Britain’s AI sovereignty is an illusion

Without more domestic capacity we risk being locked out of the future

(PhotoGranary02/Shutterstock)

Last weekend, following intense discussion and negotiation, the US government forced Anthropic, the world’s most valuable AI firm, to close down two of its most capable models for foreign nationals. Fable 5 and Mythos 5 operate at the frontier of machine cyber-capability and were switched off because Washington decided that letting a foreigner use them amounted to exporting a weapon.

Among those reportedly locked out was Rishi Sunak, who while serving as an adviser to Anthropic, is still a foreign national, and it is likely that the UK’s AI Security Institute was also blocked. This, of course, being the same internationally-renowned institute that “red teams” these models by simulating an adversary. Now we have seen how fragile Britain’s position is. Models can be, and were, rescinded overnight. Frontier capability was never ours; we were mere rentiers.

This is not really a story about Anthropic, though a company that spent two years warning that its models were too dangerous to release should not be astonished when the state finally takes it at its word. Nor is this really a story about the Trump administration; I suspect any powerful US government would have made the same decision. The story is that we have neither a frontier model of our own, nor a government with the heft – or the geopolitics in its favour – to keep our access to someone else’s open. We have enjoyed these frontier capabilities so far thanks to what remains of the transatlantic alliance but these fragile permissions are unlikely to last. That leaves us with shaky foundations in an AI-dominated future.

Britain’s strategy so far has depended on it exercising influence through the deployment and adoption of AI technologies alongside testing at the AI Security Institute (which, chaired by Ian Hogarth, has been a remarkable success story). And while this is true, and scores of AI companies are coming to call London home, we lack meaningful geopolitical influence. DeepMind may be headquartered here but, for all its immense accumulated talent and knowledge, it is not ours (having been acquired by Google in 2014).

We lack the basic infrastructure to be competitive in the global race which the proliferation of AI capabilities forces us to be. We are stuck between the US and China, and Europe does not necessarily answer our woes either – its regulatory regime has squandered much of the innovation necessary to nurture its own challengers to US titans.

We have had good policy intentions in the UK, most notably entrepreneur Matt Clifford’s AI Opportunities Action Plan, which successive governments would have been lost without. AI Growth Zones were launched to boost domestic compute capacity but have been let down by our energy prices, which are twice the average in Europe and four times higher than in the US – far too high to make data centres attractive. As long as our compute rollout is limited, we will lack leverage and a potential chokepoint.

In her Mais Lecture earlier this year, Rachel Reeves championed an “active and strategic state [that] protects British interests through AI sovereignty”, but actions speak louder than words and serious reform of the conditions that limit AI scale-ups is more critical.

Our political class has failed to internalise the real dangers of having no access to frontier capabilities. To their minds, we can have our cake and eat it by letting others burn capital on AI models, while capturing the gains from using them well. It is a tempting story for Westminster because it asks so little of our politicians. It suggests that Britain doesn’t need a new focus on cheap electricity and a far more radical approach to planning and infrastructure. As Wes Streeting noted in his speech this week on “progressive capitalism”, “[energy] grid connections can be delayed into the 2030s, and major upgrades can take more than a decade”.

There also appears to be little political will to reform the licensing and copyright regime, which inhibits commercial text and data mining. A prerequisite for any serious AI strategy will be changing this; under current laws, no general-purpose LLM can be trained in Britain. These will be tough trade-offs, but a government that takes seriously the precarious world in which we live will need to make a political argument for acceleration.

AI adoption is critical, but it cannot be our only approach. To avoid being locked out of the future, Britain must prove that internationally competitive firms can thrive here – and fast.