The Bletchley Declaration

Two days at Bletchley Park and the narrow claim

On 1 and 2 November 2023, representatives of 28 countries and the European Union met at Bletchley Park in Buckinghamshire, the wartime codebreaking site associated with Alan Turing, and issued a short joint text now known as the Bletchley Declaration. The document was the headline outcome of the first AI Safety Summit, convened by the United Kingdom government and described on the official summit page as a gathering of governments, leading AI companies, civil society and researchers. Its documented claim is modest and precise: the signatories declared a shared concern about the safety risks of "frontier" AI, agreed on a common vocabulary for describing those risks, and committed to cooperate through science and national policy. The declaration text runs to roughly 1,300 words and carries no binding obligations, no capability thresholds, no enforcement mechanism, and no budget.

The distance between what the event was and why it is remembered is unusually large here, and worth holding apart throughout. The narrow event is a diplomatic communiqué: a statement of intent signed by governments. The broad historical judgment, that AI safety became a subject of formal international coordination, rests not on the text alone but on what the text helped set in motion: a recurring summit series, a shared scientific report, and a set of national safety institutes. A declaration is a promise to keep talking, not a proof that a problem has been addressed. Reading it accurately means keeping those two levels of claim distinct.

What the text actually says: "frontier" as the unit of concern

The declaration's central move is to name its object. It focuses on "frontier AI," which it glosses as "highly capable general-purpose AI models, including foundation models," alongside relevant specific narrow systems. This framing matters because it deliberately scopes the concern to a small set of the most capable models rather than to all software that uses machine learning. The text warns that "there is potential for serious, even catastrophic, harm, either deliberate or unintentional, stemming from the most significant capabilities of these AI models," and it singles out the domains of cybersecurity and biotechnology, along with the risk that such systems could amplify disinformation.

Two other phrases carry weight. The document flags "unintended issues of control relating to alignment with human intent," which is the language of the AI safety research community entering an intergovernmental text for the first time at this level. And it structures its forward agenda around two commitments: first, "identifying AI safety risks of shared concern" and building "a shared scientific and evidence-based understanding of these risks"; second, "building respective risk-based policies across our countries." That pairing, a joint diagnosis plus nationally chosen treatments, is the whole operative content. It asks nations to converge on how to describe the problem while leaving each free to legislate as it sees fit.

Who signed, and why the guest list was the substance

The signatories are worth reading as carefully as the prose, because the composition of the room was arguably the more significant fact. The list spans the United States, the United Kingdom and the European Union alongside China, India, Brazil, Indonesia, Nigeria, Kenya, Rwanda, Saudi Arabia, the United Arab Emirates, Japan, the Republic of Korea, Singapore and others. The inclusion of both the United States and China in a single AI governance text, at a moment of sharp technological competition between them, was the outcome that observers found most striking. It did not signal agreement on export controls, model access or values; it signalled only that both were willing to co-sign a framing of catastrophic and safety risk.

That is a real but limited result. Getting rivals to endorse a shared description of a problem is a precondition for coordination, not coordination itself. The declaration explicitly commits its parties "to work together in an inclusive manner," and the breadth of the list, reaching well beyond the usual G7 and OECD memberships, was intended to give the framing legitimacy outside the wealthy West. Whether that legitimacy was genuine or ceremonial is a separate question, taken up below. New Zealand was later recorded as adding its name in 2024, but the founding cohort of the text was the 1 to 2 November 2023 group.

The gap it addressed: principles existed, a common table did not

Frontier AI governance in 2023 was not a blank page. The OECD AI Principles, first adopted in 2019 and updated in 2024, already committed dozens of countries to values such as transparency, robustness and accountability. Days before Bletchley, on 30 October 2023, G7 leaders agreed the Hiroshima Process International Code of Conduct for Advanced AI Systems, a voluntary risk-based code for developers. The same day, the United States issued its own domestic instrument, Executive Order 14110 on Safe, Secure, and Trustworthy Development and Use of Artificial Intelligence, which invoked the Defense Production Act to require reporting on the most compute-intensive frontier models.

What none of these provided was an inclusive forum in which geopolitical rivals sat at one table and endorsed a single description of frontier risk. The OECD framework does not include China; the Hiroshima code is a G7 product; the executive order was a national measure. The constraint Bletchley addressed was therefore not a shortage of principles but the absence of a broad, cross-bloc venue and a shared diagnosis. The declaration's contribution was to convene that venue and to record, in one text signed by both Washington and Beijing, that frontier AI safety was a legitimate object of collective attention. That is a coordination artifact, not a regulatory one.

How a non-binding declaration is designed to work

The mechanism here is diplomacy, not engineering. Unlike a model release or a technical safety method, the Bletchley Declaration changes nothing about how any system is built or deployed. Its causal pathway runs through commitment and expectation: by publicly co-signing, states raise the reputational cost of ignoring the agenda and create a scaffold for follow-on meetings where more concrete steps can be negotiated. The publication record on gov.uk is explicit that the output is a joint statement and a basis for further discussion, not a treaty.

This is the standard shape of soft-law international coordination, and its strengths and weaknesses are well understood. A declaration can move quickly, include parties who would never ratify a binding treaty, and establish shared language that later instruments can reference. But it cannot compel, verify or penalise. Its value is entirely contingent on what follows. A useful test, then, is not whether the November 2023 text bound anyone (it did not, by design) but whether the successor commitments it enabled produced anything measurable. Two concrete deliverables let that test be applied rather than asserted.

The concrete deliverable: a shared scientific assessment

The most tangible product traceable to Bletchley is a scientific one. The summit commissioned a "State of the Science" report to build the shared, evidence-based understanding the declaration called for. That commission became the International AI Safety Report, chaired by the Turing Award laureate Yoshua Bengio and produced by a large body of independent experts nominated by around 30 countries plus the European Union and the United Nations. An interim version appeared before the Seoul summit in May 2024, and the first full report was published in January 2025 ahead of the Paris AI Action Summit.

This is the part of the Bletchley legacy that most resembles the model of climate governance, where an intergovernmental scientific assessment (the IPCC) is kept institutionally separate from the political negotiation. The report does not set policy; it synthesises evidence on general-purpose AI capabilities and risks so that negotiators argue over a shared factual baseline. Its existence is the strongest available evidence that the declaration's first commitment, a shared scientific understanding, produced something durable rather than rhetorical. It is also honest about disagreement: the report repeatedly documents where experts diverge, which is a more credible output than false consensus.

The institutional legacy: safety institutes and a summit series

The declaration also coincided with, and helped normalise, a new institutional form: the national AI safety institute. On 2 November 2023 the UK published the overview of its AI Safety Institute, a state body created to evaluate advanced models and, in its own words, to "minimise surprise to the UK and humanity from rapid and unexpected advances in AI." The United States announced a counterpart, and a network of such institutes across several countries followed. These bodies gave governments an in-house technical capacity to test frontier models rather than relying solely on developers' self-reporting.

The summit itself became a series. The AI Seoul Summit 2024 produced the Seoul Declaration in May 2024 and, more concretely, the Frontier AI Safety Commitments, under which sixteen leading AI companies agreed to publish safety frameworks and to define thresholds of risk they would treat as intolerable. The Paris AI Action Summit followed in February 2025, notably re-weighting the agenda toward opportunity, adoption and energy rather than catastrophic risk, and ending without US and UK signatures on its main statement. That trajectory, from safety-first framing at Bletchley to a more contested emphasis at Paris, is part of the honest record of what the declaration set in motion.

What the declaration deliberately did not do, and what critics said it missed

The limitations are as important as the achievements, and several are structural rather than accidental. The text is non-binding and creates no verification, so a signatory can endorse it and change nothing. It says little about the harms that were already occurring in deployed systems in 2023, such as algorithmic bias, labour displacement, non-consensual imagery, copyright and the concentration of power in a few firms, because its explicit scope is the frontier. Critics from civil society argued that this framing over-indexed on speculative, catastrophic and long-horizon risks at the expense of present, distributed ones, and that a summit centred on the most capable models effectively centred the small number of companies and states that build them.

There was also a representation critique. Although the guest list was broad, decision-making power and technical capacity remained concentrated, and many Global South participants were present without a comparable stake in the outcomes. And the "frontier" abstraction, useful for focusing attention, can obscure that most AI harm arrives through ordinary systems rather than record-breaking ones. None of this makes the declaration empty, but it does bound the claim: it was a governance beginning aimed at a specific slice of the risk landscape, not a comprehensive settlement of AI's societal effects.

How Bletchley sits among the governance instruments of its era

Placing the declaration accurately requires comparing it with the harder instruments arriving around the same time, because they do different jobs. The EU AI Act, which entered into force on 1 August 2024 with phased application, is binding law with a tiered risk classification, prohibited practices and real penalties. Executive Order 14110 was an enforceable domestic directive, later rescinded under a new US administration in 2025, illustrating how national measures can reverse in a way a multilateral declaration is partly insulated against. In March 2024 the UN General Assembly adopted its first resolution on artificial intelligence, a US-led, consensus text co-sponsored by more than 120 states calling for safe, secure and trustworthy systems.

Against that backdrop, Bletchley occupies a specific niche: it is the inclusive, non-binding, agenda-setting layer that sits beneath the binding regional law, the reversible national orders and the universal UN resolution. Its comparative advantage is breadth of membership and speed; its comparative weakness is that it commands nothing. Reading it as a rival to the EU AI Act misunderstands both. The Act regulates; the declaration convenes.

How a reader should place the event in AI history

The defensible summary is that the Bletchley Declaration was the moment frontier AI safety acquired an inclusive intergovernmental forum, not the moment the problem was solved or even meaningfully constrained. Its narrow documented content is a shared framing and a two-part commitment. Its broader significance is inferred from its descendants: the Bengio-chaired scientific report, the network of safety institutes, and the Seoul commitments that turned a general agenda into specific corporate pledges. Those descendants, not the November 2023 prose, are what justify calling it a turning point in AI governance.

A useful way to navigate outward from this event is to keep asking three questions of everything downstream. Did a later instrument make a Bletchley commitment binding or merely restate it? Did a later body (an institute, a report) give the shared understanding real technical substance? And did a later summit narrow or widen the scope the declaration set? Holding those questions steady prevents two opposite errors: dismissing the declaration as a photo opportunity, and inflating it into a global regulatory regime it never claimed to be. It was a well-timed act of convening, and in international coordination, convening the right parties around a shared description is often the hardest and earliest step.

References

• AI Safety Summit (2023). The Bletchley Declaration by Countries Attending the AI Safety Summit, 1 to 2 November 2023.
• UK Government (2023). AI Safety Summit 2023: The Bletchley Declaration.
• UK Government (2023). AI Safety Summit 2023.
• UK Government (2023). AI Safety Institute: overview.
• Bengio and contributors (2025). International AI Safety Report.
• AI Seoul Summit (2024). Frontier AI Safety Commitments, AI Seoul Summit 2024.
• UK Government (2024). Seoul Declaration for safe, innovative and inclusive AI, AI Seoul Summit 2024.
• The White House (2023). Executive Order 14110: Safe, Secure, and Trustworthy Development and Use of Artificial Intelligence.
• European Commission (2023). Hiroshima Process International Code of Conduct for Advanced AI Systems.
• European Commission (2024). Regulatory framework on AI (the AI Act).
• OECD (2019). OECD AI Principles.
• United Nations (2024). General Assembly adopts landmark resolution on artificial intelligence.