28 Nations Sign First Pact on AI Risks at Bletchley Park

Bletchley Park, the Victorian mansion where Alan Turing and his team cracked the Nazi codes during the Second World War, was the setting today for another moment with historic ambitions: the signing of the Bletchley Declaration, the first international agreement focused specifically on the risks of the most advanced artificial intelligence.

Twenty-eight countries and the European Union signed the document on the opening day of the AI Safety Summit, hosted by Rishi Sunak's British government. The summit runs through tomorrow, Thursday, with further working sessions bringing together governments, companies and academics.

A Brief Text with a Notable Diplomatic Achievement

The Bletchley Declaration is not a binding treaty and doesn't set up any enforcement mechanisms. It's a statement of intent: it acknowledges that so-called "frontier models" — the most powerful, general-purpose AI systems — pose risks ranging from harms already visible today, such as disinformation or discriminatory bias, to more speculative scenarios of catastrophic or even existential risk. The text calls on countries to cooperate on assessing these risks and to share research on AI safety.

What matters here isn't so much the content, which is deliberately cautious, as the list of signatories. The United States and China have signed the same document — a rare occurrence at a moment of rising geopolitical tension between the two powers. Alongside them are the rest of the G7 countries, the European Union, and emerging AI economies such as India, Brazil, Saudi Arabia, the United Arab Emirates, Nigeria and Kenya. Bringing together such a diverse group behind a single text on AI, barely a year after the launch of ChatGPT, is in itself the summit's most tangible result.

Safety Institutes on Both Sides of the Atlantic

The summit also served as a showcase for concrete institutional announcements. The UK confirmed the creation of its AI Safety Institute, which will grow out of the Frontier AI Taskforce that Sunak set up months earlier. Its mission will be to evaluate the safety of the most advanced models before they're deployed to the public.

At the same time, US Vice President Kamala Harris travelled to London today to announce the creation of an equivalent institute in the United States, the US AI Safety Institute, housed within the National Institute of Standards and Technology (NIST). In her speech, Harris stressed that AI risks aren't limited to hypothetical long-term scenarios but include harms that already exist today — from election disinformation to job losses to misuse by malicious actors.

That emphasis reflects a tension that has run through the summit's preparations ever since the British government announced it this summer: while some participants — including several of the major AI companies — have focused on existential risks tied to future, far more capable systems, other governments and civil society organizations have insisted that priority should go to the harms AI is already causing today.

Voluntary Commitments from Big Tech

The summit also brought together executives from OpenAI, Google DeepMind, Anthropic, Microsoft and Meta, among other companies, alongside government officials and academics such as Yoshua Bengio, one of the most-cited scientists in deep learning. Bengio has been appointed to lead the drafting of an international scientific report on the state of AI safety — a kind of equivalent to the Intergovernmental Panel on Climate Change, but for artificial intelligence.

As part of the summit, several of these companies committed to giving governments early access to their most powerful models so they can be safety-tested before commercial release. These are, once again, voluntary commitments: there's no law requiring them, and no penalty if they aren't honoured.

What Remains Unresolved

The Bletchley Declaration doesn't create any regulatory body, nor does it specify which models should be subject to oversight or under what criteria. Nor does it resolve the underlying disagreement over whether the priority risk is the one that already exists — algorithmic discrimination, disinformation, mass surveillance — or the one that could arrive with systems far more capable than today's.

What this first day does leave behind is a precedent: governments with widely differing views on technology, geopolitics and digital rights have agreed to sit down and sign a common text on AI. Whether that gesture translates into concrete rules, institutions with real oversight power, or simply another round of well-meaning declarations will depend on what happens at the summits already announced for South Korea and France.