The AI race America and China cannot afford to stop

US-CHINA

The US still leads the global AI race, but China is closing fast. A new Sibylline report warns that fear of falling behind could drive both sides to take dangerous risks – creating an opportunity for Britain, the world’s No. 3 AI player, to help broker safeguards.

THE UNITED States remains ahead in the race for artificial intelligence. But the gap with China is narrowing – and the battle to stay ahead may itself be becoming one of the greatest dangers posed by the technology.

A new Sibylline report, Who is more likely to lose control of AI? The US, China and the Risk of an AI Catastrophe, examines which superpower would be more likely to develop a system that escaped meaningful human control.

Its conclusion is initially surprising.

For now, the US is probably the more likely source of such an event because it has more frontier AI companies, computing power and investment.

But the report’s deeper warning is that the greatest danger may ultimately come not from either country in isolation, but from the increasingly intense competition between them.

The US produced 59 notable AI models in 2025 against China’s 35 and retains the advantage at the technological frontier.

Yet that lead is shrinking rapidly. By March, the performance gap between the leading US and Chinese models had fallen to just 2.7 percentage points on Stanford University’s composite benchmark.

China already leads in AI publication volume, citations and patent grants, while more than 6,600 AI companies were operating there by June.

That growing strength creates a paradox.

Beijing possesses powers over Chinese technology companies that Washington could not easily replicate.

It can restrict access to computing power, halt deployment and compel companies to co-operate if it concludes that a model poses an unacceptable threat.

But having the power to apply the brake does not mean China would choose to use it.

Sam Olsen, Sibylline’s Chief Analyst and a co-author of the report, says technological development has become central to Beijing’s economic and strategic ambitions as China confronts slowing growth and demographic decline.

“So it is imperative in Beijing’s mind that they win the AI race,” he says.

“And even if they don’t win the AI race, to stop basically America winning it.”

That calculation could become critical if evidence emerged that a new generation of AI posed serious risks.

“The chances of there being an existential threat in China’s mind is a lot lower than the likelihood that America’s going to win the AI race if they slow down,” Olsen says.

Washington faces its own version of the same dilemma. US frontier development is driven predominantly by intensely competitive private companies, while government has less direct control over them than Beijing exercises over Chinese developers.

America’s advantage is greater transparency. Independent researchers have more opportunity to scrutinise frontier models and its AI safety ecosystem is more developed.

China’s advantage is the state’s ability to intervene directly. Neither, the report concludes, provides an obvious structural guarantee against losing control.

The result is a security dilemma in which every breakthrough increases the potential cost of restraint.

Slowing development to investigate a dangerous capability risks allowing the other side to move ahead, while continuing to accelerate could increase the very risks both are trying to contain.

Technology itself makes that problem harder.

Jake James, Sibylline’s Lead Analyst for Technology and Geopolitics and the report’s other co-author, says comparisons with the Cold War nuclear arms race only go so far.

Nuclear deterrence rested on the doctrine of mutually assured destruction (MAD) – the understanding that a nuclear attack would provoke devastating retaliation, giving both sides a powerful incentive not to strike first.

AI offers no equivalent.

“There’s no MAD in AI,” James says.

The distinction becomes particularly important with open-weight AI models, which can be downloaded, modified and redeployed by users far beyond the control of their original developer.

Chinese companies have pursued this strategy aggressively. Models originating in China now account for around 41% of downloads on Hugging Face, the world’s largest platform for sharing AI models.

That has helped Chinese technology spread rapidly despite America’s frontier advantage. But it creates another control problem: once a model has been downloaded and copied internationally, even Beijing cannot simply order it back.

“Some random person sitting in a house somewhere in the Midwest America doesn’t have any access or control to nuclear weapons,” James says.

“But they do have access to open-weight models.”

None of this means an AI catastrophe is inevitable.

Experts remain deeply divided over whether genuinely autonomous systems will ever escape human control, let alone when, and the Sibylline report explicitly avoids attempting to put a probability on such an event.

But it argues that the consequences are serious enough to make AI safety an international security issue rather than simply a regulatory or technological one.

That is where Britain enters the equation.

The UK has the world’s third-largest AI market, behind the US and China, and hosts Google DeepMind alongside substantial operations belonging to other leading AI developers.

Its AI Security Institute conducts technical research into advanced AI risks and works directly with frontier developers.

That combination gives Britain something relatively rare: technological credibility without being one of the two principal competitors.

“The UK holds a sort of really unique position in the AI sphere,” James says.

Britain cannot force Washington or Beijing to slow their AI programmes.

But the report argues it could help create the architecture needed to make restraint possible – establishing common definitions of dangerous capabilities, improving monitoring and verification, encouraging rapid disclosure of serious incidents and pursuing narrow agreements where both sides share an overwhelming interest in maintaining human control.

That role is becoming more explicit.

Britain is already using the AI Security Institute to work with leading developers on model evaluation and has positioned international co-operation over AI security as part of its wider diplomatic agenda.

Olsen believes Britain could ultimately act as an “honest broker” between Washington and Beijing.

“Britain has massive AI ambitions and the benefit of Brexit is unbelievably big,” he says.

“The Europeans have brought in rules and regulations which effectively strangle AI at birth.

“I think the UK can play an honest broker in the middle. We have the AI Security Institute and we are trusted more by China than the US is.”

The obstacle, however, is not simply finding safeguards that work. It is persuading two strategic rivals that observing them will not leave either side dangerously exposed if the other cheats.

And that, the report concludes, is the central paradox now confronting the AI race:

“The more dangerous advanced AI becomes, the stronger the case for international restraint; but the more strategically important it becomes, the harder restraint is to achieve.”