What happens if America shuts off your AI?

Yet AI and space exploration have something in common. The first country to “get there” – wherever “there” turns out to be – will enjoy clear military and possibly economic supremacy over the rest. State-backed companies that deploy it will outcompete their rivals in countries that do not have access to cutting edge AI. “Frontier AI” nations will have unprecedented powers of surveillance over their people, and the ability to build and deploy weapons in new and potentially terrifying ways.

These weapons might be lethal in their own right (for example, autonomous drones) or they might enable states to carry out cyberattacks on a hitherto unprecedented scale, or deploy disinformation to destabilise rivals. Russia is, of course, already trying to do this with the tools it has.

As the American AI company Anthropic put it in May: “Historically, the reach of authoritarian rule has been limited by its dependence on human enforcers to carry out surveillance and repression. Powerful AI systems may remove that dependency, enabling automated repression on a far greater scale.” 

To put it slightly differently, humans can refuse to obey orders. AI, if it is programmed to do so, could kill without compunction. It cannot suffer. And it can be destroyed without regret or moral scruple. This will transform policing and warfare in ways that science fiction has grappled with, but the international community has not.

Right now, America is out in front in the race for frontier AI, but China is doing all it can to catch up. That is why the enormous sums being invested in artificial intelligence have been compared to the nuclear arms race. And it is why Donald Trump’s decision to ban non-Americans from using Anthropic’s latest release, Mythos, caused such a stir. 

The decision has since been reversed, but it set a precedent. If a country hoards its cutting-edge AI, preventing others from using it, what should “middle powers” like European countries, India and Australia do? Until recently the US has been a trusted partner. But now it can no longer be relied on. 

Suggested Reading

AI job interviews are good for only one thing

Caroline Marie

The leading AI companies in the US and Europe have generally argued that their work is for the good of humankind – that claim was part of Elon Musk’s recent court case against OpenAI, which he co-founded. OpenAI had ceased to be a non-profit company, something Musk said he never envisaged or endorsed. These altruistic values have become increasingly hard to defend as AI becomes embedded in military technology and causes ever more job losses.

But the principle is a useful one for companies to endorse. Some, like Anthropic, actively seek government oversight of what they are doing, so as to create a level playing field, avoid responsibility for mass layoffs and establish international agreements and export controls. 

When a technology is this disruptive, it makes sense for the company that controls it to cooperate with authorities. In turn, AI firms know that governments will buy their models in order to stay ahead. This delicate balancing act – submit to regulation and make themselves invaluable to government – is now breaking down.

Anthropic had been working closely with the US defence department, but recently made it clear that it would not let its AI be used for “domestic surveillance and autonomous offensive weapons”. This went down badly with the Trump administration. In February Trump ordered federal agencies to stop using Anthropic: “We don’t need it, we don’t want it, and will not do business with them again!”

Anthropic pushed back, saying it was determined not to let China get hold of more of its AI expertise, but that it needed the government’s help to prevent that from happening. China has been carrying out “distillation” attacks in which it runs millions of requests on Anthropic models to harvest their methods and learn more about how they work. In May the company urged the US to tighten export controls on “compute” (the semiconductors on which models are trained).

It pointed out that China already uses AI to repress its people, and warned: “Frontier AI systems will make those capabilities cheaper to maintain, far more pervasive, and more sophisticated… A [Chinese Communist Party]-led AI frontier could dramatically strengthen repression around the world.” Anthropic also patriotically made the case for America to export home-grown AI to other countries before China has the chance to do so. 

But Trump’s move – to deny any other country access to Anthropic’s most powerful models – was an explicit attempt to hobble the company, forcing it to submit to the will of the US government. It was also an attack on European countries that are using Anthropic themselves. 

The UK had signed a memorandum of understanding with the company in February, but when Anthropic shut off access, some called the decision a “kill switch”. Al Carns, the ambitious Labour MP, posted: “This week the most advanced AI model on the planet got switched off by a foreign government. British researchers were studying it. British companies were testing it. British hospitals were piloting it. Not any more.” 

France is equally worried. “Anthropic is their Strait of Hormuz,” said French politician Gabriel Attal, referring to Iran’s ability to shut down key international shipping. Like Trump’s regular threats to quit Nato, the shutdown is a wake-up call for European democracies. America is no longer on our side – not militarily, and not technologically either.

The Trump administration claims it had become aware of a way of “jailbreaking” Anthropic or misusing it. Anthropic says other models like ChatGPT share the same vulnerability, and it is being unfairly targeted. Yet it has boasted about Mythos’s abilities, saying it had found “thousands of high-security vulnerabilities” in major operating systems. Who is right? Very few of us are in any position to know. 

Suggested Reading

Trump ‘humiliates’ Meloni – and her popularity soars

Silvia Marchetti

But as scholars like David Gantt of Harvard point out, if we distrust the Trump administration to make the right call about sovereign AI, we should be equally wary of deferring its governance to a company with no accountability except to its own board. “The more consequential a question, the more deserving it is of a democratic answer. Correct though they are, these red lines are not Anthropic’s to draw.”

Where does this leave Europe? Does it have the ability to build AI that can compete with America and Russia? No. Kanishka Narayan, the UK’s AI minister, told an FT podcast that although he saw little prospect of being able to compete with the large language models being built in America, Europe had expertise in chip manufacturing that he hoped would give the continent more leverage in gaining access to them. He also believes the British AI Security Institute is an asset because it has early access to testing – although it, too, has been denied access to Mythos.  

Chatham House sees four ways in which middle powers like Britain can try to give themselves leverage in a world where the US and China inevitably dominate. One is to “carve out limited niches in the global AI supply chain” to try to make themselves indispensable and receive favourable treatment. 

The second is to make themselves dependent on either the US or China “in exchange for access and protection”. Until now, Britain has found itself in this position, but it would prefer to avoid it. A third is to ally with other middle powers to try to build collective influence. The last is to do a little of all three in an effort to hedge their bets.

“AI sovereignty will inevitably remain partial for these countries, but efforts in this direction will not be in vain,” say Francisco Javier Varela Sandoval and Isabella Wilkinson.

Andy Burnham talked about technology in his speech about the economy, but not AI. We can assume he wants to try to build UK sovereignty in AI, but which countries does he want to ally with? Even if the US slide into authoritarianism is reversed when Trump leaves office, the country is still likely to be wary about sharing its AI expertise. 

In any case, we have already begun to see the risks of technological dependency on the US. It should be clear that allowing the systems we depend on to become embedded in American technology is a risky strategy. And AI is not the only technology to consider: quantum computing, which would fundamentally change the nature of computer processing, may be just around the corner. 

The Oxford Internet Institute (OII) compares AI companies to the myth of Perseus. They claim to be able to act alone. In fact, Perseus relied on help from four other Greek gods to slay Medusa. 

Put another way, “the public pays for AI through substantial (in)direct subsidies including tax breaks, direct public funding, privileged access to public lands, water, electricity, data, and infrastructure, deregulation of data protection, labour, environmental and intellectual property laws”. The OII argues that we deserve “socially beneficial tech” in return – not just the threat of disruption and warnings about the costs of falling behind. 

During the Brexit years, we heard a lot about the importance of sovereignty. Sceptics warned that in the modern world, middle powers like Britain can do relatively little on their own. It was true then, and the advent of frontier AI makes it even more urgent. Only by working with allies do we have much hope of resisting US-Chinese hegemony – and, perhaps, building an AI that might work for the good of society.