Trump thinks AI can police itself. So what can Britain do? - Ian Sollom MP Cambridge Independent October 2026

Originally written for the Cambridge Independent Column – October 2026

In the rush of party conference season over the last month, the day-to-day business of politics has quickly moved on from what was, at the beginning of September, a new and alarming threat to the future of humanity. And yet the resignation of a researcher from one of the world's leading AI companies, and the accompanying warning that the industry was racing towards systems that could, in his words, kill us all by the end of the decade, triggered wider warnings that should not be so quickly forgotten by those of us involved in politics.

That many of those frontier companies are based in the US should focus minds more with Donald Trump in the White House. His answer, so far, has been a friendly lunch. On 29th September six executives - from Google, Anthropic, Meta, OpenAI, xAI and Nvidia - joined him to sign a one-page document promising to monitor their own systems for dangerous capabilities, to set up internal teams to check those controls, to appoint outside auditors, and to create board committees to oversee the lot.

President Trump called it "almost like a constitution" and "morally binding” and said there would be a tremendous self-policing aspect. Hardly reassuring.

Ed Davey used his conference speech last month to argue for something considerably more serious: a global treaty on the development of super-intelligent AI, and a new international AI safety agency modelled on the International Atomic Energy Agency. The nuclear comparison is appropriate - some risks are too large to be left only to the people creating them.

However, we should not be blind to the challenges of getting there. International treaties are hard to agree even when there is alignment, never mind ratified and implemented. And the current US administration clearly is not aligned.

So, what else can we do? The answer I think is to do the work that would make a treaty worth signing in the first place.

It may sound obvious, but it isn’t always the case. The Biological Weapons Convention (BWC) entered into force in 1975 - though ‘force’ is a moot point as it had no agreed verification protocol at the time. Talks to add an inspection regime collapsed in 2001 and the BWC continues to be among the weakest agreements of its kind. An agreement nobody can check is a ceremony.

Or a lunch. The White House document is even thinner than it looks in this regard: an audit is only worth as much as the standard it audits against, and at the moment there are no agreed standards for frontier AI.

This is the part of the problem that tends to get less attention and is also the hardest. While there are parallels, safety regimes developed for other purposes don’t easily transfer. The general capabilities of frontier AI systems make it challenging to regulate as we might a new medicine or a new airframe for example. The same AI system that helps design a new antibiotic is the system that helps design a new pathogen - not two technologies, but one asked different questions. And what these systems can do is not reliably known before they are built.

That is also why simple prohibition is harder than it sounds. Ban the development of super-intelligent AI and to enforce it somebody would have to determine, before a system exists, that it could do something it has not done.

And even if it could be defined, a ban here in the UK would reach very little. Google DeepMind's headquarters is in London, but the data centres where the systems are trained are in the US, along with the other leading companies. It is unlikely our own caution would have much impact on the US.

Where we have more clout is in capability. The AI Security Institute (AISI), set up at the international Bletchley Park summit on AI convened by Rishi Sunak in 2023, was the first body of its kind anywhere in the world and has now evaluated more than thirty of the most capable systems. Its testing platform was published openly for anyone to use, and it is one of the tools other national institutes now work with. The Institute's biosecurity testing has found dozens of vulnerabilities in the models it has examined, including new ways of getting around their safeguards. Where it finds them, it can work with willing developers to get them fixed.

It is slow, unglamorous scientific work, but it is the foundation for everything else. An international agency of the kind Ed Davey proposes would need exactly this capability and AISI continues to lead the way with that international network.

And that network reaches further than people assume. Even China has established its own counterpart to AISI and has proposed an international governance body - which suggests a country that would rather be inside a process than outside one.

There is an obvious gap of course. Our Institute tests systems by invitation. It has no legal right of access, no published threshold a system would have to fail, and no power to delay a release whatever it finds. A Frontier AI Bill has been promised and not introduced. When it comes, it should close that gap. And by continuing to develop such standards we can encourage adoption globally.

The American administration will not help with any of this; Trump has called concern about AI risk a conspiracy and attacked the companies for asking to be regulated. But measuring, testing and standard-setting do not need Washington's permission, and they are what any future international agreement will have to be built on.