FaultLines
The dual-use AI governance gap that no treaty covers

Export controls were built to contain weapons of mass destruction. They are now the primary AI governance mechanism for 4.6 billion people in Asia — without safety logic, without dual-use doctrine, and without a single Asian state at the table when they were designed. The only binding international AI treaty exempts national security entirely. The gap is structural, documented, and unaddressed by any existing multilateral framework — including the thematic architecture of the UN Global Dialogue on AI Governance.

AI Diffusion Rule — rescinded May 2025 CoE Treaty Art. 3.2 — national security exempt UN Global Dialogue on AI Governance — A/RES/79/325 Apart Research · Global South AI Safety Hackathon 2026
0
Binding AI treaties with dual-use provisions
4.6B
People in Asia governed by WMD containment logic
7
Governance layers required — AI has 1, broken
2025
Year the only AI governance mechanism for Asia was rescinded — no replacement exists
Vibha Amarnath
Sources: US BIS · Federal Register · Council of Europe CETS 225 · Vietnam National Assembly · OECD AI Observatory · METR · UN A/RES/79/325
How to use this tool
Policy researchers and analysts
Start with Governance Stack — the seven-layer WMD comparison is the core research contribution. It maps exactly what is missing and why, at layer level and country level. Precedents provides the mechanism-level design specifications: six WMD governance mechanisms mapped to their AI equivalent with institutional design requirements and draft treaty language. This is the material that fills the gap analytically.
Civil society, journalists and public interest advocates
Every claim is linked to a primary source document. Start with Why the Vacuum Persists — it documents five confirmed AI deployments in Asian critical infrastructure, names the specific institutions that should be governing them and cannot, and explains the structural reason no existing body can act. The Briefing tab gives the complete evidenced timeline with every source linked.
Policymakers, legislators and treaty practitioners
Dialogue Readiness scores the thematic architecture of any multilateral AI governance process against the dual-use gap — with a structural argument per theme for why the gap is outside current scope, not merely under-addressed. Accountability maps four risk scenarios to draft treaty clauses modelled on WMD precedents. The structural failures here are not session-specific — they are gaps in the global governance architecture that the next binding instrument must address.
The multilateral process — why this moment is consequential

The UN Global Dialogue on AI Governance and the 2027 Treaty Mandate

United Nations General Assembly Resolution A/RES/79/325, adopted December 2025, established the first multilateral AI governance process with a binding treaty mandate and genuine Global South inclusion. All 193 UN member states are represented. The resolution explicitly requires that Global South voices be centred — responding to a documented critique that every major AI governance instrument to date has been designed by and for a small group of Western governments and companies.

The process runs in two sessions: an initial dialogue phase building shared understanding, followed by treaty negotiations at the 2027 New York session. The positions advanced and the language agreed in the dialogue phase directly shape the architecture of the first international AI treaty with genuine global scope. This tool documents the structural gap that the treaty must address — and provides the mechanism-level design specifications for doing so.

Seven official thematic areas — A/RES/79/325
1. Safe, secure and trustworthy AI
2. AI capacity-building
3. Social, economic, ethical and cultural implications
4. Interoperability of governance approaches
5. Protection and promotion of human rights
6. Transparency, accountability and human oversight
7. Open-source software, open data and open AI models
None of these themes explicitly addresses export controls or dual-use AI governance. This is not a gap in the dialogue process alone — it reflects the absence of dual-use doctrine from the entire international AI governance architecture. This tool documents that absence and specifies what filling it requires.
The situation

The Structural Failure — What Every Practitioner Needs to Know

Export controls were designed for weapons of mass destruction (WMD). They became the primary AI governance mechanism for Asia by default — not by design. The architecture collapsed before it came into force. No replacement exists. The only binding international AI treaty exempts the highest-risk use cases entirely. This is the origin of the governance vacuum this tool maps. The vacuum matters for AI safety specifically because ungoverned dual-use AI in critical infrastructure is a documented pathway to the harms AI safety research is designed to prevent: biosecurity uplift for non-state actors, electoral interference at scale, and lethal autonomous systems without accountability or attribution.

The core finding: The dual-use AI governance stack requires seven functional layers. For AI, one layer exists — compute export controls — and it was rescinded two days before compliance. Nuclear governance took 25 years and two weapon deployments to build the equivalent stack. AI is moving faster and has almost nothing.

The collapse timeline

October 2022
US Bureau of Industry and Security (BIS) imposes first AI chip export controls
NVIDIA A100 and H100 banned for export to China and Russia. Export Control Classification Numbers (ECCNs) 3A090 and 4A090 created under the Export Administration Regulations (EAR). First application of WMD containment logic to AI hardware — with no AI safety mandate behind it.
January 13, 2025
Biden AI Diffusion Rule published — Federal Register Vol. 90 No. 10 (90 FR 4544)
Most ambitious AI export control framework attempted. Three-tier country system covering 120+ countries. First-ever controls on AI model weights (ECCN 4E091). Compliance date: May 15, 2025.
↗ Federal Register 90 FR 4544 — Full Rule Text
May 13, 2025 — two days before compliance
Trump administration rescinds the Diffusion Rule
BIS cites that the rule "would have stifled American innovation." No replacement published. The three-tier framework never came into force. No replacement rule exists as of June 2026.
↗ BIS Official Rescission Statement
May 2025 — ongoing
Regulatory vacuum — patchwork guidance only
What currently applies: pre-2022 EAR baseline chip controls on China, Entity List restrictions (42+ Chinese entities added 2025), BIS enforcement guidance on Huawei Ascend chips, January 2026 case-by-case China chip review rule. No framework governs Asian Tier 2 states in law.
September 2024
Council of Europe (CoE) Framework Convention on Artificial Intelligence and Human Rights, Democracy and the Rule of Law opens for signature — CETS No. 225
First binding international AI treaty. Article 3.2 provides blanket exemption for "all activities within the lifecycle of AI systems related to the protection of national interests." The highest-risk dual-use AI applications are entirely outside the treaty's scope. India, Vietnam, Indonesia, Singapore — not party.
↗ CoE CETS No. 225 — Full Treaty Text
2026–2027 — the treaty window
UN Global Dialogue on AI Governance → 2027 New York Treaty Session
A/RES/79/325 establishes a two-phase process: a dialogue phase (2026) building shared understanding, followed by binding treaty negotiations in New York (2027). 193 member states. This is the first multilateral AI governance process with genuine Global South inclusion and a treaty mandate. Export controls and dual-use AI are not among the seven official thematic areas — which is the structural argument this tool makes.
↗ UN Global Dialogue — Official Site

Critical gap
No dual-use doctrine for AI

Nuclear has the Nuclear Non-Proliferation Treaty (NPT) and the Nuclear Suppliers Group (NSG). Chemical has the Chemical Weapons Convention (CWC) and the Australia Group. Biological has the Biological Weapons Convention (BWC). AI has no agreed definition of what constitutes a dual-use capability, no carve-out distinguishing offensive from defensive use, and no international verification mechanism. The same model that enables pandemic detection enables pathogen design. No governance architecture can distinguish them.

Critical gap
No Asian voice in export control design

The US BIS — operating under a WMD containment mandate with no AI safety objective — designed the only binding AI governance mechanism for Asia. India, Vietnam, Indonesia, Singapore had no input. The architecture reflects US geopolitical objectives, not Asian security needs, not dual-use safety logic, not the governance capacity of the states it governs.

Structural failure
Controls redirect, not prevent

In January 2025, Indonesia's second-largest telecom Indosat Ooredoo Hutchison signed an MoU with AIonOS to develop AI solutions built on DeepSeek for food security, tourism, and talent development. In May 2025, Malaysia's Deputy Communications Minister announced a sovereign AI infrastructure powered by Huawei Ascend GPUs and DeepSeek — then retracted the statement under US export control pressure; the episode itself documents a country caught between US restrictions and Chinese infrastructure with no multilateral framework governing the choice. Singapore's OCBC — a systemically important bank — has deployed 30+ GenAI applications and its 2024 Annual Report specifically cites DeepSeek reasoning models as a capability it is adopting, with no safety evaluation standard applying to Chinese-origin models regardless of scale. Export controls did not prevent AI development in these countries — they redirected it toward infrastructure with no safety architecture, no transparency requirements, and no incident reporting. The highest-risk countries ended up on the least-governed infrastructure.

Theory of change
If the structural gap is documented with mechanism-level specificity, treaty practitioners have a design specification rather than a general argument. Design specifications survive negotiation. General calls for better AI governance do not. The 2027 treaty negotiation is the proximate intervention point. This tool's contribution is the specification: what each missing governance layer requires institutionally, what WMD precedent it maps to, and what the political bargain that makes developing-state participation viable looks like. The gap is not inevitable. It is addressable. The architecture exists — in the NPT, the CWC, the OPCW, ICAO Annex 19. What is missing is the will to apply it to AI before a catastrophe makes the case.
Multilateral AI governance gap analysis

Dialogue Readiness — What the Thematic Architecture Misses

Resolution A/RES/79/325 identified seven thematic areas for the multilateral AI governance process. None explicitly addresses export controls or dual-use AI governance. Each theme is scored on how well the current international governance landscape addresses the dual-use dimension — and what the structural gap means for the treaty design process, regardless of which session or forum is being addressed.

Scoring methodology — sourced from two external indices
Oxford Insights Government AI Readiness Index 2024
Scores 188 countries across 40 indicators in three pillars: Government Capacity, Technology Sector, and Data & Infrastructure. Used here as the governance capacity baseline for each country. Referenced as UNESCO and G20 benchmark. Scale 0–100. ↗ Full Report
OECD AI Policy Observatory Index 2024
Tracks AI policy implementation across OECD members against the OECD AI Principles (updated 2024). Used here for policy framework maturity. Coverage note: India, Vietnam, and Indonesia are not OECD members — their absence from OECD scoring is itself a governance gap finding. ↗ OECD AI Index
How the theme scores are calculated: Each thematic score (1–10) is a weighted composite: 40% governance capacity gap (inverse of Oxford Insights Government AI Readiness Index 2024 score, normalised), 40% dual-use risk exposure (BIS tier status + documented infrastructure pivot risk), 20% OECD AI Policy Observatory framework presence (0 for non-members, which is itself a finding). The dual-use and export control exposure layer is original to this tool — no existing index captures the post-May 2025 rescission status. Scores reflect the export control and dual-use dimension within each theme only, not the theme's overall treatment in global governance.
Score: 1 = export control / dual-use dimension entirely unaddressed within this theme. 10 = fully addressed. No theme scores above 4. Click each theme to expand the gap analysis and the structural argument for any multilateral AI governance process.
Why low scores are structural, not incidental: Even if each thematic area were perfectly executed — safe AI fully realised, capacity-building fully funded, human rights fully protected — the dual-use AI governance vacuum would remain. The CoE Convention's Article 3.2 national security exemption removed dual-use capability assessment from treaty scope at the same time these thematic areas were being defined. Export controls and dual-use doctrine are not a gap in execution. They are outside the architecture of every existing multilateral AI governance framework. The next binding instrument must add them explicitly or they will not appear.
Theme 1
Safe, Secure and Trustworthy AI
2/10
Click to expand

Safety frameworks exist only for Western frontier labs. No equivalent applies to Chinese frontier models now deployed across Asian critical infrastructure. The CoE Convention exempts national security — exactly where the highest-capability AI systems are being deployed. "Trustworthy" AI cannot be assessed for models where training data, safety evaluations, and incident history are proprietary and ungoverned.

Structural argumentSafety evaluations cannot be verified for AI deployed outside Western governance. Any binding AI governance instrument must mandate that evaluations equivalent to UK AI Safety Institute pre-deployment assessments apply to all frontier models deployed in member states, regardless of origin. The CoE national security exemption (Article 3.2, CETS 225) must be narrowed in the 2027 treaty to exclude dual-use capability assessments.
↗ METR Common Elements of Frontier AI Safety Policies
Theme 2
AI Capacity-Building
1.5/10
Click to expand

Export controls create a structural paradox for capacity-building: the compute required to build AI safety evaluation capacity is the same compute Tier 2 restrictions limit. India cannot build an operational AISI without the frontier compute needed to evaluate frontier models — compute its developers are restricted from accessing. The capacity-building agenda cannot succeed while the primary governance mechanism actively constrains the hardware required for governance infrastructure.

Structural argumentAny binding AI governance instrument must establish a governance compute carve-out — a guaranteed pathway for AI Safety Institutes in developing states to access frontier compute for safety evaluation, independent of geopolitical tier classification. This mirrors the NPT's Article IV right to peaceful nuclear use: states accepting governance obligations receive guaranteed access in return.
Theme 3
Social, Economic, Ethical and Cultural Implications
3/10
Click to expand

Export controls are reshaping AI adoption across Asia with profound social and economic implications — redirecting development toward Chinese infrastructure, accelerating dual-stack adoption, and creating technology dependencies difficult to reverse. Indonesia's 277 million people face deepfake-driven electoral interference with no AI governance framework to address it. Vietnam's human-in-the-loop mandates (Decree 142/2026) cannot be operationalised without access to the models they are designed to govern.

Structural argumentExport controls must be part of the governance architecture's analysis of AI's social and economic implications for the Global South. Redirection of AI development toward ungoverned infrastructure is not a trade policy outcome — it is a safety and social harm outcome with documented consequences for 4.6 billion people.
Theme 4
Interoperability of Governance Approaches
1.8/10
Click to expand

Vietnam has a binding AI law and sits in the same restricted compute tier as Indonesia, which has no AI law at all. Governance quality is entirely disconnected from compute access tier. The EU AI Act, US voluntary frameworks, and Asian national laws operate in isolation. The CoE Convention has no mechanism for developing-state participation or recognition of equivalent national frameworks.

Structural argumentThe next binding AI governance instrument must establish a mutual recognition mechanism for national AI governance frameworks, with governance maturity linked to compute access — directly analogous to the NSG's recognition of IAEA safeguards as a prerequisite for nuclear technology transfer. Vietnam has binding AI law and the same compute access vacuum as Indonesia, which has none. The architecture cannot distinguish between them. It must.
↗ Vietnam Law No. 134/2025/QH15
Theme 5
Protection and Promotion of Human Rights
3.5/10
Click to expand

The CoE Framework Convention grounds AI governance in human rights — but exempts national security applications (Article 3.2). AI-enabled mass surveillance, autonomous targeting, and election interference — all dual-use applications with direct human rights implications — fall outside the treaty. No Asian developing state is party. The human rights framework most applicable to high-risk AI does not reach the countries where high-risk AI is being deployed without safety governance.

Structural argumentThe next binding AI governance instrument must explicitly bind dual-use AI applications to IHL and human rights law — closing the Article 3.2 national security gap. The UN High Commissioner for Human Rights should be tasked with assessing AI-enabled rights violations in the national security carve-out space, which currently has no oversight mechanism.
↗ CoE CETS No. 225 — Article 3.2
Theme 6
Transparency, Accountability and Human Oversight
2/10
Click to expand

No accountability mechanism exists for AI-enabled dual-use harm. When an AI system contributes to a biosecurity incident, electoral interference, or infrastructure attack, there is no investigation body, no attribution standard, no reporting requirement, and no liability framework that applies across jurisdictions. The OPCW has a challenge inspection mechanism. The NTSB has a no-blame investigation model. Aviation has a black box standard. AI has nothing equivalent.

Structural argumentAn international AI incident reporting standard — modelled on ICAO Annex 13 — is achievable without a treaty. ITU can develop this as a UN technical standard. It requires a no-blame reporting channel for near-miss events and a mandatory attribution protocol for AI-enabled security incidents. This addresses Layer 5 of the governance stack and does not require waiting for the 2027 treaty.
Theme 7
Open-Source Software, Open Data and Open AI Models
2.5/10
Click to expand

Open-weight models (Meta Llama, DeepSeek) partially bypass export controls — Asian states can access capable models without compute restrictions. But the safety governance that applies to closed models (AISI evaluations, Frontier Model Forum commitments) does not apply to open-weight models at all. Indosat's DeepSeek deployment and OCBC's adoption of DeepSeek reasoning models are open-weight: outside every safety governance framework that exists. This is not a temporary anomaly — it is structural. Open-weight models cannot be recalled, restricted post-release, or governed through access controls. The governance vacuum for open-weight AI is permanent under the current architecture.

Structural argumentOpen-weight models (DeepSeek, Llama) already partially bypass compute controls. A safety evaluation standard that does not depend on access control is required. The UN Scientific Panel on AI should assess open-weight dual-use capability, and the 2027 treaty must include open-weight AI in scope — the CoE Convention currently leaves this unaddressed, creating the largest near-term exploitable gap in the governance architecture.
↗ METR Common Elements — Open-Source Safety Policy Status
Why the vacuum persists

Why No Existing Actor Can Fill the Vacuum

The governance vacuum has persisted for three structural reasons — not because no one has noticed it, but because the actors who control the relevant levers have incentives and mandates that prevent them from filling it. These three findings explain why. Together they make the case that any binding AI governance instrument must create new institutional architecture rather than tasking existing actors with a mandate they structurally cannot fulfil.

Finding 1 of 3

The Same Four Companies Control Everything That Matters

The four companies with binding frontier AI safety commitments are the same four companies whose chips and cloud infrastructure Asian states are restricted from accessing. This is not a coincidence — it is the architecture. The countries excluded from compute access are excluded from safety governance by the same set of actors.

Frontier Compute Control
Hardware
NVIDIA (80–90% AI training chips) · TSMC (manufactures NVIDIA chips) · ASML (sole supplier of EUV lithography machines)
Cloud Infrastructure
AWS (33% global) · Azure (23%) · Google Cloud (12%) · Combined: 68% of global AI cloud
Safety Governance
Frontier Model Forum
Anthropic · OpenAI · Google DeepMind · Microsoft — the four companies that set voluntary safety standards for frontier AI globally
Safety Evaluators
UK AISI (evaluates Anthropic + OpenAI only) · METR (funded by Open Philanthropy, invested in Anthropic)
Export Control Design
Rule-Makers
US BIS (WMD containment mandate, no AI safety objective) · National Security Council · Semiconductor Industry Association lobby
Key Influence Event
NVIDIA CEO Jensen Huang met Trump, December 2025. Diffusion Rule rescinded May 2025. No replacement issued.
The overlap — who sits in all three circles
🟦
Anthropic
AWS cloud · Frontier Model Forum · UK AISI evaluations · Claude used in BIS policy briefings
🟦
OpenAI
Azure cloud · Frontier Model Forum · UK AISI + US AISI evaluations · NSC AI advisory access
🟦
Google DeepMind
GCP cloud · Frontier Model Forum · UK AISI evaluations · Hiroshima AI Process participation
🟦
Microsoft
Azure cloud · Frontier Model Forum · OpenAI investor · BIS Wassenaar Arrangement lobbying
Countries excluded from all three: India · Vietnam · Indonesia · Singapore — not consulted on export controls, not party to the CoE Convention, not included in Frontier Model Forum evaluations, not reached by AISI pre-deployment assessment.
Finding 2 of 3

Export Controls Redirected Asia to Ungoverned Infrastructure

The restriction of frontier US compute did not prevent AI development in Asia. It redirected it — toward Chinese infrastructure with no safety architecture, no transparency requirements, and no incident reporting. These are confirmed deployments, not projections. Each one sits entirely outside every governance framework discussed in this tool.

Country / Entity Deployment Confirmed Chinese Stack Safety Governance Applies? Incident Reporting? CoE Treaty?
🇮🇩 Indosat Ooredoo Hutchison (Indonesia)
Second-largest Indonesian telco · 100M+ subscribers
Signed MoU (Jan 2025) with AIonOS to develop solutions built on DeepSeek for food security, tourism, talent development; AI Centre of Excellence announced. By mid-2026 pursuing hybrid strategy with Nvidia H100 GPUs and own Indonesian LLM (Sahabat AI). DeepSeek + Alibaba Cloud None None Not party
🇲🇾 Malaysia — Skyvast Sovereign AI
Private initiative; MoU with Huawei Malaysia signed
Deputy Communications Minister announced sovereign AI on Huawei Ascend GPUs (May 2025); retracted next day under US export control pressure. Ministry of Digital had separately published an MoU between Huawei Malaysia and Skyvast for Ascend-powered sovereign AI cloud. Huawei denied selling chips; government denied involvement. Huawei Ascend (announced, disputed) None None Not party
🇸🇬 OCBC Bank (Singapore)
Systemically important bank · MAS regulated
30+ GenAI applications deployed (OCBC Annual Report 2024); DeepSeek reasoning models cited as capability being adopted; 300+ total AI use cases. Model-level provenance of individual applications not publicly disclosed. DeepSeek (cited); multiple providers None for Chinese models None Not party
🇹🇭 Thailand AI Sector
Government and private sector
Simultaneous investment from Microsoft and ByteDance — competing infrastructure 2024–25 ByteDance (dual-stack) Partial (Microsoft stack only) None Not party
🇻🇳 Vietnam Digital Economy
Law 134/2025 in force — best governed Tier 2
Alibaba Cloud data centres in region; growing DeepSeek deployment across enterprise sector Alibaba Cloud + DeepSeek (growing) Law 134/2025 (domestic only) Not for Chinese models Not party
Every deployment in this table sits outside every safety governance framework referenced in this tool. No AISI pre-deployment evaluation. No Frontier Model Forum safety commitment. No CoE Convention obligations. No incident reporting channel. The Diffusion Rule, had it been enforced, would not have reached a single open-weight deployment — because open-weight models bypass compute controls by design. The Malaysia episode is particularly instructive: a country announced Chinese AI infrastructure at a government event, retracted under US pressure the next day, with no multilateral body having any jurisdiction over the choice. This is not edge-case behaviour. It is the governance vacuum operating in real time.
Finding 3 of 3

Every Institution That Should Govern AI Is Structurally Prevented From Doing So

This is not a gap waiting to be filled. These institutions exist. They have mandates, budgets, and staff. The problem is structural: each one is prevented from reaching the highest-risk AI applications by its founding instrument, its membership composition, or its enforcement mechanism. Understanding why each institution cannot act is the prerequisite for knowing what the 2027 treaty must build.

Institution What it can do Why it cannot reach AI The structural barrier What would fix it
US BIS
Bureau of Industry and Security
Export controls on dual-use goods and technology WMD containment mandate — no AI safety objective. Rescinded the only comprehensive AI rule it ever published. Unilateral — no multilateral mandate. Wrong mandate Replace with multilateral mechanism with explicit AI safety objective
OPCW
Organisation for the Prohibition of Chemical Weapons
CWC implementation, chemical weapons verification, challenge inspections Jurisdiction limited to chemical weapons as defined in 1993 CWC. AI-enabled chemical weapon design is not within scope unless a new protocol extends the mandate. Wrong jurisdiction CWC protocol extending OPCW mandate to AI-enabled CBRN uplift assessment
WHO
World Health Organization
Global health governance, pandemic preparedness, biosecurity norms No mandate over AI systems. AI-enabled pathogen design sits at intersection of AI and biosecurity — WHO covers the biosecurity side only. No joint mandate with any AI body. No AI mandate Joint WHO–AI Safety Secretariat protocol for dual-use AI biosecurity evaluation
CCW GGE
Convention on Certain Conventional Weapons — Group of Governmental Experts
Autonomous weapons discussions since 2014; lethal autonomous weapons systems norms CCW mandate expires September 2026. 12 years of discussion with no binding outcome. Consensus requirement means Russia and China can block. Not a treaty — a discussion forum. Mandate expiring · Consensus blocked Meaningful human control standard in 2027 treaty, bypassing CCW consensus requirement
ITU
International Telecommunication Union
Telecommunications standards, internet governance, digital development Technical standards body only — no enforcement mandate. AI Focus Group produces guidance, not binding standards. Universal membership but no power to compel compliance. Standards, no enforcement ITU develops AI provenance and incident reporting technical standards; 2027 treaty mandates adoption
IAEA
International Atomic Energy Agency
Nuclear safeguards, verification, peaceful use promotion Mandate limited to nuclear material and technology. AI governance requires a new international body with an equivalent inspection mandate — IAEA cannot extend its remit unilaterally. Wrong domain — but strongest institutional model International AI Safety Secretariat modelled on IAEA Article 12 — the Precedents tab specifies the design
CoE Convention (CETS 225)
Only binding international AI treaty
Human rights, democracy and rule of law obligations for AI lifecycle Article 3.2 blanket national security exemption covers exactly the highest-risk dual-use applications. No Asian developing state is party. No inspection or verification mechanism. Self-exempted the highest-risk uses Narrow Article 3.2 in 2027 treaty; extend participation to Asian developing states via Article IV-equivalent bargain
UN GGE on Cybersecurity
Group of Governmental Experts — ICT security
Norms for responsible state behaviour in cyberspace; 11 agreed norms since 2015 Norms are non-binding. No enforcement mechanism. AI-enabled cyberattacks are not addressed in existing norm language. Russia and China have parallel OEWG process, fragmenting the framework. Non-binding · Fragmented Budapest Convention extension to AI-enabled attacks; 2027 treaty fills the enforcement gap
The structural pattern
Every institution with an enforcement mandate has the wrong jurisdiction. Every institution with the right jurisdiction has no enforcement mandate. This is not a coincidence — it reflects how AI governance has developed: technically capable institutions (OPCW, IAEA) were built for other threats, and the new institutions created for AI (Frontier Model Forum, AISI network) were designed by industry with no enforcement power. The 2027 treaty must either extend existing mandates or create a new body with both the right jurisdiction and enforcement authority.
Treaty design implication
Delegates should enter the Geneva Dialogue having read this table. The Dialogue's seven thematic areas implicitly assume that existing institutions can be adapted. This table demonstrates they cannot — not without structural reform. The 2027 treaty mandate should explicitly task the UN Secretary-General with a gap analysis of existing institution mandates and a proposal for the minimum new institutional architecture required.
Dual-use governance architecture

The Seven-Layer Stack

Methodology — how the seven layers were derived
Each layer corresponds to a documented functional requirement of a working WMD governance regime, derived from the primary treaty texts and institutional designs of the NPT/IAEA (Layer 1: compute/precursor control; Layer 2: capability identification via CWC Schedules; Layer 3: use-case differentiation via CWC Article VI; Layer 4: national governance capacity via CWC Article VII national authorities; Layer 5: incident attribution via OPCW Technical Secretariat; Layer 6: international verification via IAEA safeguards/OPCW inspections; Layer 7: consent and representation via NPT/CWC multilateral negotiation). The layers are not arbitrary — each maps to a specific institutional failure mode in WMD governance history that required a specific institutional solution. AI's absence on six of seven layers is assessed against this functional standard, not against a constructed benchmark.

Every functioning WMD governance regime required seven layers to work. Nuclear built all seven — it took 25 years and two weapon deployments. Chemical built six. Biological built three, and the missing verification layer is precisely why AI-enabled bioweapon design is the highest-risk ungoverned dual-use application today. AI has fragments of one layer, and that layer was rescinded before it came into force. Click any layer to compare.

Analytical boundary — open-weight models and the scope of export controls
Do open-weight models make export controls irrelevant? A common challenge: since DeepSeek is open-weight and freely downloadable, the Diffusion Rule's rescission would not have prevented Indonesia or Malaysia from adopting Chinese AI regardless. This is partially true — and is exactly the point. The rescission matters not because export controls would have prevented Chinese stack adoption, but because they were the only mechanism that created any governance obligation at all, however imperfectly designed. The governance vacuum is about the absence of safety obligations, not the absence of access. Open-weight models make the vacuum permanent for Layer 1: no access-control mechanism can govern a model that has already been released. This is why Layers 2–7 are not merely desirable — they are the only governance architecture that can reach open-weight AI. The tool's argument is strengthened, not weakened, by the open-weight reality.
Key: Functional   Partial   Absent
L
Layer
USA
UK
EU
JPN
KOR
IND
VNM
IDN
SGP
Layer 1
Compute Control
USA
UK
EU
JPN
KOR
IND
VNM
IDN
SGP
Nuclear

NSG export trigger criteria. IAEA safeguards as precondition for nuclear technology transfer. Functioning since 1970s.

Chemical

Australia Group dual-use chemical export controls. Schedule 1/2/3 precursor controls under CWC. Functioning since 1985.

Biological

Australia Group biological agent controls. Partial — gaps in dual-use equipment. BWC has no export control mandate.

AI — Status

Biden Diffusion Rule rescinded May 2025 — never enforced. No replacement. Pre-2022 chip controls on China only. Asian Tier 2 states: no binding framework. The only AI governance layer is broken.

↗ BIS Rescission
Layer 2
Capability Identification
USA
UK
EU
JPN
KOR
IND
VNM
IDN
SGP
Nuclear

IAEA defines nuclear material categories. Internationally agreed taxonomy of what constitutes weapon-relevant material.

Chemical

CWC Schedules 1/2/3 define dual-use chemicals by military utility and commercial use. OPCW Technical Secretariat maintains the taxonomy.

Biological

Australia Group pathogen control lists. Partial — gaps in dual-use equipment. No agreed international taxonomy of dual-use biological capabilities.

AI

No agreed international taxonomy of dual-use AI capabilities. UK AISI evaluates CBRN uplift — but voluntarily, for cooperating labs only, and covering no Chinese-origin models.

↗ METR Common Elements
Layer 3
Use-Case Differentiation
USA
UK
EU
JPN
KOR
IND
VNM
IDN
SGP
Nuclear

IAEA distinction between peaceful nuclear use (Article IV NPT right) and weapons-relevant activity. Enrichment level thresholds distinguish civilian from weapons-grade material.

Chemical

CWC explicitly distinguishes industrial production from weapons precursors. Challenge inspection verifies disputed use-cases. End-use certificates required for dual-use chemicals.

Biological

BWC Article I prohibits offensive weapons but has no verification mechanism. Dual-use research oversight exists in some jurisdictions, not internationally harmonised.

AI

AI pandemic detection and AI pathogen design require identical compute and similar models. Deepfake detection and generation use the same architecture. No governance mechanism can currently distinguish defensive from offensive AI use at capability level. This layer is entirely absent for AI.

Layer 4
National Governance Capacity
USA
UK
EU
JPN
KOR
IND
VNM
IDN
SGP
Nuclear analogue

National nuclear regulatory bodies, safeguards agreements, domestic enrichment oversight. Required for NPT adherence.

Chemical analogue

National Authorities under CWC Article VII. Mandatory declaration and inspection facilitation. Every CWC state party must have a designated national authority.

AI — what exists

UK AISI (strong, CBRN mandate). EU AI Office (binding). Korea KAISI (Jan 2026). Japan AISI (exists). Vietnam Law 134/2025 (binding, no eval capacity). India AISI (proposed). Indonesia: nothing.

AI — the gap

No country in this set has a dual-use AI evaluation mandate with CBRN scope. US has no federal AI law. No operational AISI in India or Indonesia. The layer exists in fragments in three Western countries only.

Layer 5
Incident Attribution
USA
UK
EU
JPN
KOR
IND
VNM
IDN
SGP
Nuclear

IAEA investigation mandate. Nuclear forensics programme for attribution. Post-incident material analysis to trace weapon origin.

Chemical

OPCW fact-finding missions. Technical Secretariat investigation capacity. Syria and Salisbury investigations demonstrated a functioning attribution mechanism.

Biological

No OPCW equivalent. Article VI BWC allows Security Council consultation but has no investigation mechanism. COVID-19 origin investigation demonstrated the gap.

AI

No investigation body for AI-enabled dual-use harm. No forensics standard. No attribution protocol. No no-blame reporting channel. Intelligence agencies can sometimes attribute — but outside any public governance framework.

Layer 6
International Verification
USA
UK
EU
JPN
KOR
IND
VNM
IDN
SGP
Nuclear

IAEA safeguards inspections. Additional Protocol. Comprehensive safeguards agreements. Verification is the backbone of the NPT.

Chemical

OPCW routine inspections. 98% of declared chemical weapons stockpiles destroyed and verified. The strongest WMD verification regime.

Biological — the failure

No verification protocol. Negotiated 1994–2001, collapsed when US withdrew citing trade secrets and technical verification limits. AI will face identical arguments. This is the warning precedent.

↗ UN BWC Documentation
AI

No international verification mechanism. CoE Convention has a Conference of the Parties — but no inspection, no technical secretariat, no challenge mechanism. Hiroshima AI Process is non-binding with no verification.

Layer 7
Consent and Representation
USA
UK
EU
JPN
KOR
IND
VNM
IDN
SGP
Nuclear

NPT is multilateral — 191 state parties consented to its terms. IAEA Board of Governors includes developing-state representation. Imperfect, but consent-based.

Chemical

CWC negotiated through the Conference on Disarmament. 193 states parties. Universal participation in rule design. OPCW has universal membership.

Biological

BWC has 183 states parties. Consent-based but weak — representation without enforcement. Implementation conferences every five years.

AI

Export controls designed by BIS (US agency) with no Asian input. CoE Convention negotiated by CoE members — India, Vietnam, Indonesia, Singapore absent. Frontier Model Forum: four companies. The UN Global Dialogue is the first inclusive forum — but non-binding in its first session. This is the most absent layer of all seven.

Stack finding: Nuclear has all seven layers. Chemical has six. Biological has three — and the missing verification layer (Layer 6) is why AI-enabled bioweapon design is the highest-risk ungoverned dual-use application. AI has fragments of Layer 4 in three Western countries and a collapsed Layer 1. This is not a governance gap. It is a governance vacuum.
Analytical boundary — scope of the WMD governance comparison
Does comparing AI to WMD overstate the risk? The WMD comparison in this tool is structural and functional, not a claim of risk equivalence. The argument is not that AI is as immediately lethal as a nuclear weapon or nerve agent. It is that dual-use AI requires the same governance architecture that dual-use chemistry required — because it shares the same structural properties: identical technology for beneficial and harmful purposes, inability to distinguish use-cases at point of access, and asymmetric harm potential. The CWC was built for chemistry that can kill in microgram quantities. The governance principles it instantiated — use-case differentiation, challenge inspection, verification, consent-based rule design — are not chemistry-specific. They are the functional requirements for any dual-use technology governance regime. That is the claim this tool makes and documents. It does not require agreement that AI risk equals nuclear risk.

Country coverage — sourced from Oxford Insights Government AI Readiness Index 2024 and OECD AI Policy Observatory

Per-Country Stack Coverage and Governance Baseline

Urgency score is a weighted composite: 40% governance capacity gap (inverse of Oxford Insights score), 40% dual-use risk exposure, 20% export control exposure post-rescission. India, Vietnam, and Indonesia are not OECD members — their absence from the OECD AI Index is itself a governance gap finding.

Country Oxford AI Readiness 2024 OECD Member Export Control Status Layers Covered (of 7) Urgency
🇮🇩 Indonesia65.85 / rank ~55 NoVacuum post-rescission0 / 79.4
🇮🇳 India~64.0 / rank 46 NoVacuum post-rescission0 / 78.1
🇻🇳 Vietnam61.42 / rank ~65 NoVacuum post-rescission1 / 77.8
🇸🇬 Singapore84.25 / rank 2 NoVacuum — same tier as Indonesia2 / 76.2
🇯🇵 Japan75.75 / rank 11 YesTier 1 — semiconductor leverage3 / 74.8
🇰🇷 South Korea79.98 / rank 3 YesTier 1 — Samsung/SK Hynix3 / 74.2
🇺🇸 United States87.03 / rank 1 YesRule designer — no replacement5 / 7Designer
🇬🇧 United Kingdom78.88 / rank 5 YesTier 1 — co-designer5 / 7Comparator
🇪🇺 European Union~76 avg / Top 10 Yes (members)Tier 1 — AI Act binding5 / 7Comparator
Sources: Oxford Insights Government AI Readiness Index 2024 · OECD AI Policy Observatory Index 2024 · BIS Rescission Statement May 2025
Institutional mechanism mapping

WMD Precedents — Design Specifications for 2027

Six specific governance mechanisms that make WMD regimes function — mapped to their AI equivalent at institutional level, with design requirements and draft treaty language. The research contribution here is not the analogy. It is the specification: what institutional architecture, what mandate, what enforcement mechanism, and what political bargain each layer requires. Every mechanism is sourced to primary treaty texts.

Each card shows: the mechanism, its WMD origin, institutional design required, and what the AI equivalent would need. Linked to primary treaty texts.
Nuclear · NPT Article III + IAEA Statute
Safeguards — Verification by a Technical Secretariat
Trigger: 1945 Hiroshima → 1957 IAEA founded → 1968 NPT Article III
What it does

IAEA inspectors verify nuclear material is not diverted from peaceful use to weapons. Comprehensive safeguards agreements are legally binding on all NPT non-nuclear states. Additional Protocol expands inspection rights to undeclared sites.

Institutional design required

Independent technical secretariat with inspection authority. Safeguards criteria agreed multilaterally. Material accountancy system. Board of Governors with enforcement referral to Security Council.

AI equivalent — 2027 treaty needs

An independent International AI Safety Secretariat with authority to conduct pre-deployment evaluations of frontier models, maintain a registry of dual-use capability assessments, and refer non-compliant systems to the UN Security Council. Evaluation criteria agreed by all member states — not set by four Western companies. Modelled on IAEA Article 12 safeguards, applied to training compute above the dual-use capability threshold.

↗ IAEA INFCIRC/153 — Comprehensive Safeguards Agreement
Treaty design requirement
An International AI Safety Secretariat with inspection authority, multilaterally agreed evaluation criteria, treaty-mandated referral to the UN Security Council for non-compliance, and a registry of evaluated AI systems accessible to all state parties.
Chemical · CWC Article VI + OPCW
Challenge Inspection — Dispute Resolution at Technical Level
Trigger: Chemical weapons in WWI (1915) → CWC negotiated 1980–92 → OPCW 1997
What it does

Any CWC state party may request an inspection of any facility in any other state party on grounds of suspected non-compliance. OPCW Technical Secretariat conducts the inspection. Executive Council may block with 3/4 majority vote.

Key strength

98% of declared chemical weapons destroyed and verified. Syria and Salisbury investigations demonstrated functioning attribution. The CWC verification regime is the model for what works.

AI equivalent — 2027 treaty needs

A challenge evaluation mechanism: any state party may request independent evaluation of an AI system suspected of dual-use capability. The International AI Safety Secretariat conducts evaluation using standardised CBRN uplift protocols. Confidentiality protections for model weights and training data (analogous to CWC trade secret protections). Results reported to an Executive Board with enforcement referral authority.

↗ CWC Article VI — OPCW Official Text
Treaty design requirement
A challenge evaluation mechanism allowing any state party to request independent assessment of any AI system, conducted within 60 days, with CWC-equivalent confidentiality protections for model weights and Executive Board enforcement authority.
Multi-domain · Australia Group (1985–present)
Control List Update Process — Dual-Use Taxonomy Maintenance
Trigger: Iraq's chemical weapons use in Iran-Iraq war → Australia Group formed 1985
What it does

42 participating countries maintain and update lists of dual-use chemicals, biological agents, and equipment subject to export controls. Annual plenary reviews and updates in response to scientific developments. Informal — not a treaty.

Key limitation

Western-dominated — no China, Russia, or major Global South participation. Created governance fragmentation rather than universal coverage. Its informality is both its strength (speed) and its weakness (legitimacy).

AI equivalent — 2027 treaty needs

An international Dual-Use AI Capability Review Process: a standing technical body that annually updates a taxonomy of AI capabilities crossing the dual-use threshold, with universal membership unlike the Australia Group. The UN Scientific Panel on AI (established 2026) is positioned to perform this function — but needs a formal mandate, a dual-use lens, and an export control linkage.

↗ Australia Group Control Lists
Treaty design requirement
A standing Dual-Use AI Capability Review Body with universal membership and an annual update cycle, formally linked to export control classifications, with a mandate to the UN Scientific Panel on AI to maintain the dual-use capability taxonomy.
Biological · BWC (1972) — The Failure Precedent
What Happens Without Verification
BWC adopted 1972. Verification protocol negotiated 1994–2001. US withdrew 2001. No protocol exists today.
What went wrong

The BWC prohibits biological weapons but has no verification mechanism. The Soviet Union ran the world's largest offensive biological weapons programme (Biopreparat) throughout the BWC's existence, in clear violation — revealed only after Soviet collapse.

Why verification failed

In 2001, the US withdrew from Protocol negotiations citing concerns about protecting commercial trade secrets and the inability to verify compliance in dual-use biological facilities. These are exactly the arguments that will be raised about AI model weights.

The warning for every treaty negotiator

The AI governance community will face identical arguments to those that killed the BWC verification protocol: model weights are trade secrets, verification would disadvantage Western companies, compliance cannot be technically verified. The BWC failure shows what happens when these arguments succeed. The 2027 treaty must have verification built in from inception — not negotiated separately, later, and unsuccessfully.

↗ UN BWC Documentation and History
Treaty design requirement
Verification must be in the 2027 treaty from the outset, not negotiated separately later. When trade secret arguments are raised against model weight inspection, the CWC confidentiality provisions are the precedent to follow, not the BWC capitulation.
Nuclear · NPT Article IV
Peaceful Use Right — The Political Bargain Precedent
NPT Article IV: "the inalienable right of all Parties to develop research, production and use of nuclear energy for peaceful purposes"
Why it made the NPT work

Non-nuclear states accepted NPT restrictions in exchange for a guaranteed right to peaceful nuclear use. Without Article IV, developing states had no incentive to join a treaty that restricted their technology access without offering anything in return.

The AI parallel

Developing states will not accept a 2027 AI treaty that restricts their AI development without an equivalent guarantee. The export control architecture's fundamental political problem is that it imposes restrictions without any guaranteed pathway to beneficial AI access.

AI equivalent — the political bargain

States accepting dual-use AI governance obligations (capacity-building requirements, incident reporting, AISI-equivalent evaluation) receive a guaranteed right to access frontier AI compute for peaceful applications — biosecurity defence, healthcare AI, agricultural AI. The governance compute carve-out proposed in Tab 2 is the operational form of this Article IV right. Without it, the treaty will fail to achieve developing-state participation.

↗ NPT Full Text — Article IV
Treaty design requirement
An explicit governance compute carve-out: states accepting dual-use AI governance obligations receive a guaranteed right to frontier compute access for safety evaluation, unconditioned on geopolitical tier. This is the political bargain that makes developing-state participation in the 2027 treaty viable.
Aviation · ICAO Annex 19 (2013)
Convergence Without Catastrophe — The Only Success Precedent
ICAO Annex 19 adopted 2013 — built through 33 years of professional consensus, not triggered by a crash
Why this is different

Every other WMD and safety governance regime was triggered by catastrophe. ICAO Annex 19 is the only major international safety standard built through proactive professional consensus. It achieved universal adoption because it was technically credible, built incrementally, and gave operators a compliance pathway rather than an immediate ban.

What made it work

Four pillars: safety policy, safety risk management, safety assurance, safety promotion. Each pillar was operationalisable. Compliance was measurable. Built by practitioners, not politicians. Extended progressively to new domains without requiring new treaties each time.

AI equivalent — the most viable path to international safety standards

An international AI Safety Management Standard modelled on ICAO Annex 19: four pillars (safety policy, dual-use risk assessment, safety assurance/evaluation, safety promotion/incident reporting). Binding on frontier AI developers above a compute threshold. Extensible over time. The UN Scientific Panel on AI should draft this standard for the 2027 New York session. It is achievable without waiting for a catastrophe.

↗ ICAO Annex 19 — Safety Management Systems
Treaty design requirement
A binding international AI Safety Management Standard with four pillars (safety policy, dual-use risk assessment, safety assurance, incident reporting), mandated for frontier AI developers above a compute threshold, drafted by the UN Scientific Panel on AI for the 2027 New York session.
Accountability gap analysis

The Attribution Gap — Layer 5 of the Governance Stack, Entirely Absent

When an AI system contributes to a biosecurity incident, electoral interference, or infrastructure attack, there is no investigation body, no attribution standard, no forensics protocol, and no liability framework that applies across jurisdictions. The OPCW can attribute chemical weapon use. The NTSB investigates aviation accidents. AI has no equivalent. These four scenarios document what the absence of Layer 5 means in practice — and what any binding governance instrument must build to fill it.

Stack layers absent: Layer 2 (Capability Identification) · Layer 5 (Incident Attribution)
The risk scenario

An AI system is used to design a novel pathogen with enhanced transmissibility. Conducted by a non-state actor using open-weight models (DeepSeek, Llama) on cloud infrastructure in a jurisdiction with no AI governance framework. A pandemic results. Who is legally accountable?

Current legal accountability

BWC prohibits state development of biological weapons but has no enforcement mechanism. Non-state actors are not directly bound. AI developer has no liability for downstream misuse of open-weight models. Cloud provider faces terms of service violations — civil, not criminal. Result: no binding accountability mechanism exists.

The attribution problem

No AI forensics standard equivalent to nuclear material fingerprinting or OPCW chemical analysis. Models used, compute accessed, and design process leave no internationally admissible evidentiary trail. Intelligence agencies may be able to attribute — but outside any public legal framework. The BWC verification gap means no treaty mechanism exists to investigate even suspected state involvement.

What the 2027 treaty needs

Mandatory compute and API access logging for frontier model providers above a capability threshold — creating an admissible evidentiary trail. An international AI forensics protocol for post-incident attribution modelled on OPCW chemical analysis. Liability provisions extending to frontier model developers for foreseeable dual-use applications, with safe harbour for developers who have conducted AISI-equivalent CBRN evaluations.

Article [X] — Dual-Use AI Incident Investigation 1. Each Party shall require providers of AI systems above the dual-use capability threshold to maintain records of: (a) training data sources and any biosecurity-relevant fine-tuning; (b) API access logs for requests triggering biological agent design capabilities; (c) safety evaluation records, including CBRN uplift assessments. 2. In the event of a suspected AI-enabled biological incident, any Party may request an investigation by the International AI Safety Secretariat. The Secretariat shall conduct a technical assessment within [90] days and report findings to the Executive Board. 3. Providers that have conducted pre-deployment CBRN evaluations pursuant to Article [Y] and implemented recommended safeguards shall not bear liability for misuse by third parties acting outside authorised use parameters. — Draft clause modelled on BWC Article VI + CWC challenge inspection mechanism
↗ BWC Documentation · UN Office for Disarmament Affairs
Stack layers absent: Layer 3 (Use-Case Differentiation) · Layer 5 (Incident Attribution)
The risk scenario

AI-generated deepfakes of political candidates distributed at scale during an Indonesian or Indian election. Produced using open-weight Chinese models (DeepSeek) hosted on domestic servers. Electoral outcome materially affected. Who is accountable?

Current legal accountability

Indonesia: no AI law. Electronic Information and Transactions Law applies to some online content but not AI-generated deepfakes specifically. India: IT Act provisions apply but enforcement against state-linked actors is effectively impossible. Model developer: no international liability exposure. Result: near-zero accountability.

The attribution problem

AI-generated deepfake attribution requires model fingerprinting and artefact analysis combined with investigation of distribution infrastructure. No international standard exists. States with technical attribution capability (US, UK, Five Eyes) may investigate — but findings are intelligence, not legal evidence, and are not shared through any multilateral mechanism.

What the 2027 treaty needs

A Content Provenance Standard: mandatory watermarking and provenance metadata for AI-generated content above a capability threshold (C2PA standard or equivalent), legally admissible in electoral interference proceedings. An international Electoral AI Integrity mechanism within the UN system, with authority to investigate AI-enabled election interference.

Article [X] — AI-Generated Content Provenance 1. Each Party shall require providers of generative AI systems capable of producing synthetic media to implement: (a) cryptographic content provenance metadata compliant with internationally agreed standards; (b) disclosure mechanisms enabling verification of AI origin by accredited third parties. 2. AI-generated content used in the context of electoral processes shall be subject to mandatory disclosure requirements under national law, with minimum standards set by this Convention. 3. Any Party alleging material interference in its electoral process through AI-generated content may request investigation by the [Electoral AI Integrity Panel], which shall have authority to examine content provenance records and issue findings within [60] days. — Draft clause modelled on ITU technical standards process + ICCPR Article 25
Stack layers absent: Layer 4 (National Governance Capacity) · Layer 5 (Incident Attribution)
The risk scenario

An AI-enabled cyberattack disrupts Singapore's financial system. The attack uses AI systems for planning, evasion, and payload generation, developed by a state-linked actor. A systemically important bank's AI systems — sourced from multiple providers including Chinese-origin models with no safety evaluation — are implicated as an attack vector. Who is accountable?

Current legal accountability

Budapest Convention on Cybercrime applies to some aspects but has no AI-specific provisions. China and Russia are not parties. UN GGE norms are non-binding. OCBC may have liability under MAS regulations for inadequate third-party risk management. DeepSeek developer: no international liability mechanism exists.

The attribution problem

AI-enabled cyberattacks are designed to be deniable. The same AI tools used for attack are used for legitimate security research. Attribution requires access to training data, deployment logs, and infrastructure — none accessible through current international legal mechanisms. Singapore AISI has no authority over Chinese-origin AI systems deployed in Singapore's financial sector.

What the 2027 treaty needs

Extension of the Budapest Convention framework to AI-enabled attacks. Mandatory security evaluation for AI systems deployed in critical financial infrastructure regardless of model origin. Mandatory incident reporting to a cross-border AI Security Incident Registry. A state responsibility doctrine for AI-enabled attacks where the state of the model developer can be held to account for foreseeable misuse.

Article [X] — AI Systems in Critical Infrastructure 1. Each Party shall require that AI systems deployed in critical infrastructure sectors — including financial services, energy, water, and healthcare — are subject to: (a) mandatory security evaluation by an accredited third party prior to deployment; (b) ongoing monitoring and incident reporting to the national AI Safety Institute; (c) disclosure of model origin and training provenance to the regulatory authority. 2. These requirements apply to all AI systems deployed in critical infrastructure, regardless of the jurisdiction of the model developer. 3. In the event of an AI-enabled attack on critical infrastructure, the affected Party may invoke this Convention's investigation mechanism. The state in which the model developer is domiciled has an obligation to cooperate and provide access to relevant technical records. — Draft clause modelled on Budapest Convention + IMF FSAP financial resilience framework
↗ Budapest Convention on Cybercrime
Stack layers absent: Layer 3 (Use-Case Differentiation) · Layer 5 (Incident Attribution)
The risk scenario

An autonomous AI targeting system deployed in active conflict makes a targeting decision resulting in civilian casualties. The AI system was developed using models fine-tuned in a jurisdiction with no AI governance framework. Who is accountable under IHL?

Current legal accountability

IHL requires human accountability for targeting decisions. If an AI system makes targeting decisions without meaningful human control, there is an "accountability gap" — no human made the specific decision, so no human can be held to account. CCW GGE has discussed autonomous weapons since 2014. No binding treaty as of mid-2026. CCW mandate expires September 2026.

The attribution problem

AI targeting systems are designed by developers, trained on datasets, deployed by militaries, and operated by combatants — with decisions distributed across this chain. IHL requires a human decision-maker accountable for targeting. AI systems create an accountability gap where no individual in the chain made the specific targeting decision that caused harm.

What the 2027 treaty needs

A Meaningful Human Control standard for autonomous AI weapons systems, with a pre-deployment validation requirement. The standard should specify: minimum human oversight required for AI-assisted targeting to meet IHL accountability requirements; a pre-deployment validation protocol run by an independent technical body; a post-incident attribution protocol that can trace targeting decisions through the AI decision chain.

Article [X] — Meaningful Human Control in AI-Assisted Targeting 1. Each Party shall ensure that AI systems used in targeting decisions in armed conflict: (a) are subject to pre-deployment validation by an accredited independent technical body, confirming compliance with IHL principles of distinction, proportionality, and precaution; (b) retain an auditable decision log enabling post-incident attribution of each targeting output to a responsible human commander; (c) are not deployed in configurations that remove meaningful human control from the final targeting decision. 2. "Meaningful human control" means that a human commander has: reviewed the AI system's targeting output; applied independent judgment informed by IHL; and retained authority to override or abort the engagement prior to execution. 3. The International AI Safety Secretariat shall maintain a registry of validated AI targeting systems. Deployment of unvalidated systems in targeting roles constitutes a violation of this Convention. — Draft clause modelled on CCW GGE discussions + IHL accountability doctrine
↗ UNIDIR Autonomous Weapons Systems Programme