What happens when access to the best AI models is at the discretion of one country’s leadership? Over the weekend, we got a taste. Anthropic, the company behind the Claude family of AI tools, said on Friday night that it had received a directive from the US government to prevent foreign nationals – including its own employees – from using its latest models Fable 5 and Mythos 5.
The implications for the UK are severe. Should blocks like this become a frequent occurrence, Britain’s tech companies that build on top of US AI models will be operating a step behind their American rivals. Meanwhile, American companies may find it less worthwhile to hire in London, triggering further brain drain towards the US.
The ramifications go well beyond the tech sector. If even US allies like Britain can be cut off with no notice (Anthropic reportedly had just 90 minutes to respond to the directive), that poses a threat to security, business operations and public services that have integrated the technology into their systems. Recall the chaos when a cloud service provider like Amazon Web Services goes down, disrupting banks and government services. Then imagine it not as a technical blip but as an ongoing policy.
This time, thankfully, Fable had only been available for a few days, so disruption was minimal. Mythos, which Anthropic says is its most powerful model, was already only being shared selectively with certain companies and institutions. And although the ban looked targeted at non-Americans, complying with that rule would be so complex that the company opted to disable those new releases for all of its customers, including in the US.
The company insists this is all a misunderstanding on the administration’s part and hopes to rectify it. But even if the decision were now reversed, the nature of the directive has demonstrated the US government’s willingness to control AI. The incident raises the prospect that public access to the most advanced technologies will now be subject to state approval. And that leaves everyone else beholden to the country where the producers of the best models are located: the United States.
It is a form of dependence we have become accustomed to in other areas. For its Trident nuclear programme, Britain is already deeply reliant on the US for both technical support and political backing. But the blocking of Fable could herald a new kind of interdependence, where the UK and other middle powers are not just locked into contracts with US firms but locked out of certain technologies entirely.
This latest chapter in the story of the AI race is dramatic but not unforeseen. Researchers have warned for years of the ways in which a lack of competitiveness could leave countries other than the US and China without benefits from the gains of AI.
Sceptics might look at the gimmicks and froth of today’s AI market and wonder if that really matters. But access to technology has serious consequences; just look at how vaccine hoarding unfolded during the pandemic.
Historic parallels and the pace of the AI race so far suggest we shouldn’t be hugely surprised by this newest development. The question is: should we panic? Inside the British government, there is no way last week’s events had not been considered as a possibility. In fact, eventualities like these explain why “sovereignty” is the buzzword du jour of Whitehall.
The Prime Minister’s own AI adviser, Jade Leung, foresaw the dynamics that would play out in her 2019 PhD thesis. While the view at the time was that tech firms had become too big to regulate, she predicted that “these firms will increasingly face legislative restrictions imposed on them by the state, particularly in the form of tools such as export controls, and particularly motivated by the state’s prioritisation of national security concerns”.
This was based on precedent: the US had used export controls as a means to exercise its power over satellites, cryptography, and other strategically important technologies. More recently, we have also seen rules about the export of semiconductors and chip making equipment. It is a well-used tool for a security state looking to protect its interests.
The timing, too, has been expected. In the influential 2024 essay series Situational Awareness, former OpenAI researcher Leopold Aschenbrenner predicted 2026 and 2027 as the years Washington would begin to wrest control of AI from Silicon Valley, as officials realised the national security implications.
Just this week, a group of European researchers and investors published the report Europe 2031. In it they argued that, under its current strategy, Europe will be left behind in the AI race without some form of leverage – owning an indispensable part of the AI supply chain, for instance – to ensure it can still access top models. The idea of being cut off from AI seems inconceivable now. But the report depicts this as an inevitability when the world’s demand for AI outstrips the amount of built computing power to fuel it.
“We can’t conceive of a world in which we’ll be able to solve disease, cure cancer, and it’s just we won’t be able to access it, we won’t be able to afford it,” Judith Dada, an investor and one of the report’s authors, told me on a call last week (before the Anthropic announcement). “It can absolutely be us.”
A day after publication, Europe 2031 already looked prescient for having explored the potential of even Western nations being cut off from the frontier. A ban on Fable for foreign nationals gave a taste of that reality.
Britain occupies a better position than the EU when it comes to the implications of the US clamping down on foreign users of AI models. The authors of Europe 2031 foresee a version of events in which countries get tiered access to AI according to American alliance. Britain gets into the top cohort by virtue of the Five Eyes intelligence group, while most European nations are in tier 2.
The UK also has other strengths, such as the existence of its AI Security Institute (AISI), which has helped set the agenda on AI safety testing globally. Set up under Rishi Sunak, it was the first state-backed organisation dedicated to conducting independent evaluations of AI models and researching AI applications for the public interest. Anthropic even mentioned in its statement on Friday night that it had been working with the institution as part of its testing for the Fable model.
And yes, there’s Brexit. Dada says that for the first time she’s hearing even staunch remainers say, behind closed doors, that they are happy not to be in the EU because it allows the UK to broker direct deals with Washington on compute, model access and other issues.
Nevertheless, the nature of the export control imposed on Mythos and Fable underscores that there is no exception for best friends. No shiny new Anthropic office in King’s Cross counters an Americans-only order. Whatever happened to the US-UK Tech Prosperity Deal that was meant to enhance cooperation between the two countries on science and technology? It’s still on ice.
And if Britain wants to build its own alternatives, it faces similar challenges to its European neighbours – sometimes worse, in the case of high energy costs that make AI infrastructure expensive. The authors of Europe 2031 suggest that a coalition of several middle powers, including the UK, will be the only plausible route to maintaining enough leverage to avoid being shut out of AI developments altogether.
It can be difficult, for those outside of the tech bubble, to get animated about moments like these. So many of them have come and gone without an immediate, obvious consequence for everyday life. Grand predictions about how the world is about to change are made several times a day by tech company CEOs, politicians and the occasional religious leader.
Those who work in AI are sounding the alarm this time, but there’s sometimes a faint ring of consumerism to their concern. For people who spend all day burning through credits, taking on more and more complex tasks with AI, a new model is like a thousand new iPhone releases. Having access withdrawn is a new and unwelcome experience for AI engineers and researchers.
They may have to get used to it. Since the launch of ChatGPT, we’ve been in an era of frequent new releases from AI companies to as broad an audience as possible. That might now be over. The sheer cost of running these models, which users often get limited access to for free or for a flat monthly subscription, perhaps made that inevitable anyway. But many anticipated getting priced out, rather than locked out by a government.
As Transformer editor Shakeel Hashim argues, “the administration has effectively established a licensing regime” via the backdoor of an export control that officials know in practice will take the models off the market for everyone, including Americans. This is what safety advocates have been hoping for. But, he says, an “arbitrary, post-hoc system” is not the way to do it.
The narrative that this is about safety is further complicated by the tetchy relationship between Anthropic and the government. Commentators are already seeing this as a continuation of the dust-up between the company and the Department of War that flared up in January. Pete Hegseth gloating on X gives credence to that idea.
It is tempting then to brush off concerns about sovereignty and see this as something unrelated to Britain: a domestic US fracas between the state and one of its home-grown companies.
But we should remain mindful of the various motivations at play: from the companies hyping up the capability of their products, to the investors pushing to break ground on new data centres, to governments justifying clampdowns with safety concerns. Panic is not helpful, because panic allows you to be swept up in one actor’s vision of the future. Yet the fact that this particular fable was so easily foretold should give us reason to think seriously about what is coming next.
[Further reading: Mythos and the insecurity avalanche]
Content from our partners
Related
Join the debate
Subscribe here to comment