Bletchley testing agreement builds evidence, not a global AI licence

Governments and leading AI developers have agreed to collaborate on testing the next generation of advanced models before and after deployment, as the Bletchley Park AI Safety Summit concluded on November 2. The UK has also launched an AI Safety Institute to build the public-sector expertise needed for that work.

File photo: Union Jack flag in London, by 117PXL (2022). CC BY-SA 4.0, via Wikimedia Commons. Cropped and resized; this adaptation and its responsive variants carry the same licence. Not a Bletchley Park or summit photograph.

The useful distinction is between three outcomes: a broad declaration of shared priorities, a more focused testing partnership, and a British research institution. They support one another, but they don’t amount to a single international licensing system for AI.

The declaration and testing partnership have different participants

The November 1 Bletchley Declaration announcement identifies 28 countries and the European Union, including China and the United States. Its agenda joins scientific work on frontier-AI risks with national risk-based policies, recognising that domestic approaches can differ.

The November 2 testing statement names a smaller group: Australia, Canada, the EU, France, Germany, Italy, Japan, South Korea, Singapore, the US and the UK. Its company participants include OpenAI, Anthropic, Google and Google DeepMind, Microsoft, Meta, Amazon Web Services, Inflection AI, Mistral AI and xAI.

That distinction matters when describing the agreement. China’s participation in the declaration doesn’t establish its participation in this separate testing statement, whose list doesn’t include it. Nor does the declaration promise that every signatory will use an identical evaluation process.

Testing supplies evidence, not a universal release certificate

The chair’s summary says participating companies will support appropriate independent testing of their next models. Governments intend to build evaluation capacity, develop shared methods and eventually common standards. Results would be shared with other countries where it can be done safely, securely and appropriately, not automatically published in full.

These are cooperative commitments, not a newly enacted global approval law. The summary separately records suggestions for putting voluntary commitments on a legal footing and creating pre-deployment safety gates as matters for further discussion. Those proposals shouldn’t be presented as rules already adopted at the summit.

The UK’s November institute paper makes the operational boundary unusually clear: the institute isn’t a regulator, won’t designate a system as safe and won’t be responsible for release decisions. Its evaluations are intended to improve understanding of capabilities, safeguards and behaviour. They aren’t a pass-or-fail deployment authorisation.

Why testing continues after a model launches

The testing statement keeps responsibilities with developers, deployers and users throughout the AI lifecycle. The institute paper explains why one pre-release examination can’t do every job: societal-impact research may draw on usage data and incident reports after deployment, while other evaluations examine capabilities or safeguards directly.

For an organisation adopting AI, the practical implication is to distinguish evidence about a model from evidence about its particular application. An external evaluation can inform a deployment decision; it doesn’t, by itself, settle how that system should operate in a specific workplace or service.

The institute will evolve from the Frontier AI Taskforce, with Ian Hogarth continuing as chair. Its first testing milestone is to prepare an evaluation process for the next generation of models, including releases expected in 2024. Separately, Yoshua Bengio will lead a State of the Science report synthesising existing research. The institute paper says that report won’t itself make policy recommendations. Bletchley’s immediate contribution is therefore a shared evidence-building programme, alongside work towards future standards.