Britain won't leave AI firms to police themselves — but hasn't signed the call
On Monday, Britain's AI minister Kanishka Narayan gives his first speech since artificial intelligence became a portfolio with a seat at the cabinet table. According to The Guardian, which previewed the speech, his central message is that the companies building the most capable models must do more on safety — and that Britain cannot rely on the technology companies to police their own models.
He will also argue that Britain needs to build more of its own AI infrastructure, and he is expected to announce a strengthening of the country's cyber defences. What that strengthening consists of was not disclosed ahead of the speech.
Narayan has been Minister of State (Minister for Artificial Intelligence) since 20 July 2026, with a brief spanning both the Cabinet Office and the business and science department, BIST.
The argument already exists in writing — without a British signature
On 21 September, on the sidelines of the UN General Assembly, Finland's President Alexander Stubb and Norway's Prime Minister Jonas Gahr Støre launched a declaration titled "A Call for Control of Frontier AI Models". It asks for broadly what Narayan is now saying: mandatory pre-deployment testing of models, independent evaluation in which qualified evaluators are given sufficient access, coordinated common standards across countries, and shared reporting of serious safety incidents. On top of that, UN member states are urged to explore an international institution that could set standards and enable verification when models cross certain capability thresholds.
31 leaders from 30 countries have signed, plus the President of the European Commission. The signatories include Mette Frederiksen, Emmanuel Macron, Friedrich Merz, the Netherlands' Rob Jetten, Ireland's Micheál Martin and Austria's Christian Stocker. The United Kingdom is not on the list. Neither is the United States.
That is the contradiction in the speech: London is making the declaration's case without having put its name to it.
What the risk is about
The Guardian quotes Geoffrey Irving, a former chief scientist at the UK's AI Security Institute, as saying that "the AI companies do not have security that would prevent an adversarial country from stealing the models". His concern is the theft of models before they are released. OpenAI told the paper it has multiple layers of protection against precisely that, including encryption and continuous monitoring.
Having protections is not the same as being able to withstand a state intelligence service. And that is the core of Narayan's point: what is in question is not the companies' good intentions, but who checks them.
Why this matters for Europe
British AI policy is no longer only British. In September the EU announced cooperation with the United Kingdom and Canada on model evaluation and verification and on early warning — see The EU wants to slow down the AI that Europe isn't building. Cooperation of that kind presupposes that the parties agree on what is to be measured and who is allowed to measure it. That is exactly the question the September declaration tries to answer — and the one Britain has so far stayed outside of.
Narayan's other message, that Britain should build more of its own AI infrastructure, is word for word the sovereignty argument Brussels has been making for a year and a half. On that point Europe and Britain agree without having to sign anything.
Sources
- A Call for Control of Frontier AI Models (Government of Norway) — the full text of the declaration, its demands and the complete list of signatories
- The Guardian (Robert Booth and Dan Milmo, via AOL) — the preview of the speech, the Geoffrey Irving quote and OpenAI's response
- GOV.UK — Kanishka Narayan's title, departments and date of appointment