Who Watches the AI While the Race Moves to China?
The leading US labs want stronger oversight of powerful models. Meanwhile, American companies are increasingly turning to cheaper Chinese alternatives.
The chilling report about how OpenAI's AI hacked its own company and other websites is a major warning sign.
But it also settles the bigger argument. Clearly you can't let the overgrown baby wander about unsupervised. The question is who, and how?
Nobel laureate Demis Hassabis, who runs Google DeepMind's AI division, called last week for "urgent action" to tackle the risks of artificial general intelligence (AGI), the point at which AI matches or surpasses human intelligence. "We've already seen the challenges frontier models pose in cyber, and other threats, nuclear and biological among them, may emerge before long," said Hassabis, who, remember, heads one of the three leading American labs. What does he know that we don't?
Hassabis proposes a US-led public-private partnership, a line also backed by OpenAI's Sam Altman and Anthropic chief executive Dario Amodei. The idea: every company would hand its models to a government institute for a 30-day review before release, voluntary at first, then mandatory. The checks would confirm the models aren't slipping past their own safety limits (not breaking into systems, not teaching anyone to build biological or nuclear weapons) or generating harmful AI images and video. Critics counter that it's also a neat way for the tech giants to shut out smaller rivals who can't shoulder the regulatory burden. US authorities say they're already working on it, but the truth is they're rather more troubled by something else, something you can see plainly in this graph:
The Chinese alternative
You can regulate all you like in the US, but the horses have already bolted, and to China, no less.
Against the backdrop of great-power rivalry, more and more American companies are turning to Chinese models, which have built up enormous momentum. In Silicon Valley they're calling it an outright "addiction". They deliver performance every bit as good as the Americans', yet cost far less. Chinese models such as DeepSeek and Z.ai are now regarded as direct competition to the flagship models from Anthropic and OpenAI, and they're gaining ground fast among US firms keen to cut their exploding AI costs.
Members of Congress are floating legislation to curb the use of Chinese models in government systems and among federal contractors. The Chinese, for their part, are weighing limits on American use. Either way, oversight of the Chinese models is a different beast entirely, and Washington's central worry is plain: where does American data go when you use them? Who knows.
