UK-US AI Safety Institute Memorandum of Understanding

Operative Mechanics

Signed 1 April 2024, the MoU is the first bilateral instrument dedicated to frontier-AI safety, committing the two AI Safety Institutes to coordinated pre-deployment evaluations, red-team data sharing, and methodological alignment on capability elicitation across at least one major frontier-model release. Its substantive scope is foundation-model evaluation (governs, foundation_models), but the operative obligations run between agencies rather than to model developers: the instrument is an information-sharing and joint-exercise compact, not a market-access rule. Catastrophic-risk coverage is implicit — joint evaluations encompass CBRN and autonomy-uplift questions without enumerating explicit capability thresholds, an omission Scholefield et al. ¹ treat as the gap a conditional treaty with compute triggers should close, and one that fits Bengio et al.'s ² warning that current initiatives 'lack the mechanisms and institutions to prevent misuse and recklessness'. Agentic-behaviour testing is likewise folded into capability evaluation rather than separately mandated.

Cross-Jurisdiction Position

The MoU sits beside, not within, hard regulatory regimes: it creates evaluation capacity rather than the binding developer duties of Regulation (EU) 2024/1689, whose GPAI provisions (Art. 51-55) impose systemic-risk obligations the MoU deliberately avoids. This division of labour — EU statute defining obligations, AISIs supplying the testing science — is sharpened by the definitional instability Fernández-Llorca et al. ³ trace across the AI Act's shifting 'foundation model'/'GPAI' terms, and by Hulok's ⁴ account of autonomous generation straining authorship and accountability categories. The MoU's transparency footprint is internal (information sharing between institutes, not public disclosure), so it does nothing for the training-data leakage problems Ruschemeier ⁵ locates under the GDPR. It is precedent infrastructure: echoed in the Seoul Declaration and its annexed Statement of Intent, and generalised into the multilateral AISI network (US, UK, JP, SG, CA, FR, KR).

Key Fault Lines

Three critiques recur. First, scope-without-threshold: by leaving CBRN and autonomy-uplift criteria unenumerated, the MoU defers the hard line-drawing that Druzin et al. argue international law's precautionary obligation actually requires, and that Kasirzadeh ⁶ complicates by distinguishing decisive from accumulative existential risk — the latter invisible to single-release joint exercises. Second, agentic blind spots: capability evaluations now lean on benchmarks like AgentHarm ⁷, yet Kolt ⁸ and the agent-infrastructure agenda ⁹ show attribution, liability, and authenticated delegation ¹⁰ sit outside an evaluation MoU's reach, as do multi-agent miscoordination and collusion ¹¹. Third, the instrument binds agencies, not labs — it cannot compel a developer to submit a model or halt a release.

Implementation Trajectory

The partnership has outlived its signatories' names: the UK AI Safety Institute became the UK AI Security Institute (14 Feb 2025) and the US AI Safety Institute became the Center for AI Standards and Innovation, CAISI (June 2025), yet the joint pre-deployment evaluation and testing partnership remains in force and has expanded under the renamed institutes. The trajectory is one of multilateralisation rather than legalisation — the bilateral became the template for the Seoul-consolidated AISI network, reflected in the Seoul Declaration's commitment to an international network of AI Safety Institutes. Whether this voluntary, agency-to-agency model matures into the compute-gated, audit-empowered treaty Scholefield et al. ¹ propose, or hardens into a UN-backed agency on the IAEA model that Robinson ¹² argues is needed for legitimate global oversight, or remains soft coordination, is the open question; biosecurity governance pathways ¹³ and human-oversight limits ¹⁴ mark where binding obligation, not just shared methodology, would have to attach.

Cross-jurisdiction comparison

How peer instruments treat the topics UK-US AI Safety Institute Memorandum of Understanding governs.

°= industry self-imposed voluntary framework. Comparing a voluntary code's "governs" tint with a binding regulation's "governs" tint flattens the legal-force distinction; use the instrument-page banner for the operative status of each.

Does this instrument’s approach work? — the social-science evidence

Aggregated over the 5 topics this instrument governs: whether each harm is empirically real, and whether the peer-reviewed evidence shows governance reduces it. The badge is the epistemic status of the evidence— “thin”/“absent” efficacy evidence is itself a finding (the “second silence”). Each epistemic-status label is Policy Window's editorial assessment of the cited evidence base (a structured classification), not a verdict any single source issues.

Of the 5 governed topics with a social-science evidence review, evidence that governance reduces the harm is established for 0, contested for 0, thin for 0, and absent for 5 — for most, no replicated study yet shows this instrument's approach works (the "second silence").

• Agentic AI Governance

  • Is the harm real?evidence: thin

    The capability that agentic governance targets — autonomous multi-step action — is real and rapidly, measurably advancing: METR finds the task length AI agents complete at 50% reliability has doubled roughly every seven months for the past six years (about 50 minutes for frontier 2025 models), and the UK AI Security Institute's first Frontier AI Trends Report (Dec 2025, >30 systems) reports models now finish hour-long software tasks >40% of the time versus <5% in late 2023. The distinct realized HARM from agency (as opposed to the underlying model) is, however, thinly documented: on consequential real-world tasks agents still fail the majority — Gemini 2.5 Pro completed only 30.3% of TheAgentCompany's 175 professional tasks (OpenHands scaffold, project leaderboard) — so the agency-specific harm magnitude is early and context-dependent rather than established at scale.

    Sources: Kwa, West, Becker et al. 2025 (METR; arXiv:2503.14499, 'Measuring AI Ability to Complete Long Tasks'); UK AI Security Institute 2025 (Frontier AI Trends Report, Dec 2025); Xu, Song, Zhou et al. 2024 (TheAgentCompany, arXiv:2412.14161); 30.3% figure per TheAgentCompany leaderboard (OpenHands)

  • Does governance work?evidence: absent

    There is no impact-evaluation evidence that agent-specific governance reduces agentic harm: the operative regimes — the EU GPAI Code of Practice (published July 2025, voluntary/non-binding), the Seoul Frontier AI Safety Commitments (2024, voluntary), and AISI agent evaluations — are 2024-25 vintage and have never been measured against an outcome. The scholarship itself has not settled the contested unit of regulation: Kolt (2025) argues for governing the agentic relationship via principal-agent and agency-law tools, while Chan, Ezell, Kaufmann et al. (2024) propose agent-specific visibility mechanisms (identifiers, real-time monitoring, activity logging) that remain proposal-stage and unevaluated — meaning the field has design proposals but, as with most frontier-AI rules, the evidence that any of them works is absent rather than merely thin.

    Sources: Kolt 2025 ('Governing AI Agents', 101 Notre Dame L. Rev., forthcoming; arXiv:2501.07913); Chan, Ezell, Kaufmann et al. 2024 ('Visibility into AI Agents', ACM FAccT 2024, pp. 958-973; DOI 10.1145/3630106.3658948); EU AI Office 2025 (GPAI Code of Practice, July 2025); Seoul Frontier AI Safety Commitments 2024

• Catastrophic & Existential Risk

  • Is the harm real?evidence: contested

    The catastrophic-uplift premise is genuinely contested: the empirical uplift studies that exist find current frontier models add little. RAND's red-team study found no statistically significant difference in the viability of bioweapon-attack plans produced with vs. without LLMs (Mouton, Lucas & Guest 2024), and OpenAI's 100-participant trial found GPT-4 gave at most a mild, non-significant accuracy uplift (mean +0.88 out of 10 for PhD experts, +0.25 for students; Patwardhan et al. 2024). Honest caveat: the harm is forward-looking, not yet observed — expert opinion on the catastrophic tail is sharply split (median AI researcher puts ~5% on extremely-bad/extinction outcomes, mean ~9-16% across differently-framed questions, n=2,778; Grace et al. 2024), and forecasters underestimated how fast risk-relevant capabilities (e.g. virology troubleshooting) actually arrived (Forecasting Research Institute 2025), so the relevant capabilities are a moving target rather than a settled magnitude.

    Sources: Mouton, Lucas & Guest 2024 (RAND RR-A2977-2, Operational Risks of AI in Large-Scale Biological Attacks: Results of a Red-Team Study); Patwardhan et al. 2024 (OpenAI, Building an Early Warning System for LLM-aided Biological Threat Creation); Grace et al. 2024 (Thousands of AI Authors on the Future of AI, arXiv:2401.02843); Forecasting Research Institute 2025 (Forecasting LLM-enabled Biorisk and the Efficacy of Safeguards)

  • Does governance work?evidence: absent

    There is essentially no impact evidence that catastrophic-risk governance reduces catastrophic risk, and structurally there cannot yet be: the harm is a low-probability civilisational tail event, so no controlled trial or before/after evaluation of a realised catastrophe is possible. The dominant instruments are recent, voluntary developer frameworks (Anthropic's Responsible Scaling Policy 2023; OpenAI's Preparedness Framework 2023) built on if-then capability thresholds the developers themselves describe as speculative and qualitative rather than validated risk thresholds. The closest evidence is adjacent and indirect: trained-in deceptive behaviours can persist through standard safety training (Hubinger et al. 2024) — a demonstration that current mitigation may be insufficient, not that any governance regime works — and Anthropic's documented loosening of earlier commitments (RSP 2025 dropped the original pledge to define higher-tier ASL evaluations before developing the corresponding models) illustrates that even the strongest voluntary regimes lack external enforcement or measured efficacy.

    Sources: Anthropic 2023 (Responsible Scaling Policy); OpenAI 2023 (Preparedness Framework); Hubinger et al. 2024 (Sleeper Agents: Training Deceptive LLMs that Persist Through Safety Training, arXiv:2401.05566); Hendrycks, Mazeika & Woodside 2023 (An Overview of Catastrophic AI Risks, arXiv:2306.12001)

• Foundation Models / GPAI

  • Is the harm real?evidence: contested

    Whether the foundation-model category maps to a coherent capability/risk tier is genuinely contested. The original case rests on scale-driven 'emergent abilities' that appear unpredictably above a size threshold (Wei et al. 2022; Ganguli et al. 2022 documented capabilities that are smoothly predictable in aggregate loss yet locally surprising), but Schaeffer, Miranda & Koyejo (2023, a NeurIPS Outstanding Paper) showed many 'emergent' jumps are artefacts of discontinuous metrics and dissolve under linear/continuous scoring — implying capability scales more smoothly than a sharp tier would suggest. Honest caveat: this is a live empirical disagreement about measurement, not a settled finding either way, and compute (the regulatory proxy) is an imperfect stand-in for capability or risk regardless of which side is right.

    Sources: Wei et al. 2022 (Emergent Abilities of Large Language Models, TMLR; arXiv:2206.07682); Schaeffer, Miranda & Koyejo 2023 (Are Emergent Abilities of Large Language Models a Mirage?, NeurIPS 2023, Outstanding Paper; arXiv:2304.15004); Ganguli et al. 2022 (Predictability and Surprise in Large Generative Models, ACM FAccT; DOI 10.1145/3531146.3533229)

  • Does governance work?evidence: absent

    There is no impact evaluation showing that GPAI/foundation-model governance reduces harm — the rules are too new (EU AI Act GPAI obligations and the 10^25-FLOP systemic-risk presumption only began binding on 2 August 2025) and the central regulatory lever is itself contested: Hooker (2024) argues compute thresholds are a shortsighted proxy because compute does not reliably track capability or risk, and the thresholds already diverge across jurisdictions (EU 10^25 vs. the now-rescinded US EO 14110's 10^26 operations, rescinded 20 January 2025). The mandated mitigation methods also lack validated efficacy: model evaluation and red-teaming face well-documented coverage limits and an 'audit gap' in the survey/position literature (behavioural testing cannot establish the absence of untested failure modes), and adversarial red-teaming repeatedly defeats deployed safeguards — the UK AI Safety Institute reports finding universal jailbreaks for every frontier system it has tested, and a large public agent-injection competition elicited policy violations across all 22 frontier models tested from ~1.8M attacks (Zou et al. 2025). Even compliant evaluation therefore cannot yet certify the safety the rules demand. (Caveat: this is an absence-of-evidence claim — no efficacy study has been done — not evidence the rules are ineffective.)

    Sources: Hooker 2024 (On the Limitations of Compute Thresholds as a Governance Strategy, arXiv:2407.05694); EU AI Act Arts. 51 & 55 (GPAI systemic-risk presumption, 10^25 FLOP; binding 2 Aug 2025); US EO 14110 (10^26-operation reporting threshold, rescinded 20 Jan 2025 by EO 14148); Zou et al. 2025 (Security Challenges in AI Agent Deployment: Insights from a Large Scale Public Competition / Gray Swan Arena, arXiv:2507.20526 — 22 frontier agents, ~1.8M attacks); UK AI Safety/Security Institute, Frontier AI Trends Report (universal jailbreaks for every system tested); METR, Common Elements of Frontier AI Safety Policies (2024)

• International Coordination

  • Is the harm real?evidence: contested

    The DESCRIPTIVE premise is well-established: IR scholarship now treats global AI governance as a fragmented 'regime complex' of partially overlapping G7/G20/OECD/GPAI/UN/standards-body arrangements with no central hierarchy (Tallberg et al. 2023 — verified verbatim: 'the emerging governance architecture for AI can be described as a regime complex'; Cihon, Maas & Kemp 2020). But the implied HARM — that forum-shopping and regulatory arbitrage cause a measurable race-to-the-bottom or relocate AI development to lax jurisdictions — is largely theorized/anticipated rather than empirically demonstrated for AI; Tallberg et al. explicitly flag forum-shopping as a dynamic whose presence in the AI regime complex is an open empirical question ('Establishing whether these patterns and dynamics are key features also of the AI regime complex stand out as important priorities in future research'). Honest caveat: the strongest empirical arbitrage evidence comes from analogue footloose digital markets (e.g., ICO reallocation after US securities enforcement) — itself a mixed/contested literature — not from AI firms, so the magnitude of coordination-failure harm in AI specifically remains contested and under-measured.

    Sources: Tallberg, Erman, Furendal, Geith, Klamberg & Lundgren 2023 (International Studies Review 25(3): viad040); Cihon, Maas & Kemp 2020 (Should AI Governance be Centralised?, AIES '20: 228-234); Lancieri, Edelson & Bechtold 2025 (AI Regulation: Competition, Arbitrage & Regulatory Capture, Theoretical Inquiries in Law 26(1): 239-262)

  • Does governance work?evidence: absent

    There are essentially no impact evaluations showing that the negotiated-coordination mode (AI Safety Institute network MoUs, forum-shifting, multilateral declarations) actually produces regulatory convergence or reduces arbitrage — the AISI Network began only as a statement of intent at the Seoul Summit (Seoul Statement of Intent, 21 May 2024) and held its first operational meeting in November 2024, with no defined metrics or outcome studies, so these soft-law instruments are too new to have measurable effects. The closest analogue evidence is mixed and works through DIFFERENT mechanisms than this topic describes: Bradford's Brussels Effect documents de-facto convergence driven by market access rather than negotiated coordination, and the FATF transgovernmental-network literature shows peer-review mutual evaluation can drive AML convergence — but neither evaluates voluntary AI MoU networks, and FATF's effects come with well-documented unintended consequences (de-risking, financial exclusion). The plain finding: the evidence that AI-governance coordination 'works' is itself missing.

    Sources: Bradford 2020 (The Brussels Effect: How the European Union Rules the World, Oxford University Press); Nance 2018 (The regime that FATF built: an introduction to the Financial Action Task Force, Crime, Law and Social Change 69(2): 109-129; cf. Slaughter 2004, A New World Order, Princeton University Press); International Network of AI Safety Institutes — Seoul Statement of Intent toward International Cooperation on AI Safety Science (21 May 2024; network's first meeting San Francisco, Nov 2024)

• Transparency Obligations

  • Is the harm real?evidence: contested

    Documentation artifacts (model cards, datasheets) are well-specified as proposals and are genuinely adopted, but the empirical premise that mandated disclosure produces meaningful transparency is contested. Selbst & Barocas (2018) argue inscrutability and non-intuitiveness are distinct problems and that disclosing rules does not resolve the latter, and large-scale audits find documentation is sparsely and unevenly completed: a systematic analysis of 32,111 Hugging Face model cards (Liang et al. 2024) found environmental-impact, limitations and evaluation sections least often filled, and Bhat et al. (2023, 45 practitioners) found a substantial gap between the documentation proposal and actual practice. Honest caveat: the documentation frameworks themselves are real and adopted, so the dispute is about whether disclosure conveys decision-relevant information, not whether the artifacts exist.

    Sources: Selbst & Barocas 2018 (Fordham Law Review 87:1085-1139); Liang et al. 2024 (Nature Machine Intelligence, s42256-024-00857-z, 'Systematic analysis of 32,111 AI model cards'); Bhat et al. 2023 (CHI '23, 'Aspirations and Practice of ML Model Documentation', DOI 10.1145/3544548.3581518); Mitchell et al. 2019 (FAccT, Model Cards for Model Reporting); Gebru et al. 2021 (CACM 64(12):86-92, Datasheets for Datasets)

  • Does governance work?evidence: absent

    There is no rigorous impact evaluation showing that AI transparency mandates (model cards, training-data summaries) measurably reduce bias, misuse or accidents — the central regulatory assumption is empirically untested, partly because flagship mandates like EU AI Act Art. 53(1)(d) GPAI training-data summaries are only subject to AI Office enforcement/verification from 2 August 2026 (the obligation itself began 2 August 2025 for new models). The closest analogue, mandated consumer disclosure, shows small and context-dependent effects: Bollinger, Leslie & Sorensen (2011) found mandatory calorie posting cut average calories per transaction by about 6%, while Loewenstein, Sunstein & Golman (2014) review evidence that disclosure effects are frequently diminished or even reversed by limited attention and often change provider rather than recipient behavior. These are analogues, not AI studies; no study demonstrates that AI transparency disclosure achieves its stated downstream safety aims.

    Sources: Bollinger, Leslie & Sorensen 2011 (AEJ: Economic Policy 3(1):91-128); Loewenstein, Sunstein & Golman 2014 (Annual Review of Economics 6:391-419, 'Disclosure: Psychology Changes Everything'); EU AI Act Art. 53(1)(d) GPAI training-data summary (obligation from 2 Aug 2025; AI Office enforcement from 2 Aug 2026)

Related instruments

• Bletchley Declaration on AI Safety · global
• Seoul Declaration on Safe, Innovative and Inclusive AI · global
• Singapore Model AI Governance Framework for Generative AI · SG
• Japan METI AI Guidelines for Business · JP