enforceable remedies

The table illustrates that the five propositions cannot be implemented through a single regulatory instrument. Some can be incorporated into technical assurance and authorisation processes, while others require legislative allocation of duties or international agreements. The principal institutional risk is formal compliance without functional effect: an operator may appear responsible without possessing genuine authority, a review process may exist without access to relevant evidence, and an international institution may possess a broad mandate without the resources or legal powers needed to exercise it.

The European regulatory comparison: human oversight as a defined legal obligation

The European Union’s Artificial Intelligence Act provides a useful comparison because it translates human oversight into operationally specific requirements for high-risk AI systems within its scope. Article 14 requires such systems to be designed and developed so that natural persons can exercise effective oversight during use. Its provisions address understanding system capabilities and limitations, detecting anomalies, recognising automation bias, interpreting outputs, disregarding or overriding system recommendations and interrupting operation where appropriate. The importance of the provision is not that it resolves every ethical question raised by the papal propositions. It is that it demonstrates how an abstract concept such as human oversight can become a set of identifiable system-design and organisational requirements.

However, that comparison must be legally qualified. The AI Act’s scope excludes relevant activities carried out exclusively for military, defence or national-security purposes under the conditions specified in Article 2. The regulation therefore cannot be presented as an EU-wide legislative prohibition on autonomous military weapons. Its relevance lies in illustrating a regulatory technique: the relationship between system capabilities, human understanding, intervention authority and documented responsibility. Whether comparable requirements should govern military systems is a separate question involving international humanitarian law, military procurement, state security responsibilities and potential arms-control instruments. Regulation (EU) 2024/1689, Articles 2 and 14 — European Parliament and Council of the European Union — June 2024, consolidated text.

Article 14 also demonstrates why the concept of human oversight is not equivalent to attaching a nominal approval step to an automated process. The provision addresses the tendency of operators to rely excessively on system outputs, the need to interpret results correctly and the practical ability to disregard or interrupt a system. These obligations concern organisational behaviour as well as software functionality. Their underlying logic supports the proposition that human authority depends on the design of the entire decision process, not simply on whether the final interface contains a confirmation button.

Chapter 1 — Key judgments

High confidence: The five propositions are best understood as a hierarchy of ethical and institutional requirements rather than as five equivalent legal commands. Human responsibility and non-delegation impose the strongest constraints on permissible authority, while common-good distribution and institutional strengthening require positive policies whose content must be defined through legislation, agreements and resource allocation.

High confidence: A credible operational translation must establish explicit decision authority, verifiable technical limitations, identifiable duty-bearers and procedures for reviewing consequences. Systems that provide formal human involvement without meaningful intervention or independent assessment do not necessarily satisfy the substantive demands of the framework.

Medium-high confidence: Existing legal instruments provide partial mechanisms for implementation, particularly weapons reviews, risk management and human oversight, but they do not collectively establish a universal legal regime matching the entire ethical framework.

What would change the assessment: Adoption of new domestic laws explicitly prohibiting particular forms of autonomous decision-making, agreement on an internationally binding definition of prohibited autonomous weapons, or establishment of international institutions with enforceable inspection and remedial powers would materially narrow the gap between normative authority and legal constraint.

Open official record: The decisive unresolved documents concern the precise legal scope of future autonomous-weapons restrictions, the allocation of responsibility across integrated AI supply chains and the technical standards by which meaningful human control could be assessed consistently across jurisdictions.

Chapter 2 — Strategic Competition, Autonomous Weapons and International Humanitarian Law

Principal judgment — High confidence: The strategic incompatibility between a comprehensive prohibition on lethal autonomous weapons and contemporary military development is rooted less in an explicit rejection of humanitarian law than in disagreement over the permissible distribution of human judgment and machine execution. States can accept that international humanitarian law governs every attack while disagreeing over whether an autonomous weapon may lawfully select and engage targets after a human commander has authorised its operating parameters. This distinction creates substantial space for technological development under national legal frameworks, even where governments publicly endorse responsible AI and human accountability. The resulting international negotiations face three interconnected obstacles: the absence of a universally agreed functional definition, competing interpretations of acceptable human involvement and the difficulty of verifying restrictions on software-intensive capabilities that may be integrated into otherwise lawful military systems.

The operational transformation: autonomy is a function, not a synonym for artificial intelligence

A rigorous military assessment must distinguish artificial intelligence, autonomous navigation, automated decision support and autonomy in the selection and application of force. These capabilities may coexist within the same platform, but they have different operational and legal consequences. A remotely piloted aircraft can use sophisticated algorithms for navigation, image processing or route optimisation while a human operator independently identifies the target and authorises an attack. Conversely, a weapon may use relatively simple deterministic software to detect and engage objects without requiring a separate human decision for each target. The legal and humanitarian problem associated with autonomous weapons therefore cannot be identified solely by asking whether a system incorporates machine learning, nor can it be resolved by distinguishing systems marketed as intelligent from those described as conventional automation.

The International Committee of the Red Cross defines autonomous weapon systems by their ability, once activated, to select and apply force to targets without further human intervention. Its October 2026 assessment identifies certain sensor-fused munitions, loitering munitions and defensive systems as relevant examples, while emphasising that many existing applications employ relatively narrow, rule-based functions. This distinction matters because an international prohibition drafted around the word artificial intelligence could fail to cover dangerous autonomous behaviour produced by conventional software, while inadvertently restricting military uses of AI that do not independently select or engage targets. The regulatory object should therefore be the function performed by the system, the consequences of that function and the degree of control retained by human authorities. Frequently Asked Questions: Autonomous Weapon Systems — International Committee of the Red Cross — October 2026.

This classification introduces a critical relationship between the operating environment and the predictability of engagement. A system configured to intercept incoming projectiles within a defined defensive perimeter operates under different conditions from a system that searches a populated urban area for individuals matching a behavioural or visual profile. In the former case, the range of relevant objects, expected trajectories and permitted engagement conditions may be comparatively constrained. In the latter, target status can depend on context, intention, conduct and changes in circumstances that are difficult to reduce to stable technical indicators. This difference does not establish the automatic lawfulness of defensive systems or the automatic unlawfulness of every autonomous system deployed in a complex environment. It demonstrates why the relationship between the weapon, the target class and the operational environment is indispensable to legal assessment.

The military attraction of autonomous functionality arises from several operational pressures. Faster engagement can be valuable against incoming threats. Automated processing can reduce the burden of monitoring large quantities of sensor information. Distributed systems may continue operating when communication links are disrupted, and multiple platforms may conduct coordinated tasks with fewer personnel directly controlling each action. These advantages can support defensive operations, reconnaissance and other military functions without necessarily requiring autonomous lethal targeting. Nevertheless, as operational advantages become associated with reduced human intervention, pressure emerges to extend autonomy from supporting functions into decisions with progressively greater consequences.

The distinction between a system’s declared capability and demonstrated battlefield performance is essential. A manufacturer’s assertion that a platform can identify targets autonomously is evidence of an advertised function, not proof of reliable identification in contested environments. A military test demonstrating successful classification is not proof that the system can distinguish lawful targets in all relevant operational conditions. A publicly announced procurement programme does not establish operational deployment, and operational deployment does not establish compliance with international humanitarian law. The available official record supports a general conclusion that military organisations are investing in autonomous and AI-enabled capabilities. It does not support assigning universal rates of adoption, verified battlefield accuracy or precise future deployment dates across major powers.

The applicable legal framework: existing obligations and proposed restrictions

The foundational legal question is whether existing international humanitarian law already prohibits lethal autonomous weapons as a category. The answer requires separating the legality of a weapon from the legality of its use. International humanitarian law restricts weapons that are prohibited by specific treaty rules or by generally applicable principles, including restrictions on indiscriminate means and methods of warfare. Separately, the law governing the conduct of hostilities applies to individual attacks and imposes obligations concerning distinction, proportionality and precautions. A weapon that is not prohibited in every circumstance may nevertheless be used unlawfully in a particular attack.

Additional Protocol I is particularly relevant. Article 36 establishes an obligation for its High Contracting Parties to examine new weapons, means and methods of warfare in light of applicable international law. Article 48 establishes the basic rule of distinction between civilian populations and combatants, and between civilian objects and military objectives. Article 51 addresses protection of civilians, including prohibitions relating to indiscriminate attacks. Article 57 requires feasible precautions in attack, including target verification, minimisation of incidental civilian harm and cancellation or suspension when the relevant legal conditions are no longer satisfied. The legal effect of these provisions must be interpreted according to their treaty scope, applicable customary law and the obligations binding the state concerned. They do not collectively amount to an express universal prohibition of every autonomous weapon.

The relevant primary provisions are Additional Protocol I, Article 36: New Weapons — ICRC Treaty Database — June 1977, Additional Protocol I, Article 48: Basic Rule — ICRC Treaty Database — June 1977, Additional Protocol I, Article 51: Protection of the Civilian Population — ICRC Treaty Database — June 1977 and Additional Protocol I, Article 57: Precautions in Attack — ICRC Treaty Database — June 1977.

The operational difficulty lies in the contextual nature of these obligations. Distinction is not merely a demand for accurate object recognition. In many circumstances, identifying a lawful target requires information about the person’s status, conduct or relationship to hostilities. Proportionality requires an assessment of whether expected incidental civilian harm would be excessive in relation to the concrete and direct military advantage anticipated. Precautions require decisions about available information, alternative methods of attack and changing circumstances. A system may achieve high statistical accuracy on a restricted identification task while lacking the information necessary to make these broader legal assessments. It follows that an improvement in classification accuracy does not, by itself, establish that greater autonomous engagement authority is legally justified.

The distinction between advance authorisation and subsequent autonomous execution is especially consequential. A commander may review an intended attack and establish parameters for a system’s operation. Whether this is sufficient depends on the degree to which the future targets, environment and potential incidental effects can be anticipated. If the system is authorised to operate in circumstances that change materially after activation, the initial assessment may no longer correspond to the conditions governing an eventual attack. Conversely, where targets and effects can be adequately constrained and anticipated, advance human decisions may retain substantial legal significance. The legal assessment must therefore focus on whether the relevant duties can actually be discharged, rather than imposing an abstract assumption that all real-time automation is either lawful or unlawful.

The ICRC’s position illustrates an approach designed to address these differences. It advocates prohibiting autonomous systems that are unpredictable in their effects and systems designed to target persons directly, while restricting other autonomous weapons through limits on target categories, geographic scope, duration, scale and means of neutralisation or deactivation. The organisation expressly distinguishes that position from a comprehensive prohibition on every autonomous weapon or all military applications of artificial intelligence. This is a policy recommendation for additional legally binding rules, not a statement that its entire proposed framework already constitutes universally applicable law. Frequently Asked Questions: Autonomous Weapon Systems — International Committee of the Red Cross — October 2026.

The Convention on Certain Conventional Weapons: negotiation without an agreed prohibition

The Convention on Certain Conventional Weapons provides an established diplomatic venue for discussing restrictions on weapons considered excessively injurious or indiscriminate. Its Group of Governmental Experts on emerging technologies in the area of lethal autonomous weapons has examined legal, military, technical and humanitarian questions. The April 2026 Chair’s summary is particularly revealing because it records detailed negotiations over the content of a possible instrument while maintaining the qualification that its nature had not been predetermined. The document is an account submitted under the Chair’s sole authority, rather than a consensus-adopted treaty or a definitive statement of every delegation’s position.

The mandate for the 2026 discussions was to formulate, by consensus, elements of an instrument while considering existing CCW protocols, proposals from High Contracting Parties and alternative measures. Such a mandate permits negotiation of technical provisions without requiring agreement in advance on whether the final result will be a legally binding protocol, another form of instrument or a less demanding set of commitments. This procedural openness makes dialogue possible among states with different priorities. It also creates the possibility that negotiations may produce increasingly detailed language without resolving the decisive question of enforceability.

One important issue concerns the definition of a lethal autonomous weapon system. A functional approach may focus on systems capable of identifying, selecting and engaging targets without human intervention in those activities. However, defining autonomy in relation to particular functions does not automatically establish which functions are prohibited. A system may operate autonomously within narrow constraints and still satisfy the functional definition. States must therefore negotiate not only whether a system is autonomous but which combinations of autonomy, target type, operating environment and human involvement constitute unacceptable conduct.

A second issue concerns the meaning of human judgment. The Chair’s summary records that some delegations regarded context-appropriate human judgment and control as essential for implementing international humanitarian law, while others objected to terminology that could create uncertainty about the relationship between existing legal obligations and proposed new requirements. The disagreement is not merely semantic. If continuous human intervention is required, some systems that automatically engage incoming threats may become difficult to accommodate. If advance parameter-setting is sufficient in every circumstance, a broad range of targeting functions could remain technically autonomous despite formal human authorisation. An agreement that fails to distinguish these situations may be either operationally unacceptable or too permissive to deliver meaningful restrictions.

A third issue concerns the difference between lifecycle control and engagement control. Human supervision can occur during design, validation, acquisition, deployment planning and post-operation review. These activities can reduce risk and support accountability. Yet they do not necessarily establish that an identifiable person exercised the necessary judgment over an individual engagement when the circumstances demanded such judgment. The CCW debate therefore concerns how lifecycle governance interacts with decisions made at the point of applying force. This problem has no universal technical solution because the significance of real-time intervention varies with the operational context.

The authoritative record for these distinctions is the Chair’s Summary of the First 2026 Session of the Group of Governmental Experts on Lethal Autonomous Weapons Systems, CCW/GGE.1/2026/WP.2 — United Nations Office for Disarmament Affairs — April 2026, especially its discussion of operational characterisation, human involvement and possible regulatory approaches.

National doctrine and observed military practice: why positions cannot be treated as uniform

State practice demonstrates that support for responsible military AI does not necessarily imply support for prohibiting all autonomous weapons. The United States, United Kingdom, France and Germany provide useful comparisons because their official positions distinguish the desirability of human responsibility from the permissible technical arrangements through which that responsibility is exercised. These positions are not identical, nor do public policy documents establish the full extent of classified development, procurement or operational capabilities.

The United States Department of Defense’s Directive 3000.09, reissued on 25 January 2023, establishes policy and responsibilities for autonomous and semi-autonomous functions in weapon systems. Its stated purpose includes minimising the probability and consequences of failures that could result in unintended engagements. The directive is significant because it treats autonomy as a category requiring governance and authorisation rather than establishing a universal prohibition. It addresses the development and use of autonomous and semi-autonomous capabilities through requirements relating to appropriate human judgment, system performance, testing and review. The official document demonstrates that the United States has established a formal internal governance framework for such systems, but its existence does not establish that every application has been independently verified or that all military AI capabilities are publicly documented. Department of Defense Directive 3000.09: Autonomy in Weapon Systems — United States Department of Defense — January 2023.

The United Kingdom’s published defence AI policy provides a particularly explicit example of context-dependent human involvement. The Ministry of Defence states that it does not rule out incorporating AI into weapon systems and recognises that some applications may be impossible to deliver safely and responsibly. It requires context-appropriate human involvement in weapons that identify, select and attack targets, but accepts that such involvement may take the form of real-time supervision or control through parameters established before operation. This policy illustrates the gap between an ethical principle requiring human authority and a categorical prohibition on autonomous engagement. The United Kingdom’s position is framed around responsible application, operational context and military effectiveness rather than an unconditional rejection of automated targeting functions. Ambitious, Safe, Responsible: Our Approach to the Delivery of AI-Enabled Capability in Defence — United Kingdom Ministry of Defence — June 2022.

British operational experimentation also demonstrates why autonomous capability must be analysed by function and maturity. In June 2025, the Ministry of Defence and Defence Science and Technology Laboratory reported a trial in which one operator controlled three uncrewed air and land vehicles used to detect and classify potential threats. The announcement also identified plans to advance experimentation with coordinated unmanned systems for intelligence, surveillance and reconnaissance. This is evidence of experimentation with human-machine teaming, control efficiency and reconnaissance functions. It is not proof that the tested platforms were authorised to select and engage targets lethally without human intervention. Nevertheless, it illustrates the military incentives to expand the number of platforms and tasks overseen by a single operator. Successful Trial Paves the Way for Improved Reconnaissance on Army Operations — United Kingdom Ministry of Defence and Defence Science and Technology Laboratory — June 2025.

France’s official position must be interpreted according to its national definition of lethal autonomous weapons. The French Ministry for Europe and Foreign Affairs states that France would not develop weapons falling within the category it identifies as lethal autonomous weapons operating without human intervention and describes its support for multilateral rules on autonomous systems. This position is relevant evidence of an official restraint commitment, but it should not be interpreted as proof that France rejects all forms of automatic engagement, autonomous navigation or AI-enabled military targeting support. The scope of the national category is decisive. A state may reject fully autonomous lethal decision-making while retaining systems that conduct particular functions automatically under defined human authority. Systèmes d’armes létales autonomes : l’action de la France — Ministère de l’Europe et des Affaires étrangères — Updated August 2026.

Germany provides an instructive comparison because its official statements distinguish political commitment to international regulation from the operational requirements of certain defensive systems. At a federal government press briefing on 28 August 2026, a Foreign Office representative confirmed support for international rulemaking concerning lethal autonomous weapons. A Defence Ministry representative stated that decisions on whether an attack takes place remain under human authority and that the ministry did not commission the development of the type of independently lethal systems being discussed. These statements establish the government’s declared policy; they are not evidence of every element of national defence research or procurement. Federal Government Press Conference — German Federal Foreign Office — August 2026.

A subsequent briefing on 25 September 2026 clarifies the operational distinction. A German Defence Ministry representative discussed a hypothetical ship-defence environment in which a human authorises the use of a system against a defined threat, but the system may subsequently need to react faster than a human could provide a second engagement authorisation. The representative distinguished that constrained defensive environment from an urban setting where comparable delegation would raise different concerns. The significance is not evidence of a specific combat deployment. It is the public articulation of an operational doctrine in which the appropriateness of autonomous action depends on the threat, environment, initial human authorisation and available reaction time. Federal Government Press Conference — German Federal Foreign Office — September 2026.

Italy must be treated separately from these positions. Its obligations under applicable international humanitarian law, its participation in European and Atlantic security structures, and its role in the CCW process do not establish the content of an independent national policy on every category of autonomous weapon. The evidence examined for this chapter does not provide a sufficiently specific, current Italian government statement to support a categorical assertion that Italy either endorses or rejects all forms of autonomous lethal engagement. This is a consequential documentary limit, not grounds for assimilating the Italian position to that of France, Germany or the United Kingdom. The relevant future assessment requires Italy’s formal CCW interventions, defence-policy documents and authoritative procurement or weapons-review records.

Comparative military-policy matrix

Jurisdiction	Documented position or practice	Human-control approach	Operational implication	Principal qualification
United States	Formal DoD directive governing autonomous and semi-autonomous weapons	Appropriate human judgment, governance, testing and review	Permits regulated development subject to national requirements	Directive does not prove universal compliance or disclose classified programmes
United Kingdom	AI-enabled weapons not categorically excluded	Context-appropriate human involvement, including pre-set parameters where justified	Preserves substantial flexibility for military AI integration	Does not establish unrestricted permission for autonomous attacks
France	Official commitment not to develop systems falling within its defined fully autonomous lethal category	Retention of human authority; regulation through multilateral principles	Restrains defined classes while permitting other forms of military automation	Definition determines which capabilities are covered
Germany	Supports international regulation and publicly maintains human authority over attack decisions	Differentiates prior human authorisation from rapid defensive execution	Allows analysis of narrowly constrained defensive automation	Public explanations do not constitute a complete programme inventory
Italy	Applicable international legal obligations and participation in relevant diplomatic frameworks	Precise current national operational position requires further official documentation	Procurement and engagement doctrine cannot be inferred from allied positions	Insufficient verified current national record for a categorical comparison
European Union	AI Act establishes a civilian regulatory framework with relevant scope exclusions	Detailed human-oversight requirements for covered high-risk systems	Regulatory techniques may inform national assurance processes	The AI Act does not itself prohibit autonomous military weapons

The comparison demonstrates why declarations of support for human control cannot be counted as equivalent commitments to a comprehensive prohibition. The relevant variables are the functions subject to human authorisation, the operational circumstances in which automated engagement is permitted, the legal consequences of deviating from those limits and the existence of independent procedures for assessing compliance.

Verification: the difference between prohibiting a capability and proving its absence

The technical characteristics of AI-enabled weapons create distinct arms-control verification problems. Traditional restrictions on physically identifiable weapons may be supported by inspections of inventories, production facilities, delivery systems or declared stockpiles. Restrictions on software-enabled autonomy may require examination of system configuration, source code, model behaviour, operating parameters and the interfaces through which humans exercise authority. Many of these characteristics can change without a conspicuous physical modification to the platform. A system that ordinarily operates under direct human authorisation may possess configurable modes capable of reducing that intervention. A system whose operational range is restricted during declared testing may be technically capable of functioning across a wider environment. These possibilities complicate efforts to verify compliance through visual inspection or national declarations alone.

A meaningful verification regime would have to establish what constitutes the prohibited behaviour. Prohibiting particular autonomous functions requires the ability to determine whether a system can perform them, under which conditions and through which configurations. Prohibiting only operational use would require different evidence from prohibiting development, acquisition or possession. Limiting autonomous systems to specified target categories would require confidence that the technical characteristics and operating rules can be tested. Demanding permanent human approval for every engagement would require evidence about command interfaces, authorisation records and the system’s ability to operate independently. Each approach imposes different burdens on national security, commercial confidentiality and technical assurance.

There is also a distinction between declared restrictions and irreversible limitations. A software prohibition may be represented through configurable controls that can be changed by authorised personnel. Hardware-based limitations may be more difficult to alter but can impose greater development costs and reduce operational flexibility. Cryptographically protected configuration records can improve traceability but cannot establish lawful target selection by themselves. Independent testing can provide evidence about behaviour within defined conditions but cannot prove the absence of every undesirable response in an unbounded operational environment. Consequently, verification depends not on a single technical mechanism but on the combination of legal obligations, controlled configuration, testing, auditable authorisation and consequences for deliberate circumvention.

The international dimension is especially difficult. Governments may resist intrusive inspections of systems integrated into sensitive command networks. Companies may resist disclosure of proprietary code or model parameters. States may disagree about whether external inspectors should have access to training data, operational logs or security-classified targeting information. Even when inspection is accepted, evidence obtained from a system at one moment may not establish that its configuration remains unchanged after the inspection. A robust treaty would therefore need arrangements governing declarations, approved configurations, change control, inspection access, confidentiality, disputes and responses to noncompliance.

The legal capacity to impose these obligations remains distinct from the technical capacity to verify them. A treaty that establishes an ambitious prohibition but lacks agreed verification procedures may still influence behaviour and create legal responsibility for parties. Its enforcement effectiveness, however, would depend more heavily on national implementation, intelligence, whistleblower disclosures, post-incident investigations and diplomatic pressure. Conversely, a narrower treaty with technically demonstrable restrictions may support more reliable compliance assessment even if it leaves some ethically contested uses outside its prohibitions.

Escalation and strategic stability: autonomous speed can reduce decision time

The strategic risks of autonomous weapons are not confined to unlawful individual engagements. A broader question concerns whether increasing the speed of detection, target selection and response changes the opportunities for human intervention during crises. Automated defensive functions can reduce vulnerability to fast-moving attacks, but interaction between multiple automated systems may also compress decision cycles. If a system interprets an ambiguous event as hostile and initiates a rapid response, an opposing system may classify that response as a further threat. The resulting sequence can develop faster than senior authorities can assess its political consequences. Whether such a sequence occurs depends on the specific systems, command arrangements, rules of engagement and surrounding military situation; it should not be asserted as a demonstrated universal effect of autonomy.

The ICRC identifies possible escalation and reduced space for human restraint among the risks associated with autonomous weapons. These concerns provide a credible mechanism for examining strategic instability but do not supply a reliable numerical estimate of escalation probability. The relevant question for defence establishments is whether automation creates additional opportunities to avoid a harmful engagement or removes opportunities for correction once an erroneous process begins. A system that improves early warning but leaves the engagement decision with informed human authorities may increase available decision time. A system that automatically initiates force after detecting a poorly specified indicator may reduce it. The net effect depends on system architecture and doctrine rather than the presence of AI alone. Frequently Asked Questions: Autonomous Weapon Systems — International Committee of the Red Cross — October 2026.

These dynamics also influence military competition. A state may believe that restricting autonomous engagement will reduce its ability to respond to threats from an adversary that does not accept comparable restrictions. Another state may consider limits on autonomous targeting valuable because they reduce risks to its own forces, civilian populations and strategic stability. Such incentives can coexist within the same government, particularly where defensive, offensive and intelligence applications differ. Negotiated restrictions must therefore address asymmetric security concerns rather than assuming that all states receive identical benefits and costs from the same prohibition.

Conditions under which states may accept or resist a restrictive regime

A government considering binding restrictions must assess the consequences for military capability, verification, reciprocity and national sovereignty. Restrictions are more readily accommodated when they concern capabilities that the government does not intend to develop, when prohibited conduct can be defined without affecting essential defensive functions, when compliance can be assessed without unacceptable disclosure of sensitive information and when noncompliance by other states can be detected or otherwise addressed. Conversely, a state possessing or pursuing capabilities that would be substantially restricted may demand broader exceptions, weaker verification arrangements or a narrower definition of prohibited autonomy.

This is why the distinction between prohibition and regulation is important. A two-tier approach can prohibit specific systems considered unacceptable while regulating other applications through operational limits, human-control requirements and weapons reviews. Such an approach may preserve more military flexibility than a categorical ban. However, it introduces difficult boundary questions. States may disagree over what constitutes direct targeting of persons, which degrees of unpredictability are unacceptable, how reliably operational limits can be enforced and whether civilian-populated environments should trigger additional restrictions. If the permitted category becomes excessively broad, the regime could formally satisfy the demand for regulation without preventing the conduct that generated the humanitarian concern.

A further complication is the problem of nonparticipation. Treaty obligations ordinarily bind the states that accept them, subject to the relevant rules of international law. A prohibition supported by a significant group of states may influence international expectations and create political costs for nonparticipation, but it does not automatically acquire universal treaty force. Conversely, participation by technologically influential states can shape procurement markets and supplier behaviour even where no global inspection authority exists. The effective reach of a restrictive regime may therefore depend partly on export licensing, procurement conditions and access to controlled technological inputs, though these mechanisms raise their own distributive and geopolitical questions.

The decisive military-policy distinction

The analysis supports a conclusion that is narrower than the claim that autonomous weapons are inherently incompatible with international humanitarian law and broader than the claim that existing law eliminates the need for additional rules. Applicable law already constrains weapons and attacks, while the technical characteristics of autonomous systems can make compliance substantially more difficult to assess and demonstrate. Proposals for additional restrictions respond to those difficulties, but their scope remains contested because states differ in how they evaluate operational necessity and human involvement.

A government can therefore endorse the continued application of international humanitarian law, implement national weapons reviews, regulate autonomous systems and oppose a comprehensive prohibition without necessarily contradicting its stated interpretation of existing law. Equally, humanitarian organisations and states can argue that the same legal framework requires or justifies stronger categorical restrictions because certain autonomous functions cannot be controlled adequately in realistic operational conditions. The dispute concerns the relationship between legal obligations, technical feasibility and acceptable risk, not merely a difference in rhetorical commitment to human dignity.

Chapter 2 — Key judgments

High confidence: Autonomous weapons cannot be defined adequately by reference to AI alone. Selection and application of force, operating context, target categories and human authority are the decisive variables.

High confidence: International humanitarian law applies to autonomous weapons, but it does not currently establish a universal express prohibition of all systems capable of autonomous engagement. Proposed additional rules must be distinguished from existing legal obligations.

High confidence: The CCW negotiations have identified the central issues of definition, human judgment, operational restrictions and accountability but have not, on the documentary record examined, resolved the legal form or substantive scope of a universal restrictive instrument.

Medium-high confidence: Governments will evaluate restrictions according to their effects on operational flexibility, defensive capability, reciprocity, verification and classified information. This mechanism explains why support for responsible military AI does not necessarily produce agreement on categorical prohibition.

Medium confidence: Software configurability, operational secrecy and the difficulty of reproducing real-world combat conditions will remain significant obstacles to verifying broad restrictions on autonomous targeting.

What would change the assessment: An agreed CCW negotiating mandate for a binding instrument, a shared operational definition of prohibited autonomous functions, acceptance of credible verification arrangements by leading military powers or published national doctrine materially narrowing autonomous engagement authority would alter the current balance between ethical restriction and military flexibility.

Open official record: Current complete national position papers, independently verifiable weapons-review methodologies, sufficiently detailed public operational doctrines and final decisions from the November 2026 CCW Review Conference remain essential for a more conclusive assessment of the negotiating trajectory.

Chapter 3 — Human Responsibility, Machine Personhood and the Allocation of Liability

Principal judgment — High confidence: Maintaining human responsibility for AI-enabled harm requires more than denying legal personality to machines. Effective accountability depends on the ability to establish which persons or organisations possessed authority, which legal duties applied to them, which technical and organisational decisions contributed to the outcome, and which judicial or administrative mechanisms can provide an effective remedy. The principal governance risk is not that an AI system will formally become an independent legal person in the immediate future, but that complex technological processes will make responsibility practically difficult to establish even while human and corporate legal responsibility remains formally intact. This problem takes different forms in civilian regulation, state responsibility, military command and individual criminal liability, and it cannot be resolved by treating these regimes as interchangeable.

Responsibility without personhood: the legal subject and the technological instrument

The second papal proposition rejects the attribution of personhood to artificial systems because such recognition could obscure the responsibility of the people and institutions responsible for their development and deployment. The argument raises an important distinction between moral responsibility and legal attribution. Moral responsibility concerns the capacity to understand and answer for choices in relation to human values. Legal responsibility concerns obligations established under a particular legal system and the consequences of violating those obligations. These concepts may overlap but are not identical. A corporation can be subject to legal liability without possessing human consciousness or moral experience, while an individual may bear moral responsibility for conduct that does not satisfy the legal elements of a particular offence.

The existing legal significance of this distinction is that AI systems generally operate as products, services, software components or instruments used by identifiable organisations and individuals. Their behaviour may be complex, adaptive or difficult to explain, but these characteristics do not automatically create a new legal subject. The law must instead determine how conduct and consequences are attributed to the actors who designed the relevant capabilities, selected their intended purposes, authorised deployment, imposed operational limits or failed to respond to known problems. A legal regime can recognise that an AI system generated an output without accepting that the system itself became the appropriate defendant, criminal accused or bearer of an independent duty of care.

A specific source of confusion concerns the term autonomy. In engineering, autonomy describes a system’s capacity to perform functions without continuous external intervention. In law, autonomy can refer to legally protected decisional independence or the authority of a recognised legal subject. These meanings should not be transferred between domains without justification. An autonomous weapon can perform functions independently without possessing legal authority to decide whether a person may lawfully be attacked. An AI model can generate recommendations without acquiring the legal competence of the professional or public official responsible for their use. An automated commercial agent can execute transactions under delegated authority without necessarily becoming an independent legal person.

The corresponding governance principle is that technical independence does not establish legal independence. A system may act without immediate human instruction while remaining subject to obligations borne by its provider, owner, operator or deploying institution. The difficulty is determining which obligations attach to which actor, particularly when system behaviour is shaped by multiple organisations. If legal responsibility is allocated solely by formal ownership, liability may fall on an entity that lacks sufficient technical control. If it is allocated solely by technical expertise, it may overlook the authority exercised by the deploying institution. If it is allocated entirely through contractual arrangements, injured third parties may face limitations arising from agreements to which they were not parties.

The papal position therefore supports a principle of non-substitution: no attribution of technological agency should remove access to an otherwise responsible legal actor. However, this principle must be translated into specific rules governing legal duties, evidence, causation, jurisdiction and available remedies. The Address to the Participants in “The Future of AI” — Holy See — October 2026 provides the normative foundation, while the legal implementation depends on the relevant national and international regimes.

Civil liability: responsibility must follow the relevant duty and the mechanism of harm

Civil liability for AI-enabled harm cannot be reduced to a universal rule that the developer or operator is automatically responsible whenever an adverse outcome occurs. The result depends on the applicable legal framework, the nature of the claim, the identity of the defendant and the relationship between the system and the harm. Product-liability rules may apply where defective software or an AI-enabled product falls within the relevant statutory definition. Negligence or comparable fault-based rules may apply where an actor fails to exercise legally required care. Contractual liability may arise between parties that have agreed to performance, security or service obligations. Public-law remedies may be available where an administrative authority acts unlawfully through an automated process. These mechanisms can overlap, but each has distinct requirements.

The difficulty of establishing causation becomes especially important when a system is probabilistic, relies on changing external inputs or combines several independently developed components. An injured person may establish that an AI-enabled decision preceded harm without being able to demonstrate which design choice, input, configuration change or operational decision caused it. The system provider may possess evidence about model development but lack access to the deployment environment. The deploying organisation may possess operational records but be unable to examine proprietary model internals. An integrator may have modified the system without fully documenting how those changes altered its behaviour. These information asymmetries can make accountability practically inaccessible even where substantive legal rights exist.

The problem is not solved simply by requiring an explanation of the final output. A post-hoc explanation may describe factors associated with a decision without establishing the full causal process that generated it. Conversely, complete disclosure of source code may not explain a failure caused by inadequate training, erroneous operational assumptions or the use of a system outside its validated environment. Effective legal investigation may require several types of evidence: the system’s intended purpose, model or software version, available input records, testing results, known limitations, deployment instructions, operator actions and changes made before the relevant event.

This creates a direct relationship between technical documentation and access to justice. Documentation requirements are often described as regulatory compliance burdens, but they also determine whether a subsequent claimant, regulator or court can reconstruct a decision sufficiently to assess legal responsibility. Where relevant information is not generated or retained, the absence of evidence may benefit the organisation best positioned to prevent that absence. A governance regime concerned with human responsibility must therefore address record creation and preservation before harm occurs, subject to lawful limits concerning privacy, security, proportionality and confidential information.

The European AI Act provides a relevant example of regulatory obligations concerning the generation and retention of technical evidence. Its high-risk AI framework includes lifecycle risk management under Article 9, technical documentation under Article 11, logging capabilities under Article 12, information requirements under Article 13 and human oversight under Article 14. These obligations are not themselves a comprehensive civil-liability code, and their applicability depends on the regulation’s scope and temporal provisions. Their institutional significance is that they establish operational requirements capable of supporting later investigation and accountability. Regulation (EU) 2024/1689, Articles 9–14 — European Parliament and Council of the European Union — June 2024, consolidated text.

Organisational accountability: why identifiable individuals are necessary but insufficient

An organisation can formally designate an accountable official without giving that person adequate authority over the system. Such arrangements create a gap between assigned responsibility and practical control. A compliance officer may be required to approve deployment but lack access to relevant technical records. An operational supervisor may be responsible for monitoring AI outputs but lack the authority to suspend a production system. A military commander may be held responsible for an operational capability whose technical characteristics are not adequately communicated through acquisition and training processes. In each case, formal allocation of responsibility can become ineffective if the designated person cannot exercise the powers necessary to discharge the assigned duty.

A credible accountability architecture must therefore connect authority with information, competence and intervention capability. The individual or institution responsible for authorisation must understand the system’s intended use and material limitations. The organisation responsible for monitoring must possess access to performance information and incident reports. The person authorised to suspend operation must be able to exercise that authority without depending on discretionary cooperation from an external provider during a critical incident. Where responsibility is divided, escalation procedures must identify which decisions require higher-level approval and how conflicts between operational, technical and legal personnel are resolved.

These requirements are especially important for organisations that procure AI services rather than develop their own systems. Procurement contracts may allocate responsibilities between suppliers and customers, but contractual provisions do not necessarily determine the rights of injured third parties or eliminate statutory obligations. A public authority cannot assume that purchasing an externally developed model transfers its own legal duties to the supplier. Equally, a supplier may remain subject to obligations arising from defects or inadequate disclosures even where the deploying institution makes independent operational decisions. The central governance issue is whether contractual responsibility matches the actual distribution of control and whether applicable law provides remedies when it does not.

The most demanding cases arise when several actors possess partial control over a system and no actor can independently reconstruct its behaviour. A model developer may control training and architecture, a platform provider may control access and updates, an integrator may control connections to external databases, and a public authority may control consequential decisions. The resulting chain can be entirely composed of legally recognised persons and organisations while still producing a serious accountability gap. The absence of machine personhood does not prevent this outcome. It makes the need for explicit organisational duties more apparent.

Responsibility across the AI lifecycle: an allocation model

Lifecycle stage	Relevant controlling actor	Principal responsibility question	Evidence required	Potential accountability failure
Model conception	Developer and commissioning organisation	Were intended applications and foreseeable risks identified?	Design specifications, requirements and documented risk decisions	Harmful uses ignored during initial design
Data selection and development	Developer, data controller where applicable, and technical suppliers	Were relevant data limitations, errors and foreseeable biases addressed?	Data provenance, evaluation records and development documentation	Inadequate data governance or inaccessible records
Model validation	Developer, testing body and approving authority	Was performance evaluated under conditions relevant to intended use?	Testing protocols, limitations and validation results	Claims exceeding demonstrated performance
Integration	System integrator and deployment owner	Did integration introduce new functions, dependencies or risks?	Configuration records, interface documentation and integration tests	Changes not reflected in the original approval
Procurement and authorisation	Purchaser, regulator or competent public authority	Was the intended use lawful and adequately assessed?	Approval decisions, contractual conditions and risk reviews	Formal approval without sufficient technical understanding
Operational deployment	Deploying organisation and authorised operators	Were limitations observed and decisions taken by competent persons?	Operational procedures, training and system logs	Unauthorised use or ineffective human oversight
Monitoring and modification	Provider, system owner and maintenance authority	Were changes, failures and emerging risks detected and managed?	Update records, incident reports and monitoring evidence	Undocumented modification or continued use after known failures
Harm investigation	Competent investigative authority, organisation and relevant suppliers	Can the conduct, causal chain and applicable legal duties be reconstructed?	Preserved technical and operational evidence	Missing records, conflicting accounts and jurisdictional obstacles

This model is an analytical allocation of questions rather than a statement that every actor bears identical duties under all legal systems. It demonstrates why human responsibility is most credible when legal obligations are assigned throughout the lifecycle, rather than imposed only after an adverse outcome. The most significant institutional failure can occur before harm is visible: a system may enter operational service without complete records of its intended limitations, independent evaluation or a clearly established chain of authorisation. Subsequent legal proceedings then inherit an evidentiary deficit created by earlier governance decisions.

Military responsibility: separate the state, the commander and the operator

Military AI introduces additional legal complexity because the relevant responsibility framework differs from ordinary civilian product use. States bear obligations under applicable international law concerning the conduct of their armed forces and other attributable conduct. Individual persons may incur criminal responsibility where the legal elements of an international or domestic offence are satisfied. Military commanders and other superiors may possess specific responsibilities arising from command authority and their relationship to subordinate personnel. These levels of responsibility cannot be combined into a single claim that a human commander is automatically criminally liable for every harmful action involving an autonomous system.

At the level of state responsibility, the central question concerns whether conduct is attributable to the state and whether it breaches an applicable international obligation. The use of automated or autonomous equipment does not ordinarily alter the fact that a state’s armed forces act as organs of that state. The machine is not a separate sovereign actor capable of absorbing the state’s obligations. Nevertheless, identifying a breach requires examination of the applicable legal rule and the circumstances of the relevant conduct. Harm to civilians, for example, does not by itself prove a violation of international humanitarian law, although it may trigger a duty to investigate or otherwise require legal examination depending on the circumstances.

Individual criminal responsibility requires a different analysis. Under the Rome Statute, criminal liability depends on the jurisdiction of the International Criminal Court and the applicable elements of the offence, including relevant mental and contextual requirements. Article 25 addresses individual criminal responsibility, while Article 28 governs responsibility of commanders and other superiors under specified conditions. These provisions do not establish that every operational error involving autonomous equipment constitutes a war crime. Nor do they permit a commander to avoid legal responsibility simply by attributing conduct to the machine. The investigation must determine the acts and omissions of relevant persons, their knowledge or required mental element, their authority and the relationship between their conduct and the alleged offence.

The source for the governing distinctions is the Rome Statute of the International Criminal Court, Articles 25, 28 and 30 — International Criminal Court — July 1998.

A critical issue concerns the application of superior responsibility to harm caused through autonomous systems. Article 28 addresses crimes committed by forces under the effective command and control, or effective authority and control, of a superior, subject to the relevant legal conditions. An autonomous weapon is not itself a human subordinate in the ordinary sense contemplated by command responsibility. It would therefore be legally imprecise to assume that a machine can simply replace a subordinate within the provision without further analysis. The more defensible approach is to examine the conduct of the human personnel involved in acquisition, authorisation, deployment, supervision and the use of force, and then determine whether the applicable statutory requirements are satisfied.

This distinction matters because military accountability may fail through a succession of individually authorised decisions rather than a single unlawful command. A weapon may pass a legal review under particular assumptions, be approved for deployment under a specified doctrine and later be used in circumstances that differ materially from those assumptions. A system update may change performance without being adequately assessed by the responsible authority. A commander may authorise use without receiving relevant information about operational limitations. An operator may comply with formal procedures while lacking an effective opportunity to intervene before engagement. Each possibility raises different questions about the duties and knowledge of the persons involved.

Consequently, maintaining human accountability requires a chain of responsibility that extends from the acquisition authority to operational command. Legal review, procurement, testing, deployment authorisation, rules of engagement, operator training and incident investigation must function as connected processes. The importance of this architecture is reflected in national military policy documents that emphasise appropriate human judgment, rigorous testing and the mitigation of unintended engagements. Department of Defense Directive 3000.09: Autonomy in Weapon Systems — United States Department of Defense — January 2023.

Meaningful human control: four distinct institutional conditions

The term meaningful human control is frequently used as if it described a single technical configuration. In practice, it combines several requirements that can fail independently. Human authority may exist without adequate understanding. Understanding may exist without sufficient time for intervention. Intervention capability may exist without independent judgment. Independent judgment may exist without records capable of demonstrating what actually occurred. The institutional assessment must therefore distinguish authority, competence, practical control and accountability.

Authority concerns whether an identifiable person possesses the legal and organisational power to authorise, restrict, suspend or reject use. A supervisor who can express disagreement but cannot alter the outcome does not exercise effective decision authority. A commander who can authorise deployment but cannot define the system’s operating constraints may possess incomplete control. A regulator whose approval is formally required but whose access to technical evidence is inadequate may lack the practical basis for exercising independent authority.

Competence concerns the information and expertise necessary to evaluate the system’s behaviour. Operators need not understand every mathematical operation performed by an AI model, but they must understand the relevant capabilities, limitations and conditions under which system outputs should not be relied upon. In military environments, this includes the relationship between system behaviour and the rules governing target identification, expected incidental effects and precautions. Training must therefore address operational judgment rather than merely the use of interface controls.

Practical control concerns the ability to influence the system’s behaviour within the time available. A cancellation mechanism is not meaningful if it cannot be activated before the relevant consequence occurs, unless another justified control arrangement satisfies the applicable legal requirements. Similarly, continuous human observation does not guarantee control if the system operates too quickly for informed intervention. The appropriate design may differ according to the operating environment, but that difference must be established through analysis and testing rather than assumed from a generic description of human supervision.

Accountability concerns the capacity to reconstruct and assess the decision after the event. A system may have been lawfully authorised, but that assertion must be capable of verification where accountability is required. Relevant records may include the operational configuration, defined limitations, person authorising deployment, time and location of activation, available sensor information, subsequent changes and actions taken in response to unexpected behaviour. The exact records required should be proportionate to the application, compatible with applicable law and protected against inappropriate disclosure or manipulation.

The European AI Act’s Article 14 provides a civilian regulatory example of how these requirements can be specified. Its human-oversight provisions address operator understanding, automation bias, interpretation of outputs, the ability to disregard recommendations and intervention or interruption. These requirements apply within the regulation’s scope and should not be misrepresented as automatically governing military weapons. They nevertheless demonstrate that effective oversight requires system capabilities and institutional procedures to be designed together. Regulation (EU) 2024/1689, Article 14: Human Oversight — European Parliament and Council of the European Union — June 2024.

Human-control architecture: comparing operational arrangements

Control arrangement	Human authority	Principal benefit	Principal limitation	Necessary accountability evidence
Human selects each target and authorises each engagement	Direct engagement-level decision	Clear location of immediate decision authority	May be incompatible with extremely compressed reaction times	Engagement authorisation, supporting information and operational records
Human authorises a defined operating area, duration and target class	Advance operational decision	Permits rapid execution within pre-established limits	Changing conditions may exceed original assumptions	Parameter records, legal review, activation and termination logs
Human supervises autonomous operation with intervention authority	Ongoing supervisory control	Allows monitoring and possible corrective action	Intervention may be too slow or information insufficient	Operator display records, alerts, intervention capability and training
System operates under technical restrictions established during development	Design and authorisation control	Can constrain prohibited behaviour before deployment	Restrictions may be incomplete, altered or invalid outside testing conditions	Configuration integrity, validation and change-management records
Human reviews completed system actions	Retrospective accountability	Supports investigation, learning and remedial action	Cannot prevent an already completed harmful event	Complete audit records and independent investigative access

The arrangements are not mutually exclusive. A system may incorporate design restrictions, prior human authorisation, real-time monitoring and post-operation review. Their combined effectiveness depends on whether the control measures address the specific risks associated with the system’s authorised use. However, retrospective review should not be treated as a substitute for required precautions before or during an attack. Similarly, a nominal human approval mechanism cannot compensate for an inability to understand or challenge the underlying system output when the law requires informed judgment.

Corporate behaviour: the economics of liability avoidance and compliance

The allocation of responsibility also affects corporate development incentives. Companies developing high-capability AI systems may face competing pressures to reduce operational risk, expand market access, protect proprietary information and minimise potential legal exposure. These incentives do not necessarily produce deliberate attempts to evade responsibility. They can, however, create organisational structures in which technical complexity and contractual fragmentation make accountability more difficult to establish. The challenge for regulators is to distinguish legitimate allocation of responsibilities from arrangements that separate commercial benefit from legally relevant control.

A provider may have incentives to define the intended purpose narrowly, thereby limiting the uses for which it accepts responsibility. A system integrator may seek contractual warranties from the original developer while retaining authority over deployment-specific configuration. A customer may expect the supplier to assume responsibility for system errors while exercising substantial control over operating conditions. These arrangements are not inherently improper. They become problematic when contractual allocations do not correspond to the actor’s actual capacity to prevent harm or when injured persons cannot identify a legally responsible party.

The economics of compliance also depend on market structure. Large developers may be able to finance extensive testing, legal review and documentation as part of normal operations. Smaller firms may face proportionately higher burdens when required to demonstrate comparable levels of assurance. A regulatory regime that demands complex certification without providing clear, accessible technical requirements could increase barriers to market entry. Conversely, weak obligations can favour firms willing to deploy inadequately tested systems, transferring the costs of resulting harm to customers, public authorities or affected individuals. Effective governance therefore requires requirements proportional to the application and its consequences, while preserving sufficient documentation and oversight for accountability.

A further issue concerns the use of third-party models and general-purpose systems. An organisation deploying a model into a high-impact context may add decision logic, connect it to external information sources or allow it to initiate consequential actions. The original provider may not have validated that specific combination. The deploying organisation may lack access to all information about the underlying model. Responsibility must therefore account for material modification, foreseeable integration risks and the difference between general capability and authorised use. A statement that a system is safe in ordinary applications does not establish its suitability for military targeting, detention decisions or other highly consequential functions.

Liability and control in civilian and military systems: a comparative legal matrix

Legal regime	Potentially responsible actor	Governing question	Relevant evidence	Principal limitation
Product liability	Producer and other actors covered by the applicable statute	Was there a legally relevant defect connected to the harm?	Product characteristics, design, warnings and causation evidence	Different statutory scope and causation requirements
Fault-based civil liability	Persons or organisations subject to an applicable duty of care	Was a duty breached, causing legally recoverable harm?	Foreseeability, conduct, standards and causal connection	Information asymmetry and difficulties proving breach
Contractual liability	Parties bound by relevant agreements	Were contractual obligations violated?	Service agreements, specifications and performance records	Limited reach to nonparties and contractual exclusions
Administrative or public-law responsibility	Competent public authority	Was a consequential public decision lawful, reviewable and procedurally valid?	Decision records, authority, reasons and review mechanisms	Jurisdictional limits and restrictions on access to evidence
International state responsibility	State to which conduct is attributable	Was an applicable international obligation breached?	Attribution, relevant conduct and legal obligations	Jurisdiction, enforcement and evidentiary constraints
Individual criminal responsibility	Natural persons	Are the applicable offence elements and mental requirements established?	Individual conduct, intent or knowledge and contextual evidence	High evidentiary thresholds and limited jurisdiction
Military superior responsibility	Commanders or other qualifying superiors	Are the statutory conditions concerning subordinates, control, knowledge and failure to act satisfied?	Effective authority, communications, warnings and reasonable measures	Cannot assume the autonomous system itself is a human subordinate

This comparison demonstrates why the demand to maintain human responsibility is institutionally meaningful but cannot be implemented through a single universal liability provision. Legal responsibility depends on the specific duty violated. The evidence needed to establish defective design differs from that needed to establish unlawful public administration or individual criminal responsibility. A governance framework should therefore ensure that technical evidence is available to the competent authority without presuming that all harmful outcomes have the same legal character.

The specific problem of explainability and evidentiary reconstruction

Explainability has become a central theme in AI governance, but the term covers several different objectives. An operator may need an explanation sufficient to understand a recommendation. A regulator may need evidence that the system meets prescribed performance and risk-management requirements. A court may require records supporting a finding of causation or breach of duty. A military investigator may need to reconstruct the sequence of authorisations, inputs, system actions and operational circumstances surrounding a particular engagement. These needs overlap but cannot necessarily be satisfied through the same technical explanation.

In consequential systems, the most useful evidence may include a combination of contemporaneous records and independent testing rather than a natural-language explanation generated after an event. A system can produce a plausible description of its decision without that description constituting reliable evidence of the computational process. Conversely, technical logs can establish which inputs were processed and which outputs were generated while leaving unresolved whether the system’s design was appropriate for the task or whether the relevant human authority understood the consequences. Evidentiary sufficiency therefore requires a broader reconstruction of design, configuration, operational context and human conduct.

A defensible governance framework would establish record-retention duties according to the sensitivity and consequences of the application. High-impact deployments require greater traceability than applications with negligible effects on rights or safety. The retention period, access rules and security protections must comply with applicable law and account for legitimate confidentiality interests. In military systems, classified information can complicate independent scrutiny, creating a further need for competent investigative mechanisms capable of handling protected evidence without abandoning accountability altogether.

The key limitation is that complete technical transparency is neither always possible nor necessarily sufficient. Even extensive model documentation may not permit accurate prediction of every output in every environment. The objective should be to preserve the evidence needed to determine whether reasonable design, authorisation, testing and operational obligations were fulfilled. This is a more precise and legally meaningful goal than asserting that every complex AI decision must be fully explicable in an absolute technical sense.

Remedies: accountability is incomplete without an accessible procedure

The existence of an identifiable responsible actor does not guarantee that an injured person can obtain a remedy. Effective accountability also depends on jurisdiction, access to evidence, procedural standing, applicable liability rules and the ability to enforce a resulting decision. Cross-border AI services can create difficulties where the developer, infrastructure provider, deployer and injured person are located in different legal systems. Military applications present additional constraints involving jurisdictional immunities, armed-conflict rules, national security and the conditions under which international adjudication is available.

A credible remedial framework should distinguish preventive measures, administrative enforcement and compensation or other legal remedies. Preventive obligations may require an authority to suspend or restrict a system when unacceptable risks become apparent. Administrative sanctions may apply where regulated entities breach specified duties. Civil remedies may provide compensation or other relief when the applicable legal requirements are satisfied. Criminal proceedings may arise in the narrower circumstances where an individual offence and the relevant jurisdictional conditions can be established. None of these mechanisms is a complete substitute for the others.

The principle of maintaining human responsibility therefore becomes operational only when the affected person or competent investigating authority can move from a harmful outcome to a legally assessable claim. This requires identifying the relevant actors, obtaining sufficient records, establishing the applicable duties and providing a procedure for determining whether those duties were violated. Without these elements, formal statements of accountability may remain politically significant but practically ineffective.

State and corporate incentives: where responsibility becomes contested

States and companies may both support accountable AI while disagreeing over disclosure obligations, liability standards and oversight powers. Governments have incentives to preserve national-security information and protect operational flexibility. Technology companies may seek to protect trade secrets, limit legal exposure and avoid burdens that reduce competitiveness. Regulators require information to evaluate compliance, while injured parties may need access to evidence that is controlled by those same institutions. These competing interests create an institutional conflict that cannot be resolved by agreement on the abstract desirability of responsibility.

The central design challenge is to provide sufficient access to relevant evidence without assuming that all proprietary or classified information must become public. Independent confidential review, judicially supervised disclosure, protected reporting mechanisms and appropriately accredited technical assessment can potentially reconcile some of these interests. Each mechanism has costs and limitations. Confidential inspection may improve technical access but reduce public transparency. Broad disclosure can support accountability while creating security and commercial risks. Narrow disclosure may protect legitimate interests but leave courts and regulators unable to reconstruct harmful conduct. The appropriate balance depends on the sensitivity of the system, the legal proceeding and the authority of the institution requesting access.

A further concern arises when public authorities rely heavily on a small number of technology suppliers. Technical dependence may reduce the government’s ability to independently evaluate a system or modify it without supplier assistance. Contracts can require cooperation, testing and information-sharing, but the effectiveness of such provisions depends on technical feasibility and enforcement. Procurement policy therefore becomes an important mechanism for ensuring that public authorities retain enough knowledge and control to meet their legal obligations. This does not require public ownership of every underlying model or software component. It requires that essential accountability functions remain exercisable by the competent authority.

Recommended accountability architecture for consequential AI systems

The following is an analytical governance model, not a representation of universally established legal requirements. Its purpose is to identify the mechanisms necessary to preserve accountable human decision-making across complex systems.

Control function	Operational requirement	Responsible authority	Verification evidence	Failure consequence
Legal classification	Determine applicable duties before deployment	Competent legal and regulatory authority	Recorded legal assessment	Unclear or incorrectly applied legal regime
Purpose authorisation	Approve defined use and operational boundaries	System owner or designated public authority	Authorisation and permitted-use conditions	Use beyond approved purpose
Technical assurance	Validate relevant capabilities and limitations	Developer, testing authority and deployment owner	Test results and documented limitations	Unsupported assumptions about reliability
Human decision authority	Identify who may approve consequential action	Competent official, operator or commander	Authority matrix and decision records	Nominal supervision without effective power
Operational monitoring	Detect deviations and emerging risks	Deploying organisation	Monitoring, alerts and incident records	Continued operation after material warning
Suspension and withdrawal	Enable effective interruption or cessation	Designated operational authority	Tested procedures and decision records	Inability to prevent further harm
Incident reconstruction	Preserve evidence needed for investigation	System owner and relevant suppliers	Logs, configuration and chain-of-custody evidence	Inability to establish relevant conduct
Independent review	Permit competent scrutiny of disputed conduct	Regulator, court or authorised investigator	Accessible records and procedural safeguards	Accountability dependent on self-reporting
Remedy	Determine and enforce consequences for proven violations	Competent judicial or administrative authority	Decisions, corrective measures and enforcement	Formal rights without practical redress

The value of this architecture lies in the relationship among its elements. Technical assurance can reduce the likelihood of failures but cannot replace lawful authorisation. Human decision authority can support accountability but becomes ineffective without sufficient information. Incident records can facilitate investigation but cannot compensate for the absence of reasonable preventive measures. Independent review is credible only when the reviewing authority can examine relevant evidence. Remedies are meaningful only when resulting decisions can be implemented against legally responsible actors.

Chapter 3 — Key judgments

High confidence: Rejecting machine personhood preserves the conceptual distinction between autonomous technical performance and legal responsibility, but it does not resolve the evidentiary and jurisdictional problems associated with complex AI systems. The decisive accountability problem is the attribution of conduct, duties and causal responsibility among identifiable persons and organisations.

High confidence: Civil liability, international state responsibility, individual criminal responsibility and military superior responsibility are separate legal regimes. Their conditions must be evaluated independently. In particular, the Rome Statute does not establish automatic criminal liability for commanders whenever an autonomous system causes harm.

High confidence: Effective human control requires an identifiable decision-maker with relevant authority, information, competence and practical intervention capability. Formal human participation is insufficient where the person cannot meaningfully assess or influence the decision.

Medium-high confidence: Documentation, version control, testing, traceable authorisation and operational records are central to practical accountability because they determine whether competent authorities can reconstruct a system’s behaviour and the conduct of responsible actors.

Medium confidence: Greater regulatory requirements may increase compliance costs and create disproportionate burdens for smaller providers, while insufficient requirements may transfer the social and legal costs of failures to affected persons and public institutions. The resulting market effects will depend on the scope, proportionality and implementation of the applicable rules.

What would change the assessment: Judicial decisions clarifying responsibility for autonomous behaviour, harmonised statutory treatment of AI-related defects and causation, enforceable requirements for access to technical evidence or new military accountability rules specifically addressing autonomous engagement would materially affect the present legal analysis.

Open official record: The decisive gaps concern jurisdiction-specific liability decisions involving consequential AI systems, the content of national military weapons-review procedures, enforceable access to confidential technical evidence and evidence demonstrating whether designated human supervisors possess effective operational authority in practice.

Pillar I — Integrated Strategic Assessment

The combined findings of the three chapters establish a structured distinction between ethical legitimacy, permissible military capability and legally attributable responsibility. The papal propositions identify substantive limits on the treatment of human life, liberty and technological authority, but these limits operate differently within national regulation, international humanitarian law and institutional enforcement. The ability to translate them into effective rules depends on precision in three areas: the conduct that is restricted, the authority responsible for compliance and the evidence required to establish whether compliance occurred. Without those elements, broad normative agreement may coexist with divergent military practice and unresolved liability.

The central legal distinction is that a human decision-maker cannot be replaced by a claim that a machine independently possesses responsibility, yet existing law does not necessarily require an individual human to issue a separate command at every stage of every automated military process. The relevant legal duties depend on the weapon’s characteristics, the operational environment and the applicable rules governing its use. Future arms-control negotiations may introduce additional categorical restrictions, but those restrictions cannot be assumed to exist before the relevant legal instrument is adopted and becomes binding on the state concerned.

The central military distinction is between technological autonomy and the delegation of legally consequential judgment. States may employ automated systems in narrowly defined defensive environments while rejecting systems that independently determine whether human beings should be targeted. The degree to which this distinction can be maintained depends on the specificity of operational restrictions, the predictability of system behaviour and the availability of competent human oversight. As military systems become increasingly dependent on software and data, national authorisation procedures and international verification mechanisms will have to evaluate functional behaviour rather than relying solely on conventional descriptions of weapons platforms.

The central accountability distinction is between formal responsibility and responsibility that can actually be exercised and enforced. A legal system may identify a developer, operator or commander as responsible while denying investigators access to the evidence necessary to examine their conduct. A military doctrine may declare that humans retain control while authorising operational arrangements in which relevant intervention is impracticable. A corporation may accept regulatory obligations while distributing essential technical information across several entities. These are institutional design problems that require specific legal powers, operational procedures and evidentiary standards.

For governments, the practical consequence is that ethical AI governance cannot be treated as a single programme implemented through general declarations of responsible innovation. Defence ministries must assess weapon functions, authorisation arrangements and operational constraints. Civil regulators must address system safety, documentation, human oversight and enforceable legal duties. Legislators and courts must clarify responsibility for harm where existing rules are insufficiently precise. Diplomatic authorities must decide which restrictions should become international obligations and whether the institutions responsible for supervision will possess credible verification and enforcement powers.

The principal unresolved strategic question is whether states will accept constraints that remain binding when technological competition creates incentives to depart from them. The evidence examined supports the existence of both substantial humanitarian concerns and legitimate differences in the interpretation of permissible military autonomy. It does not establish that the principal military powers are prepared to accept a comprehensive prohibition supported by intrusive external inspection. Nor does it establish that national regulatory measures alone can resolve the cross-border and military dimensions of AI accountability.

Final Pillar I judgment — High confidence: The effectiveness of the five propositions will depend on whether governments can preserve meaningful human authority throughout increasingly complex technological systems while establishing legal obligations that are sufficiently precise to be implemented, investigated and enforced. The decisive distinction through 2030 will not be between states that publicly endorse ethical AI and those that reject it. It will be between governance arrangements that can demonstrate lawful authority, controlled deployment and attributable responsibility, and arrangements in which those requirements remain declarations without corresponding operational capacity.

The subsequent analysis of institutional competence, industrial concentration, technological access and international distribution must therefore assess the resources and political authority necessary to sustain these obligations. Those questions form the separate analytical functions of Pillar II and are not presumed resolved by the legal and military findings established here.

PILLAR II — Institutional Capacity, Industrial Competition and Distributional Power

Chapter 4 — International AI Governance: Institutional Mandates, Oversight and Enforcement

Principal judgment — High confidence: International artificial-intelligence governance has progressed beyond the stage of general ethical declarations, but the resulting institutional architecture remains substantially more capable of coordinating scientific knowledge, establishing regulatory expectations and facilitating diplomatic participation than of compelling compliance across sovereign jurisdictions. The fundamental institutional weakness is not the absence of international organisations. It is the fragmentation of authority among bodies whose mandates differ in legal character, geographic reach, technical competence, enforcement powers and access to information. The United Nations can convene governments and establish scientific assessment mechanisms; the Council of Europe can develop treaty obligations concerning human rights and democratic governance; the European Union can impose directly applicable regulatory requirements within its jurisdiction; and technical standardisation organisations can establish criteria that public authorities and commercial institutions may incorporate into binding requirements. None of these functions, separately or collectively, creates a universal authority competent to supervise all civilian and military AI systems, inspect sensitive national infrastructure, adjudicate cross-border violations and impose enforceable remedies. The strategic challenge through 2030 is therefore to connect existing institutional capabilities without assuming that coordination, regulatory convergence and coercive jurisdiction are equivalent.

The institutional transformation: scientific legitimacy, regulatory authority and the problem of enforcement

The establishment of new international AI mechanisms has changed the institutional environment in which governments negotiate technological governance. On 26 August 2025, the United Nations General Assembly adopted Resolution A/RES/79/325 without a vote, establishing the Independent International Scientific Panel on Artificial Intelligence and the Global Dialogue on AI Governance. The Panel comprises 40 members, serving in their personal capacities for three-year terms, and is mandated to produce evidence-based assessments, including an annual policy-relevant but non-prescriptive report. Its founding resolution provides for conflict-of-interest disclosures, geographic representation and procedures intended to preserve scientific independence. These characteristics are institutionally important because they attempt to reduce the disparity between governments able to commission sophisticated national assessments and governments lacking comparable scientific resources. They do not, however, give the Panel power to compel private companies to disclose proprietary model architectures or states to reveal classified military programmes. Terms of Reference and Modalities for the Establishment and Functioning of the Independent International Scientific Panel on Artificial Intelligence and the Global Dialogue on Artificial Intelligence Governance, A/RES/79/325 — United Nations General Assembly — August 2025.

The distinction between assessment and enforcement is made particularly clear by the structure of the Global Dialogue. Its inaugural meeting took place in Geneva on 6–7 July 2026, with delegations from 163 countries, together with participants from industry, academia, civil society and the technical community. The Dialogue addresses subjects including trustworthy AI, capacity-building, social and economic consequences, interoperability, human rights, transparency, human oversight and the development of open technologies. The second meeting is scheduled for 3–4 May 2027 in New York, while the independent scientific assessment process provides a recurring evidentiary basis for international discussion. These arrangements expand participation and the availability of shared information, but the United Nations explicitly identifies the Dialogue as a forum rather than a regulatory or enforcement authority. Attendance by 163 countries is consequently evidence of broad diplomatic participation, not proof of agreement on binding obligations, international inspection or common liability standards. Global Dialogue on Artificial Intelligence Governance — United Nations — 2026; Independent International Scientific Panel on Artificial Intelligence: Mandate and Institutional Arrangements — United Nations — 2026.

This difference has practical consequences for institutional credibility. A scientific body can identify emerging risks, compare research findings and publish assessments that inform national policy, but it cannot independently ensure that a government acts on those findings. A diplomatic platform can encourage common terminology, identify conflicting interests and facilitate political commitments, but it cannot enforce those commitments without an additional legal basis. A domestic regulatory authority may possess powers to request documentation, conduct inspections and impose sanctions, yet those powers may not extend to a foreign military programme or to an organisation operating beyond the jurisdictional reach of the relevant legal instrument. The result is a governance architecture in which knowledge production, political negotiation, conformity assessment and enforcement remain divided among institutions that may not share information, resources or legal authority.

The significance of this fragmentation becomes greater as AI systems are developed and operated across borders. A model may be trained in one jurisdiction, hosted in another, integrated into a commercial service elsewhere and used to support decisions affecting individuals in several countries. The developer, infrastructure provider and deploying authority may be subject to different obligations. Even where all relevant governments endorse general principles of safety and accountability, they may disagree over which law governs the conduct, which authority can demand technical evidence, whether information can be transferred abroad and which remedies are available. International coordination can reduce those difficulties by encouraging common definitions and mutual recognition of appropriate safeguards, but it cannot eliminate conflicts over legal jurisdiction or national-security exceptions without additional political agreement.

The Council of Europe: a treaty framework with a distinct legal function

The Council of Europe’s Framework Convention on Artificial Intelligence and Human Rights, Democracy and the Rule of Law provides a more expressly legal approach than the UN scientific and dialogue mechanisms. Opened for signature in Vilnius on 5 September 2024, the Convention establishes a framework through which participating parties address risks arising from AI activities in relation to human rights, democratic institutions and the rule of law. Its institutional importance derives from the use of an international treaty rather than a purely voluntary declaration. Its practical effect nevertheless depends on entry into force, participation, applicable scope and implementation through national or regional legal systems. The treaty’s entry-into-force condition requires five ratifications, including at least three Council of Europe member states. The treaty office’s status record must therefore be consulted rather than treating signature as equivalent to binding participation. Council of Europe Framework Convention on Artificial Intelligence and Human Rights, Democracy and the Rule of Law, CETS No. 225 — Council of Europe — September 2024.

A significant development occurred on 15 May 2026, when the European Union deposited its instrument of approval and declarations concerning the Convention. The EU declaration addressing private actors specifies that the relevant principles and obligations will be implemented through the EU Artificial Intelligence Act, together with other applicable elements of Union law. This is important because it illustrates a method of converting international treaty commitments into an already existing regulatory structure rather than creating an entirely separate supranational compliance apparatus. It also demonstrates that treaty obligations and regional legislation must be interpreted together according to their respective scopes. The declaration does not establish that the Convention governs every military or national-security AI system, nor does it eliminate the need to examine the treaty’s exclusions and provisions governing such activities. Reservations and Declarations for Treaty No. 225 — Council of Europe Treaty Office — May 2026.

The Convention’s institutional model creates a distinction between international legal commitment and centralised supervision. A framework convention can establish principles and require parties to maintain measures consistent with its objectives without creating an international inspectorate empowered to examine every relevant AI system. Monitoring, reporting and cooperation arrangements may support compliance, but their effectiveness depends on the precision of underlying obligations and the domestic institutions responsible for carrying them out. This approach can be politically more acceptable than granting extensive coercive powers to a new international authority because it preserves substantial national implementation discretion. Its limitation is that national governments may adopt measures differing in coverage, enforcement intensity and available remedies, producing formally compatible but operationally uneven regimes.

The central institutional lesson is that treaty status must be assessed through three separate questions. First, has the treaty entered into force under its own provisions? Second, which states or organisations are bound by it, and with what reservations or declarations? Third, what powers and duties does the instrument actually establish? The existence of a treaty text is not proof of universal applicability, and the use of legally binding language is not proof of centralised enforcement capacity.

The European Union: directly applicable regulation within defined jurisdictional limits

The European Union has developed a more operationally enforceable regulatory architecture through Regulation (EU) 2024/1689, as amended. Its importance lies not only in establishing substantive obligations but in assigning responsibilities to providers, deployers, conformity-assessment bodies, market-surveillance authorities and EU-level governance structures. The regulation uses risk-based requirements, documentation obligations, transparency rules, conformity mechanisms and sanctions within the legal scope established by the instrument. Unlike an international discussion forum, the EU regime can create obligations directly applicable to regulated economic actors. Unlike a global treaty, however, its jurisdiction derives from EU law, and its exclusions and transitional provisions delimit the activities to which it applies. Regulation (EU) 2024/1689, consolidated version of 27 July 2026 — European Parliament and Council of the European Union — July 2026.

The legal timeline is particularly important because it changed during 2026. The AI Omnibus, which entered into force on 27 July 2026, extended the application of requirements for certain high-risk systems and introduced measures intended to simplify implementation and support innovation. Under the amended framework, high-risk AI systems within Annex III are scheduled to become subject to their corresponding requirements from 2 December 2027, while the applicable high-risk requirements for AI systems associated with products within Annex I are scheduled from 2 August 2028. Other provisions began applying earlier, including rules concerning general-purpose AI models and governance from August 2025, and many transparency and regulatory provisions from August 2026. The staggered timeline means that descriptions of the AI Act as either entirely operational or not yet operational are misleading. Different provisions have different effective dates and enforcement consequences. AI Omnibus Enters into Force — European Commission — July 2026; Timeline for the Implementation of the EU AI Act — European Commission AI Act Service Desk — 2026.

The Omnibus amendments also reveal an important tension between regulatory ambition and industrial competitiveness. The Commission extended transition periods, broadened certain simplification measures for smaller enterprises and expanded opportunities for supervised experimentation, including an EU-level regulatory sandbox. These measures reflect the practical difficulty of imposing sophisticated technical and organisational requirements on companies that differ substantially in capital, expertise and access to compliance infrastructure. A regulation may improve safety and legal accountability while simultaneously raising fixed costs that smaller developers find more difficult to absorb. Delaying the application of selected requirements can reduce immediate compliance pressure, but it can also prolong the interval before mandatory safeguards apply to particular systems. This is not evidence that regulation and competitiveness are inherently incompatible. It demonstrates that their interaction depends on implementation cost, the availability of standards, supervisory capacity and the economic structure of the regulated market.

The EU AI Office occupies a significant position in this institutional architecture, particularly concerning general-purpose AI models and selected forms of EU-level supervision. National competent authorities retain important functions relating to market surveillance, enforcement and the application of the regulation within Member States. The result is a multi-level regulatory system rather than a single central authority exercising exclusive jurisdiction over all AI applications. That design can combine EU-wide consistency with national administrative capacity, but it also creates coordination demands concerning investigations, technical expertise, information-sharing and the interpretation of obligations across borders.

Institutional comparison: mandate, jurisdiction and enforcement capacity

Institution or instrument	Established function	Geographic or legal scope	Binding effect	Independent oversight capacity	Principal limitation
UN Independent International Scientific Panel on AI	Scientific assessment and policy-relevant research synthesis	Global scientific and diplomatic participation	Assessments are non-prescriptive	Independent scientific assessment within its mandate	No general compulsory inspection or sanctioning powers
UN Global Dialogue on AI Governance	Political discussion, coordination and participation	Open to all UN Member States and stakeholders	No direct regulatory obligations	No independent enforcement mandate	Consensus-building does not establish legal compliance
Council of Europe AI Framework Convention	Treaty framework for human rights, democracy and rule of law	Parties and activities within treaty scope	Treaty obligations depend on entry into force and party status	Cooperative treaty arrangements	Implementation depends heavily on parties’ legal systems
EU Artificial Intelligence Act	Risk-based AI regulation and market governance	Covered AI activities within the regulation’s jurisdiction	Directly applicable EU legal obligations	EU and national supervisory structures	Scope exclusions, phased implementation and administrative capacity
National AI regulators	Domestic supervision, investigation and enforcement	National jurisdiction and powers defined by applicable law	Binding where statutory authority exists	Varies by jurisdiction	Cross-border reach and technical resources
CCW Group of Governmental Experts	Negotiation of possible autonomous-weapons rules	Participating CCW states	Negotiating discussions do not themselves create new treaty prohibitions	Diplomatic scrutiny rather than a standing inspectorate	Consensus requirements and unresolved legal architecture
International technical standards bodies	Develop terminology, technical and organisational standards	International standard-setting participation	Generally voluntary unless incorporated into law or contracts	Technical evaluation frameworks	Standards do not independently confer enforcement authority

The underlying sources are UN General Assembly Resolution A/RES/79/325 — August 2025, Council of Europe Framework Convention on AI, CETS No. 225 — September 2024, Regulation (EU) 2024/1689, consolidated text — July 2026 and CCW Group of Governmental Experts, Chair’s Summary CCW/GGE.1/2026/WP.2 — United Nations Office for Disarmament Affairs — April 2026.

The comparison establishes that institutional influence cannot be measured through a single hierarchy of organisations. Scientific independence, democratic participation, treaty-making competence and coercive enforcement are different institutional capabilities. A global scientific body may provide information of high international value without possessing regulatory authority. A national regulator may enforce sophisticated requirements within its jurisdiction while having little power to influence a foreign military programme. A treaty institution may have broad legitimacy among participating states but lack technical access to the systems it seeks to supervise. A technical standards body may produce detailed assurance requirements that become influential through procurement and legislation even though its standards are not independently binding.

The enforcement problem: why institutional design must begin with jurisdiction

The establishment of genuinely effective international oversight requires governments to resolve questions that precede the design of technical inspection procedures. The first is the source of legal authority. An international organisation cannot generally impose regulatory duties on private entities or compel access to sovereign installations merely because its mandate concerns a subject of global importance. Such powers must arise from an applicable legal instrument or another recognised source of authority. The second question concerns the object of supervision. Governments might agree to monitor safety incidents, commercial models, high-risk deployments, prohibited weapons functions or particular categories of cross-border conduct. Each object creates different requirements for access, confidentiality and verification. The third concerns the consequences of noncompliance. Reporting irregularities to participating governments is institutionally different from suspending market access, ordering corrective measures, imposing penalties or establishing state responsibility.

The distinction between information-sharing and compulsory disclosure is particularly relevant. A company may voluntarily contribute evaluation data to an international scientific process without granting regulators access to internal security controls or proprietary development records. A government may publish its AI strategy while withholding military procurement details. An international body could therefore possess extensive published information but remain unable to determine whether particular high-impact systems comply with legally defined requirements. Expanding institutional powers would require agreement on access conditions, protection of sensitive information, investigative procedures and rights of review. Those arrangements are politically difficult because the information necessary for credible oversight may also contain commercial secrets, security vulnerabilities or classified operational capabilities.

An additional problem concerns the capacity to enforce findings. International monitoring can identify concerns, but implementation often depends on national authorities that may have different levels of technical expertise and political willingness. Cross-border cases can also require cooperation between regulators whose legal frameworks impose incompatible confidentiality rules or evidence standards. Where an activity is lawful in one jurisdiction but prohibited in another, information-sharing may improve factual understanding without resolving the underlying conflict of laws. Mutual recognition can reduce duplication only if participating authorities have sufficient confidence in each other’s assessment procedures. Otherwise, recognition may create opportunities for regulated entities to select the jurisdiction with the least demanding practical supervision.

The emerging institutional structure should consequently be understood as a network of partly overlapping jurisdictions rather than a prospective universal AI regulator. Its effectiveness depends on whether obligations can be carried from international principles into national law, from national law into organisational behaviour, and from organisational behaviour into evidence that competent institutions can independently examine.

Regulatory fragmentation and the political economy of compliance

Fragmentation is not simply a consequence of poor coordination. It also reflects material differences in national interests, economic structures and constitutional arrangements. States possessing substantial domestic AI industries may seek to avoid rules that impose disproportionately large costs on local developers or limit their international competitiveness. States primarily importing AI services may favour stronger supplier obligations, access to technical information and protection against discriminatory outcomes. Governments with extensive national-security applications may demand exclusions for sensitive activities. Countries with limited regulatory budgets may support international standards but lack the personnel, laboratories and legal infrastructure necessary to implement them.

These differences can create asymmetric compliance costs. A multinational company may maintain legal teams, technical evaluation facilities and documentation systems capable of serving several jurisdictions. Smaller developers may need to purchase external audits, adapt products separately for different markets and obtain specialised legal advice. Public administrations in less-resourced countries may be unable to assess complex general-purpose models without external assistance. The economic consequence is that formal regulatory equality can coexist with unequal implementation capacity. Requirements that apply uniformly may impose different effective burdens depending on access to capital, expertise and independent technical infrastructure.

The July 2026 EU amendments provide an identifiable example of institutional adaptation to these concerns. By extending selected transition periods and broadening measures for smaller and mid-cap enterprises, the EU acknowledged the importance of implementation conditions, not merely the existence of substantive safeguards. These modifications should not be interpreted as evidence that all regulatory burdens have been resolved. Their significance is that compliance cost and industrial capacity have become explicit elements of regulatory design. AI Omnibus Enters into Force — European Commission — July 2026.

A related concern is the use of regulatory influence by technologically powerful jurisdictions. Large markets can establish product requirements that international suppliers may adopt beyond the jurisdiction where those requirements are legally binding. This can produce beneficial convergence when standards improve safety and interoperability. It can also generate competitive advantages for firms already able to meet complex obligations, particularly where technical standards reflect infrastructure or organisational resources concentrated in advanced economies. The resulting distributional effects cannot be inferred from the formal content of regulations alone. They require evidence about implementation costs, market-entry conditions, conformity-assessment capacity and the availability of accessible alternatives.

The institutional conditions for independent oversight

A credible international oversight arrangement would need several distinct institutional capabilities. It would require an agreed legal mandate specifying the systems and conduct covered. It would need qualified technical personnel capable of assessing model behaviour, software architecture, operational safeguards and relevant information-security constraints. It would require procedures for obtaining evidence, maintaining its integrity and protecting confidential information. It would need mechanisms for coordination with national authorities and defined consequences when regulated actors fail to cooperate. Where the institution is expected to address individual harm, it would additionally require access to competent adjudicatory procedures or arrangements for referring claims to authorities able to provide remedies.

These capabilities are not interchangeable. Scientific expertise cannot compensate for the absence of legal jurisdiction. Legal jurisdiction cannot ensure reliable findings without technical competence. Technical competence may be ineffective without access to the system or its operational records. Inspection authority may have limited deterrent value where findings cannot be acted upon. Strong enforcement powers can also create legitimacy concerns if the institution lacks procedural safeguards and clear limits on discretionary action. Effective governance therefore requires a balance of authority, competence, independence and accountability.

Institutional function	Minimum operational capability	Current institutional analogue	Missing or uneven element	Strategic consequence
Scientific assessment	Independent expertise, research synthesis and conflict-of-interest safeguards	UN Independent Scientific Panel	Limited compulsory access to non-public technical information	Knowledge can improve without corresponding compliance
International negotiation	Inclusive participation and procedures for negotiating commitments	UN Global Dialogue and CCW discussions	No guaranteed agreement on binding rules	Diplomatic convergence may remain nonbinding
Technical standardisation	Defined methods for testing, measurement and documentation	International and regional standards organisations	Inconsistent adoption and unequal technical capacity	Different jurisdictions may interpret conformity differently
Regulatory supervision	Statutory authority, qualified inspectors and evidence access	EU AI Office and national regulators	Jurisdictional boundaries and uneven resources	Comparable risks may receive different levels of scrutiny
Military compliance assessment	Agreed definitions, access rules and verification procedures	National weapons reviews and arms-control negotiations	No general global autonomous-weapons inspectorate	Military activities remain difficult to verify externally
Investigation and remedies	Procedural jurisdiction, preserved evidence and enforceable decisions	National courts and administrative authorities	Cross-border access, secrecy and inconsistent liability rules	Legal responsibility may be difficult to establish in practice
Capacity-building	Sustainable funding, training and independent infrastructure	UN and regional development initiatives	Unequal resources and technological dependence	Formal participation may exceed actual governance capacity

The table identifies a practical division of institutional labour rather than advocating a single centralised structure. Some functions could be strengthened through existing institutions, including interoperable reporting, recognised testing methods and support for national regulators. Others would require new agreements, especially where external inspection of sensitive military capabilities is contemplated. The political feasibility of these arrangements depends partly on whether states can obtain reliable oversight without surrendering information or decision-making authority they consider indispensable to national security.

Chapter 4 — Key judgments

High confidence: The UN scientific and dialogue mechanisms materially expand international participation and shared assessment, but they do not possess the general regulatory, inspection or enforcement powers required for compulsory global AI governance.

High confidence: The Council of Europe Convention and EU AI Act provide more specific legal pathways, but their effectiveness depends on legal scope, participation, national implementation and the powers assigned to competent authorities.

High confidence: The 27 July 2026 AI Omnibus amendments changed the EU implementation timetable, with major high-risk obligations scheduled for December 2027 and August 2028 according to category. Analyses relying exclusively on the original 2024 timetable would misstate the regulatory baseline.

Medium-high confidence: Regulatory fragmentation will continue to affect technical assurance, cross-border evidence access and corporate compliance costs unless common standards are supported by compatible supervisory procedures and adequate national capacity.

Medium confidence: Strengthening existing institutions through targeted information-sharing, technical assistance and recognised conformity procedures may encounter fewer sovereignty barriers than establishing an international authority with broad compulsory inspection powers. This is a judgment about the institutional burdens associated with each model, not a claim that all governments prefer the same solution.

What would change the assessment: New treaties conferring compulsory oversight authority, enforceable cross-border evidence-sharing arrangements, independently funded international inspection capabilities or substantial harmonisation of national enforcement procedures would alter the balance between coordination and effective supervision.

Open official record: The most consequential evidence concerns implementation budgets, staffing and technical capacity of national regulators; the practical operation of cross-border enforcement arrangements; ratification and implementation of the Council of Europe Convention; and any formally agreed international mandate extending beyond scientific assessment and political dialogue.

Chapter 5 — Computing Infrastructure, Export Controls, Corporate Incentives and Equitable Access

Principal judgment — High confidence: The global distribution of AI capabilities is determined increasingly by control over a connected industrial system comprising advanced semiconductors, high-performance computing, specialised manufacturing equipment, electricity infrastructure, cloud platforms, proprietary software, technical expertise and capital. The strategic significance of this system extends beyond ownership of individual models. States and corporations controlling its critical inputs can influence which countries develop independent AI capabilities, which organisations obtain access to advanced computational resources and which applications remain commercially or technically feasible. Ethical commitments to equitable access cannot overcome these dependencies unless translated into financing, infrastructure provision, licensing conditions, technical assistance and sustained local capability development. Export controls can serve legitimate national-security objectives, but their implementation also affects competitive relationships and may reinforce structural dependencies. The central distributive question is therefore not simply whether AI services become available to more users, but whether countries and organisations outside the leading technology centres obtain sufficient control over the resources needed to develop, evaluate, adapt and govern AI independently.

The industrial structure of AI power: control of infrastructure precedes control of applications

AI competition is frequently described in terms of model performance, algorithmic innovation and commercial adoption. Those variables matter, but they represent only part of the underlying industrial structure. Advanced model development requires access to computing hardware, interconnected servers, suitable facilities, reliable energy, specialised software and personnel capable of managing training and deployment infrastructure. The production and operation of those resources depend on industrial systems extending from semiconductor design and fabrication to data-centre construction, power generation, telecommunications and cybersecurity. A country may possess strong academic research capabilities while lacking affordable access to large-scale computing. Another may host substantial data-centre capacity without controlling the models or underlying intellectual property. A third may have promising AI companies but depend heavily on foreign cloud infrastructure and accelerator suppliers. These configurations produce different forms of technological dependence and different degrees of political autonomy.

The economic consequences emerge through the relationship between fixed capital requirements and access to scale. Large computing installations require considerable upfront investment, access to advanced components and the ability to finance electricity, cooling, networking and maintenance. Model development additionally requires specialised engineering personnel and organisational capacity to conduct evaluations, manage data and maintain reliable deployment environments. Companies controlling several elements of this stack can reduce integration costs and capture economic value across infrastructure, platforms and applications. Those advantages do not necessarily imply anticompetitive conduct. They do, however, create entry barriers that cannot be removed solely through the publication of ethical principles or the expansion of consumer access to AI services.

A further distinction is necessary between the ability to use AI and the ability to control its development. Commercial APIs and hosted services can make sophisticated models available to organisations that do not possess their own training infrastructure. This can accelerate adoption, particularly for smaller businesses and public administrations. Yet access may remain conditional on pricing, licensing, technical restrictions, network availability and the continuing willingness of an external provider to supply the service. For high-dependence applications, these conditions can become strategic rather than merely commercial. Governments need to consider whether critical functions can continue if external access is restricted, prices rise sharply, model behaviour changes or foreign legal requirements affect the provider.

The European Commission’s AI Continent Action Plan explicitly connects competitiveness, productivity and technological sovereignty. Its strategy combines computing infrastructure, access to data, model development, skills and adoption. This is an acknowledgment that regulatory capacity alone cannot establish technological independence. A jurisdiction may possess sophisticated AI legislation while depending on external providers for advanced accelerators, large-scale training facilities and proprietary AI systems. Closing that gap requires physical infrastructure and sustained productive capability, not simply additional compliance requirements. AI Continent Action Plan — European Commission — 2025.

The electricity constraint: AI competitiveness depends on physical infrastructure

The energy dimension illustrates why AI development cannot be evaluated as an exclusively digital industry. Data centres require continuous electricity supplies, transmission capacity, cooling equipment and reliable connections to local power networks. Their expansion can be constrained by electricity prices, grid-connection delays, available generating capacity and the procurement of essential electrical equipment. These conditions affect both the economics of AI investment and the geographic distribution of computing infrastructure. Regions able to offer affordable, reliable electricity and suitable industrial sites may attract investment, while regions facing grid congestion or uncertain connection schedules may encounter delays irrespective of their research capabilities.

The International Energy Agency’s 2026 assessment provides an important quantitative baseline. It estimates global data-centre electricity consumption at approximately 485 terawatt-hours in 2025, with a central projection of approximately 950 TWh in 2030. These figures concern all data centres rather than AI systems alone, although the IEA identifies AI-focused facilities as a major driver of growth. Its updated assessment also identifies physical bottlenecks in electricity equipment, advanced chips and data-centre development, limiting the pace at which announced projects can become operational. The central projection is therefore not a committed construction schedule or guaranteed future outcome. It is an analytical scenario subject to technological efficiency, investment, electricity infrastructure and demand conditions. Key Questions on Energy and AI: Executive Summary — International Energy Agency — 2026.

The scale of the projection is economically significant. The difference between the IEA’s 2025 estimate and its 2030 central projection is approximately 465 TWh annually. That difference represents additional electricity consumption, not the amount of generating capacity that would need to be constructed. Converting annual energy demand into installed capacity would require assumptions about generation technology, operating hours, capacity factors, network losses and the composition of the electricity supply. Such a conversion cannot be made responsibly without specifying those variables. The relevant policy conclusion is that AI infrastructure development must be coordinated with energy planning, including the availability of generation, transmission, substations, cooling systems and the financial arrangements required to deliver them.

The geographic consequences may become substantial even where the global share of electricity consumption remains relatively limited. A large data centre can impose a concentrated load on a regional network whose expansion requires years of planning and construction. By contrast, the development of computing hardware and software may proceed on shorter commercial cycles. This mismatch between technological investment and electricity infrastructure creates a potential scheduling constraint that governments cannot eliminate through faster model development. Public authorities that announce large-scale AI infrastructure programmes without confirming power availability risk treating nominal investment commitments as future operational capacity.

The IEA further reports that the financing of data-centre development is becoming increasingly dependent on broader capital markets rather than solely on technology companies’ internal resources. This creates exposure to interest rates, expected returns, financing conditions and investor assessments of future AI demand. A project can possess technical feasibility but remain financially vulnerable if the cost of capital rises or utilisation falls below expectations. Similarly, substantial announced investment can coexist with delayed construction, revised project specifications or changing commercial arrangements. These distinctions are crucial when comparing national strategies: announced financing, committed financing, equipment procurement, installed hardware and realised computing output are separate stages of industrial development.

The global infrastructure baseline: documented figures and interpretation

Indicator	Official value or status	Reference period	What the figure measures	What it does not establish
Global data-centre electricity consumption	Approximately 485 TWh	2025	Estimated annual electricity use across data centres	AI-only electricity consumption
Global data-centre electricity demand	Approximately 950 TWh	2030 central projection	Expected consumption under the IEA’s updated central outlook	Guaranteed future demand or committed generating capacity
Projected increase from 2025 to 2030	Approximately 465 TWh per year, calculated from the two IEA figures	2025–2030	Difference between estimated and projected annual consumption	A measured change already realised
World population using the Internet	Approximately 74%	2025	Estimated basic Internet participation	Access to advanced AI computing or model-development capability
Population remaining offline	Approximately 2.2 billion people	2025	Estimated population without Internet use	Number of people unable to use all AI services
Internet use in high-income countries	94%	2025	Share of individuals using the Internet	Independent advanced AI production capacity
Internet use in low-income countries	23%	2025	Share of individuals using the Internet	Equivalent measure of AI adoption or skills
EU EuroHPC and AI Factory investment	€10 billion over 2021–2027	Programme envelope	Overall planned investment by the Commission, Member States and associated countries	Fully deployed or utilised AI computing capacity
EU InvestAI AI gigafactory initiative	€20 billion mobilisation objective	Announced programme	Proposed financing mobilisation for advanced AI infrastructure	Completed investment or operational facilities
UK AI Research Resource	21 to 420 AI ExaFLOPS	2025 baseline and 2030 target	Government-stated objective for public AI computing capacity	A verified 2030 installation or comparability with every other FLOPS metric

Sources: Key Questions on Energy and AI — International Energy Agency — 2026; Measuring Digital Development: Facts and Figures 2025 — International Telecommunication Union — November 2025; AI Factories — European Commission; AI Continent Action Plan — European Commission; UK Compute Roadmap — United Kingdom Department for Science, Innovation and Technology — July 2025.

The figures require several methodological qualifications. The IEA’s 2026 central projection supersedes the earlier 2025 estimate of approximately 945 TWh for 2030, but the difference should not be interpreted as a meaningful revision in realised consumption. Both are projections generated from different information vintages. Internet-use statistics measure connectivity rather than AI capability. The EuroHPC and InvestAI figures concern different funding mechanisms and should not be added without examining whether they overlap. The UK computing target uses an officially defined AI ExaFLOPS measure; it cannot be compared directly with the aggregate peak performance of unrelated systems without reconciling precision, workloads and measurement conventions.

These distinctions prevent a common analytical error: treating announced investment, computing performance, installed infrastructure and user access as interchangeable evidence of technological power.

Export controls: national-security restrictions as instruments of industrial influence

Export controls affect AI competition because access to certain advanced computing components and related technologies can be restricted through national licensing regimes. These measures may seek to prevent the transfer of sensitive capabilities, limit military applications or protect national-security interests. Their economic consequences extend beyond the immediate transaction. Licensing requirements can influence where data centres are constructed, which suppliers companies select, how customers structure procurement, and whether foreign governments pursue alternative industrial capabilities. They can also create incentives for companies to develop products that fall within permitted technical thresholds or to seek alternative sources of computing capacity.

The United States provides a concrete example of how export-control policy can change without abandoning security-based licensing. On 13 January 2026, the Department of Commerce’s Bureau of Industry and Security announced revised treatment of specified advanced semiconductor exports to China. The policy provides for case-by-case licence review involving products such as the NVIDIA H200 and AMD MI325X, subject to defined security conditions. The conditions described by BIS include requirements relating to the availability of semiconductor production capacity for US customers, compliance procedures undertaken by Chinese purchasers and independent testing in the United States concerning product performance and security. This is evidence of a conditional licensing regime, not unrestricted authorisation of the products concerned. It also demonstrates that export policy cannot be described accurately through a simple division between complete prohibition and unrestricted trade. Department of Commerce Revises License Review Policy for Semiconductors Exported to China — US Bureau of Industry and Security — January 2026.

The strategic mechanism is straightforward. A state controlling important technological inputs can condition access on requirements that reflect its security objectives. Companies wishing to serve restricted markets must then consider whether licences are available, what compliance procedures are required and how regulatory uncertainty affects commercial planning. Recipient countries must assess the reliability of supply, the cost of alternative technologies and the desirability of investing in domestic capabilities. Such controls can create substantial leverage when substitute products are unavailable or expensive. That leverage can diminish as alternatives emerge, suppliers relocate activities or technical innovation changes the importance of the restricted component.

However, the effectiveness of export controls cannot be established solely by observing that restrictions have been imposed. A complete assessment requires evidence of licensing decisions, actual deliveries, substitution, domestic production, diversion risks and the performance of alternative technologies. It must also distinguish between preventing access to specific high-performance components and preventing all meaningful AI development. Countries with limited access to the most advanced accelerators may still develop specialised models, use existing infrastructure or pursue applications requiring less computation. Conversely, possession of advanced chips does not automatically establish the technical, industrial and organisational capability to develop globally competitive models.

Export controls therefore create an unavoidable tension with universalist concepts of technological benefit. From a national-security perspective, restrictions may be justified by concerns over military or strategic use. From a distributional perspective, those same restrictions can affect scientific access, commercial development and the ability of governments to build independent capabilities. The two objectives cannot be reconciled through an unqualified demand for unrestricted technology transfer or an assumption that every restriction is protective. The practical question is whether the scope, conditions and consequences of restrictions correspond to identifiable risks, and whether alternative arrangements can preserve legitimate scientific and developmental uses.

Intellectual property, cloud dependencies and contractual control

Control over AI capabilities extends beyond physical hardware. Software licences, proprietary model weights, technical interfaces, training methods and contractual restrictions can determine how organisations use and modify AI systems. Intellectual-property law protects particular forms of innovation and investment, while contracts can establish additional conditions relating to access, redistribution, deployment and commercial use. The resulting legal framework can encourage investment but can also limit the ability of customers to evaluate, adapt or transfer critical systems. The strategic relevance depends on whether those limitations affect ordinary commercial substitution or create dependencies in functions essential to public administration, national infrastructure or security.

One form of dependency emerges when a customer can access a model only through a hosted interface. Such access may reduce capital requirements and accelerate deployment, but it can limit independent testing, modification and continuity planning. A government that relies on an external provider for a critical administrative function may need contractual guarantees concerning service availability, data handling, model changes and access to evidence. Those guarantees are especially important when the provider retains discretion to update or withdraw the model. A public authority can remain legally responsible for consequential decisions even where essential technical elements are controlled by a commercial supplier.

Another form concerns the distinction between open-source software, open data and open model weights. These terms do not automatically describe equivalent permissions or operational freedoms. A model whose weights are available may remain subject to licence restrictions, dependencies on proprietary components or limitations arising from access to training data and evaluation infrastructure. Open publication can improve scientific inspection and local adaptation, but it may also raise security concerns concerning the dissemination of particular capabilities. The institutional task is to distinguish the benefits of openness for competition and research from situations in which unrestricted distribution creates material risks.

A distributional policy should therefore avoid treating open access as a binary condition. The relevant variables include whether users can examine system behaviour, modify the model, deploy it independently, transfer it between infrastructure providers and obtain the computing resources needed to operate it. These capabilities determine whether access produces durable local competence or continuing dependence on external suppliers.

Corporate incentives and the concentration of market power

Corporate investment decisions influence the distribution of AI capacity because private firms control much of the infrastructure, software and expertise necessary for advanced development. Companies may pursue vertical integration to reduce costs, improve performance, protect intellectual property and secure reliable supply. They may enter long-term relationships with cloud providers, accelerator manufacturers, energy suppliers and governments. These arrangements can create efficiencies and support investment at scales that would otherwise be difficult to finance. They can also increase switching costs and give established providers leverage over customers requiring access to specialised infrastructure.

A critical distinction is required between concentration and unlawful exclusionary conduct. A market may be highly concentrated because of substantial fixed costs, technological advantages or economies of scale without every dominant firm’s conduct constituting an infringement of competition law. Conversely, contractual restrictions, discriminatory access conditions or practices that make it unnecessarily difficult for customers to switch suppliers may raise competition concerns under applicable legal rules. The economic assessment must therefore examine relevant markets, barriers to entry, actual conduct and the availability of substitutes rather than inferring illegality from scale alone.

For governments, the principal strategic issue is the relationship between public dependency and commercial bargaining power. A public administration may procure AI services to improve efficiency while remaining dependent on the provider for updates, security support and technical explanations. A defence establishment may rely on commercial technology while requiring operational resilience and protection of classified information. A research institution may obtain subsidised computing access but face restrictions on allocation or continued use. Each arrangement creates a different balance between the benefits of specialisation and the costs of dependency.

The distributive consequences are particularly important for smaller companies. Access to advanced infrastructure through shared computing facilities or public research programmes can lower entry barriers. Nevertheless, nominal eligibility does not ensure effective access if application procedures, technical requirements or allocation mechanisms favour organisations already possessing substantial resources. Evaluating whether public infrastructure genuinely broadens participation therefore requires evidence about actual allocations, user categories, utilisation, technical assistance and the results achieved by beneficiaries.

The global development divide: connectivity is necessary but insufficient

The International Telecommunication Union’s 2025 estimates provide a measurable starting point for assessing global digital inequality. Approximately 74% of the world’s population, or about 6 billion people, used the Internet in 2025, leaving approximately 2.2 billion people offline. Internet use reached 94% in high-income countries, compared with 23% in low-income countries. The ITU also reported that approximately 96% of people remaining offline lived in low- and middle-income countries. These are substantial disparities in basic connectivity, which remains an important enabling condition for accessing many AI services. Measuring Digital Development: Facts and Figures 2025 — International Telecommunication Union — November 2025.

The implications extend beyond the immediate absence of Internet access. Countries with lower connectivity may face disadvantages in digital education, online commercial services, research collaboration and access to remote computing infrastructure. They may also confront higher relative costs for reliable broadband, cloud services and technical equipment. However, the figures do not directly measure AI adoption, compute availability or the sophistication of national AI research. A connected population may still lack the infrastructure necessary for independent model development. Conversely, a country with relatively low overall connectivity may possess selected research institutions or industrial facilities with substantial computing capabilities. Connectivity should therefore be treated as one part of a broader capability assessment.

The relevant developmental question is how countries can progress from consuming AI services to exercising independent technological and institutional choices. This requires the ability to evaluate systems, train specialists, adapt models to local languages and needs, participate in international standards and enforce domestic legal requirements. A country that imports AI services without acquiring such capabilities may experience productivity benefits while remaining dependent on external decisions about pricing, model availability and technological development. More equitable distribution consequently requires investment in institutional capacity as well as commercial access.

Distributional policy mechanisms: implementation conditions and limitations

Policy mechanism	Intended economic effect	Implementation requirement	Main limitation	Observable indicator
Public computing allocations	Reduce barriers for researchers and smaller developers	Funded infrastructure, transparent eligibility and technical support	Scarcity of advanced computing and unequal ability to use it	Actual allocated computing hours and beneficiary categories
Shared AI research facilities	Improve independent testing and development	Equipment, staff, data access and sustained maintenance	Capital and operating costs	Operational capacity and research utilisation
Concessional financing	Enable investment where commercial financing is insufficient	Eligible projects, financial controls and performance conditions	Fiscal exposure and project viability	Disbursements, completed facilities and economic results
Targeted technology transfer	Develop domestic productive capabilities	Legally permitted access, training and technical integration	Export controls, licensing and absorption capacity	Verified operational capability transferred
Open and interoperable technologies	Reduce dependence on particular vendors	Accessible technical standards, usable licences and compatible infrastructure	Security risks and insufficient local expertise	Portability, independent deployment and switching outcomes
Differential regulatory assistance	Reduce compliance disadvantages among countries and smaller firms	Training, testing infrastructure and institutional support	Uneven supervisory capacity	Successful conformity assessments and enforcement capability
Public procurement conditions	Improve accountability, continuity and access	Enforceable contracts and technical evaluation	Supplier concentration and contract-management capacity	Audit rights exercised and supplier performance
Local-language development	Extend benefits to underserved populations	Data resources, linguistic expertise and evaluation methods	Limited datasets and commercial incentives	Validated local-language applications
Energy and connectivity investment	Expand the physical basis for digital participation	Electricity supply, networks and sustainable financing	Construction delays, energy prices and geographic inequality	Reliable service availability and completed infrastructure

These mechanisms should not be evaluated as universally beneficial in isolation. Public computing without sufficient expertise may result in low utilisation. Technology transfer without a sustainable industrial ecosystem may create dependence on imported maintenance and replacement components. Open-source publication without appropriate security evaluation may create new risks. Public procurement requirements that demand unrealistic supplier guarantees may reduce competition. The relevant measure of distributive success is therefore the creation of durable capabilities that participating institutions can operate, evaluate and maintain, rather than the announcement of programmes or the number of formal beneficiaries.

Chapter 5 — Key judgments

High confidence: Advanced AI capability depends on a capital-intensive industrial system extending beyond model development into semiconductors, computing infrastructure, electricity, software, financing and technical expertise.

High confidence: The IEA’s updated central outlook places worldwide data-centre electricity consumption at approximately 950 TWh in 2030, compared with approximately 485 TWh in 2025. The figures concern all data centres and should not be misrepresented as AI-only demand.

High confidence: Export controls and technology licensing give states controlling key inputs practical influence over access to advanced AI capabilities. The January 2026 US semiconductor licensing revision demonstrates that such controls may be conditional and product-specific rather than absolute.

High confidence: The global digital divide remains significant, with ITU estimates showing 2.2 billion people offline in 2025 and a substantial gap between Internet use in high-income and low-income countries.

Medium-high confidence: Policies designed to distribute AI benefits will be more effective when they combine access to infrastructure with technical assistance, financing and local capacity-building. Access to commercial applications alone does not establish technological independence.

Medium confidence: Corporate concentration, contractual dependencies and regulatory costs may reinforce the advantages of established firms, although the net competitive effects depend on market structure, specific conduct and available alternatives.

What would change the assessment: Sustained evidence of lower capital requirements for advanced model development, substantial expansion of affordable independent computing access, reduced dependence on restricted semiconductor inputs, or demonstrable narrowing of capability gaps would weaken the conclusion that industrial concentration remains a central determinant of AI power.

Open official record: The decisive evidence concerns realised rather than announced computing investment, actual allocation of publicly supported infrastructure, model-development and deployment costs, export-licence outcomes, corporate switching conditions and consistent international indicators of AI research and evaluation capacity.

Chapter 6 — European and Transatlantic Implications: Italy, France, Germany, the United Kingdom and the European Union

Principal judgment — High confidence: Europe possesses substantial scientific, industrial, regulatory and institutional capabilities relevant to artificial intelligence, but these assets are distributed unevenly across national economies and do not yet constitute a unified system of technological sovereignty. The European Union has established a legally significant regulatory framework and is expanding publicly supported computing infrastructure, while individual governments pursue different strategies reflecting their industrial structures, electricity systems, fiscal resources, military requirements and relationships with non-European technology suppliers. Italy’s priorities centre on institutional implementation, industrial adoption and the expansion of public computing capability. France is pursuing large-scale infrastructure investment and the development of nationally anchored AI capabilities. Germany is integrating AI into a broader strategy for industrial competitiveness and advanced manufacturing. The United Kingdom is developing a substantial national computing programme outside the EU regulatory framework, while maintaining its own approach to AI governance and defence applications. These strategies create possibilities for cooperation, but they also generate competition for investment, computing resources, skilled personnel and influence over regulatory standards. The principal challenge through 2030 is to distinguish the existence of European technological assets from the ability to control the complete industrial chain on which advanced AI development depends.

The European strategic problem: regulatory influence and productive capacity are not equivalent

The European approach to AI governance combines two ambitions that are related but institutionally different. One is the establishment of rules governing the lawful and responsible development and deployment of AI systems. The other is the construction of sufficient technological capacity to ensure that European governments, companies and research institutions can participate competitively in advanced AI development. The first ambition can be pursued through legislation, supervision and legal remedies. The second requires investment, industrial capacity, reliable electricity, advanced computing, technical personnel and access to critical technological inputs. Success in one does not automatically produce success in the other. A jurisdiction can exercise considerable influence over the regulatory conditions under which AI services enter its market while remaining dependent on external firms for the infrastructure and models that provide those services.

The European Commission’s AI Continent Action Plan attempts to address this distinction by combining regulatory governance with investment in AI Factories, AI Gigafactories, data resources, skills and adoption. The Commission’s current description identifies 19 AI Factories and 13 associated antennas being established, with at least nine new AI-optimised supercomputers planned for procurement and deployment. The network is intended to provide computing resources and supporting services to researchers, public institutions, startups, small and medium-sized enterprises and industrial users. The distinction between being selected, being established and being operational remains important. Announced facilities and procurement programmes are evidence of infrastructure policy, not necessarily evidence that all planned computing resources are already available to users. AI Factories — European Commission — 2026.

The EU’s more ambitious Gigafactory programme introduces another layer of industrial strategy. The Commission describes an initiative involving up to seven AI Gigafactories, supported by potential EU and national financing and intended to mobilise substantial private investment. These facilities are designed to support more demanding AI development than the existing research-oriented infrastructure alone. Their strategic significance lies in the attempt to overcome the capital and scale disadvantages facing European model developers. However, projects at this scale depend on commercial commitments, procurement, physical construction, electricity supply and long-term utilisation. They cannot be evaluated through nominal financing targets alone. The same requirement applies to the European Chips Act, which seeks to strengthen semiconductor production, research and supply-chain resilience. AI Factories — European Commission — 2026; European Chips Act — European Commission — 2026.

The strategic relationship with the United States is correspondingly complex. European organisations rely on technologies developed within global supply chains, including computing accelerators, software platforms and cloud infrastructure associated with US companies. At the same time, the United States and European countries participate in overlapping research, commercial and security relationships. Their interests may converge on reliable technology, resilience and appropriate safeguards while diverging over regulatory burdens, export licensing, market access and the location of high-value investment. European technological sovereignty should therefore not be interpreted as complete technological separation. A more operationally meaningful definition concerns the capacity to make and implement critical choices without unacceptable dependence on decisions taken elsewhere.

Italy: national AI governance, industrial adoption and computing infrastructure

Italy’s strategic position combines an established industrial base, substantial public research capabilities and a regulatory framework developed to complement the European AI Act. The government published its Italian Strategy for Artificial Intelligence 2024–2026 on 22 July 2024. The strategy was prepared by a committee of 14 experts, supported by the Agency for Digital Italy, and addresses the development and responsible adoption of AI across research, public administration, enterprises and related institutional functions. The document includes an emphasis on monitoring implementation, an aspect that is particularly important because national strategies frequently describe objectives without establishing sufficiently detailed mechanisms for determining whether those objectives have been achieved. Italian Strategy for Artificial Intelligence 2024–2026 — Department for Digital Transformation, Presidency of the Council of Ministers — July 2024.

A subsequent institutional development occurred in September 2025, when the Italian Parliament completed the approval of national AI legislation. The government’s official account identifies the National Cybersecurity Agency, ACN, and the Agency for Digital Italy, AgID, as competent national authorities with differentiated responsibilities. ACN is assigned functions relating to the adequacy and security of AI systems, including inspection powers within its competence, while AgID has responsibilities concerning notifications and the promotion of secure applications. The legislation also establishes a framework for updating the national AI strategy every two years and providing annual monitoring information to Parliament. These provisions are significant because they connect national governance to identifiable administrative bodies and recurring institutional procedures rather than relying solely on general policy objectives. Final Approval of Italian Artificial Intelligence Legislation — Department for Digital Transformation, Presidency of the Council of Ministers — September 2025.

The official announcement also describes a €1 billion programme intended to support startups and small and medium-sized enterprises operating in AI, cybersecurity and related emerging technologies. The distinction between a legislative investment programme and realised expenditure is essential. The announced amount should not be presented as proof that the entire sum has been disbursed or that it has generated equivalent productive capacity. Its economic significance will depend on the financing mechanisms, the projects selected, the availability of private co-investment and the degree to which funded activities strengthen domestic capabilities rather than merely finance the acquisition of externally supplied technologies.

Italy’s national governance structure also illustrates the complexity of implementing European requirements through existing domestic authorities. Under Article 77 of the AI Act, authorities responsible for fundamental-rights protection can obtain relevant documentation and request technical examination where necessary within the applicable legal framework. The Italian government’s implementation record identifies competent bodies including the Data Protection Authority, the Competition and Market Authority, the Communications Authority and authorities concerned with the rights of persons deprived of liberty and persons with disabilities. This arrangement distributes oversight across institutions possessing different statutory mandates. The resulting advantage is access to existing sector-specific expertise; the challenge is maintaining effective coordination when an AI system raises simultaneous issues of privacy, discrimination, competition, public administration and technological safety. Italian Authorities for the Protection of Fundamental Rights under Article 77 of the AI Act — Department for Digital Transformation.

The infrastructure dimension is represented by IT4LIA, the Italian AI Factory located within the computing ecosystem associated with Bologna’s Tecnopolo and operated by CINECA. On 22 April 2026, the European High Performance Computing Joint Undertaking announced a procurement contract with E4 Computer Engineering and Dell Technologies for a new AI-optimised supercomputer. The planned infrastructure incorporates NVIDIA GB200 NVL4 technology and is intended to complement existing EuroHPC capabilities. The importance of this arrangement lies in its combination of public computing infrastructure, technical support and industrial access, particularly for smaller organisations that cannot independently finance large computing facilities. The contract confirms a procurement milestone; it should not be described as proof that every planned service or unit of future computing capacity is fully deployed. EuroHPC JU Signs Contract to Boost AI Capabilities with IT4LIA AI Factory — European High Performance Computing Joint Undertaking — April 2026.

The national strategic consequence is that Italy possesses instruments to combine regulatory supervision, public computing access and industrial adoption. However, the existence of these instruments does not establish that they function as an integrated national industrial policy. The critical issue is whether companies can move from experimental adoption toward sustained productivity improvements and whether public investment strengthens independent technical competence. For manufacturing enterprises, AI can support process optimisation, quality control, predictive maintenance, logistics and product development. The economic value depends on reliable industrial data, integration with existing systems, workforce skills and the capacity to maintain deployed solutions. Public computing facilities may assist research and model development, but industrial productivity also requires organisational investment at enterprise level.

Italy faces a particular policy challenge in translating research capability into applications accessible to smaller industrial firms. These companies may benefit from specialised AI systems without needing to develop frontier general-purpose models. Their requirements include affordable integration, cybersecurity, trustworthy data management, measurable performance and support from technically competent suppliers. A policy focused exclusively on large-model development could therefore overlook a substantial part of the potential economic value available through industrial applications. Conversely, a strategy limited to purchasing imported AI services could increase adoption while doing relatively little to strengthen domestic technological independence. The relevant policy balance must be evaluated through realised investment, industrial uptake and the availability of national technical capabilities.

France: infrastructure investment, energy advantage and national technological capacity

France has placed substantial emphasis on attracting investment in AI infrastructure and strengthening domestic technological capabilities. At the Paris AI Action Summit in February 2025, the French government announced €109 billion in prospective AI-related investment, including infrastructure and deployment projects. The figure is an announcement concerning future private investment and should not be treated as completed construction or realised capital expenditure. The Élysée identified several factors supporting France’s industrial strategy, particularly access to relatively low-carbon electricity, the availability of suitable data-centre locations, transmission infrastructure and measures intended to facilitate investment. Making France an AI Power — Presidency of the French Republic — February 2025.

The infrastructure strategy reflects the importance of electricity in determining the geography of AI investment. France’s nuclear generation system and other electricity resources can provide an industrial advantage where large, continuous power supplies are commercially attractive. However, national electricity generation is not equivalent to immediately available grid capacity at every proposed data-centre site. Investment remains dependent on local connection conditions, transmission infrastructure, equipment procurement, planning procedures and commercial demand. The relevant measure is therefore not the aggregate national electricity supply alone but the ability to deliver reliable power to specific facilities under financially sustainable conditions.

The French government has also linked technological sovereignty to cybersecurity, semiconductor supply chains, sensitive data and trusted cloud services. In its account of the June 2025 Council of Ministers meeting, the government emphasised the need to strengthen control over infrastructure and data, including the role of ANSSI’s SecNumCloud qualification framework in establishing requirements for trusted cloud offerings serving sensitive public and commercial uses. This approach treats sovereignty as a combination of industrial capability and the capacity to impose security requirements on essential service providers. It also demonstrates that sovereignty does not necessarily require every technological component to originate domestically; rather, public authorities seek to establish conditions under which dependence on external suppliers remains manageable. Council of Ministers Meeting of 12 June 2025 — Government of France — June 2025.

France’s strategy nevertheless faces the same distinction between announced investment and completed productive capacity that applies elsewhere. Data-centre projects must secure power connections, financing, land, specialised equipment and long-term customers. Domestic model developers require access to computing infrastructure at costs compatible with global competition. Research capabilities must translate into commercial products and industrial applications. The availability of low-carbon electricity can improve investment conditions, but it does not eliminate dependencies relating to advanced semiconductors, specialised software or global capital markets.

The strategic consequence is that France is pursuing a combination of infrastructure attraction, national technological capacity and regulatory influence within the EU. Its policy creates opportunities for European cooperation but can also generate competition with other Member States for the location of major facilities and associated investment. The degree to which French projects strengthen broader European sovereignty depends on whether access to computing capacity, research infrastructure and technological benefits extends beyond the immediate beneficiaries of nationally located investment.

Germany: industrial AI, technological sovereignty and the manufacturing economy

Germany’s AI strategy differs from the French emphasis on large-scale infrastructure announcements because its official industrial programme places particular weight on translating technological development into manufacturing productivity and broader economic value. The federal government’s Hightech Agenda Deutschland, adopted in July 2025, identifies artificial intelligence among six strategic key technologies, alongside quantum technologies, microelectronics, biotechnology, fusion and climate-neutral energy technologies, and sustainable mobility technologies. The agenda is explicitly linked to economic competitiveness, domestic value creation and technological sovereignty. Hightech Agenda Deutschland — Federal Government of Germany — July 2025.

The AI roadmap establishes several measurable objectives through 2030. It sets an ambition for 10% of German economic output to be AI-based by 2030, seeks to increase the number of AI startups to four times the 2025 level, and identifies the goal of raising AI adoption among manufacturing SMEs to more than 50% by 2029. The government also identifies industrial AI, robotics, public-sector procurement and the development of domestic model capabilities as strategic priorities. These are policy objectives, not measured achievements, and their interpretation depends on the definitions used in the roadmap. In particular, the objective that 10% of economic output should be AI-based cannot be treated as a forecast of a 10% increase in GDP attributable to AI. It describes a targeted classification of economic activity, not an independently estimated causal growth effect. Hightech Agenda Deutschland: Artificial Intelligence Roadmap — Federal Government of Germany — 2026.

This distinction matters because industrial AI adoption can take many forms. A manufacturing company may use machine learning for quality control, production scheduling or predictive maintenance without replacing its core production process. Another may develop AI-enabled industrial products, embedded systems or robotic capabilities that become part of its commercial offering. The economic contribution of these activities cannot be measured simply by counting firms that report using AI. Assessing productivity effects requires evidence concerning production costs, output quality, labour productivity, capital expenditure and the persistence of gains after implementation. The German roadmap’s adoption targets therefore provide useful monitoring indicators, but they do not establish the magnitude of future economic benefits.

Germany also faces the challenge of connecting advanced computing infrastructure to industrial applications across a geographically distributed manufacturing sector. Shared research infrastructure, industrial testing facilities, software integration capabilities and technical personnel can support adoption, but firms must still overcome difficulties relating to data quality, legacy machinery, cybersecurity and organisational change. The strategic objective is not solely to develop frontier models. It is to create a domestic and European ecosystem capable of integrating AI into industrial products and productive processes.

The German approach consequently has implications beyond national competitiveness. Germany’s industrial relationships with Italy, France and other European economies create potential spillovers through machinery, automotive components, industrial software and research cooperation. However, these benefits depend on access conditions and the compatibility of systems across firms and countries. A fragmented European AI infrastructure could reduce the efficiency of such cooperation, while interoperable computing and evaluation services could facilitate the development of shared industrial applications.

United Kingdom: sovereign computing outside EU regulatory jurisdiction

The United Kingdom’s AI strategy is distinguished by its position outside the EU’s directly applicable regulatory framework and by the scale of its public-computing objectives. The AI Opportunities Action Plan, published in January 2025, established a programme focused on expanding infrastructure, encouraging investment, improving public-sector adoption and supporting economic growth. The government’s response committed to substantially expanding sovereign computing capacity and developing new facilities serving researchers and enterprises. AI Opportunities Action Plan: Government Response — United Kingdom Department for Science, Innovation and Technology — January 2025.

The subsequent UK Compute Roadmap provides more specific quantitative targets. It describes an investment programme of up to £2 billion for public computing infrastructure, including more than £1 billion intended to expand the AI Research Resource and up to £750 million for a national supercomputing service in Edinburgh. The government aims to increase the AI Research Resource’s computing capacity from approximately 21 AI ExaFLOPS in 2025 to 420 AI ExaFLOPS by 2030, a twentyfold expansion. These figures are expressed using the government’s AI computing performance metric and represent a programme target rather than an independently verified future result. They should not be equated automatically with the performance figures used for other supercomputing systems unless the precision and measurement methods are comparable. UK Compute Roadmap — United Kingdom Department for Science, Innovation and Technology — July 2025.

The strategic purpose of this investment is to reduce barriers for organisations requiring access to advanced computing resources, including research institutions and smaller businesses. Public computing can strengthen domestic research and reduce reliance on purely commercial allocation mechanisms. However, its effectiveness depends on equipment procurement, operational funding, technical support and access conditions. A twentyfold expansion objective is significant, but the availability of computing capacity must be measured through deployed hardware, actual usage and the quality of services delivered to beneficiaries.

The United Kingdom’s institutional position also differs from that of EU Member States. It can develop domestic AI regulatory arrangements without applying the EU AI Act as domestic law, while companies operating in the EU market may still become subject to the regulation when their activities fall within its territorial and material scope. This creates a dual consideration for British technology firms: domestic policy conditions and the regulatory requirements of major export markets. The existence of these overlapping frameworks can impose additional compliance obligations but may also create opportunities for technical cooperation and compatible assurance standards.

In security policy, the United Kingdom maintains close relationships with the United States and European allies while preserving national authority over defence procurement and military AI doctrine. Its approach to AI-enabled weapons, discussed in Pillar I, demonstrates that civilian technology governance and military operational requirements remain separate legal and institutional domains. The relevant implication for Pillar II is that international cooperation in AI research or regulation does not automatically produce a common position on autonomous-weapons restrictions.

The European Union: regulation, public infrastructure and semiconductor policy

At EU level, the central strategic challenge is the relationship between industrial investment and regulatory coherence. The AI Act establishes an important common legal framework for covered activities, while the AI Continent Action Plan and EuroHPC initiatives aim to expand the infrastructure available to European organisations. The European Chips Act addresses a related dependency by seeking to strengthen semiconductor research, manufacturing and supply-chain resilience. These policies operate through different legal and financial mechanisms. Their effectiveness cannot be assessed by simply aggregating headline investment figures.

The European Commission states that the Chips Act framework is associated with public investment expectations exceeding €43 billion, while a broader total of more than €100 billion in policy-driven public and private investment is identified through 2030. The Commission also records a €2 billion Italian state-aid measure, approved in May 2024, to support a semiconductor manufacturing facility associated with STMicroelectronics. These figures concern policy-driven investment and support measures, not a verified increase of equivalent value in completed European chip production. The distinction matters because announced investment can be delayed, revised or implemented over several years, and because different financing instruments may overlap. European Chips Act — European Commission — 2026.

The Commission’s June 2026 proposal for a further Chips Act initiative also demonstrates that the semiconductor strategy remains subject to legislative and policy development. A proposal is not equivalent to an enacted legal instrument, and stated industrial objectives are not evidence that domestic production has achieved the targeted market share or technical capability. The appropriate assessment must track actual facilities, equipment installation, commercial production, qualified personnel and the relationship between European semiconductor output and the requirements of advanced AI computing.

The same discipline applies to AI Factories. The network is intended to expand access to computing resources across Europe, but geographical distribution alone does not ensure equitable utilisation. Users may face differences in available services, technical support, application procedures and the suitability of particular computing systems for their workloads. The EuroHPC Joint Undertaking’s published service catalogue provides evidence of differentiated sectoral support, including manufacturing, healthcare, scientific research and smaller enterprises. The quality of the network should therefore be evaluated through utilisation, the diversity of beneficiaries, the ability to support independent development and the resulting research and industrial outputs. AI Factories Services — European High Performance Computing Joint Undertaking — 2026.

Comparative European industrial and regulatory indicators

Jurisdiction	Verified official figure or programme	Date / horizon	Strategic objective	Evidentiary qualification
Italy	€1 billion programme for startups and SMEs in AI and related technologies	Announced September 2025	Domestic innovation and adoption	Programme commitment, not verified expenditure
Italy	IT4LIA AI Factory procurement contract	April 2026	Expand public AI computing capability	Procurement milestone, not proof of full operation
France	€109 billion in AI-related investment announcements	February 2025 onward	Data centres, infrastructure and deployment	Announced private investment, not completed capital expenditure
Germany	10% of economic output targeted as AI-based	2030 target	AI-related economic value creation	Policy objective; not an estimated 10% GDP increase
Germany	Manufacturing SME AI adoption above 50%	2029 target	Industrial diffusion	Target, not measured adoption
Germany	Fourfold increase in AI startups relative to 2025	2030 target	Expand domestic innovation ecosystem	Target dependent on definitions and baseline
United Kingdom	Up to £2 billion public-computing programme	Roadmap published 2025	Sovereign research and computing infrastructure	Programme envelope, not completed spending
United Kingdom	AIRR expansion from 21 to 420 AI ExaFLOPS	2025–2030 target	Twentyfold public AI compute expansion	Official objective using UK-specific performance measure
European Union	19 AI Factories and 13 antennas being established	Status reported in 2026	Shared computing and research infrastructure	Establishment does not mean every facility is fully operational
European Union	At least nine new AI-optimised supercomputers planned	Programme status in 2026	Expand EuroHPC AI capacity	Planned procurement and deployment
European Union	More than €43 billion anticipated public investment associated with Chips Act	Programme horizon through 2030	Semiconductor resilience	Investment expectation, not realised output
European Union	More than €100 billion in wider policy-driven semiconductor investment	Through 2030	Strengthen public and private semiconductor investment	Broader figure; not additive to the €43 billion

Sources: Italian AI Legislation — Department for Digital Transformation — September 2025; IT4LIA Procurement Contract — EuroHPC JU — April 2026; French AI Infrastructure Strategy — Presidency of France — February 2025; German AI Roadmap — Federal Government of Germany — 2026; UK Compute Roadmap — UK Department for Science, Innovation and Technology — July 2025; AI Factories — European Commission — 2026; European Chips Act — European Commission — 2026.

The figures are not directly comparable measures of national technological power. France’s €109 billion represents investment announcements, Italy’s €1 billion concerns an investment programme, the UK’s £2 billion concerns public computing plans, and Germany’s percentage targets concern policy objectives for economic activity and adoption. EU semiconductor investment figures have a different sectoral and financial scope. Combining these amounts into a single European investment total would be methodologically unsound without reconciling overlapping projects, public versus private financing, programme periods and actual disbursement.

European industrial competition and the problem of internal fragmentation

The national strategies create potential complementarities because European economies possess different industrial specialisations. Italy can contribute applied industrial capabilities and public research infrastructure. France can attract large computing facilities and support nationally anchored model-development activity. Germany can connect AI research to manufacturing systems, industrial software and robotics. The United Kingdom can provide public research computing and participate in scientific cooperation through arrangements distinct from EU membership. The EU can support common technical infrastructure, legislative harmonisation and cross-border research programmes.

However, complementarities do not automatically become integrated industrial capability. Governments may compete for the same data-centre investments, researchers and technology companies. National funding programmes may impose different eligibility rules or require particular activities to be located domestically. Public procurement may favour nationally anchored suppliers, while industrial policy may seek to retain economic value within national borders. These preferences can conflict with the wider objective of creating a sufficiently large European ecosystem to compete internationally.

The resulting problem is not simply political disagreement. It concerns the economics of scale. Advanced AI infrastructure benefits from large capital investments, specialised technical teams and high utilisation. Fragmented investments can duplicate costs or create incompatible systems if they are not coordinated. Conversely, excessive centralisation may disadvantage regions and countries outside the principal technology centres. Effective European cooperation must therefore balance the efficiency of shared infrastructure with the need for broad access and geographically distributed industrial benefits.

One important mechanism is the allocation of publicly supported computing resources. Shared infrastructure can allow research organisations and smaller companies in different countries to access capabilities that they could not finance independently. However, eligibility and allocation conditions must be sufficiently transparent to prevent the concentration of practical benefits among a small group of established institutions. Monitoring should therefore distinguish nominal access from computing time actually allocated, technical assistance provided and projects successfully completed.

Another mechanism concerns public procurement. European governments purchase digital services and technology for public administration, healthcare, education, infrastructure and defence. Procurement requirements can influence the technical standards, cybersecurity arrangements and accountability obligations adopted by suppliers. Yet uncoordinated national procurement can produce incompatible systems and reduce opportunities for economies of scale. A common set of technical assurance expectations could support interoperability, provided national authorities retain the ability to address genuine differences in legal and operational requirements.

Transatlantic dependence and the limits of technological sovereignty

European and British AI strategies must also account for the international supply chains on which advanced computing depends. Access to high-performance accelerators, manufacturing equipment, cloud software and specialised components is influenced by companies and governments outside Europe. This creates a distinction between sovereignty over the location of infrastructure and sovereignty over the technologies operating within it. A data centre built in Europe may still depend on foreign-made accelerators, external software and global supply chains. Domestic legal jurisdiction over the facility does not eliminate those dependencies.

The April 2026 IT4LIA procurement contract illustrates this structure. A European public computing initiative can use European integration and operational arrangements while incorporating hardware and architectural technologies supplied by international companies. Such arrangements can provide economically valuable capability without constituting complete technological independence. The relevant policy question is whether external dependencies are sufficiently diversified, legally manageable and technically substitutable to preserve continuity under adverse conditions.

The transatlantic relationship also involves potential regulatory divergence. The EU AI Act establishes obligations for covered activities, while US policy combines federal and sectoral approaches with national-security controls and industrial objectives. UK regulation develops through its own domestic institutions. Companies operating across these markets must therefore consider different legal requirements and compliance expectations. These differences may create costs but can also provide incentives for common technical standards and contractual practices. Whether convergence occurs depends on the willingness of governments and regulators to recognise compatible approaches without surrendering their own statutory responsibilities.

Security cooperation adds another layer. European NATO members and the United States may share military objectives while adopting different national approaches to autonomous weapons and AI-enabled systems. Joint operations require interoperability, reliable command arrangements and compliance with applicable law, but they do not require every participating state to possess identical domestic procurement rules. A system accepted by one military may raise legal or policy concerns for another. The relevant institutional response may involve agreed operational restrictions, joint review arrangements or mission-specific conditions rather than complete harmonisation of national doctrine.

The competitive consequences of European regulatory policy

The EU’s ability to impose regulatory requirements on products and services entering its market gives it a form of influence distinct from control of global computing infrastructure. Large international suppliers may adapt their systems to satisfy EU requirements where access to the market is commercially important. This can encourage more consistent documentation, transparency and risk management. However, the economic benefits of such influence depend on whether European enterprises can participate effectively in the resulting market.

If compliance costs are disproportionately difficult for smaller companies, regulation may unintentionally reinforce the position of firms possessing extensive legal and technical resources. If requirements are too weak or poorly enforced, firms may compete through reduced safeguards while the costs of harmful outcomes fall on users and public institutions. The regulatory challenge is therefore to establish requirements that correspond to identifiable risks, supported by accessible standards, effective supervisory institutions and proportionate compliance mechanisms.

The July 2026 AI Omnibus demonstrates an attempt to adjust that balance by extending selected deadlines and simplifying some obligations. Its effect on European competitiveness cannot yet be determined from the amendments themselves. The relevant evidence will concern implementation costs, market entry, investment, innovation outcomes, enforcement activity and the performance of high-risk systems after the revised provisions become applicable. AI Omnibus Enters into Force — European Commission — July 2026.

The same uncertainty applies to public infrastructure programmes. EuroHPC expansion may improve access to computing, but the resulting competitive effect depends on whether the infrastructure enables additional domestic research, stronger industrial applications and the emergence of firms capable of sustaining commercial operations. Announced facilities and public funding are inputs to industrial development, not proof of the outcomes that policy seeks to achieve.

Comparative national strategic implications

Strategic dimension	Italy	France	Germany	United Kingdom	European Union
Principal documented policy emphasis	Institutional implementation and industrial adoption	Infrastructure investment and national AI capability	Industrial AI and manufacturing competitiveness	Sovereign research computing and economic adoption	Common regulation and shared technological infrastructure
Relevant public authority	AgID, ACN and sectoral regulators	National ministries and cybersecurity authorities	Federal government and relevant national authorities	DSIT and sectoral authorities	European Commission, AI Office and national competent authorities
Infrastructure priority	IT4LIA and research computing	Large data-centre and computing investment	Computing access linked to industrial innovation	AIRR expansion and national supercomputing	EuroHPC AI Factories and Gigafactories
Main industrial opportunity	Manufacturing applications and research-to-industry transfer	National model development and infrastructure services	Industrial software, robotics and production systems	AI research, services and innovation	Cross-border scale, standards and public infrastructure
Principal dependency	Advanced hardware, integration and investment execution	Hardware supply, facility delivery and external technology	Computing inputs and enterprise implementation capacity	Hardware supply and long-term infrastructure funding	Global semiconductor and technology supply chains
Central implementation risk	Fragmented adoption and insufficient commercial scale	Announced projects not becoming fully operational	Targets not translating into measurable productivity gains	Expansion targets not matched by realised capacity	Regulatory and industrial programmes developing at different speeds
Relevant monitoring evidence	Funding disbursement, industrial uptake and computing utilisation	Completed facilities, power connections and actual investment	Verified SME adoption and productivity outcomes	AIRR deployment and user access	Factory utilisation, investment execution and regulatory compliance

The comparison shows that European technological sovereignty is not a single measurable attribute. A country may exercise relatively strong regulatory control while possessing limited independent computing infrastructure. Another may attract substantial physical investment but remain dependent on external semiconductor suppliers. A third may have a strong industrial application base without possessing internationally competitive frontier models. The effectiveness of European cooperation will depend on whether these differentiated strengths can be combined into reliable, interoperable and economically sustainable capabilities.

The diplomatic consequences: standards, market access and strategic autonomy

Europe’s international influence in AI governance derives partly from its market size, regulatory institutions and participation in multilateral negotiations. Yet this influence does not automatically translate into the ability to determine global technological development. States controlling critical computing inputs can shape supply conditions. Companies possessing leading technical capabilities can influence commercial standards. Governments with substantial defence programmes may resist constraints that affect military flexibility. European institutions must therefore operate within an international environment in which legal, industrial and military forms of power are distributed differently.

The distinction creates three diplomatic consequences. First, regulatory dialogue with major technology-producing states cannot be separated entirely from trade, export controls and industrial policy. Second, international cooperation on AI safety may advance in some sectors even while competition over advanced computing intensifies. Third, global governance arrangements that lack meaningful participation by technologically less-resourced countries may face legitimacy problems, particularly if they impose obligations without providing access to the infrastructure needed for compliance.

European policymakers consequently face a trade-off between pursuing national technological champions and developing shared infrastructure accessible across the continent. National investment may strengthen domestic firms, but insufficient coordination can limit regional scale. Shared European capacity may improve efficiency and resilience, but its allocation must be designed to avoid reproducing concentration within the European market. The central policy issue is not whether national or European investment is intrinsically preferable. It is whether financing, procurement, technical standards and access conditions are coherent enough to deliver the objectives attached to each programme.

Chapter 6 — Key judgments

High confidence: Italy, France, Germany and the United Kingdom pursue materially different AI industrial strategies. Treating them as a single European model would obscure differences in regulatory authority, investment priorities and technological dependencies.

High confidence: The EU AI Act establishes significant regulatory influence within its scope, while EuroHPC, AI Factories and semiconductor initiatives seek to strengthen the productive infrastructure necessary for technological competitiveness. These functions are complementary but not equivalent.

High confidence: Italy has established national AI governance responsibilities and announced a €1 billion programme for startups and SMEs; France has announced €109 billion in prospective AI-related investment; Germany has adopted measurable industrial AI targets; and the United Kingdom plans a twentyfold expansion of its AI Research Resource by 2030. These figures concern different financial and operational categories and must not be aggregated as equivalent achievements.

Medium-high confidence: Shared computing infrastructure, interoperable standards and coordinated public procurement can improve European technological capacity where they reduce unnecessary duplication and expand effective access. Their benefits depend on implementation and actual utilisation.

Medium confidence: Transatlantic technological dependencies will continue to influence European AI development because infrastructure investment does not eliminate reliance on internationally supplied hardware and software. The degree of strategic exposure depends on supply diversification, technical substitutability and applicable trade restrictions.

What would change the assessment: Verified large-scale deployment of European AI computing facilities, substantial independent model-development capabilities, measurable industrial productivity improvements and reduced dependence on critical external suppliers would support a stronger assessment of European technological autonomy. Conversely, repeated project delays, insufficient computing utilisation or persistent fragmentation of national investment would weaken the expected benefits of existing programmes.

Open official record: The critical documents concern realised public and private expenditure, operational computing availability, national and European procurement outcomes, industrial AI adoption statistics, verified energy connections for major facilities, and the extent to which public infrastructure supports smaller companies and institutions outside established technological centres.

Pillar II — Integrated Strategic Assessment

The findings of Chapters 4–6 establish that the governance of artificial intelligence is inseparable from the distribution of technological and institutional power. International organisations can develop scientific knowledge, facilitate negotiations and establish legal frameworks, but the effectiveness of those frameworks depends on the capacity of states and regulated organisations to implement them. Computing infrastructure, semiconductor supply chains, electricity, financial capital and technical expertise determine which actors can develop advanced systems and which remain primarily consumers of externally supplied technologies. These industrial conditions also influence the willingness of governments and companies to accept regulatory restrictions. A state that controls a scarce technological input may be able to impose access conditions extending beyond its territory, while a state lacking the corresponding capacity may have limited bargaining power even when formally represented in international governance institutions.

The institutional asymmetry is visible in the coexistence of expanding global dialogue and fragmented enforcement. The UN’s scientific and diplomatic mechanisms provide a wider evidentiary and participatory foundation, while regional and national legal instruments establish enforceable obligations within more limited jurisdictions. This division is not inherently defective: different institutions may be suited to different functions. The strategic problem arises when an internationally endorsed principle requires compliance that no competent authority can verify, or when a national regulator possesses legal powers but lacks the technical resources to exercise them. Effective governance therefore depends on connecting scientific assessment, legal obligations, technical standards, independent supervision and remedies without assuming that a single institution can perform all these functions.

The industrial analysis demonstrates that equitable access is similarly dependent on operational conditions. The existence of AI services does not establish independent technological capability. Countries require infrastructure, expertise, financing and institutional authority to develop and govern systems according to their own lawful priorities. Public computing initiatives can reduce barriers, but their contribution must be assessed through actual availability and utilisation rather than announced budgets. Export controls can protect national-security interests while simultaneously shaping the distribution of advanced computing resources. Intellectual-property protections can support innovation while contributing to commercial dependency where alternatives are limited. These tensions cannot be resolved through universal declarations alone; they require specific policies whose costs, benefits and restrictions can be evaluated.

Europe provides a particularly relevant example because it combines advanced legal institutions with substantial but uneven industrial capabilities. Its governments are investing in computing infrastructure, research and adoption while confronting competition for capital, energy, technical personnel and advanced components. The EU’s regulatory influence is significant, but its productive capacity remains dependent on successful implementation of industrial programmes. Italy, France, Germany and the United Kingdom demonstrate that national strategies differ in the assets they seek to develop and the economic outcomes they prioritise. Their potential complementarities are substantial, but cooperation must overcome differences in funding, procurement, infrastructure planning and market incentives.

Final Pillar II judgment — High confidence: The practical influence of ethical AI governance through 2030 will depend less on the breadth of international declarations than on the ability of institutions to supervise compliance and the willingness of technologically powerful states and corporations to accept enforceable obligations. The countries and organisations that control advanced computing, critical industrial inputs and major markets possess forms of influence that advisory institutions do not. Consequently, a credible common-good framework must address both the enforceability of standards and the distribution of the material capacity required to meet them.

The unresolved strategic issue is whether international cooperation can create effective constraints without requiring states to surrender advantages they regard as essential to security and competitiveness. The evidence establishes substantial progress in institutional coordination, regional regulation and public computing investment, but it does not establish a universal enforcement authority or an equitable global distribution of advanced AI capability. These limitations constitute the institutional and industrial baseline for the subsequent assessment of verification, strategic trajectories and policy choices through 2030.

---

PILLAR III — Strategic Trajectories, Verification and Decisions Through 2030

Chapter 7 — Three Governance Scenarios: Normative Influence, Partial Institutionalisation and Restrictive Arms Control

Principal judgment — Medium-high confidence: The development of international AI governance through 2030 will depend on whether the political convergence already achieved around human responsibility can be converted into obligations that materially restrict autonomous military capabilities, establish verifiable controls and preserve identifiable legal accountability. Three distinct governance trajectories remain supportable: continued normative influence without substantial new international constraints; partial institutionalisation through technical assurance, human-control requirements and national enforcement; and the emergence of a restrictive international arms-control regime supported by more demanding verification and oversight. These trajectories differ principally in the authority they confer on institutions and the extent to which participating states surrender operational discretion. They cannot be assessed solely through the number of agreements adopted or the diplomatic prominence of their sponsors. Their practical significance will be determined by the functions prohibited, the systems covered, the enforceability of national implementing measures, the treatment of military exceptions and the consequences of noncompliance. The most important development already established in 2026 is the adoption by consensus of a substantive set of elements for a possible autonomous-weapons instrument. That development strengthens the legal and diplomatic foundation for further negotiations but does not establish that a new treaty has been concluded or that participating governments have accepted a universal prohibition.

The September 2026 diplomatic breakthrough: agreement on elements, not agreement on a treaty

The strategic assessment must begin with the final outcome of the Convention on Certain Conventional Weapons Group of Governmental Experts, rather than with the earlier negotiation drafts. On 4 September 2026, the Group adopted by consensus its final report covering the 2024–2026 sessions. The report, issued as document CCW/GGE.1/2026/3, contains a set of elements for a possible instrument addressing lethal autonomous weapons. The agreement is significant because it moves the discussion from competing national proposals toward a common negotiating reference. Its importance must nevertheless be qualified: the Group formulated elements without predetermining the eventual instrument’s nature. The report itself does not create new treaty prohibitions or establish a verification organisation. The Seventh Review Conference, scheduled for 16–20 November 2026 in Geneva, is expected to consider the agreed elements and determine the mandate and modalities of future work. As of 5 November 2026, that conference has not occurred. Its decisions cannot be anticipated as established fact. Report of the 2024–2025–2026 Sessions of the Group of Governmental Experts, CCW/GGE.1/2026/3, paragraphs 23–25 and 45–46 — United Nations Office for Disarmament Affairs — September 2026; Seventh Review Conference of the High Contracting Parties to the CCW — United Nations Office at Geneva — November 2026.

The report’s operational characterisation is especially consequential. Paragraph 25 addresses a combination of weapons and technological components capable of identifying, selecting and engaging a target in a manner resulting in loss of life without intervention by a human operating the system. It clarifies that human preprogramming of potential target characteristics does not necessarily remove a system from the characterisation. Conversely, the text distinguishes situations in which personnel determine the specific target or group of targets and the timing of an attack. This language creates a more precise foundation for future regulation than a general reference to artificial intelligence or autonomous machines. It also reveals the continuing importance of technical and legal boundary questions: the characterisation is explicitly open to future modification and does not determine which systems will ultimately be prohibited or restricted. CCW/GGE.1/2026/3, paragraph 25, pages 4–5 — United Nations Office for Disarmament Affairs — September 2026.

The most politically relevant provisions concern human judgment, operational restrictions and responsibility. Paragraph 34 recognises the need for control and human judgment to ensure compliance with applicable international humanitarian law, while allowing for direct or indirect human control through measures taken before or during employment. Paragraph 35 identifies a responsible human chain of command, limits on operational scale, target categories, duration and geography, and means of timely intervention or deactivation where possible or feasible. Paragraphs 38 and 39 address testing under anticipated operational conditions, training, unintended bias and automation bias. Paragraphs 40–44 address continuing human and state responsibility, individual accountability, operational guidance and mechanisms for investigating potential violations. These provisions are important because they provide a potential basis for measurable national obligations. However, their wording also preserves discretion, particularly where compliance measures depend on contextual judgments and technical feasibility. CCW/GGE.1/2026/3, paragraphs 34–44, pages 5–6 — United Nations Office for Disarmament Affairs — September 2026.

The diplomatic significance is further demonstrated by the joint statement delivered to the UN General Assembly First Committee on 22 October 2026 on behalf of 50 states, including Italy, France and Germany. The statement welcomed the agreed elements as a basis for negotiations and called for a negotiating mandate within the CCW. The document also recognised that some substantive elements had not been retained in the final report. This matters because consensus on a negotiating foundation can be achieved partly by excluding language on which states remain divided. The agreement is evidence of the ability to negotiate common language; it is not proof that participating states share identical positions concerning prohibitions, verification, human intervention or the legal form of the eventual outcome. United Nations General Assembly First Committee 2026: Joint Statement on Lethal Autonomous Weapon Systems — New Zealand Ministry of Foreign Affairs and Trade — October 2026.

The International Committee of the Red Cross has separately reported that more than 70 states have explicitly supported beginning negotiations, while continuing to advocate a treaty prohibiting unpredictable autonomous weapons and systems designed or used to target persons directly. The ICRC’s position illustrates both the scope of diplomatic support and its limitations: agreement that negotiations should begin is not equivalent to agreement on the exact treaty provisions, particularly where states differ over acceptable autonomous functions. The humanitarian demand for additional restrictions also remains distinct from the existing legal obligations that apply to weapons and attacks under international humanitarian law. Statement to the United Nations General Assembly First Committee — International Committee of the Red Cross — October 2026; Frequently Asked Questions: Autonomous Weapon Systems — International Committee of the Red Cross — October 2026.

The diplomatic baseline therefore supports neither the claim that arms control has failed nor the conclusion that a restrictive regime is now assured. The agreed report establishes a concrete negotiating foundation. The subsequent outcome depends on whether governments authorise negotiations, whether they agree on the legal character of the instrument and whether operational restrictions can be made sufficiently precise to satisfy both humanitarian objectives and national-security requirements.

Scenario construction: the variables that distinguish genuine institutional change

The three scenarios are defined according to the legal and operational characteristics of the resulting governance arrangements, rather than according to broad descriptions of political cooperation. The first assumes that international norms continue to influence domestic policy but that significant new international restrictions do not become binding. The second assumes that states establish more detailed assurance obligations, including requirements relating to testing, documentation, authorisation and human control, while retaining substantial discretion over military applications. The third assumes that a group of states adopts a legally binding restrictive instrument addressing specified autonomous-weapons functions, accompanied by more credible national or international compliance mechanisms.

These scenarios are not assigned numerical probabilities. No defensible base-rate series exists for forecasting the adoption of international autonomous-weapons restrictions under sufficiently comparable conditions. Historical experience with arms-control treaties can illuminate institutional mechanisms but does not provide a reliable numerical model for an emerging technology whose military capabilities, software characteristics and strategic applications differ from previous categories of weapons. In addition, the scenarios may coexist geographically: one group of states could operate under a binding treaty while nonparties remain outside it, and civilian AI governance could become more demanding even where military arms-control negotiations remain inconclusive. Their use is therefore to identify conditional pathways and decision consequences, not to produce artificial statistical precision.

The most diagnostic variables are the scope of prohibited conduct, the status of human-control obligations, the availability of independent verification, the reach of national enforcement, the treatment of proprietary and classified information, and the degree of participation among states possessing significant military AI capabilities. A framework can become legally binding without creating intrusive inspections. A technically detailed assurance regime can influence procurement without requiring a new treaty. A universal political declaration can acquire substantial normative authority while remaining unenforceable as a legal obligation. These distinctions make it possible to evaluate institutional progress without confusing diplomatic visibility with the effective restriction of conduct.

Scenario A — Normative Influence Without Binding International Constraint

Under the first scenario, the ethical and legal language surrounding human responsibility becomes increasingly prominent, while the international system fails to establish a materially restrictive new framework for autonomous weapons. Governments continue to endorse the application of international humanitarian law, appropriate human oversight, responsible innovation and the need to prevent unacceptable harm. National defence ministries develop internal authorisation procedures and technical assurance practices, but the extent of autonomous functionality permitted remains primarily a matter of national doctrine, procurement decisions and applicable existing law. International institutions continue to publish scientific assessments, convene governments and identify common risks, yet they acquire no general authority to inspect military systems or compel compliance with newly defined prohibitions.

The principal mechanism sustaining this trajectory is the divergence between reputational benefits and operational costs. Governments can obtain diplomatic recognition for supporting responsible AI without accepting restrictions that significantly reduce military flexibility. Corporate developers can publish governance frameworks, support technical standards and participate in voluntary evaluation initiatives without necessarily accepting public disclosure of sensitive information or liability obligations exceeding applicable law. International organisations can expand participation and develop increasingly sophisticated recommendations without acquiring independent enforcement authority. Such developments may improve governance in particular areas, but the absence of common restrictions preserves substantial discretion over how autonomy is integrated into military and security functions.

The September 2026 CCW agreement does not eliminate this possibility. Consensus may support continuation of technical discussions without producing agreement on a binding treaty. Governments may endorse the agreed characterisation and the general requirement for human judgment while interpreting those requirements in ways compatible with existing national programmes. Negotiations may concentrate on best practices, reporting, operational recommendations or voluntary commitments. In that event, the principal institutional outcome would be greater consistency in the language used to describe responsible military AI, not a universal limitation on the autonomous functions states may develop or deploy.

The effect on military adoption would be differentiated. Functions considered to offer operational advantages would remain subject to national security assessments and applicable legal obligations, but there would be no additional universal treaty prohibition constraining deployment. The absence of new restrictions would not imply unrestricted deployment or automatic lawfulness. Existing international humanitarian law, national weapons-review obligations and domestic procurement rules would continue to apply. Nevertheless, military organisations would retain greater discretion in deciding whether advance authorisation, continuous supervision or other forms of control were appropriate for particular applications.

Corporate compliance expenditure under this scenario would remain driven predominantly by domestic regulation, contractual obligations, procurement standards and market access. Firms supplying high-risk civilian systems in regulated jurisdictions would continue to incur costs associated with documentation, evaluation and oversight. Defence suppliers could face demanding requirements imposed by particular government customers. The critical characteristic would be the absence of a common international military assurance framework capable of creating consistent obligations across participating states. Firms operating in several jurisdictions could therefore face different expectations regarding system testing, disclosure, incident reporting and the permissible degree of autonomous functionality.

The distributional consequences would also be uneven. Large technology suppliers could benefit from the continuation of nationally differentiated markets and the absence of additional international restrictions on certain applications. Smaller companies might benefit from avoiding expensive new universal certification requirements, but could remain dependent on dominant infrastructure providers and fragmented procurement systems. Developing economies could participate in international discussions while possessing limited influence over technological supply conditions. Normative commitments to equitable access would not automatically change the allocation of advanced computing resources, intellectual-property rights or export-controlled technologies.

The principal strategic risk is that a growing body of responsible-AI language could create the appearance of institutional convergence without reducing the operational behaviours that generate the most serious concerns. This is not an argument that voluntary standards have no value. They can improve organisational practices, establish common terminology and support professional accountability. The limitation is that voluntary compliance is least reliable when the costs of restraint rise and the immediate benefits of noncompliance appear substantial. A system whose effectiveness depends entirely on the willingness of powerful actors to accept disadvantageous restrictions cannot be treated as equivalent to independently enforceable law.

Scenario B — Partial Institutionalisation Through Transparency, Assurance and Human-Control Requirements

The second scenario assumes that governments convert parts of the emerging international consensus into more detailed institutional obligations without adopting a comprehensive prohibition on lethal autonomous weapons. National authorities develop operationally specific requirements concerning predeployment review, system testing, approved operating environments, human decision authority, training, configuration management and incident investigation. These practices may be supported by international standards, procurement conditions, technical cooperation and selective reciprocal commitments. International institutions continue to coordinate evidence and policy, while enforceable obligations are implemented predominantly through domestic law, national military procedures and contractual arrangements.

The September 2026 CCW report provides an unusually relevant foundation for this trajectory because it identifies measures that can be operationalised without first resolving every dispute over categorical prohibition. The report addresses understandable human-machine interfaces, testing under anticipated operational conditions, responsible human command structures, controls on operational scope and procedures for managing unintended engagements. These measures can become elements of domestic procurement and assurance systems even if states remain divided over the desirability or scope of a restrictive treaty. Their institutional value lies in making the exercise of human authority more observable and the operation of autonomous systems more susceptible to investigation. CCW/GGE.1/2026/3, paragraphs 35–44 — United Nations Office for Disarmament Affairs — September 2026.

Under this scenario, the concept of human control would become more detailed but would not necessarily require direct human approval of every engagement. National authorities could distinguish applications according to target type, operating environment, expected effects and the ability to intervene. They could establish approval criteria that require documented evidence of system performance under relevant conditions and demonstrate that operators understand the system’s limitations. Such measures would preserve a degree of national operational discretion, including the possibility of tightly constrained autonomous execution after advance authorisation. Their effectiveness would depend on the integrity of the underlying tests and whether operators have a genuine ability to comply with the restrictions imposed.

The military adoption consequences would therefore concern the timing and conditions of deployment rather than necessarily the elimination of particular capabilities. Programmes requiring extensive validation could face additional stages before operational approval. Existing systems could require reassessment when modified. Suppliers might need to demonstrate compliance with defined technical and organisational requirements, including operating constraints, update controls and evidence preservation. Defence establishments would incur additional administrative and technical burdens, but these costs could be partly offset by reduced failure risks, improved procurement discipline and more reliable system integration. No defensible universal cost estimate can be derived from the current record because testing requirements, programme complexity and national procurement practices differ substantially.

This scenario has a stronger potential effect on corporate behaviour. Government procurement contracts could require access to technical records, disclosure of relevant limitations, assistance with incident investigations and controls over material system changes. Civilian regulators would continue implementing applicable high-risk AI obligations. Technical standards could provide common methods for managing risk and documenting performance, reducing some duplication across markets. At the same time, compliance could favour firms with established technical assurance capabilities unless governments provide proportionate procedures and shared testing infrastructure for smaller suppliers.

The National Institute of Standards and Technology’s AI Risk Management Framework provides an example of a structured but voluntary approach that can inform such practices. Its four functions—govern, map, measure and manage—are intended to support continuous assessment and treatment of AI risks. The framework does not itself create a binding military regime, and its requirements cannot simply be transferred into international humanitarian law without legal and operational adaptation. Its relevance is that it illustrates how governance can move from general principles to systematic organisational procedures, including documentation, testing, independent review and reassessment as deployment conditions change. AI Risk Management Framework Core — United States National Institute of Standards and Technology — January 2023.

The European regulatory timetable also makes partial institutionalisation observable within the 2030 horizon. Following the July 2026 amendments, major high-risk AI requirements apply from 2 December 2027 for relevant Annex III systems and from 2 August 2028 for the relevant Annex I product categories. The consolidated regulation also contains transitional requirements extending to 2 August 2030 for certain high-risk systems intended for use by public authorities. These dates provide specific milestones against which implementation can be examined. However, the civilian regulatory timetable should not be presented as a schedule for the introduction of autonomous-weapons restrictions, because the AI Act’s relevant scope exclusions remain legally significant. AI Omnibus Enters into Force — European Commission — July 2026; Regulation (EU) 2024/1689, consolidated text of 27 July 2026, Articles 111 and 113 — European Parliament and Council of the European Union — July 2026.

The distributional effects would be mixed. Common assurance requirements could improve safety, support procurement interoperability and make it easier for smaller jurisdictions to rely on recognised evaluation procedures. Yet highly demanding certification obligations could exclude organisations lacking testing infrastructure, specialised personnel or access to relevant technical information. A governance regime designed to reduce risk could therefore increase dependence on a small group of companies able to meet expensive requirements. This consequence would not be inevitable. Public testing facilities, shared evaluation methods and technical assistance could reduce compliance inequalities, provided they offer genuine access to infrastructure and do not simply reproduce commercial dependency.

The principal limitation of Scenario B is the possibility of procedural compliance without substantive restriction. A weapon may undergo extensive testing while remaining authorised to conduct functions that some states or humanitarian institutions consider unacceptable. A procurement system may document human oversight without establishing that operators can meaningfully intervene in realistic conditions. A company may meet formal reporting requirements while withholding information relevant to independent assessment under lawful confidentiality restrictions. The scenario therefore represents stronger institutional governance but not necessarily a solution to every ethical concern associated with autonomous lethal targeting.

Scenario C — Restrictive Arms Control and Stronger International Oversight

The third scenario assumes that states negotiate and adopt a legally binding instrument establishing prohibitions on defined autonomous-weapons functions, restrictions on other autonomous systems and mechanisms for implementing and reviewing compliance. A restrictive regime could emerge within the CCW framework or through another legally valid negotiating process if participating governments were unable to reach agreement within existing institutional arrangements. Its geographic reach would depend on participation, and its practical effect would depend on the precision of obligations, implementation timelines, verification methods and consequences of violations. A treaty could be legally binding on its parties without becoming universal, and states outside it would remain subject to otherwise applicable international law rather than automatically acquiring the treaty’s new obligations.

The most important substantive decision would concern the scope of prohibited systems. A treaty might prohibit weapons whose effects cannot be adequately predicted, systems designed or used to autonomously target persons directly, or particular forms of engagement lacking specified human decision-making authority. It might also restrict other systems through requirements governing target categories, geographic boundaries, duration, scale and intervention mechanisms. These elements correspond to the ICRC’s proposed approach, but that proposal remains distinct from the narrower language on which CCW participants achieved consensus. The difference matters because a binding instrument cannot be evaluated without knowing what it actually prohibits. A treaty that excludes numerous existing military applications through broad definitions could produce limited operational change, whereas one that imposes stringent restrictions on functions considered essential to national defence could encounter greater resistance and implementation costs. Frequently Asked Questions: Autonomous Weapon Systems — International Committee of the Red Cross — October 2026.

A restrictive regime would also need to determine whether obligations concern development, acquisition, possession, deployment or use. These distinctions affect verification and the costs of implementation. Prohibiting a system’s development or possession requires the ability to identify prohibited capabilities before deployment and assess relevant facilities or software configurations. Prohibiting particular uses focuses attention on operational circumstances, authorisation and incident evidence. Requiring the removal of certain functions from existing weapons could impose modification costs and require verification that the prohibited functions cannot be restored without detection. Each approach has different implications for commercial suppliers, defence procurement and national secrecy.

The transition from an agreed negotiating foundation to a restrictive treaty would require several political decisions. Governments must establish a mandate that permits substantive negotiations, agree on definitions, identify prohibited conduct, determine the treatment of existing systems and negotiate the relationship between international oversight and national security. They must then adopt the instrument through the relevant procedures, complete applicable domestic approval processes and establish implementing measures. Even a successful diplomatic negotiation would therefore be followed by a potentially lengthy period of domestic legislation, procurement adjustment and technical implementation. No reliable basis exists for specifying a universal completion date before 2030.

The military consequences would depend on the relationship between prohibited functions and existing programmes. States that do not possess or intend to acquire covered capabilities could face relatively limited adjustment costs. States with extensive development programmes might need to modify design requirements, suspend procurement, restrict deployment or withdraw certain capabilities. If treaty provisions apply to existing systems, the financial and operational consequences could be greater than if obligations apply only to future development. Transitional arrangements would therefore become politically significant. A state may support a prohibition in principle but demand sufficient time to adapt its defence architecture and maintain functions it regards as essential.

Corporate compliance costs would similarly depend on the scope of the treaty and the level of assurance demanded. Defence suppliers might be required to maintain evidence that prohibited functions are absent or disabled, support independent evaluation, preserve configuration records and disclose relevant modifications to competent authorities. Where products are distributed internationally, firms could encounter different requirements depending on whether the customer state is a party and how the treaty has been implemented domestically. A common restrictive framework could eventually reduce some regulatory uncertainty among participating states, but its initial implementation could require substantial changes to contracts, testing facilities, development procedures and software governance.

The most difficult question would concern verification. If an instrument requires external inspection of sensitive systems, governments must agree on access to potentially classified equipment, software and operational records. If verification is left predominantly to national authorities, the international system must determine how much confidence can reasonably be placed in declarations and domestic reviews. If the treaty includes confidential reporting or protected technical assessments, it must define who receives the information, how it is secured and what consequences follow from disputed findings. These problems are not unique to AI, but software configurability and the dependence of system behaviour on operational conditions make them particularly demanding.

The distributional consequences would depend on how the regime interacts with technological access. Restrictions on autonomous weapons could reduce particular proliferation and humanitarian risks, but they would not independently alter the global distribution of advanced computing infrastructure. A treaty accompanied by assistance for national weapons reviews, technical training and compliance infrastructure could improve the ability of less-resourced states to implement their obligations. Conversely, a complex inspection and certification regime could impose costs that are difficult for smaller states and suppliers to meet. Equitable participation would require attention to implementation resources rather than assuming that identical treaty language imposes an identical practical burden on all parties.

The residual space for unilateral action would remain significant even under a restrictive regime. States not party to the treaty would not automatically become bound by its new provisions. Parties might retain discretion in areas not prohibited or restricted. Military AI functions outside the treaty’s definition of autonomous weapons could continue to develop under otherwise applicable law. The practical significance of the regime would therefore depend on participation by states possessing relevant capabilities, the extent of supplier compliance and the credibility of monitoring. Its success should not be measured exclusively by the formal existence of a treaty.

Comparative scenario matrix through 2030

Strategic variable	Scenario A — Normative influence	Scenario B — Partial institutionalisation	Scenario C — Restrictive arms control
International legal outcome	No substantial new binding autonomous-weapons restrictions	More specific obligations or national measures; international framework may remain incomplete	Binding restrictions for participating states
Defining institutional feature	Declarations, dialogue and voluntary practices	Documentation, assurance and enforceable domestic controls	Prohibitions or restrictions combined with implementation mechanisms
Military adoption	National programmes retain broad discretion under existing law	Additional authorisation and testing requirements	Covered capabilities restricted, modified or withdrawn according to treaty terms
Human-control standard	National and organisational interpretation	More specific technical and procedural requirements	Treaty-defined requirements for covered functions
Corporate compliance cost	Primarily existing jurisdictional and procurement burdens	Increased assurance, testing, training and record-keeping	Additional treaty implementation and verification obligations
North–South technology gap	Persists absent separate industrial measures	May narrow through shared assurance, or widen through compliance costs	Depends on accompanying financial and technical assistance
Independent verification	Limited beyond national or voluntary mechanisms	Selected technical audits and national oversight	Potentially stronger treaty-related procedures
Treatment of military secrecy	Primarily national discretion	Controlled disclosure to competent authorities	Negotiated balance between inspection and confidentiality
Unilateral state action	Broad within existing law	Restricted in regulated contexts	Restricted for parties within treaty scope; remains possible outside it
Principal governance failure	Normative endorsement without operational change	Formal compliance without substantive control	Legal prohibitions without credible verification or sufficient participation
Main condition for transition	Political willingness to define obligations	Common enforceable controls and technical capacity	Agreement on prohibited conduct, participation and compliance

The matrix is conditional and does not describe completed developments. Its purpose is to isolate which decisions would alter military, commercial and institutional behaviour. An agreement to negotiate is a diplomatic milestone; a legally binding instrument is a legal milestone; implementation by national authorities is an administrative milestone; and independently verified compliance is an operational milestone. The four must not be treated as simultaneous achievements.

Military adoption timelines: why a single global forecast would be misleading

Military adoption through 2030 cannot be represented by one international timeline because autonomous capabilities are heterogeneous and subject to different procurement, security and legal conditions. A defensive application intended to operate in a constrained environment may follow a different approval process from a system designed for extended operation in a complex civilian-populated environment. The development cycle of a software component may be shorter than the acquisition cycle of the military platform that incorporates it. A defence establishment may authorise experimentation without approving operational deployment. Conversely, operational urgency may lead to accelerated fielding of capabilities that remain subject to subsequent review or modification.

Under Scenario A, adoption would remain principally shaped by national procurement programmes and operational requirements. This does not imply that deployment will necessarily accelerate everywhere. Budget constraints, technical failure, integration difficulties and existing legal review processes can delay systems independently of international arms control. Under Scenario B, additional testing and documentation requirements would become a more systematic part of procurement, potentially lengthening some approval processes while improving reliability and traceability. Under Scenario C, prohibited functions would encounter legal restrictions, while permitted systems could continue to develop subject to the treaty’s conditions.

The relevant forecast is therefore qualitative. International rules can influence the conditions under which systems are deployed, but their effect on adoption speed depends on the underlying capability, the legal scope of restrictions and the stage of national development programmes. A claim that a restrictive treaty would delay every military AI programme by a specified number of years would require detailed programme-level evidence unavailable in the public record. A more defensible measure would be whether the introduction of new requirements causes documented changes to procurement specifications, operational approvals, system configurations or deployment authority.

Corporate compliance expenditure: fixed costs, recurring costs and industrial concentration

The corporate consequences of the three scenarios should be distinguished by cost category rather than described through a universal percentage increase. The introduction of new obligations can create fixed costs associated with legal interpretation, system documentation, testing infrastructure and the establishment of internal governance procedures. It can also create recurring costs involving evaluation, monitoring, record retention, incident reporting, personnel training and external audit. Some costs depend on the number of systems developed, while others depend on the number of jurisdictions served or the complexity of deployment environments.

The distribution of these costs matters. A large developer may spread fixed compliance expenditure across numerous customers and products. A smaller supplier may face a proportionately greater burden when equivalent documentation or evaluation infrastructure must be developed for a limited number of systems. Public procurement arrangements can influence this outcome by providing common technical requirements or supporting shared evaluation facilities. However, procurement standards that are too complex or inconsistent can restrict competition and favour established firms. The relationship between stronger governance and industrial concentration is therefore conditional: common standards can lower uncertainty, but expensive compliance procedures can increase entry barriers.

Under Scenario A, firms would continue to face the costs associated with existing regulation and procurement. Under Scenario B, a more consistent set of assurance procedures could increase immediate compliance spending while reducing some duplication and improving the predictability of requirements. Under Scenario C, firms supplying covered military capabilities would incur additional costs arising from the need to demonstrate that prohibited functions are absent, restricted or otherwise compliant. The magnitude of those costs cannot be calculated credibly without the substantive treaty provisions and programme-level information.

Distributional outcomes: different forms of inequality require different remedies

The three scenarios also differ in how they affect countries with limited technological capacity. Normative influence alone does not provide computing infrastructure, specialised personnel or independent evaluation resources. Partial institutionalisation can improve the quality of governance but may increase compliance costs unless assistance is available. A restrictive treaty can establish common limits on particularly sensitive military applications, but less-resourced states may lack the technical expertise required to evaluate weapons, verify supplier claims and investigate incidents.

The resulting inequality is not confined to access to technology. It also concerns the ability to understand, regulate and challenge technological systems supplied by external actors. A country may possess legal authority over domestic deployment while lacking technical access to model internals, relevant testing facilities or personnel trained to evaluate complex systems. Its formal sovereignty may therefore coexist with substantial practical dependence. International governance can reduce that asymmetry through technical assistance, shared testing infrastructure and compatible reporting procedures, but only if such programmes are adequately financed and accessible.

This creates an important distinction between equality of legal obligation and equality of implementation capacity. A treaty may impose identical requirements on participating states without ensuring that those states possess comparable resources for compliance. A policy of differentiated assistance need not imply weaker substantive safeguards. It may be necessary to make those safeguards practically achievable while maintaining common legal obligations.

Scenario transition conditions and diplomatic milestones

Transition	Required political development	Required legal development	Required technical development	Observable evidence
Normative influence to partial institutionalisation	States agree to implement specific assurance requirements	Domestic legislation, binding procurement rules or other enforceable measures	Standardised testing, documentation and review procedures	Published requirements and evidence of implementation
Partial institutionalisation to restrictive negotiations	States accept a mandate addressing substantive prohibitions	Formal negotiating mandate and agreed treaty architecture	Working definitions and preliminary compliance methods	Adopted conference decisions and negotiating texts
Restrictive negotiations to adopted treaty	Agreement on scope, exceptions and implementation	Adoption of a legally binding instrument	Defined systems, prohibited functions and verification procedures	Authenticated treaty text and adoption record
Treaty adoption to effective national implementation	Political approval and administrative commitment	Ratification or accession, entry into force and implementing law	Inspection, audit and enforcement capabilities	Depositary records, legislation and funded authorities
National implementation to credible compliance	Sustained willingness to enforce obligations	Procedures for investigation and remedies	Independent testing, protected evidence and corrective measures	Inspection findings, enforcement records and remedial action

These transitions are sequential in legal logic but not necessarily identical across jurisdictions. Domestic assurance measures can advance before a treaty is negotiated. National prohibitions can be adopted independently. Some governments may participate in international negotiations without subsequently ratifying the resulting instrument. The decisive monitoring issue is whether each stage produces an observable change in legal obligations or operational conduct.

Chapter 7 — Key judgments

High confidence: The consensus report adopted on 4 September 2026 establishes an important common negotiating foundation, including an operational characterisation of lethal autonomous weapons and provisions concerning human judgment, operational restrictions and responsibility. It does not itself create a binding international prohibition.

High confidence: The Seventh CCW Review Conference of 16–20 November 2026 is the next decisive diplomatic milestone. Its actual decisions, rather than the existence of the consensus report alone, will determine the immediate negotiating mandate.

Medium-high confidence: Partial institutionalisation can occur through national procurement, technical assurance and domestic regulation even if negotiations on binding autonomous-weapons prohibitions remain incomplete. Its capacity to improve accountability depends on substantive implementation rather than procedural declarations.

Medium confidence: A restrictive international regime is institutionally possible, but its geographic reach and operational effect will depend on definitions, ratification, implementation and participation by states possessing relevant military capabilities.

Low confidence: The public record does not support a reliable numerical probability for any of the three scenarios or a universal timetable for the entry into force of a comprehensive autonomous-weapons treaty before 2030.

What would change the assessment: A formal negotiating mandate adopted at the November 2026 Review Conference, agreement on specific prohibited functions, credible proposals for independent verification and subsequent national ratifications would strengthen the basis for Scenario C. Conversely, a decision to continue nonbinding technical discussions without a negotiating mandate would preserve more of the institutional conditions associated with Scenarios A and B.

Open official record: The most consequential missing evidence concerns the November 2026 conference decisions, authenticated negotiating texts, binding national implementation measures, detailed military procurement adjustments and the willingness of technologically significant states to accept external compliance scrutiny.

Chapter 8 — Verification Problems, Strategic Indicators, Policy Options and Final Net Assessment

Principal judgment — High confidence: The decisive test of international AI governance through 2030 will be whether states can demonstrate compliance with agreed obligations under conditions in which technological capability, military secrecy, commercial incentives and institutional fragmentation make independent verification difficult. Formal agreement on human responsibility is a necessary but insufficient condition for effective control. A credible regime must establish what conduct is prohibited, how compliance can be assessed, which authorities possess access to relevant evidence, how violations are investigated and what consequences follow when obligations are breached. The problem is especially demanding for autonomous weapons because the relevant capabilities may depend on software configuration, operational parameters, sensor inputs and changes in the environment rather than on readily identifiable physical characteristics alone. At the same time, a universal inspection regime cannot be assumed politically feasible simply because it would improve transparency. The most defensible policy approach is to distinguish measures that can be implemented under existing authority from measures requiring new international agreements, and to define observable thresholds that reveal whether governance is becoming more substantive or remains predominantly declaratory.

Verification as a strategic problem: proving compliance is different from establishing a rule

A government may announce that it prohibits certain autonomous-weapons functions, yet an external observer may remain unable to determine whether its systems possess those capabilities, whether they have been technically disabled or whether operational doctrine permits their use under another classification. This does not establish that the government is violating its commitment. It demonstrates the difference between the existence of a rule and the availability of evidence sufficient to determine compliance. The distinction is critical because the deterrent effect of a legal prohibition depends partly on the prospect that violations can be detected, attributed and subjected to consequences. Where compliance cannot be assessed independently, governments and companies may retain broad discretion over how their conduct is described.

The September 2026 CCW consensus report provides an unusually concrete starting point. It identifies human command structures, operational restrictions, testing, training, measures against automation bias and mechanisms for preventing and addressing violations. The report also acknowledges the sensitivity of national-security information and proprietary commercial technology when encouraging states to exchange relevant review practices. These provisions demonstrate that states have recognised the importance of operational controls and evidence, but they do not yet establish a compulsory international system capable of verifying compliance with future prohibitions. CCW/GGE.1/2026/3, paragraphs 34–44 — United Nations Office for Disarmament Affairs — September 2026.

A useful analytical distinction separates four questions. The first concerns design compliance: whether a system possesses functions that are prohibited or restricted. The second concerns configuration compliance: whether the deployed system operates within an approved configuration. The third concerns operational compliance: whether the particular use of the system satisfies applicable law and authorised restrictions. The fourth concerns accountability compliance: whether the evidence and institutional procedures necessary to investigate potentially unlawful conduct exist and are accessible to competent authorities. A system may satisfy one of these conditions while failing another. A legally reviewed design may be modified before deployment. A properly configured system may be used outside its authorised environment. An apparently lawful engagement may nevertheless reveal deficiencies in training, supervision or documentation. Conversely, an adverse outcome does not by itself prove that the relevant legal duties were violated.

These distinctions matter because a future treaty may regulate different stages of the system lifecycle. A prohibition on development requires evidence concerning research and technical capabilities. A restriction on acquisition requires controls over procurement and transfer. A restriction on deployment requires evidence about approved operating configurations and intended environments. A prohibition on particular uses requires operational records and procedures for investigating incidents. A treaty attempting to regulate all these activities would impose broader evidentiary demands than an instrument limited to particular forms of operational employment.

Software-defined capability and the limits of physical inspection

Software-based autonomy complicates verification because the relevant functionality may be distributed across several components rather than contained in a single physical object. The platform may incorporate sensors, processors, navigation systems, targeting software, communications interfaces and external decision-support services. Its behaviour can depend on the interaction among those elements and the environment in which it operates. A physical inspection can establish that particular components are installed, but it may not establish how the complete system behaves under all permitted or prohibited conditions.

The distinction between installed functionality and activated functionality is particularly important. A system may contain software capable of different operating modes, with access controlled through configuration settings or authorisation procedures. A compliant configuration at the time of inspection does not necessarily demonstrate that a prohibited function cannot later be enabled. Conversely, the theoretical possibility of changing software does not establish that an organisation has developed or deployed an unlawful weapon. Verification must focus on the requirements of the applicable legal instrument, the actual technical characteristics of the system and the evidence necessary to determine whether prohibited conduct has occurred.

Technical controls can improve confidence but have limitations. Digitally signed software releases and protected configuration records can support evidence that a system has not been modified without authorisation. Controlled update procedures can establish when software changes occurred and who approved them. Independent testing can assess behaviour within specified operational conditions. Audit logs can record actions, settings and relevant human decisions. These mechanisms may make violations more difficult to conceal and investigations more reliable, but they do not prove the absence of every undesirable behaviour. Their effectiveness depends on the integrity of the underlying software, the protection of recording systems, the quality of testing and the independence of the authority assessing the evidence.

A further problem arises from the difference between software identity and operational behaviour. Two systems using nominally identical software may behave differently because of sensors, data inputs, hardware characteristics or environmental conditions. An update that appears minor may alter the conditions under which a system selects targets or responds to unexpected events. A system tested under controlled conditions may encounter situations not adequately represented during evaluation. Consequently, conformity cannot be established through a static examination of source code or a single performance demonstration. A credible assurance process must define the intended operating domain, identify relevant limitations and establish when changes require reassessment.

The International Committee of the Red Cross’s guidance on weapons reviews is relevant because it treats the legality of weapons in relation to their normal or expected use, including circumstances in which particular methods of employment may be prohibited. Its approach recognises that a weapon cannot be assessed solely through its physical design. The legal review must consider the intended operating environment and the effects reasonably anticipated from its use. This logic is especially important for autonomous systems whose behaviour depends on interaction with external conditions. A Guide to the Legal Review of New Weapons, Means and Methods of Warfare — International Committee of the Red Cross — June 2020.

The distinction between technical confidence and legal confidence

A testing programme can demonstrate that a system satisfies defined performance criteria, but the legal significance of that result depends on whether the criteria are relevant to the obligations being assessed. High accuracy in recognising a particular object does not establish that all relevant target-selection requirements are satisfied. Reliable operation within a restricted test environment does not demonstrate that the same system can lawfully operate in substantially different circumstances. The absence of observed failures during testing does not establish that failures are impossible. Equally, the existence of technical uncertainty does not automatically render every use unlawful. The appropriate legal assessment depends on the nature of the uncertainty, the consequences of failure and the applicable restrictions.

This creates a need for documented relationships between legal requirements and technical tests. A reviewing authority should be able to identify the legal question addressed by each evaluation, the assumptions underlying the test and the circumstances in which its conclusions remain valid. If the legal concern involves the ability to discriminate between lawful and unlawful targets, the assessment must examine whether the system has access to the information necessary for that distinction under expected operational conditions. If the concern involves the risk of unintended engagements, testing must consider circumstances in which sensors, communications or software components behave unexpectedly. If a system depends on human intervention, the evaluation must examine whether intervention is operationally achievable, not merely whether the interface contains the relevant command.

A credible process must also address uncertainty explicitly. Technical testing produces evidence within defined conditions; it does not guarantee behaviour across an unlimited set of possible circumstances. The legal authority responsible for approving deployment must therefore determine whether the remaining uncertainty is compatible with the intended use and applicable law. That decision cannot be delegated entirely to the manufacturer because the manufacturer may lack operational information or the authority to determine permissible military conduct. Nor can it be made responsibly by legal personnel without sufficient technical expertise. Multidisciplinary review is necessary because the legal, technical and operational dimensions affect one another.

The CCW consensus report recognises these dependencies by addressing testing under anticipated operational conditions, operator training and the risk of unintended engagements. These measures create a possible foundation for more precise international standards, but the report does not prescribe a universal quantitative reliability threshold. Such a threshold could not be adopted responsibly without specifying the relevant function, operating conditions, consequences of failure and evidence supporting the selected value. CCW/GGE.1/2026/3, paragraphs 37–39 — United Nations Office for Disarmament Affairs — September 2026.

Verification architecture: evidence, authority and institutional independence

Verification domain	Evidence required	Responsible or competent actor	What the evidence can establish	What it cannot establish alone
System identification	Hardware inventory, software versions and component specifications	Procurement authority and technical assessor	Identity of the reviewed system and its documented configuration	Absence of all undeclared capabilities
Autonomous functionality	Functional tests, operating modes and interface documentation	Independent testing authority and relevant developer	Behaviour within specified test conditions	Behaviour under every conceivable operational condition
Human authorisation	Approval records, authority assignments and operational restrictions	Command authority or competent deploying organisation	Who approved use and under which recorded conditions	Whether every subsequent action complied with law
Operational limits	Target restrictions, geographic boundaries, time limits and configuration records	Deploying authority and technical assurance body	Whether defined constraints existed and were configured	Whether constraints remained effective throughout operation
Software integrity	Signed releases, change logs and access-control records	System owner and authorised technical authority	Evidence of approved software configuration and modifications	Complete protection against compromise or unauthorised change
Training and competence	Qualification records, exercises and operating instructions	Military command or deploying organisation	Whether relevant personnel received prescribed preparation	Whether judgment was adequate in a specific incident
Incident reconstruction	Sensor records, event logs, communications and operational context	Competent investigative authority	Sequence of recorded events and relevant decisions	Legal responsibility without further factual and legal analysis
Independent oversight	Inspection findings, evidence-access procedures and corrective orders	Legally empowered supervisory body	Compliance with requirements within the inspection’s scope	Universal compliance beyond the systems and activities examined
Treaty compliance	State declarations, national implementing measures and agreed verification results	Treaty parties and designated mechanisms	Implementation of specified obligations and assessed conduct	Compliance by nonparties or with obligations outside treaty scope

The table illustrates why verification must be designed around specific legal obligations. There is no universal evidentiary document proving that a complex AI system is responsible, safe or lawful under every conceivable use. Different obligations require different evidence, and the competence of the authority examining that evidence affects the reliability of the conclusion.

The national-security dilemma: transparency can improve verification while exposing sensitive capabilities

Military verification encounters a conflict between the information needed to assess compliance and the information governments seek to protect. Technical records may reveal sensor performance, operational limits, software architecture, communications arrangements or capabilities relevant to military advantage. States may be unwilling to provide unrestricted access to such information, particularly where inspectors represent governments with competing security interests. These concerns are not merely rhetorical. Disclosure of sensitive information can create risks even when the underlying purpose is legitimate arms control.

A credible verification framework must consequently distinguish public disclosure from confidential inspection. Governments may be able to publish legal review procedures, general control principles and aggregate implementation information without revealing operational details. More sensitive technical evidence could potentially be examined by appropriately authorised national bodies or independent institutions operating under agreed confidentiality arrangements. The effectiveness of these mechanisms would depend on the legal authority of inspectors, secure handling of information and safeguards against misuse.

The institutional design problem becomes more difficult when a state disputes the findings of an external assessor. The regime must determine whether findings can be challenged, what evidence may be disclosed to support the challenge and which body is competent to resolve the dispute. A verification arrangement that lacks credible dispute procedures may generate political conflict even when technical evidence is strong. Conversely, excessively broad confidentiality protections may prevent other parties from assessing whether compliance findings are reliable. The legitimacy of a restrictive regime therefore depends on procedural as well as technical design.

A further problem concerns asymmetric participation. States with advanced technical capabilities may be better able to satisfy complex verification requirements and protect sensitive information. Less-resourced states may depend on external technical assistance, potentially raising concerns over independence and confidentiality. If verification depends heavily on a small group of technologically dominant countries or corporations, the resulting institutional structure could reinforce the inequalities it is intended to mitigate. A credible international regime would therefore need mechanisms for developing independent technical competence among a broader group of participating states.

Failure modes: how a formally compliant system can remain institutionally ineffective

Failure mode	Observable manifestation	Primary consequence	Appropriate institutional response
Formal approval without effective control	Human supervisor lacks time or authority to intervene	Accountability becomes procedural rather than substantive	Operational testing of actual intervention capability
Incomplete functional definition	Prohibited capability remains outside legal terminology	Regulatory loopholes and divergent interpretation	Functional and contextual definitions
Undocumented software changes	Operational configuration differs from approved version	Review conclusions become unreliable	Mandatory reassessment and protected change records
Limited inspector access	Authorities depend entirely on declarations	Reduced independent verification confidence	Legally defined access and confidentiality procedures
Inadequate incident records	Harm cannot be reconstructed reliably	Difficulty establishing conduct and responsibility	Proportionate logging and evidence preservation
Weak national implementation	Treaty exists without effective domestic enforcement	Formal obligations lack practical effect	Implementing legislation, competent authorities and resources
Unequal compliance capacity	Smaller states or suppliers cannot meet technical requirements	Concentration and reduced participation	Shared evaluation infrastructure and technical assistance
Excessive reliance on voluntary disclosure	Sensitive failures remain unreported	Incomplete institutional information	Binding reporting where legally justified
Unclear consequences for violations	Noncompliance produces limited corrective action	Reduced deterrence	Defined investigation and enforcement procedures

These mechanisms are prospective analytical categories, not findings that particular governments or companies have committed violations. Their value lies in identifying weaknesses that can exist even where formal policies and declared commitments appear comprehensive.

Monitoring governance through 2030: indicators must distinguish outputs from outcomes

A recurring weakness in technology-governance assessments is the use of institutional activity as a proxy for institutional effectiveness. The number of meetings, declarations, standards and published strategies can indicate political attention, but it does not establish that harmful conduct has been prevented, accountability improved or independent compliance assessment made possible. The same distinction applies to financial and industrial policy. A large investment announcement is not completed infrastructure; a procurement contract is not operational computing capacity; a training initiative is not verified professional competence; and a regulatory deadline is not evidence of compliance.

Decision-grade monitoring must therefore distinguish inputs, institutional outputs, implementation outcomes and observed consequences. Inputs include budgets, personnel, equipment and legal authority. Institutional outputs include enacted rules, published procedures, signed agreements and completed evaluations. Implementation outcomes include functioning inspection systems, enforced requirements and demonstrated operational controls. Observed consequences include changes in deployment practices, successful corrective action, independently assessed compliance and the ability of affected persons to obtain remedies. The higher the level of claimed effectiveness, the stronger the supporting evidence must be.

The monitoring framework should also identify indicators capable of weakening the principal assessment. A study that collects only evidence of regulatory progress will exaggerate institutional capacity. Evidence of continuing nonparticipation, inadequate technical resources, inconsistent national interpretation or inability to investigate relevant incidents must carry weight. Similarly, a study focused exclusively on regulatory failure may overlook improvements in testing, procurement and national accountability that occur without a universal treaty. Indicators should be selected according to their diagnostic value, not their consistency with a preferred political conclusion.

Strategic monitoring matrix: observations, thresholds and decision consequences

Indicator	Verified baseline at 5 Nov 2026	Decision-relevant threshold or signpost	Record to monitor	Implication
CCW negotiating mandate	Consensus elements adopted 4 Sep 2026; Review Conference pending	Formal mandate to negotiate an instrument	CCW Seventh Review Conference final decisions	Distinguishes technical agreement from treaty negotiation
Legal character of future instrument	Not predetermined	Draft treaty text establishing binding obligations	Official CCW negotiating documents	Determines enforceability
Prohibited autonomous functions	Agreed working characterisation; no new universal treaty ban	Specific prohibited functions and covered systems	Authenticated instrument text	Determines operational impact
Human-control obligations	Context-dependent measures recognised in consensus report	Mandatory testable requirements adopted nationally or internationally	National legal and procurement records	Measures substantive institutionalisation
Independent verification	No general global autonomous-weapons inspectorate established by the agreed report	Agreed external access, reporting or inspection procedures	Treaty implementation and oversight documents	Determines independent compliance assessment
National implementing legislation	Existing national legal frameworks differ	Enactment of specific implementing laws after treaty adoption	National gazettes and parliamentary records	Determines domestic enforceability
Public military compliance evidence	Public records are incomplete	Published review procedures, verified restrictions or independent findings	Defence ministries and competent regulators	Reduces uncertainty about actual practice
EU high-risk AI implementation	Revised timetable adopted July 2026	Relevant rules commence 2 Dec 2027 and 2 Aug 2028; transitional duties tracked to 2030	EU legislation and Commission implementation reports	Measures civilian regulatory implementation
UN global AI dialogue	First meeting July 2026; 163 country delegations	Substantive 2027 outcomes linked to national action	UN Co-Chairs’ summary and national records	Measures translation of discussion into policy
Scientific assessment capacity	UN Panel and annual assessment process established	Published assessments with transparent evidence and independent participation	UN scientific reports	Measures shared knowledge, not enforcement
Corporate AI incident reporting	Requirements depend on applicable law and commencement dates	Verified reporting, investigation and corrective action	Competent market-surveillance authorities	Measures operational accountability
Public computing access	National and regional programmes announced or under implementation	Verified available capacity and published allocation results	Programme operators and public audit records	Tests distributive implementation
Regulatory capacity	Authorities possess differing mandates and resources	Published budgets, staffing, technical capability and enforcement results	National budgets and regulator reports	Measures practical supervision
Technology-access restrictions	Export licensing and other controls remain jurisdiction-specific	Material changes to licensing scope or actual supply	Official export-control decisions	Measures geopolitical access conditions

The monitoring thresholds are institutional events or observable conditions rather than fabricated numerical targets. They allow governments to evaluate whether the governance trajectory has changed without pretending that diplomatic complexity can be reduced to an arbitrary index.

Near-term decision calendar: 2026–2030

The final months of 2026 and the first half of 2027 contain several relevant institutional milestones. The most immediate is the Seventh CCW Review Conference in Geneva on 16–20 November 2026. The conference will consider the September expert-group report and determine the modalities of future work. Its outcome is particularly diagnostic because the preceding expert process has already produced agreed instrument elements. A decision to establish negotiations would constitute a different political commitment from a decision to continue technical discussion or maintain a general mandate for further work. The distinction should be assessed from the adopted conference documents, including the precise wording of any mandate and whether it permits negotiations toward a legally binding instrument. Seventh Review Conference of the High Contracting Parties to the CCW — United Nations Office at Geneva — November 2026.

The next identifiable international governance milestone is the 3–4 May 2027 meeting of the UN Global Dialogue on AI Governance in New York. The Dialogue is designed for inclusive discussion rather than treaty negotiation. Its results can therefore influence national priorities and international coordination without producing binding legal obligations. The relevant monitoring question is whether participating governments use its scientific and policy findings to support concrete national or international measures, particularly in areas concerning technical capacity, human oversight and accountability. The UN’s official institutional record identifies the 2027 Co-Chairs as the permanent representatives of Saint Kitts and Nevis and the Republic of Korea, appointed on 16 October 2026. Global Dialogue on AI Governance — United Nations — 2026; Appointment of the Co-Chairs of the 2027 Global Dialogue on AI Governance — United Nations General Assembly — October 2026.

The European regulatory calendar provides additional measurable dates. The revised high-risk AI implementation schedule requires attention to 2 December 2027 and 2 August 2028, depending on the system category. These dates concern the relevant provisions of the EU AI Act rather than autonomous-weapons prohibitions. Their analytical significance lies in the development of practical supervision, conformity procedures and corporate documentation across a major regulatory market. Evidence from those processes may inform broader discussions about assurance and accountability, although it cannot automatically determine the legality of military applications. The regulation also contains transitional provisions relevant to certain systems used by public authorities extending to 2 August 2030. AI Omnibus Enters into Force — European Commission — July 2026; Regulation (EU) 2024/1689, consolidated version of 27 July 2026 — European Parliament and Council of the European Union — July 2026.

Beyond these established dates, a restrictive military governance timeline cannot be specified credibly. A negotiating mandate could be adopted without guaranteeing treaty completion. Treaty adoption could occur without rapid ratification. Ratification could be followed by lengthy implementation, especially where existing military systems require modification or new inspection capabilities must be established. The relevant forecast must therefore use conditional milestones rather than assigning invented completion dates.

Diplomatic decision options: strengthening governance without conflating mandates

Governments possess several possible courses of action that differ in their legal authority, implementation burden and strategic consequences. They can pursue negotiations on binding restrictions, strengthen domestic assurance procedures, cooperate on technical standards, develop confidential evidence-sharing mechanisms or invest in the capacity of less-resourced states to implement agreed obligations. These measures are not mutually exclusive, but they should be evaluated separately because their legal foundations and implementation risks differ.

A government supporting restrictive arms control may still need to strengthen its national weapons-review procedures. A government reluctant to accept external military inspections may nevertheless support common testing methods or more transparent national policy. A government prioritising technological access may invest in public computing while maintaining export controls on sensitive equipment. The existence of these combinations is important because political agreement on every element of global AI governance is unlikely to be a prerequisite for progress in narrower areas.

The following policy options are assessed according to authority, expected effect, implementation burden, time-to-effect, reversibility and principal risks. Their inclusion is intended to support ministerial decision-making rather than to assign political merit to individual governments.

Policy option A — Negotiate a legally binding autonomous-weapons instrument

The competent authority would be the participating states acting through an authorised international negotiating process. The immediate objective would be to establish legally precise restrictions on defined autonomous-weapons functions and to determine the obligations applicable to development, procurement, deployment and use. Such negotiations would require agreement on the legal character of the instrument, its scope, the treatment of existing systems, implementation procedures and verification arrangements. A treaty could produce stronger legal constraints for participating states, but its operational impact would depend on the prohibitions adopted and on participation by states possessing relevant capabilities.

The implementation burden would be substantial where governments must modify procurement programmes, create compliance authorities or provide access to sensitive technical information. The time-to-effect would extend beyond the adoption of a negotiating mandate because a treaty must be concluded, become binding according to its terms and be implemented. Reversibility would be limited once a state becomes bound by a treaty, subject to the instrument’s provisions and applicable international law. Principal risks include insufficient participation, ambiguous definitions, extensive exceptions and verification mechanisms that prove politically unacceptable or technically inadequate.

The appropriate initial decision threshold is whether the November 2026 CCW Review Conference adopts a mandate permitting substantive negotiations. The diplomatic foundation is already documented in the September consensus report and the subsequent October joint statement by 50 states. CCW/GGE.1/2026/3, paragraphs 45–46 — United Nations Office for Disarmament Affairs — September 2026; Joint Statement on Lethal Autonomous Weapon Systems — New Zealand Ministry of Foreign Affairs and Trade — October 2026.

Policy option B — Establish mandatory national assurance and weapons-review requirements

The relevant authority would be national defence ministries, military acquisition organisations and other competent bodies acting under applicable domestic and international law. The objective would be to make national approval processes more rigorous by requiring systematic documentation of intended functionality, foreseeable operating conditions, technical limitations, human-control arrangements and material system changes. This approach could be implemented without awaiting a new international treaty, although the precise obligations would differ according to national law.

The expected effect would be stronger evidence supporting deployment decisions and greater capacity to investigate failures. Its implementation burden would include technical personnel, independent review capability, testing infrastructure, documentation systems and training. Some requirements could be introduced through procurement rules and administrative directives, while others might require legislative authority. The time-to-effect would depend on existing institutional capabilities and whether systems already in service must undergo reassessment.

The principal risk is creating a formal approval structure that lacks substantive independence or sufficient technical access. Such a system could generate extensive documentation without improving operational control. Its effectiveness should therefore be assessed through actual review decisions, the conditions imposed on deployment and the ability of competent authorities to reject or suspend applications when evidence is insufficient.

The ICRC’s guide to weapons reviews offers an established basis for developing national procedures, while the September CCW report identifies measures that states have accepted as part of a common negotiating foundation. A Guide to the Legal Review of New Weapons, Means and Methods of Warfare — International Committee of the Red Cross — June 2020; CCW/GGE.1/2026/3, paragraphs 36–39 — United Nations Office for Disarmament Affairs — September 2026.

Policy option C — Develop recognised technical assurance and confidential review procedures

This option would involve national technical authorities, defence establishments, recognised testing organisations and, where legally authorised, international partners. Its objective would be to reduce uncertainty concerning system behaviour and facilitate compatible assurance procedures without requiring unrestricted public disclosure of sensitive information. Participating institutions could develop common evaluation terminology, evidence-preservation methods, software-integrity requirements and procedures for assessing operator intervention capability.

The expected effect would be greater comparability of technical evidence and reduced duplication where states recognise compatible procedures. Implementation costs would arise from specialised testing infrastructure, technical training and secure information-handling systems. The time-to-effect could be shorter for common terminology and evaluation methods than for the establishment of an international inspectorate with compulsory powers. However, the institutional arrangements would remain limited by national security restrictions and the authority granted to participating bodies.

The principal risk is that recognised procedures may become substitutes for independent assessment rather than instruments supporting it. Certification can create confidence only when the scope of testing and the limitations of conclusions are clear. A system approved for one operational environment should not automatically be presumed compliant in another. Technical cooperation therefore needs explicit mechanisms for revising assessments when system capabilities or intended uses change.

The NIST AI Risk Management Framework provides an example of systematic lifecycle governance, including the importance of testing, documentation and reassessment. It is a voluntary framework and cannot be interpreted as a military treaty or universal legal standard. AI Risk Management Framework Core — United States National Institute of Standards and Technology — January 2023.

Policy option D — Strengthen incident reporting, investigation and access to remedies

The competent authority would include national legislators, regulators, courts and other investigative institutions acting within their jurisdiction. The objective would be to ensure that significant AI-related incidents can be identified, investigated and addressed through legally appropriate procedures. This could involve mandatory reporting for specified categories of harm, preservation of relevant records, authority to require technical information and mechanisms for corrective action.

The expected benefit would be a stronger connection between declared accountability and actual legal consequences. Implementation burden would fall on providers, deployers and competent authorities, requiring incident-management systems, investigation capacity and procedures for protecting confidential information. Some mechanisms are already present within existing regulatory regimes. For example, Article 73 of the EU AI Act establishes reporting obligations for certain serious incidents involving high-risk AI systems within its scope, including differentiated reporting periods according to the nature of the incident. Its legal significance depends on the applicable provisions and commencement dates, and it does not establish a universal international reporting regime for military operations. Regulation (EU) 2024/1689, Article 73: Reporting of Serious Incidents — European Commission AI Act Service Desk — 2026.

The principal risk concerns inadequate evidence access or procedures that encourage superficial reporting without meaningful investigation. Excessively broad disclosure obligations can also create legitimate security and privacy concerns. A proportionate regime should therefore specify which incidents trigger reporting, what evidence must be preserved, which authorities may examine it and how corrective measures are enforced.

Policy option E — Build independent technical capacity in less-resourced jurisdictions

This option would involve national governments, international development institutions, research organisations and infrastructure providers. Its objective would be to improve the capacity of states and smaller organisations to evaluate AI systems independently and comply with applicable governance requirements. Relevant activities could include training technical assessors, establishing shared evaluation facilities, supporting public computing access and assisting regulators in developing reliable incident-investigation procedures.

The expected effect would be to reduce the disparity between formal participation in governance and practical capacity to implement it. The implementation burden would involve sustained financing, equipment, trained personnel and institutional continuity. Unlike a single technology transfer, durable capacity-building requires arrangements for maintenance, security, knowledge retention and independent operation.

The principal risk is creating externally dependent institutions whose technical resources or expertise remain controlled by their sponsors. Capacity-building should therefore be assessed according to whether recipient institutions acquire the ability to make independent technical and legal judgments. The UN Global Dialogue identifies AI capacity-building and participation by less-resourced countries as explicit institutional objectives, but the Dialogue’s mandate does not itself finance or enforce every national implementation measure. Global Dialogue on AI Governance: Frequently Asked Questions — United Nations — 2026.

Comparative policy-options matrix

Policy option	Competent authority	Intended effect	Implementation burden	Time-to-effect	Reversibility	Principal risk
Binding autonomous-weapons instrument	Participating states and authorised negotiating bodies	Prohibit or restrict specified functions	Negotiation, ratification, legislation, verification	Depends on treaty process and entry into force	Limited after binding commitment, subject to treaty law	Nonparticipation and ineffective verification
Mandatory national weapons reviews	Defence ministries and legally competent authorities	Improve lawful deployment decisions	Technical assessors, testing and documentation	Can begin under existing authority; varies by state	Procedures can be amended, subject to legal duties	Formal approval without substantive scrutiny
Recognised technical assurance	Standards, testing and competent regulatory bodies	Improve evidence quality and interoperability	Specialist infrastructure and secure procedures	Progressive implementation	Technical methods can be revised	Certification exceeding demonstrated assurance
Incident reporting and remedies	Legislators, regulators, courts and investigators	Improve accountability after significant harm	Reporting, evidence retention and investigation	Depends on applicable legal commencement	Legal rules amendable through competent procedures	Inadequate evidence access and superficial compliance
International capacity-building	Governments and development institutions	Reduce implementation inequality	Sustained financing, expertise and equipment	Multi-year institutional development	Programmes adjustable; infrastructure commitments less reversible	Continuing external dependence
Public procurement safeguards	Government purchasers and acquisition authorities	Align supplier obligations with public responsibility	Contract management, testing and oversight	Can affect new acquisitions directly	Contract-dependent	Supplier concentration and incomplete technical control
Confidential cross-border cooperation	Competent national authorities and treaty partners	Improve access to sensitive compliance evidence	Legal agreements, security controls and trust	Depends on bilateral or multilateral arrangements	Agreements may be amended under their terms	Information exposure and disputed findings

The comparison indicates that different measures address different failure points. Treaty negotiations establish potential new restrictions. National reviews assess the legality of specific capabilities. Technical assurance produces evidence about behaviour. Incident procedures support investigation and remedies. Capacity-building affects the ability of less-resourced states to participate. Procurement controls connect public authority with commercial suppliers. A coherent governance architecture may require several of these mechanisms, but their adoption should be evaluated through the legal powers and practical resources actually available.

The role of deterrence: enforcement must change incentives

International governance is effective only to the extent that it influences behaviour under conditions in which actors possess incentives to depart from the relevant rules. A government may comply because it considers the rule legitimate, because compliance supports military cooperation, because violation would create legal consequences or because noncompliance would damage diplomatic and commercial relationships. A company may comply because of statutory duties, contractual requirements, reputational considerations or the risk of losing market access. These incentives differ, and their effectiveness depends on the surrounding institutional environment.

A restrictive treaty may create strong legal obligations but weak practical deterrence if violations are difficult to detect and responses uncertain. A national procurement framework may influence suppliers significantly because failure to comply can result in rejection of a product or loss of contractual eligibility. A regulatory system may exercise substantial influence through the ability to demand information and impose corrective measures within its jurisdiction. Voluntary standards may become commercially important when customers require conformity as a condition of purchase. The strength of governance therefore cannot be inferred solely from whether a rule is labelled binding or voluntary. Binding status matters legally, while practical influence depends on verification, enforcement and the incentives facing the regulated actor.

In military contexts, the relevant considerations include the consequences of unilateral restraint. A state may perceive compliance with restrictive rules as valuable for humanitarian protection, legal legitimacy and strategic stability, but may also fear that a noncompliant adversary could exploit the resulting operational disadvantage. Effective arms control must address this concern through sufficiently precise obligations, reciprocal commitments, verification and mechanisms for responding to violations. The absence of complete reciprocity does not make legal obligations irrelevant. It does, however, affect the political willingness of states to accept additional constraints and the military planning necessary to implement them.

Key uncertainties through 2030

The first uncertainty concerns the legal outcome of multilateral negotiations. The September 2026 consensus report creates a stronger foundation for future discussions, but the legal form of any resulting instrument remains unresolved. The November Review Conference may clarify the negotiating trajectory, although subsequent agreement would still be necessary.

The second concerns the military feasibility of proposed restrictions. States may agree that human responsibility must be maintained while disagreeing over whether particular automated engagement functions are compatible with applicable law. The final scope of any prohibition will determine whether it materially affects existing programmes or primarily codifies restrictions that participating states already accept.

The third concerns verification. Software configurability, the sensitivity of military information and differences between tested and operational environments complicate compliance assessment. The development of credible testing and evidence-preservation methods could reduce uncertainty, but independent international verification would also require legal authority and political agreement.

The fourth concerns the distribution of industrial capability. Advanced computing resources, specialised technical personnel and evaluation infrastructure remain unevenly distributed. Common regulatory obligations could improve safety while imposing unequal compliance costs. The extent to which capacity-building alters that distribution will depend on actual infrastructure access and sustained institutional development.

The fifth concerns the relationship between national regulatory regimes and international coordination. The EU’s implementation timetable provides specific opportunities to observe the practical effects of high-risk AI obligations, while other jurisdictions continue developing different approaches. Divergence may create costs and uncertainty, but some differences reflect legitimate distinctions in legal authority and national priorities.

The sixth concerns accountability after harm. More extensive documentation and incident reporting may improve investigations, yet legal responsibility will continue to depend on jurisdiction-specific duties, causation, evidence and procedural access. A system can be technically complex without becoming legally responsible in place of the relevant persons or institutions.

Confidence assessment of the principal strategic judgments

Principal judgment	Confidence	Supporting basis	Main uncertainty
Existing IHL applies to autonomous weapons	High	Applicable legal instruments and CCW consensus language	Interpretation in specific operational circumstances
The September 2026 CCW agreement is not itself a treaty prohibition	High	Express qualifications in the adopted final report	Future decisions of the Review Conference
New binding restrictions require additional political agreement	High	Present negotiating mandate and treaty requirements	Scope and speed of future negotiations
Independent verification remains technically demanding	High	Software configurability, operational context and evidence requirements	Effectiveness of future assurance methods
Partial institutionalisation can proceed under national authority	Medium-high	Existing regulatory, procurement and weapons-review mechanisms	National resources and political willingness
A restrictive treaty could influence covered military programmes	Medium	Legal mechanism and conditional effects of treaty restrictions	Unknown treaty scope and participation
Compliance costs may reinforce market concentration	Medium	Fixed-cost and scale mechanisms	Actual regulatory design and market responses
Technology-access inequalities will persist without specific capacity policies	Medium-high	Unequal infrastructure and implementation resources	Future investment, efficiency and access arrangements
Comprehensive global restrictions will be operational by 2030	Low	Insufficient evidence for a reliable forecast	Negotiations, ratifications and verification
Common technical assurance can reduce some governance fragmentation	Medium	Existing technical and institutional mechanisms	Adoption, recognition and enforcement

Confidence levels refer to the strength of the analytical judgments, not to numerical probabilities of future events. High confidence does not mean that every future development is known; it means that the judgment is strongly supported by the available legal, institutional and technical evidence. Low confidence indicates that decisive information is unavailable or that the outcome depends on political choices not yet made.

Final net assessment: the difference between ethical authority and enforceable power

The strategic implications of the five propositions concerning artificial intelligence can now be assessed across the full institutional, military and industrial framework. Their principal significance lies in establishing a clear distinction between technological capacity and the legitimate authority to determine how that capacity is used. The propositions insist that human beings remain responsible for consequential decisions, that technology serve broader social interests and that international institutions possess sufficient capacity to translate common principles into practice. These claims have substantial normative coherence. Their enforceability, however, depends on legal authority, technical verification, institutional competence and the willingness of states and corporations to accept obligations that may conflict with immediate operational or commercial interests.

The military domain exposes the sharpest tension. International humanitarian law already constrains weapons and attacks, and the September 2026 CCW consensus report recognises human judgment, operational restrictions and continuing responsibility as central elements of governance. The unresolved question is whether these principles will produce additional prohibitions on specified autonomous capabilities and whether those prohibitions can be implemented and verified. The agreement on instrument elements demonstrates that substantive multilateral convergence is possible. It does not establish that the principal military powers have accepted a restrictive treaty or the external scrutiny necessary for robust compliance.

The institutional domain presents a different limitation. International scientific assessment and political dialogue can improve the availability of shared knowledge and broaden participation, but they do not automatically create enforcement powers. Regional and national legal regimes can impose more specific obligations, yet their jurisdiction remains limited by applicable law. The resulting architecture can become more effective through cooperation, common technical standards and improved national implementation without necessarily producing a central global regulator. Such progress would be meaningful, but it should not be mistaken for universal legal control over AI development.

The industrial domain adds a further constraint because technological capability is distributed through markets and supply chains rather than exclusively through public institutions. Access to advanced computing, software, electricity, capital and technical expertise affects the capacity of countries and organisations to develop and govern AI. States controlling critical inputs possess leverage that can shape technological access independently of multilateral ethical commitments. Corporations controlling widely used platforms can influence commercial conditions and technical standards. Equitable governance therefore requires attention to industrial capacity and access, not merely the formal inclusion of countries in diplomatic discussions.

The European and transatlantic evidence demonstrates the coexistence of these forms of power. The EU possesses significant regulatory authority within its legal scope and is investing in technological infrastructure, while national governments pursue differentiated industrial strategies. The United States and other major technology-producing states influence access to advanced inputs and participate in competing regulatory and security frameworks. These arrangements create opportunities for cooperation but do not eliminate disagreements over trade, national security, industrial competitiveness or the degree of autonomy permissible in military systems.

The three governance scenarios illustrate how these tensions may evolve through 2030. Normative influence can continue without substantial new international restrictions. Partial institutionalisation can strengthen technical assurance, human-control requirements and accountability under national authority. A restrictive regime can establish binding prohibitions for participating states, but its practical effectiveness will depend on scope, verification and participation. None of these trajectories can be assigned a defensible numerical probability from the available record, and elements of more than one scenario may coexist across jurisdictions.

The most consequential near-term indicator is the outcome of the November 2026 CCW Review Conference. A formal negotiating mandate would establish a new diplomatic stage. Subsequent agreement on precise prohibitions, effective implementation and independent compliance assessment would provide stronger evidence that the international system is moving beyond normative influence. Failure to establish such arrangements would not eliminate existing legal obligations or domestic governance progress, but it would preserve substantial national discretion over autonomous military development.

Final strategic judgment — High confidence: The central limitation of global AI governance is not the inability to formulate ethical principles. It is the difficulty of converting those principles into obligations that remain effective when they constrain military advantage, commercial freedom or national technological sovereignty. The decisive measure of progress through 2030 will therefore be the extent to which governments and institutions can demonstrate that human responsibility remains legally attributable, operational controls remain effective, prohibited conduct can be identified, and violations can be investigated and addressed. International agreement on values can support these outcomes, but it cannot substitute for the legal authority, technical capacity and political commitments required to achieve them.

The final distinction is between the legitimacy of a rule, the authority to enforce it and the capacity to verify compliance. Ethical authority may influence what governments regard as acceptable. Legal authority determines which obligations bind particular actors. Technical and institutional capacity determines whether those obligations can be implemented and assessed in practice. A governance regime that possesses only the first element can shape debate but cannot reliably constrain conduct. A regime possessing the first two elements but lacking credible verification may create enforceable obligations whose compliance remains difficult to establish. A regime combining all three would constitute a materially different form of international AI governance, although its geographic reach would still depend on participation and the scope of its legal instruments.

As of 5 November 2026, the international system has developed important elements of the first two conditions but has not established a comprehensive global mechanism satisfying all three across civilian AI and autonomous military weapons. Whether that institutional gap narrows by 2030 will depend on decisions that remain open, particularly the adoption of binding autonomous-weapons restrictions, the development of credible national and international compliance procedures, and the distribution of the technical resources necessary to implement them.

The report’s final conclusion is that moral consensus can establish the objectives of AI governance, but only politically accepted restrictions, competent institutions and verifiable implementation can determine its practical limits.

---

Copyright of debuglies.com – Even partial reproduction of the contents is not permitted without prior authorization – Reproduction reserved