UK Minister Says Copyright Already Bars Frontier AI at Home. Who Decides What Britain Gets to Test?

On 29 September, in Liverpool, Kanishka Narayan answered the people who want a law against AI systems that a bill by MP Sobel defines by what they could do: neutralise or circumvent human authorities. His answer, as reported by Politico, was that copyright already stands in the way of building such a system in the UK. A reply like that deserves two questions. What does he command, and is the claim true?

Start with the post. The UK has had a junior AI minister since 2023. Narayan, who had been the under-secretary for AI and online safety since September 2025, became Minister for AI on 20 July. He attends Cabinet, though no source I found says whether as a full member. Nothing public specifies a budget or a legal power to compel a lab to hand over a model. And the binding rules for frontier AI that Labour promised in its manifesto still don't exist.

On the claim, lawyers are not convinced the law works this way. In the Getty v Stability case, the main UK ruling on AI and copyright so far, Getty dropped its training claim, and its secondary claim was rejected. Permission to appeal has been granted, and I found no hearing date. The next day, Politico had copyright as an obstacle, not as a ban. Those are different claims, and only one of them is a minister's reading of settled law.

Sobel's bill is at an earlier stage than its supporters sometimes suggest. It is a Ten Minute Rule bill, a procedure that lets an MP present a proposal in a short speech, and most such bills never become law. It has the backing of the campaign group ControlAI and more than 70 signatories, and it would require monitoring for "precursors" abroad. The Ada Lovelace Institute, in its four regulatory scenarios, calls the status quo "largely symbolic". The minister and his critics share that premise and draw opposite conclusions from it.

Now the part the minister's reply doesn't cover. Anthropic withheld Mythos 5.1 from the AI Security Institute (AISI), the UK body that has tested frontier models before release since 2023, on a voluntary basis. On 24 September the White House asked OpenAI and Anthropic to route frontier models to US testers first. On 2 October, Google skipped AISI for the first release of Argon. A law is written in London; an inspection is granted elsewhere.

Narayan's answer, in his words, is that the UK is not "first dibs" for anyone and that its role isn't diminished. He has a point. GPT-6 Astra was tested, Google says it will share Argon later, and The Next Web asks whether this is one decision or a policy. That is a fair question, and the same mechanism question applies to his own claim about copyright. A law is written in London; an inspection is granted elsewhere.

One hypothesis, and only that: releases now go out in concentric rings (a small group first, then hundreds of members, then everyone), which makes "pre-deployment" elastic. If it is so, a test before release could mean a test before the last ring, and Britain might be in a ring it can't see. The next model to reach AISI before any closed ring would count against this.

The questions that stay open. If the Getty appeal goes against the UK's reading, does the copyright argument turn from a shield into a problem for what Britain itself is allowed to use? And is the gap between what the UK can ban and what it can see a failure to fix, or the price of a relationship it chose? A law is written in London; an inspection is granted elsewhere.