AI Policy Wiki
Dashboard

Bletchley Declaration (AI Safety Summit, 2023)

high confidence · updated 2026-07-26

Non-binding declaration agreed by 28 countries and the EU at the UK's November 2023 AI Safety Summit, and the founding document of the summit series. Established 'frontier AI' as the object of international coordination and set a two-pillar agenda separating a shared evidence base from nationally divergent policy — the split that let the US, EU, UK, and China sign the same text, and that the later Seoul and Paris summits progressively lost.

The declaration agreed at the AI Safety Summit held at Bletchley Park, 1–2 November 2023, and the founding document of the international AI-summit series. Primary text: The Bletchley Declaration (AI Safety Summit, 1–2 November 2023).

Status

A political declaration, not a treaty or statute. It creates no obligations, sets no thresholds, and has no enforcement mechanism. Its instruments are hortatory — signatories "resolve," "affirm," and "encourage." Its significance is as a coordination point: it established a shared vocabulary and a standing process, and it is the reference against which later instruments in the series are measured.

Signatories

Twenty-eight countries plus the European Union, with New Zealand acceding on 23 October 2024. The signatory list included both the United States and China — the feature most often cited in assessing the declaration, and the one the subsequent summits did not reproduce.

Key provisions

The frontier-AI definition. "Highly capable general-purpose AI models, including foundation models, that could perform a wide variety of tasks — as well as relevant specific narrow AI that could exhibit capabilities that cause harm — which match or exceed the capabilities present in today's most advanced models." The threshold is relative rather than absolute, so it tracks the frontier rather than fixing a capability level.

Two risk sources: "potential intentional misuse or unintended issues of control relating to alignment with human intent," with cybersecurity, biotechnology, and disinformation named as domains of particular concern.

Developer responsibility. Frontier developers have "a particularly strong responsibility," discharged "through systems for safety testing, through evaluations, and by other appropriate measures," with "context-appropriate transparency and accountability."

The two-pillar agenda. A shared scientific and evidence-based understanding of risks, and risk-based national policies that "may differ based on national circumstances and applicable legal frameworks."

Comparison with the rest of the series

The declaration sits at the head of a lineage that fragmented successively:

InstrumentDateParticipantsCharacter
[[legislation/hiroshima-code-of-conduct\Hiroshima Code of Conduct]]Oct 2023G711 voluntary actions for developers
Bletchley DeclarationNov 202328 nations + EUState-level declaration; US and China both signed
[[legislation/seoul-frontier-ai-safety-commitments\Seoul Frontier AI Safety Commitments]]May 202416 companiesShifted the commitment-bearer from states to firms
[[legislation/paris-ai-action-summit-declaration\Paris AI Action Summit Declaration]]Feb 202564 signatoriesUS and UK declined to sign; reframed around "Inclusive and Sustainable AI"

The trajectory is the point: Bletchley obtained the broadest state-level agreement on the narrowest substantive commitment, Seoul obtained more specific commitments from a narrower set of private actors, and Paris obtained more signatories while losing two of the states whose participation gave Bletchley its significance.

What followed from it

The evidence-base pillar produced durable institutions: the International AI Safety Report under Yoshua Bengio's chairmanship, and the network of AI Safety Institutes — including the UK body later renamed the AI Security Institute (UK AI Safety Institute (AI Security Institute)). The policy pillar produced no convergence, which the declaration's own drafting anticipated by permitting divergence.

Key tensions

The declaration's breadth of agreement was purchased by its lack of specificity, and the two readings it supports have since diverged: the frontier-risk framing that dominates its central paragraphs, and the present-harms list — bias, privacy, transparency, human oversight — that it also affirms as urgent. Instruments later in the series have tended to pick one.

Relationships