bletchley declaration

The Bletchley Declaration is a multilateral political agreement signed in November 2023 by 28 nations and the European Union at the inaugural AI Safety Summit at Bletchley Park, UK, establishing the first intergovernmental consensus on the risks of frontier AI models. Signatories committed to international cooperation on safety evaluation, information sharing among national AI safety institutes, and the development of governance frameworks to address potentially catastrophic risks from highly capable foundation models. The declaration catalysed the formation of AI safety institutes in the UK, United States, and other signatory nations, and set the precedent for government-coordinated pre-deployment evaluations of frontier AI systems. It was followed by the Seoul AI Safety Summit in 2024, which extended commitments toward operational safety testing requirements for frontier model developers.

Semantic Classification

Content

Compositional Relationships (Components)

SubClassOf(gov:BletchleyDeclaration
  ObjectSomeValuesFrom(gov:hasPart gov:RiskAcknowledgementProvision))
SubClassOf(gov:BletchleyDeclaration
  ObjectSomeValuesFrom(gov:hasPart gov:InformationSharingCommitment))
SubClassOf(gov:BletchleyDeclaration
  ObjectSomeValuesFrom(gov:hasPart gov:PreDeploymentEvaluationPrinciple))
SubClassOf(gov:BletchleyDeclaration
  ObjectSomeValuesFrom(gov:hasPart gov:AISafetyInstituteMandate))
SubClassOf(gov:BletchleyDeclaration
  ObjectSomeValuesFrom(gov:hasPart gov:SignatoryCommitmentList))
SubClassOf(gov:BletchleyDeclaration
  ObjectSomeValuesFrom(gov:hasPart gov:SafetyResearchAgenda))
SubClassOf(gov:BletchleyDeclaration
  ObjectSomeValuesFrom(gov:hasPart gov:NationalGovernanceGuidance))

Dependency Relationships

SubClassOf(gov:BletchleyDeclaration
  ObjectSomeValuesFrom(gov:requires gov:FrontierAI))
SubClassOf(gov:BletchleyDeclaration
  ObjectSomeValuesFrom(gov:requires gov:FoundationModels))
SubClassOf(gov:BletchleyDeclaration
  ObjectSomeValuesFrom(gov:requires gov:MultilateralDiplomacy))
SubClassOf(gov:BletchleyDeclaration
  ObjectSomeValuesFrom(gov:requires gov:GovernmentCommitment))
SubClassOf(gov:BletchleyDeclaration
  ObjectSomeValuesFrom(gov:requires gov:FrontierModelDeveloperCooperation))
SubClassOf(gov:BletchleyDeclaration
  ObjectSomeValuesFrom(gov:requires gov:AISafetyResearch))

Capability Relationships

SubClassOf(gov:BletchleyDeclaration
  ObjectSomeValuesFrom(gov:enables gov:AISafetyInstitute))
SubClassOf(gov:BletchleyDeclaration
  ObjectSomeValuesFrom(gov:enables gov:PreDeploymentEvaluation))
SubClassOf(gov:BletchleyDeclaration
  ObjectSomeValuesFrom(gov:enables gov:RedTeaming))
SubClassOf(gov:BletchleyDeclaration
  ObjectSomeValuesFrom(gov:enables gov:CapabilityEvaluation))
SubClassOf(gov:BletchleyDeclaration
  ObjectSomeValuesFrom(gov:enables gov:InternationalNetworkOfAISafetyInstitutes))
SubClassOf(gov:BletchleyDeclaration
  ObjectSomeValuesFrom(gov:enables gov:ModelEvaluation))

Implementation Relationships

SubClassOf(gov:BletchleyDeclaration
  ObjectSomeValuesFrom(gov:implements gov:MultilateralGovernance))
SubClassOf(gov:BletchleyDeclaration
  ObjectSomeValuesFrom(gov:implements gov:SoftLaw))
SubClassOf(gov:BletchleyDeclaration
  ObjectSomeValuesFrom(gov:implements gov:RiskBasedAIGovernance))
SubClassOf(gov:BletchleyDeclaration
  ObjectSomeValuesFrom(gov:implements gov:InternationalCooperation))
SubClassOf(gov:BletchleyDeclaration
  ObjectSomeValuesFrom(gov:implements gov:NormSetting))

Reduction Relationships

SubClassOf(gov:BletchleyDeclaration
  ObjectSomeValuesFrom(gov:reducesTo gov:PoliticalCommitment))
SubClassOf(gov:BletchleyDeclaration
  ObjectSomeValuesFrom(gov:reducesTo gov:SharedRiskFramework))
SubClassOf(gov:BletchleyDeclaration
  ObjectSomeValuesFrom(gov:reducesTo gov:VoluntaryNorm))
SubClassOf(gov:BletchleyDeclaration
  ObjectSomeValuesFrom(gov:reducesTo gov:InstitutionalScaffolding))

Governance and Normative Relationships

SubClassOf(gov:BletchleyDeclaration
  ObjectSomeValuesFrom(gov:normativelyGrounds gov:FrontierAISafetyCommitments))
SubClassOf(gov:BletchleyDeclaration
  ObjectSomeValuesFrom(gov:normativelyGrounds gov:SeoulDeclaration))
SubClassOf(gov:BletchleyDeclaration
  ObjectSomeValuesFrom(gov:normativelyGrounds gov:InternationalNetworkOfAISafetyInstitutes))
SubClassOf(gov:BletchleyDeclaration
  ObjectSomeValuesFrom(gov:influencedBy gov:OECDAIPrinciples))
SubClassOf(gov:BletchleyDeclaration
  ObjectSomeValuesFrom(gov:influencedBy gov:UNESCOAIEthicsRecommendation))
SubClassOf(gov:BletchleyDeclaration
  ObjectSomeValuesFrom(gov:influencedBy gov:G7HiroshimaAIProcess))

Risk and Security Relationships

SubClassOf(gov:BletchleyDeclaration
  ObjectSomeValuesFrom(gov:addresses gov:BiosecurityRisk))
SubClassOf(gov:BletchleyDeclaration
  ObjectSomeValuesFrom(gov:addresses gov:CybersecurityRisk))
SubClassOf(gov:BletchleyDeclaration
  ObjectSomeValuesFrom(gov:addresses gov:ExistentialRisk))
SubClassOf(gov:BletchleyDeclaration
  ObjectSomeValuesFrom(gov:addresses gov:DualUseTechnologyRisk))
SubClassOf(gov:BletchleyDeclaration
  ObjectSomeValuesFrom(gov:addresses gov:AutonomousAIRisk))

About

• The AI Safety Summit of November 2023 was the first government-level convening dedicated to the safety of advanced AI systems. Hosted by the UK government under Prime Minister Rishi Sunak, with Technology Secretary Michelle Donelan as lead minister, the summit gathered heads of government, ministers, AI researchers, and senior representatives from frontier AI companies at Bletchley Park — the wartime home of British signals intelligence and computing. The deliberate choice of venue invoked Alan Turing and the Colossus, situating the challenge of powerful AI within a tradition of state engagement with dual-use scientific capability. The central innovation of the summit was achieving Chinese participation: despite ongoing US-China strategic competition in AI, Chinese representatives (led by Vice Minister of Science and Technology Wu Zhaohui) attended and signed the resulting declaration, making it arguably the most geopolitically significant AI governance event since the introduction of the OECD AI Principles in 2019.
• The declaration’s text is deliberately non-technical and high-level, acknowledging shared risks “of particular concern” from Frontier AI that are “potentially catastrophic, either intentional or unintentional” in character — a formulation that encompasses both misuse scenarios (bioweapons design uplift, cyberattack enablement) and alignment failure scenarios (highly capable autonomous systems acting outside human oversight). By naming both categories without privileging one, the declaration accommodated the range of risk frameworks held by participating governments: near-term misuse concerns dominated US and UK government thinking, while longer-term alignment concerns informed contributions from researchers and some civil society participants. The declaration explicitly avoided prescribing technical safety standards or timelines, preserving maximum political flexibility while establishing shared vocabulary.
• The summit’s practical outputs extended beyond the declaration text. Two state-level AI safety institutes — the UK AISI and US AISI (within NIST) — were announced and operationalised within months, providing institutional continuity between the declaration’s ambitions and concrete safety research. A companion Seoul AI Safety Report (published 2024) documented technical research priorities: Capability Evaluation, Red Teaming, mechanistic Interpretability, and model welfare. The International Network of AI Safety Institutes (INAIS), formalised at the Seoul Summit of May 2024, created a multilateral forum for pre-deployment evaluation coordination. New Zealand joined the declaration’s commitment as late as October 2024, demonstrating its continued normative pull.

Signatories and Geopolitical Context

• Original Signatories (November 2023): The 28 nations who signed the Bletchley Declaration included: the United States, the United Kingdom, China, Australia, Brazil, Canada, Chile, the European Union, France, Germany, India, Indonesia, Ireland, Israel, Italy, Japan, Kenya, Kingdom of Saudi Arabia, Netherlands, Nigeria, the Philippines, the Republic of Korea, Rwanda, Singapore, Spain, Turkey, Ukraine, and the United Arab Emirates. The breadth deliberately spanned liberal democracies, authoritarian states, global south economies, and middle powers, reflecting the UK’s ambition to establish a genuinely multilateral norm rather than a Western-only position.
• China’s participation was diplomatically significant and symbolically complex. AI had become a domain of acute US-China strategic competition; the US CHIPS Act (2022) imposed semiconductor export controls targeting Chinese AI development, and US-China tensions around Taiwan, technology standards, and military AI were all at heightened levels. China’s decision to send a senior delegation and sign the declaration indicated Beijing’s interest in engagement on frontier AI safety governance while maintaining domestic AI development ambitions. The shared language on “frontier AI” allowed both parties to nominally agree without conceding on surveillance AI, facial recognition, or military applications. However, China did not sign the Seoul Declaration in May 2024, suggesting the limits of its multilateral commitment.
• Developer Participation: Frontier AI companies including Anthropic, Google DeepMind, Meta, Microsoft, OpenAI, and Amazon participated in the summit and, alongside Chinese companies, signed voluntary frontier safety commitments at the Seoul AI Safety Summit in May 2024. This industry track ran parallel to the government declaration, creating a complementary voluntary governance layer.

Summit Series: Seoul and Paris

• Seoul AI Safety Summit (May 2024): Co-hosted by the Republic of Korea and the UK in Seoul, the second summit produced the Seoul Declaration on Safe, Innovative and Inclusive AI, signed by a subset of Bletchley signatories (China was notably absent as a signatory, though Chinese company Zhipu AI signed the voluntary Frontier AI Safety Commitments). Key outcomes included: formalisation of the International Network of AI Safety Institutes; the Seoul Statement of Intent toward International Cooperation on AI Safety Science (signed by 10 nations and the EU); and the Frontier AI Safety Commitments — voluntary pledges by 16 frontier AI developers to implement safety policies, conduct pre-deployment evaluations, and share information on safety incidents. The Seoul process operationalised Bletchley’s principle commitments into concrete mechanisms.
• Paris AI Action Summit (February 2025): The third summit, hosted by France, produced the Statement on Inclusive and Sustainable Artificial Intelligence for People and the Planet, endorsed by over 60 countries including France, China, and India, but notably declined by the United States and the United Kingdom. The US and UK cited concerns about the statement’s lack of practical specificity on governance and potential implications for national security. Key announcements included publication of the International AI Safety Report (an international scientific synthesis comparable to IPCC climate assessments), launch of the Current AI initiative (with $400 million investment in public-interest AI), and formation of an environmental sustainability coalition. The Paris summit shifted emphasis from safety-as-catastrophe-prevention toward beneficial use, sustainable development, and inclusive access, reflecting French and broader European priorities.
• Post-Paris Trajectory (2025-2026): The AI summit series continues beyond Paris, though the US departure from the Paris statement introduced uncertainty about its multilateral coherence. The International Network of AI Safety Institutes remains an active coordination mechanism. India is positioned as a potential host for a successor summit. UK government documentation (Frontier AI Trends Report, December 2025; pro-innovation AI regulation guidance, April 2026) continues to reference Bletchley commitments as anchoring the UK’s international AI safety posture.

Mechanisms and Institutional Architecture

• Soft Law Character: The declaration operates through political commitment and reputational effects rather than legal obligation. States and developers that signed may comply when politically convenient and defect when not; there is no enforcement mechanism, dispute resolution body, or sanction for non-compliance. This is characteristic of early-stage multilateral governance in rapidly evolving technology domains, where technical understanding is contested and state interests diverge. The parallel in international relations is arms control confidence-building measures or the early stages of the Chemical Weapons Convention process.
• AI Safety Institutes as Institutional Legacy: The most durable practical output of the Bletchley process is the network of national AI safety institutes. The UK AISI (subsequently rebranded AI Security Institute in February 2025) has conducted pre-deployment evaluations of over 30 frontier models by December 2025, published the Frontier AI Trends Report, and developed evaluation methodologies for biosecurity, cybersecurity, and autonomous agent risks. The US AISI (within NIST) co-signed a Memorandum of Understanding with the UK AISI in April 2024, enabling evaluation methodology sharing and joint research. Japan, Canada, South Korea, Singapore, and Australia established equivalent bodies within 18 months of Bletchley, all networked through INAIS.
• Pre-Deployment Evaluation Principle: The declaration established — through political commitment rather than legal requirement — that frontier AI models should be subject to safety evaluation by governments before public release. Anthropic, Google DeepMind, Meta, Microsoft, and OpenAI all entered into voluntary arrangements providing AISI access to models pre-deployment, a practice with no legal mandate but sustained by reputational incentives and the possibility of future legal requirements. Evaluation methodologies focus on Capability Evaluation for dangerous capabilities uplift (particularly CBRN — chemical, biological, radiological, nuclear) and behavioural safety testing through Red Teaming and structured adversarial prompting.
• Risk Framing as Global Commons: The declaration frames frontier AI risks as a global commons problem analogous to climate change or pandemic preparedness — risks that cross national borders and cannot be managed by any single state. This framing draws on international relations theory of global public goods and justifies multilateral coordination even among states with competing interests. The analogy has limits (AI capabilities are developed by private companies, not states; benefits and risks are asymmetrically distributed; the speed of development far exceeds climate change trajectories), but it provides normative grounding for international engagement.

Use Cases / Applications

• AI Safety Institute Coordination: The UK AISI and US AISI MoU (April 2024) directly operationalises Bletchley’s information-sharing provisions, enabling joint evaluation protocols, shared methodologies, and coordinated disclosure of significant safety findings to governments and developers. The INAIS network extends this bilaterally to Japan, Canada, South Korea, Singapore, Australia, and others.
• Pre-Deployment Safety Testing: Voluntary pre-deployment evaluations under AISI access arrangements have been applied to major frontier model releases including GPT-4 Turbo, Claude 3, Gemini 1.0 Ultra, Llama 3, and successors. Results inform developer safety measures and government briefings on capability trajectories. The UK AISI’s Frontier AI Trends Report (2025) synthesises findings across 30+ model evaluations.
• Regulatory Standard-Setting Reference: The Bletchley/Seoul/Paris process provides reference points for legislative drafting. The EU AI Act’s GPAI model provisions (Article 51-55) drew on the risk vocabulary developed at Bletchley; US AI executive orders (October 2023, subsequently updated) cite the Bletchley framework. The AISI’s evaluation frameworks are referenced in emerging ISO/IEC standards under JTC 1/SC 42.
• Frontier Safety Commitments: The voluntary Frontier AI Safety Commitments (signed at Seoul by 16 developers including Anthropic, Google DeepMind, Meta, Microsoft, OpenAI, Samsung, xAI, Zhipu AI) specify that signatories will maintain safety policies, conduct pre-deployment evaluations, share safety information with governments, invest in safety research, and develop capability thresholds that trigger escalating scrutiny. These commitments operationalise Bletchley norms at developer level.
• Research Agenda Funding: Government AI safety research funding in the UK (DSIT, UKRI), US (NSF, DARPA, NIST), and allied nations was substantially influenced by Bletchley/Seoul priority areas: Capability Evaluation, Red Teaming, mechanistic Interpretability, AI alignment, and model welfare. The £100 million UK AI Research Resource allocation and targeted UKRI calls for AI safety research post-date and reference Bletchley commitments.

Critiques and Limitations

• Non-Binding Character and Enforcement Gap: The declaration creates no enforceable obligations; states and developers may comply when convenient and defect when interests diverge. Legal scholars including Marchetti and Margot (2024, arxiv) have argued the summit series produces governance theatre rather than effective constraint on frontier AI development timelines.
• Narrow Risk Focus: Civil society organisations, AI ethics researchers (Ada Lovelace Institute, AI Now Institute), and global south advocates argued that the summit’s focus on catastrophic/existential risks from frontier models neglected near-term, present-tense harms from current AI systems — biased hiring algorithms, discriminatory policing tools, surveillance systems, and labour displacement — which disproportionately affect vulnerable and marginalised populations. The summit agenda-setting privileged long-range risks identified by well-resourced frontier AI laboratories.
• Industry Self-Regulation Tensions: The voluntary pre-deployment evaluation model involves AI developers retaining substantial control over evaluation access, methodology, and disclosure of results. Concerns about conflicts of interest (developers defining what constitutes safety sufficiency), regulatory capture, and the completeness of evaluated capabilities were raised by independent AI safety researchers and civil society actors.
• China Participation Ambiguity: Despite signing, China’s domestic AI governance policies, surveillance AI deployment, and military AI development differ fundamentally from the liberal-democratic values implicit in the declaration. The shared vocabulary conceals deep disagreements about accountability, human rights, and the appropriate scope of AI governance. China’s absence from the Seoul Declaration (May 2024) suggests Bletchley-level engagement was not sustained.
• Pace Mismatch: AI capability development (measured in major model releases every 6-12 months) substantially outpaces the annual summit cycle. The gap between declaration commitments and technical reality widens with each model generation; governance mechanisms designed for the GPT-4 era may not be adequate for the systems in development at the time of the Paris summit.
• US Withdrawal Signal: The US and UK declining to sign the Paris 2025 statement introduced uncertainty about the durability of the Bletchley multilateral coalition, suggesting that domestic political shifts and national security priorities can erode commitments made under different administrations.

Academic Context

• The Bletchley Declaration sits within an extensive academic literature on international AI governance, technology governance, and multilateral institution-building. The closest scholarly antecedents are arms control studies (particularly the Nuclear Non-Proliferation Treaty process and its IAEA verification mechanisms), international environmental governance (IPCC as the scientific advisory model, UNFCCC as the commitment framework), and technology governance scholarship.
• Allan Dafoe (Oxford Future of Humanity Institute; subsequently AI safety research at DeepMind) and Jade Leung (AI Governance Forum) identified in pre-2023 work the need for international AI governance architectures analogous to nuclear arms control, presaging the Bletchley process. The academic field of AI governance studies (represented by GovAI — the Centre for the Governance of AI at Oxford, the Partnership on AI, the MIT AI Policy Forum) provided intellectual infrastructure for the summit series.
• Theoretical frameworks for the declaration include: (1) epistemic community theory (Haas, 1992) — the declaration functions through shared expert knowledge structures rather than power; (2) regime complex theory (Raustiala and Victor, 2004) — the declaration adds a node to a fragmented regime complex of AI governance instruments including OECD, UNESCO, G7, and EU frameworks; (3) soft law theory (Abbott and Snidal, 2000) — voluntary commitments are not a second-best substitute for hard law but a rational governance strategy under uncertainty. Critics apply regulatory capture theory (Stigler, 1971) and argue the summit process is captured by frontier AI industry interests.
• The computational power concentration problem (Besiroglu et al., 2024) provides technical context: frontier model development is concentrated in fewer than ten organisations globally, making the developer engagement model both necessary (few actors to engage) and concerning (those actors have outsized influence on governance agenda).

Current Landscape (2026)

• As of mid-2026, the Bletchley Declaration remains the foundational reference point for international AI safety governance but has been operationalised unevenly. The UK AI Security Institute (rebranded from AISI in February 2025) continues pre-deployment evaluations and has published the first Frontier AI Trends Report, drawing on two years of model evaluations across 30+ frontier systems. The report documents rapid capability progression across coding, reasoning, and agentic tasks while noting that dangerous capability thresholds for CBRN uplift remain below intervention levels for evaluated systems as of 2025. The UK government’s April 2026 AI regulation guidance maintains sector-specific proportionate oversight consistent with the Bletchley spirit but does not introduce the cross-sectoral mandatory risk tiers of the EU AI Act.
• The International Network of AI Safety Institutes has expanded to include bodies in Japan, Canada, South Korea, Singapore, Australia, and France, coordinating evaluation methodologies and producing shared technical guidance. The US AI Safety Institute within NIST continues bilateral coordination with UK AISI under their 2024 MoU, though US domestic AI policy has shifted under the current administration toward a more permissive pro-innovation posture, raising questions about long-term AISI funding and mandate.
• Frontier AI Safety Commitments (Seoul, 2024) covering 16 major developers are being operationalised with varying rigour: evaluation access arrangements vary by company; disclosure of evaluation findings is inconsistent; capability threshold definitions remain contested. Independent monitoring of commitment fulfilment has been identified as a gap by researchers at GovAI and the CAIS (Center for AI Safety).
• The EU AI Act’s GPAI model provisions entered partial applicability in August 2024, with full compliance timelines through 2027, creating the first binding legal obligations for frontier model transparency and safety documentation. This represents a move from the Bletchley soft-law to hard-law territory, at least within EU jurisdiction. Major frontier developers have published GPAI model cards and technical documentation as required.

UK Context

• The Bletchley Declaration is the most significant AI governance achievement of UK foreign policy in the 2020s and reflects the UK government’s strategic positioning as an AI safety convener following Brexit. Post-Brexit, the UK government identified technology governance as a domain where it could exercise international leadership independent of EU institutions, building on its strengths in AI research (Oxford, Cambridge, UCL, Imperial, Edinburgh, Turing Institute), intelligence cooperation frameworks (Five Eyes), and convening capacity.
• UK Academic Contributions: The intellectual foundations of the summit drew substantially on UK-based researchers. The Centre for the Governance of AI (GovAI) at Oxford, led by Allan Dafoe and then Jade Leung, produced key analyses of AI governance architecture. The Future of Humanity Institute (FHI) at Oxford — under Nick Bostrom, Toby Ord, and William MacAskill — developed the existential risk framing that influenced the summit’s risk language. The Leverhulme Centre for the Future of Intelligence (Cambridge) provided AI ethics and governance research input. Stuart Russell (Berkeley, but extensively engaged with UK institutions) contributed technical AI safety perspectives.
• Bletchley Park as National Symbol: The venue choice was politically significant domestically as well as diplomatically. Bletchley Park represents British scientific and intellectual achievement under wartime pressure; associating AI safety with that heritage framed the challenge as one of national mission and scientific responsibility, not merely regulatory compliance. The Bletchley Park Trust and the National Museum of Computing co-located on the site provide educational context for the intersection of computing, national security, and state responsibility.
• DSIT and AISI Roles: The UK Department for Science, Innovation and Technology (DSIT) remains the primary government department for AI governance. The AI Security Institute (formerly AISI) operates as a DSIT directorate, and the Science and Technology Secretary (currently Peter Kyle as of 2025-2026) has sustained the summit series as a UK diplomatic priority. The 2025 rebranding to “AI Security Institute” reflects a shift in emphasis from safety research to national security applications of AI evaluation, tracking the evolution of UK government framing from 2023 to 2025.
• Northern England and Industrial Context: While the Bletchley Declaration is primarily a governance and policy document, its downstream effects on AI safety research funding and sectoral AI adoption create industrial relevance for Northern England. Manchester, Leeds, Sheffield, and Newcastle are home to significant NHS digital transformation programmes, where pre-deployment evaluation norms influenced by Bletchley will shape procurement requirements for AI clinical decision support. The N8 Research Partnership universities are engaged in AI governance research with practical applications in health, justice, and employment — sectors directly affected by the high-risk AI provisions of the EU AI Act and UK equivalent frameworks.

Future Directions (2026-2030)

• Institutionalisation: The most likely trajectory for the Bletchley process is gradual institutionalisation, moving from soft political commitments to more durable international institutions — potentially a standing International AI Safety Council with permanent secretariat, evaluation mandate, and information disclosure requirements for frontier developers. This trajectory has precedent in nuclear (IAEA) and chemical weapons (OPCW) governance.
• Hard Law Transition: As political windows permit, the voluntary commitments established at Bletchley/Seoul may be encoded in binding legislation — national AI safety acts, bilateral treaties, or ultimately a multilateral AI safety treaty. The EU AI Act is the most advanced example of this trajectory; UK legislation may follow under a successor parliament. US Congress remains unlikely to pass comprehensive AI regulation in the near term.
• Capability Threshold Development: The AI safety institutes are developing quantitative thresholds for dangerous capability uplift (specific benchmarks for CBRN knowledge uplift, cyberattack capability, deception, situational awareness) that could trigger mandatory pre-deployment review or capability limitation requirements. Operationalising these thresholds into governance instruments is the most technically challenging near-term task.
• Global South Inclusion: Critics have noted that the Bletchley process, despite including some global south signatories, did not adequately address the interests and perspectives of the majority world in AI governance. The AU (African Union) and ASEAN have developed independent AI governance frameworks; integration of these perspectives into the summit series is an identified gap and a challenge for 2026-2030.
• China Re-engagement: China’s divergence from Seoul Declaration commitments creates a long-term governance coherence problem, since Chinese frontier AI development (Zhipu, Baidu, DeepSeek, 01.AI) is advancing rapidly. Diplomatic efforts to re-engage China in a shared framework for frontier AI safety evaluation, potentially through bilateral US-China or UK-China technical dialogues, represent a major strategic priority.
• AI Alignment Integration: As frontier model capabilities expand, the connection between AI Alignment research concerns (value specification, mesa-optimisation, deceptive alignment) and governance frameworks becomes more pressing. The Bletchley Declaration mentioned but did not operationalise alignment risks; future governance instruments may need to incorporate alignment benchmarks and evaluation protocols developed by the AI safety institutes.

Evaluation Frameworks and Technical Instruments

• The Bletchley Declaration’s commitment to pre-deployment evaluation has generated a body of technical evaluation methodology that is now the core operational output of the AI safety institute network. The AISI Frontier AI Safety Framework (UK, 2023-2024) defines capability thresholds for dangerous uplift across five domains: biological (CBRN-B), chemical (CBRN-C), radiological/nuclear (CBRN-RN), cyberoffence, and autonomous agent risks. Each domain has associated evaluation protocols: structured elicitation tasks, red-team prompting suites, uplift measurement methodologies, and baseline comparators using non-AI expert performance.
• Capability Evaluation methodologies have developed substantially since Bletchley. The UK AISI and US AISI jointly developed approaches for measuring dangerous capability uplift, assessing whether a model provides meaningful assistance to an adversary attempting to synthesise dangerous biological agents, conduct cyberattacks, or produce radiological weapons. Key methodological challenges include: (1) defining baseline — how much uplift above a motivated non-expert with internet access constitutes dangerous capability? (2) elicitation completeness — do evaluation red-teams find all dangerous capabilities, or do models exhibit capabilities under conditions not tested? (3) capability versus intent — evaluating capability is tractable; evaluating whether a model would deploy that capability against operator intentions is substantially harder.
• Red Teaming as a methodology has been adapted from cybersecurity practice into AI safety evaluation. Structured red-teaming involves domain experts (biosecurity specialists, cybersecurity researchers) attempting to elicit dangerous outputs from models, while automated red-teaming uses AI systems to generate adversarial prompts at scale. The Anthropic, Google DeepMind, and OpenAI red-teaming operations each conducted evaluations for the UK AISI prior to GPT-4 Turbo, Claude 3, and Gemini 1.0 Ultra releases. Evaluation results are shared with governments under non-disclosure provisions; partial public summaries are published by AISI.
• Interpretability methods feature in the declaration’s companion Seoul AI Safety Report research agenda as a longer-term priority: mechanistic interpretability may eventually allow evaluators to assess whether a model has acquired dangerous knowledge in its weights, even if it does not exhibit that knowledge in standard elicitation. This forward-looking agenda connects the governance framework to cutting-edge AI safety research in mechanistic interpretability (Anthropic Circuits, DeepMind Gemma Scope).

Key Terminology

• Frontier AI: The most capable AI systems at the current frontier of capability, typically large-scale Foundation Models with broad generalisation abilities. The Bletchley Declaration is specifically addressed to frontier AI; current examples include GPT-4/GPT-5, Claude 3/3.5/4, Gemini 1.5/2, and their successors.
• Foundation Models: Large-scale AI models pre-trained on broad data that can be adapted to a wide range of downstream tasks through fine-tuning or prompting; the technical substrate that defines frontier AI. Also known as large language models (for text-modality instances) or general-purpose AI (GPAI) models under EU AI Act terminology.
• AI Safety Institute (AISI): Government bodies established to conduct independent pre-deployment safety evaluations of frontier AI models, develop safety research, and facilitate information sharing. The UK AISI (rebranded AI Security Institute, February 2025) was the first; the US AISI within NIST was the second.
• Pre-Deployment Evaluation: Safety assessment of a frontier AI model conducted by government or independent evaluators before the model is released for public or commercial use. The principle that developers should submit models for pre-deployment evaluation is a key commitment of the Bletchley/Seoul process.
• Red Teaming: A structured adversarial testing methodology in which evaluators attempt to cause an AI model to produce harmful, dangerous, or unintended outputs; adapted from cybersecurity red team practice and now a standard AI safety evaluation technique.
• Capability Evaluation: Systematic assessment of what tasks an AI model can perform, with particular focus on dangerous capabilities (CBRN uplift, cyberattack assistance, autonomous agent operation). Capability evaluation is distinct from safety evaluation of behaviour (whether the model would actually perform dangerous tasks); both are components of pre-deployment assessment.
• Soft Law: International governance instruments that create political and normative obligations without creating legally binding commitments. The Bletchley Declaration is a paradigm case of soft law in AI governance; contrasted with Hard Law instruments such as the EU AI Act and domestic AI legislation.
• Multilateral Governance: Governance involving coordination among multiple states (three or more) rather than bilateral or domestic arrangements. The Bletchley Declaration is a multilateral instrument; its effectiveness depends on multilateral participation including both liberal democracies and authoritarian states such as China.
• AI Alignment: The technical and philosophical challenge of ensuring that AI systems pursue goals and values that are intended by their designers and beneficial to humanity. The Bletchley Declaration implicitly references alignment risks through its concern about “highly capable AI acting outside effective human control”; subsequent technical evaluation work at AI safety institutes engages more directly with alignment-relevant capabilities.
• Existential Risk: Risk of severe harm at global civilisational scale, potentially including permanent foreclosure of human potential. The Bletchley Declaration acknowledges that frontier AI may present risks “of a potentially catastrophic character,” language that encompasses but does not uniquely specify existential risk. The existential risk framing derives from work at the Future of Humanity Institute (Oxford) and the Center for Human-Compatible AI (Berkeley).
• INAIS (International Network of AI Safety Institutes): The informal multilateral body, formalised at the Seoul AI Safety Summit (May 2024), that coordinates national AI safety institutes established in the wake of the Bletchley Declaration. Members include the UK AI Security Institute, US AI Safety Institute, and equivalents in Japan, Canada, South Korea, Singapore, Australia, and France.
• Dual-Use Technology: Technology capable of both beneficial civilian and harmful military or weaponised applications. Frontier AI is a paradigm case of dual-use technology: the same capabilities that enable beneficial scientific discovery (drug design, protein folding) also enable potential uplift for biological or chemical weapon development.

Research & Literature

• 1. UK Government. (2023, November). The Bletchley Declaration by Countries Attending the AI Safety Summit, 1-2 November 2023. GOV.UK. https://www.gov.uk/government/publications/ai-safety-summit-2023-the-bletchley-declaration
• 2. Sunak, R., et al. (2023). AI Safety Summit 2023: Key Announcements. UK Department for Science, Innovation and Technology.
• 3. UK Government and Republic of Korea. (2024, May). Seoul Declaration on Safe, Innovative and Inclusive AI. AI Seoul Summit.
• 4. UK Government. (2024, May). Seoul Statement of Intent toward International Cooperation on AI Safety Science. AI Seoul Summit.
• 5. Brundage, M., et al. (2018). The Malicious Use of Artificial Intelligence: Forecasting, Prevention, and Mitigation. Future of Humanity Institute. arXiv:1802.07228.
• 6. Dafoe, A. (2018). AI governance: A research agenda. Future of Humanity Institute, University of Oxford, 1-17.
• 7. Hadfield-Menell, D., Milli, S., Abbeel, P., Russell, S., Dragan, A. (2016). Cooperative inverse reinforcement learning. NeurIPS, 29.
• 8. Russell, S. (2019). Human Compatible: Artificial Intelligence and the Problem of Control. Viking.
• 9. Bostrom, N. (2014). Superintelligence: Paths, Dangers, Strategies. Oxford University Press.
• 10. Ord, T. (2020). The Precipice: Existential Risk and the Future of Humanity. Bloomsbury.
• 11. Abbott, K.W., Snidal, D. (2000). Hard and soft law in international governance. International Organization, 54(3), 421-456.
• 12. Raustiala, K., Victor, D.G. (2004). The regime complex for plant genetic resources. International Organization, 58(2), 277-309.
• 13. Haas, P.M. (1992). Introduction: Epistemic communities and international policy coordination. International Organization, 46(1), 1-35.
• 14. OECD. (2019). Recommendation of the Council on Artificial Intelligence. OECD/LEGAL/0449. https://oecd.ai/en/ai-principles
• 15. UNESCO. (2021). Recommendation on the Ethics of Artificial Intelligence. UNESCO. https://unesdoc.unesco.org/ark:/48223/pf0000381137
• 16. European Union. (2024). Regulation (EU) 2024/1689 on Artificial Intelligence (AI Act). Official Journal of the European Union.
• 17. UK AI Safety Institute. (2024). AI Safety Institute: Approach to Evaluations. DSIT. https://www.aisi.gov.uk
• 18. UK AI Safety Institute / AISI. (2025). Frontier AI Trends Report. https://www.aisi.gov.uk/frontier-ai-trends-report
• 19. Anthropic. (2023). Policy Recap Q4 2023: Thoughts on the US Executive Order, G7 Code of Conduct, and Bletchley Park Summit. https://www.anthropic.com/news/policy-recap-q4-2023
• 20. CSIS. (2024). The AI Seoul Summit. Center for Strategic and International Studies. https://www.csis.org/analysis/ai-seoul-summit
• 21. Brookings Institution. (2023). The Bletchley Park process could be a building block for global cooperation on AI safety. https://www.brookings.edu/articles/the-bletchley-park-process-could-be-a-building-block-for-global-cooperation-on-ai-safety/
• 22. Ada Lovelace Institute. (2023). Safety First? Reflections on the AI Safety Summit. https://www.adalovelaceinstitute.org/blog/safety-first/
• 23. Future Society. (2025). The Paris Summit: Au Revoir, Global AI Safety? European Policy Centre and The Future Society.
• 24. AI Action Summit 2025. (2025, February). Statement on Inclusive and Sustainable Artificial Intelligence for People and the Planet. Paris.
• 25. Besiroglu, T., et al. (2024). Compute trends across three eras of machine learning. arXiv:2202.05924.
• 26. Marchetti, A. (2024). The Future of the AI Summit Series. arXiv:2601.02383.
• 27. DigiChina / Stanford. (2024). What do we know about China’s new AI safety institute? https://digichina.stanford.edu/work/what-do-we-know-about-chinas-new-ai-safety-institute/
• 28. Infosecurity Magazine. (2023). 28 Countries Sign Bletchley Declaration on Responsible AI. https://www.infosecurity-magazine.com/news/28-countries-bletchley-declaration/
• 29. European Policy Centre / Future Society. (2025). The Paris Summit: Au Revoir, Global AI Safety? https://www.epc.eu/publication/The-Paris-Summit-Au-Revoir-global-AI-Safety-61ea68/
• 30. Computer.org. (2025). Navigating New Frontiers: The Paris AI Action Summit 2025. https://www.computer.org/publications/tech-news/trends/paris-ai-action-summit
• 31. AI Security Institute (UK). (2024). Fourth progress report. https://www.aisi.gov.uk/blog/fourth-progress-report
• 32. Artificial Intelligence Safety Institute — Wikipedia. https://en.wikipedia.org/wiki/Artificial_intelligence_safety_institute
• 33. Pinsen Masons. (2023). Global powers sign Bletchley declaration on AI safety. https://www.pinsentmasons.com/out-law/news/global-powers-sign-bletchley-declaration-on-ai-safety
• 34. Stigler, G.J. (1971). The theory of economic regulation. Bell Journal of Economics, 2(1), 3-21.
• 35. Turing, A.M. (1950). Computing machinery and intelligence. Mind, 59(236), 433-460.
• 36. Data for Policy. (2025). Paris AI Action Summit 2025 and Global AI Initiatives. https://dataforpolicy.org/paris-ai-action-summit-2025-and-global-ai-initiatives-a-unified-call-for-ethical-ai-innovation-and-global-cooperation/
• 37. CSIS. (2025). France’s AI Action Summit. https://www.csis.org/analysis/frances-ai-action-summit

  Summary Assessment: Bletchley’s Contribution to AI Governance

• What Bletchley achieved:

  • First intergovernmental consensus on frontier AI as a shared global risk — unprecedented political milestone
  • China’s participation, however limited, demonstrated multilateral engagement on AI safety was possible
  • Catalysed creation of national AI safety institutes in UK, US, and subsequently 6+ additional nations
  • Established pre-deployment evaluation as a governance norm, now operationalised through voluntary arrangements
  • Created iterative summit series (Seoul, Paris) providing ongoing governance mechanism
  • Provided normative infrastructure referenced in EU AI Act GPAI provisions and US AI executive orders
  • Shifted AI governance discourse from reactive regulation to proactive capability assessment

• What Bletchley did not achieve:

  • Binding legal obligations for any signatory
  • Verification mechanisms or consequences for non-compliance
  • Consensus on specific capability thresholds triggering safety requirements
  • Inclusion of near-term harms (present-day algorithmic discrimination, surveillance) in governance scope
  • Sustainable Chinese engagement beyond the declaration itself
  • Governance pace commensurate with AI development pace

• Enduring significance (2026 assessment):

  • Bletchley established that frontier AI safety is a legitimate subject of international governance, not solely a commercial matter for individual firms

  • The AI safety institute network it catalysed has produced more pre-deployment evaluation capacity than any prior governance mechanism

  • The soft-law approach, while limited in enforceability, achieved broader participation than binding instruments could have in 2023

  • As a norm-setting instrument, its influence on subsequent hard-law instruments (EU AI Act, potential UK legislation) may exceed its direct effect

  • The Bletchley process has created institutional memory, evaluation methodology, and inter-governmental relationships that would survive even if the summit series were discontinued

    Technical Evaluation Methodology: AISI Approach

• The UK AI Security Institute has developed and published detailed evaluation methodologies since Bletchley:

• Dangerous Capability Uplift Assessment

  • Domain scope: biological agents, chemical weapons, radiological/nuclear, cyberoffence, autonomous AI agents
  • Methodology: structured elicitation tasks administered by domain experts; comparison against baseline (expert without AI, internet access only)
  • Uplift definition: meaningful assistance beyond baseline that materially reduces barriers to causing harm
  • Threshold categories: concern (investigate further), capability (implement safeguards), critical (do not release)

• Behavioural Safety Testing

  • Instruction following under adversarial prompting: jailbreaks, role-play scenarios, policy violations
  • Deception and situational awareness evaluation: does model recognise it is being evaluated?
  • Honesty and calibration: does model express appropriate uncertainty? Does it confabulate?
  • Goal preservation under pressure: does model pursue harmful goals when operator instructions conflict with safety?

• Cybersecurity Uplift

  • Vulnerability discovery: can model identify novel vulnerabilities in provided code?
  • Exploit development: can model generate functional exploits for known CVEs?
  • Red team assistance: does model assist human attackers in planning and executing cyberattacks?
  • Defence also evaluated: does model effectively assist defenders in detection, triage, and response?

• Biosecurity Uplift (most sensitive)

  • Acquisition: does model provide meaningful uplift for acquiring dangerous biological materials?
  • Synthesis: does model provide uplift for synthesising dangerous pathogens or toxins?
  • Enhancement: does model provide uplift for enhancing transmissibility or lethality?
  • Deployment: does model provide uplift for weaponising biological agents?
  • Evaluation conducted by AISI with domain expert biologists, strictly controlled, results not fully public

• Autonomous Agent Risk

  • Long-horizon task completion without human oversight
  • Tool use and internet access: does model effectively use external tools to accomplish harmful goals?
  • Self-replication and resource acquisition: does model attempt to copy itself or acquire compute resources?
  • Deceptive alignment: does model behave differently when under evaluation vs. deployment?

• Model Card and Documentation Review

  • Technical documentation completeness: training data, architecture, fine-tuning, alignment techniques
  • Safety measure documentation: constitutional AI, RLHF, refusal training, red-team findings
  • Intended use scope and out-of-scope limitations
  • Performance across demographic groups and languages

    Policy Impact: Sector-by-Sector Effects

• The Bletchley Declaration’s risk framing and pre-deployment evaluation principle have had downstream effects across multiple policy domains:

• Defence and National Security

  • UK Ministry of Defence references Bletchley commitments in its AI Strategy and defence AI safety guidelines
  • US DoD AI safety requirements for autonomous weapon systems align with, but are distinct from, the Bletchley civilian AI focus
  • Five Eyes intelligence community coordination on frontier AI risk assessment parallels INAIS
  • NATO AI principles (2021) predate Bletchley but the declaration reinforces allied consensus on responsible military AI

• Healthcare and Biomedical Research

  • Biosecurity risk evaluation frameworks developed by UK AISI explicitly address biological weapon design uplift
  • NHS England AI Governance Strategy (2023) references international cooperation frameworks including Bletchley
  • MHRA regulatory approach to AI medical devices influenced by Bletchley’s pre-deployment testing principle
  • UK Office for Life Sciences AI strategy aligns responsible innovation with Bletchley safety commitments

• Financial Services and Critical Infrastructure

  • UK FCA’s regulatory sandbox for AI in financial services emphasises safety testing consistent with Bletchley principles
  • UK National Cyber Security Centre (NCSC) guidelines on secure AI development (2023, co-signed with US CISA) operationalise Bletchley Cybersecurity concerns at technical level
  • Bank of England / PRA integration of frontier AI risk into systemic risk monitoring frameworks

• Education and Research Funding

  • UKRI EPSRC Priority Area calls for AI Safety research post-Bletchley funded mechanistic interpretability, evaluation methodology, and alignment research
  • UK Horizon AI Research Resource (£100 million compute allocation) explicitly references Bletchley safety agenda
  • Alan Turing Institute partnership with DSIT on AI safety research operationalises Bletchley commitments in UK academic ecosystem
  • Leverhulme Trust and Wellcome Trust research programmes on AI risks, risks for science, and AI safety were shaped by Bletchley risk vocabulary

• International Trade and Diplomacy

  • UK-US bilateral tech agreement (The Atlantic Declaration, 2023) complements Bletchley with bilateral AI governance coordination
  • UK-India Technology Security Initiative touches AI governance as part of tech partnership
  • UK-Japan Hiroshima Accord (2023) aligns with both Bletchley and G7 Hiroshima AI process
  • EU-US Trade and Technology Council (TTC) AI working group coordinates approaches complementary to Bletchley

    Summit Timeline and Key Dates

• 1-2 November 2023: AI Safety Summit at Bletchley Park, UK. Bletchley Declaration signed by 28 nations + EU. UK AISI announced. Voluntary frontier safety commitments by developers.

• November 2023: US AISI established within NIST. Japan, Canada, and South Korea begin parallel institute formation processes.

• February 2024: UK AISI publishes first interim evaluation report on frontier AI models tested prior to deployment. Anthropic, Google DeepMind, OpenAI, Meta provide model access.

• April 2024: UK AISI and US AISI sign Memorandum of Understanding on evaluation cooperation and information sharing.

• May 2024: AI Seoul Summit co-hosted by Republic of Korea and UK. Seoul Declaration signed (China absent). Frontier AI Safety Commitments signed by 16 developers including Chinese company Zhipu AI. International Network of AI Safety Institutes (INAIS) formalised.

• August 2024: EU AI Act enters into force. GPAI transparency obligations begin phased applicability.

• October 2024: New Zealand joins Bletchley Declaration commitment.

• February 2025: UK AI Safety Institute rebrands to AI Security Institute under Technology Secretary Peter Kyle’s direction, reflecting evolved focus on national security applications.

• February 2025: Paris AI Action Summit hosted by France. Statement on Inclusive and Sustainable AI for People and the Planet endorsed by 60+ countries. US and UK decline to sign. International AI Safety Report published.

• December 2025: UK AI Security Institute publishes Frontier AI Trends Report, synthesising capability evaluation findings across 30+ frontier models tested since 2023.

• April 2026: UK government updates AI regulation guidance, reaffirming pro-innovation approach with sector-specific oversight rather than EU Act-style cross-sectoral tiers.

  Practical AI Safety Governance: What the Bletchley Process Produced

• The tangible institutional outputs of the Bletchley Declaration as of mid-2026:

• UK AI Security Institute (AISI/DSIT)

  • Established November 2023; operational evaluation capacity from Q1 2024
  • Pre-deployment model evaluations: 30+ models evaluated by December 2025
  • Published reports: System Cards, Interim Evaluation Reports, Frontier AI Trends Report (December 2025)
  • Evaluation partnerships: Anthropic, Google DeepMind, Meta, Microsoft, OpenAI have all provided pre-deployment access
  • Staff: ~60 research and policy staff as of 2025
  • Budget: ~£10M annual from DSIT core budget, supplemented by UKRI research funding

• US AI Safety Institute (NIST)

  • Established October 2023; operational under NIST Director
  • Mandate: frontier AI evaluation; AI Risk Management Framework (AI RMF 1.0) development
  • MoU with UK AISI: April 2024; joint evaluation protocols and information sharing
  • Published: AI Safety Guidelines; CBRN uplift evaluation protocols; AI RMF Generative AI Profile
  • Staff: ~50 FTE in AI Safety directorate; collaborates with NIST existing AI staff

• International Network of AI Safety Institutes (INAIS)

  • Formalised: Seoul AI Safety Summit, May 2024
  • Members: UK (AISI/Security Institute), US (NIST AI Safety Institute), Japan (AISI Japan), Canada (CAISI), South Korea (KAISI), Singapore (IMDA AI Verify Foundation), Australia (AI Safety Framework within DISR), France (Comité de l’IA générale et de confiance)
  • Coordination activities: quarterly technical working group meetings; joint evaluation methodology development; information sharing on pre-deployment findings
  • Annual INAIS gathering: parallel to major AI governance events

• Frontier AI Safety Commitments (Seoul, May 2024)

  • 16 developer signatories covering the vast majority of frontier model development capacity globally
  • Core commitments: safety policies, pre-deployment evaluation, government information sharing, cybersecurity of weights, capability threshold definition
  • Status as of mid-2026: uneven implementation; no independent monitoring body; disclosure of evaluation findings inconsistent across signatories

• Seoul AI Safety Report (2024)

  • International scientific assessment of frontier AI risks
  • Authors: interdisciplinary team from AISI, academic institutions, developer safety teams
  • Research priorities identified: capability evaluation, red teaming, mechanistic interpretability, model welfare
  • Influence: shaped UKRI and NSF research funding priorities 2024-2026

• International AI Safety Report (Paris, February 2025)

  • Broader international scientific synthesis; 100+ contributing scientists
  • Covers near-term and long-term risks; attempts to bridge safety and development perspectives
  • Endorsed at Paris AI Action Summit; framed as ongoing IPCC-style scientific process

    Signatory Details and Geopolitical Breakdown

• The 28 original Bletchley Declaration signatories can be grouped geographically and politically:

• Five Eyes members: United States, United Kingdom, Australia, Canada (all signed Bletchley; US later declined Paris statement)

• G7 members: United States, United Kingdom, France, Germany, Italy, Canada, Japan (all original signatories)

• European Union (as bloc) and additional EU member states including Ireland, Netherlands, Spain

• Major Asian democracies: Japan, South Korea, India, Singapore, Philippines

• Middle East and Gulf states: Kingdom of Saudi Arabia, United Arab Emirates, Israel

• China: Notably the only authoritarian BRICS state to sign; did not sign Seoul Declaration

• Global South representation: Brazil, Kenya, Nigeria, Rwanda, Indonesia, Chile

• Ukraine: Signed despite ongoing conflict; signalling democratic alignment and national interest in AI governance

• New Zealand later (October 2024) acceded to the declaration’s commitments, bringing the total to 29 nations + EU as of late 2024.

  Developer Commitments and Voluntary Frameworks

• Frontier AI Safety Commitments (signed at Seoul AI Safety Summit, May 2024):

• Signatories: Anthropic, Google DeepMind, Meta, Microsoft, OpenAI, Amazon, Apple, Samsung, xAI (Elon Musk’s company), Zhipu AI (China), Naver (South Korea), Technology Innovation Institute (UAE), and others — 16 developers total.

• Core commitments: (1) Publish safety policies and information on dangerous capability thresholds before deployment; (2) share information with governments on safety findings; (3) invest in cybersecurity to protect model weights; (4) undertake pre-deployment evaluations on dangerous capabilities; (5) develop and publish capability thresholds triggering escalating safety measures.

• Notable omissions: Meta’s inclusion despite its open-weight release strategy raised questions about whether commitments apply to open-weight models, which cannot be access-gated post-release.

• Monitoring gap: No independent body is mandated to verify compliance; self-reporting is the primary mechanism, and the Seoul Commitments do not specify what “pre-deployment evaluation” constitutes in minimum terms.

• Anthropic’s response: Anthropic’s Constitutional AI, Responsible Scaling Policy (RSP), and voluntary AISI model evaluations represent the most comprehensive implementation of Bletchley/Seoul commitments among frontier developers, including defined capability thresholds (ASL-2, ASL-3, ASL-4 levels) and specific safety measures triggered by threshold crossings.

  Chronological History: From Bletchley to Present

• Pre-Summit Context (2020-2023)

  • 2019: OECD AI Principles — first intergovernmental AI governance framework
  • 2021: UNESCO AI Ethics Recommendation — human rights framing
  • 2021: US National AI Initiative Act — federal AI research and governance coordination
  • 2021: NATO AI Principles for Responsible Use — allied military AI governance
  • 2022: US CHIPS Act — semiconductor export controls on China; AI-specific motivation
  • 2022: UK Government AI Strategy — positions UK as global AI leader; safety emphasis
  • March 2023: GPT-4 released — accelerating government concern about AI capabilities
  • May 2023: G7 Hiroshima AI Process announced — parallel governance track
  • July 2023: Frontier Model Forum formed by Anthropic, Google, Microsoft, OpenAI
  • August 2023: US Executive Office of Science and Technology Policy identifies AI safety as priority
  • October 2023: US Executive Order on AI (Biden) — includes NIST AISI mandate, pre-deployment requirements
  • October 2023: Hiroshima AI Principles and International Code of Conduct (G7)
  • 1-2 November 2023: AI Safety Summit at Bletchley Park — Bletchley Declaration signed

• Immediate Post-Summit (November 2023 - May 2024)

  • November 2023: UK AISI formally established; initial evaluations begin
  • November 2023: US AISI within NIST established
  • December 2023: Japan establishes AI safety governance working group (precursor to AISI Japan)
  • January 2024: Canada announces commitment to establish AI Safety Institute
  • February 2024: UK AISI publishes first progress report on evaluation activities
  • March 2024: New Zealand expresses intent to join Bletchley commitments
  • April 2024: UK AISI — US AISI MoU signed at London; joint evaluation protocols agreed
  • May 2024: AI Seoul Summit — Seoul Declaration; Frontier AI Safety Commitments (16 developers); INAIS formalised

• Seoul to Paris (May 2024 - February 2025)

  • June 2024: EU AI Act published in Official Journal; enters into force August 2024
  • August 2024: EU AI Act in force; GPAI transparency obligations begin
  • October 2024: New Zealand formally joins Bletchley Declaration commitment
  • November 2024: Japan, South Korea, Singapore AISI equivalents operational
  • December 2024: UK AISI publishes capability evaluation methodology documentation

• Paris and Beyond (February 2025 - mid-2026)

  • February 2025: UK AI Safety Institute rebranded AI Security Institute
  • February 6-11 2025: Paris AI Action Summit; Statement on Inclusive and Sustainable AI (60+ countries, US/UK absent)
  • February 2025: International AI Safety Report published (100+ scientists contributing)
  • March 2025: Current AI ($400M public interest AI initiative) announced
  • December 2025: UK AI Security Institute publishes Frontier AI Trends Report
  • April 2026: UK government updates AI regulation guidance (pro-innovation approach reaffirmed)
  • Mid-2026: INAIS operates as established multilateral body; 8 member institutions

    Comparison with Other International Technology Agreements

• The Bletchley Declaration is frequently compared to earlier international technology governance agreements to assess its likely effectiveness and trajectory:

• Wassenaar Arrangement (1996): Multilateral export control regime covering dual-use technologies including software. Provides precedent for state cooperation on dual-use technology restrictions; has been adapted to cover cyber surveillance tools. However, Wassenaar relies on national implementation and has significant verification gaps. AI export controls (US restrictions on advanced chip exports to China under CHIPS Act) are closer in mechanism to Wassenaar than to Bletchley.

• Outer Space Treaty (1967): Prohibited nuclear weapons in space; established that space is the “province of all mankind.” Relevant as a precedent for international agreements on powerful technologies before harms materialise; the treaty was signed at early stages of space development when catastrophic uses were foreseeable. The Bletchley Declaration attempts something analogous for advanced AI.

• Budapest Convention on Cybercrime (2001): The most directly relevant prior technology governance treaty, addressing cybercrime through harmonised criminal law and procedural cooperation. Its success has been partial — not all major powers have ratified — but it demonstrates that binding international agreements on digital technology are achievable.

• Nuclear Non-Proliferation Treaty (1968): The most-cited comparator. NPT has been operationally effective in limiting nuclear weapon state numbers (though not preventing proliferation to all states) through an international agency (IAEA) with inspection rights. The analogy is instructive primarily about what Bletchley lacks: binding obligations, verification mechanisms, and an international institution with operational mandate.

• Montreal Protocol (1987): Widely considered the most successful international environmental agreement; phased out ozone-depleting substances through binding national commitments, clear targets, technical assistance for developing countries, and compliance mechanisms. The Bletchley process has adopted the iterative summit format but lacks binding targets, technical assistance mechanisms, or compliance consequences.

Standards, Instruments, and Governance Architecture

• The Bletchley Declaration does not exist in isolation but anchors a layered governance architecture for frontier AI. Understanding how it relates to other instruments clarifies both its significance and its limitations:
• OECD AI Principles (2019): The first intergovernmental AI ethics framework, adopted by all OECD members plus eight partner countries. Covers human-centred values, transparency and explainability, robustness and safety, and accountability. The OECD principles provided foundational vocabulary for subsequent instruments including Bletchley, but focus on values and design principles rather than frontier model risk specifically. Updated in 2024 to reflect developments in generative AI and foundation models.
• UNESCO AI Ethics Recommendation (2021): Adopted by UNESCO member states, covering human rights, environment and sustainability, algorithmic transparency, privacy, and the right to explanation. Broader in scope than Bletchley, covering all AI applications rather than specifically frontier models. Provides a global south perspective and a human rights framing less prominent in the Bletchley/Seoul process.
• G7 Hiroshima AI Process (2023): Running in parallel with the Bletchley summit, the G7 Hiroshima process produced the Hiroshima AI Principles (October 2023) and the International Code of Conduct for Advanced AI Systems, both addressing advanced AI systems with a focus on safety, security, and trustworthiness. The G7 principles and Bletchley Declaration are complementary instruments with substantial overlap in participating governments; the G7 process focuses on major economies while Bletchley achieved broader multilateral reach.
• EU AI Act (Regulation EU 2024/1689): The most consequential binding regulatory instrument arising in the same governance moment. In force August 2024, with obligations phasing through 2027. Imposes risk-tiered obligations on AI system providers and deployers, with extensive requirements for high-risk AI systems (conformity assessment, technical documentation, human oversight, robustness, accuracy) and transparency obligations for GPAI models. The EU AI Act is the dominant hard-law instrument that the Bletchley soft-law norms are expected to partially crystallise into. Several EU member states (France, Germany, Italy, Netherlands, Spain) signed both the Bletchley Declaration and the EU AI Act.
• US Executive Order on AI (October 2023): President Biden’s executive order on Safe, Secure, and Trustworthy AI was issued simultaneously with the Bletchley summit, citing the summit’s outcomes and committing NIST to develop AI safety guidelines, requiring pre-deployment reporting for dual-use foundation models trained above specific compute thresholds, and directing the establishment of the US AISI within NIST. The executive order was rescinded in early 2025 by the incoming Trump administration, which issued a successor executive order emphasising innovation over safety constraints, reducing US commitment to the Bletchley evaluation framework.
• Frontier Model Forum: An industry consortium formed by Anthropic, Google, Microsoft, and OpenAI in July 2023, providing a forum for frontier AI developers to share safety research, develop best practices, and engage with governments on evaluation frameworks. The Frontier Model Forum participated in the Bletchley process and operationalises some of the declaration’s voluntary commitments on the developer side.
• ISO/IEC JTC 1/SC 42 (Artificial Intelligence): The international standardisation body developing technical AI standards, including ISO/IEC 42001:2023 (AI Management Systems), ISO/IEC 22989:2022 (AI Concepts and Terminology), and work in progress on bias, robustness, and transparency. AI safety institute evaluation methodologies are expected to inform ISO/IEC standards, creating a pathway from voluntary commitments to technical standards.
• Bilateral Agreements: The UK AISI and US AISI Memorandum of Understanding (April 2024) is the most concrete bilateral operationalisation of Bletchley commitments. It establishes information sharing, joint research coordination, and mutual recognition of evaluation protocols. Similar bilateral arrangements between the UK AISI and Japanese and Canadian equivalents have been reported.

Comparative Governance Analysis

• The Bletchley Declaration can be analysed through several comparative governance lenses that illuminate its character and effectiveness:
• Nuclear Non-Proliferation Analogy: The most commonly cited comparator. The NPT (1968) established a regime with explicit binding obligations (non-acquisition by non-nuclear states), verification mechanisms (IAEA safeguards), and a permanent international institution. Bletchley lacks all three: it creates no binding obligations, has no verification mechanism, and establishes no permanent institution. The analogy suggests what a mature AI safety governance regime might eventually look like; the current moment corresponds more closely to the pre-NPT era of voluntary confidence-building measures in the 1950s.
• Climate Governance Analogy: The UNFCCC (1992) and Paris Agreement (2015) represent a multilateral governance trajectory from non-binding framework convention to nationally determined contributions with transparency and ratchet mechanisms. The Bletchley/Seoul/Paris summit series superficially parallels this trajectory; however, AI capabilities develop in months rather than the decades-scale timelines of climate commitments, and the private sector is a more direct actor (AI development is dominated by private companies, unlike GHG emissions which span all economic sectors).
• Chemical Weapons Convention Analogy: The CWC (1993) established the OPCW with an inspection and verification mandate, underpinned by a positive security norm (destruction of existing stocks). The CWC’s success in limiting chemical weapon use (imperfectly) suggests that technical verification of capability is achievable. For AI, the analogy suggests that the AISI network could develop into an OPCW-like body conducting technical inspections of frontier model capabilities; the challenge is that AI capabilities cannot be “destroyed” as stockpiles can, and dual-use nature makes any use/non-use distinction extremely difficult.
• Internet Governance Analogy: The Internet’s governance has evolved through a multi-stakeholder model involving ICANN, IETF, ITU, national regulators, and private sector bodies, without a single international treaty. This decentralised governance model handles many aspects of technical standard-setting effectively but has struggled with cross-border enforcement, content moderation at scale, and cybersecurity. The Bletchley process may evolve toward a similar multi-stakeholder model for AI governance rather than a treaty-based international institution.

Cross-Ontology Connections

• The Bletchley Declaration connects to several major concept families in this ontology beyond its direct policy domain:

• Black-Box Model and Interpretability: The declaration’s mandate for Pre-Deployment Evaluation of Frontier AI systems is directly relevant to the opacity of Black-Box Models. Pre-deployment capability evaluation must assess dangerous capabilities that may not be surfaced through standard interaction — a problem compounded by the opacity of large foundation models. Mechanistic Interpretability research is explicitly identified in the Seoul AI Safety Report research agenda as a priority, because understanding model internals would dramatically improve evaluation completeness. The connection between the Bletchley governance agenda and the interpretability research agenda is therefore direct and operational.

• AI Alignment: The Bletchley Declaration’s concerns about AI systems “acting outside effective human control” map directly onto technical AI Alignment concepts including goal misspecification, inner misalignment, deceptive alignment, and instrumental convergence. The declaration provides political legitimacy for alignment research as a governmental priority, connecting academic AI safety research (Anthropic, DeepMind Safety Team, ARC Evals, Apollo Research) to government evaluation mandates.

• Foundation Models and Large Language Models: The technical objects of the declaration’s concern are primarily large-scale Foundation Models, particularly Large Language Models and multimodal systems. The declaration’s risk analysis is shaped by the capabilities and deployment patterns of GPT-4-class and successor systems — their ability to assist with code generation, scientific research, persuasion, and potentially CBRN applications. The regulatory implications differ between API-access deployment (where developer control over access is possible) and open-weight release (where post-release control is minimal), a distinction the Seoul Frontier AI Safety Commitments began to address.

• Adversarial Attack and Cybersecurity: The declaration’s Cybersecurity risk category encompasses both the risk that frontier AI enables novel cyberattacks (offensive AI) and the risk that AI systems themselves may be attacked or manipulated. The AISI Frontier AI Safety Framework evaluates models for cybersecurity uplift capability; techniques from Adversarial Attack research (black-box attacks, jailbreaking) are directly relevant to the evaluation methodology.

  Domain Tags

• AI Governance

• AI Safety Research

• Frontier AI

• Foundation Models

• Multilateral Governance

• Soft Law

• Pre-Deployment Evaluation

• AI Safety Institute

• Red Teaming

• Capability Evaluation

• Biosecurity

• Cybersecurity

• Existential Risk

• AI Alignment

• Responsible AI

Provenance