The off switch abroad

In the quiet corridors of British research institutions, a new kind of anxiety has taken hold. It is not about whether artificial intelligence will become too powerful, but about who holds the key to the power it already has. George Williamson, the new head of the Alan Turing Institute, recently gave voice to a concern that has been simmering in policy circles for months: the United Kingdom is sleepwalking into a dependency on foreign AI systems that could, under the right political or commercial pressures, be restricted, withheld or simply turned off.

The immediate trigger for this warning was not hypothetical. In June, Washington temporarily blocked the release outside the United States of the latest models from the AI firm Anthropic. Around the same time, American authorities reportedly asked leading AI labs not to submit their systems for testing by the UK's own AI Security Institute until they had first been examined by its American equivalent. For a country that has prided itself on a special relationship and a seat at the top table of technological governance, the message was sobering: access is a privilege, not a right, and privileges can be revoked.

Williamson's argument rests on a concept that has been gaining traction far beyond Westminster. The term 'digital sovereignty' once belonged to the lexicon of Brussels bureaucrats and privacy activists. Today it is being invoked by central bank governors and presidents of the European Commission, who have begun to speak openly about the danger of being cut off from US and Chinese technology. The fear is not merely one of commercial disadvantage. It is the realisation that AI is becoming embedded in the sinews of the state: defence logistics, national security analysis, critical infrastructure management and healthcare diagnostics. To run these on systems that cannot be fully inspected, whose update schedules are decided elsewhere and whose very availability depends on the goodwill of a foreign power or a corporate board is, in Williamson's words, an uncomfortable place for a nation to find itself.

There is a second dimension to this unease, one that touches on values rather than valves and switches. Williamson pointed to the experience of asking certain Chinese-made models about events in China's recent past and receiving answers that would be of little use to a student preparing for a history examination. The issue here is not just censorship but the subtle shaping of what is knowable. When a society relies on tools that have been trained on curated data and fine-tuned to align with another state's official narratives, it imports not just software but a frame of reference. Over time, the boundaries of acceptable inquiry can shift without anyone having voted on the matter.

The resilience argument

The counterpoint to these warnings is rarely stated as outright opposition. Few would argue that total dependence on foreign-controlled infrastructure is wise. Instead, the pushback tends to be pragmatic. Building sovereign AI capability is expensive, and the financial firepower of the major American and Chinese labs dwarfs anything the UK can realistically muster. Williamson himself conceded that matching the resources of frontier labs was not the goal. The alternative he sketched was more modest: nationally owned intellectual property, smaller models that could be made available to domestic startups, and a focus on public-good applications such as weather forecasting, agriculture and disinformation monitoring.

This is where the resilience argument gains its force. Resilience is not the same as self-sufficiency. It does not require a country to produce everything it needs; it requires that no single point of failure can bring the whole edifice down. A nation that maintains its own capacity to train, inspect and deploy AI models, even if those models are not the most advanced on the planet, retains a degree of agency. It can keep critical services running when diplomatic tensions flare. It can audit the systems it uses for bias and security flaws. It can shape the technology to accord with its own legal traditions and democratic norms, rather than accepting whatever arrives in a software update from a data centre in a different jurisdiction.

The geopolitical backdrop makes this more than an academic debate. The prime minister, Andy Burnham, announced a new national centre to tackle state-backed disinformation only last month, while the AI company Anthropic has publicly estimated a greater than ten percent chance that the technology could wipe out humanity by 2030. Whether one finds that figure alarmist or sobering, it underscores a broader point: the stakes around AI governance are rising, and the margin for strategic error is narrowing.

Williamson's own career trajectory lends weight to his intervention. Before leading the Turing Institute, he ran His Majesty's Government Communications Centre, the secretive body that builds surveillance and security equipment for British intelligence agencies. His predecessor in that organisation's lineage was Alan Turing himself. When someone with that background speaks about national resilience and the need for agency over technology, the words carry a particular resonance. They suggest that the concern is not theoretical but operational, rooted in a professional lifetime of thinking about what happens when critical systems fail or are compromised.

Yet the argument for digital sovereignty also contains an unresolved tension. The very openness that has allowed British universities and startups to thrive in AI research depends on the free flow of ideas, talent and code across borders. A push for national capability, if pursued too zealously, could slide into a techno-nationalism that isolates the UK from the global scientific community. The challenge, which Williamson's remarks acknowledge but do not fully resolve, is to build resilience without building walls. It is a balancing act that will define the next decade of technology policy, and one for which there is no off-the-shelf solution.