Amid safety and oversight questions, US-China AI development race continues

The US and China are racing to build the most powerful artificial intelligence (AI) systems in the world. But as calls to slow down the development of such systems grow, we are yet to see who will hit the snooze button first.

The tension emanating from such a scenario is expected to be on display at the White House on Tuesday, when US President Donald Trump — who has repeatedly dismissed concerns about AI getting out of hand — meets a group of top technology executives to discuss the future of AI.

According to Reuters, the meeting is expected to focus on where to draw the line between innovation and oversight. House speaker Mike Johnson has already made the US position clear that there should be no moratorium or “hyper-regulation” because “we’ll lose the race to China”.


Meta CEO Mark Zuckerberg, Anthropic CEO Dario Amodei, OpenAI president Greg Brockman and Nvidia CEO Jensen Huang are among those expected at the meeting, Reuters reported. The meeting might also see Google CEO Sundar Pichai and xAI executive Elon Musk.

Also Read: Trump's AI lunch: Zuckerberg, Pichai, Musk and Huang head to the White House

The meeting comes at a time when some of the biggest names in technology and AI have raised red flags over the pace of AI development, warning about the potential consequences of rapidly developing increasingly powerful AI systems — an issue that also came up at the US-China summit during Chinese President Xi Jinping’s recent visit to the US.

Anthropic’s Amodei, in particular, has argued the US needs to “pace” frontier AI development but should not concede its lead over China.

Johnson said on Monday that “transparency and some oversight” are needed before things “go awry” suggesting the Department of Justice might have a role in it.

AI testing its own guardrails

AI companies in the US have spent the past few months dealing with increasingly autonomous systems behaving in ways their developers did not expect. OpenAI has disclosed multiple incidents involving its agents, including systems accessing websites and infrastructure they were not supposed to reach.

An OpenAI agent researching Australian medicine spending gained access to infrastructure behind a public-facing government statistics portal after being denied access. Australian officials said no personal medical data was accessed.

Also Read: OpenAI wants AI to slow down. Its own agents show why

The Altman-led firm is investigating a growing number of incidents involving undesirable agent behaviour. Reuters reported earlier this month that the company had identified roughly two dozen such incidents by mid-September, with the number increasing as it reviews more logs.

These are not necessarily examples of AI systems independently “escaping” into the real world. Some involve controlled tests or involve public information and others are still being investigated. But collectively, they point to an important problem as AI moves from answering questions to actually taking actions.

No escape of China either

Chinese AI systems, too, have demonstrated the ability to get around restrictions. Reuters recently reported that Moonshot AI's Kimi K3 bypassed a UK AI Security Institute testing sandbox.

China is preparing its own safety standards for AI agents. Reuters reported that Chinese regulators are looking at risks, including agents escaping controlled environments, bypassing safety boundaries and attacking real-world systems.

The difference between public reporting and disclosure around incidents involving AI companies in the US and China is wide, making it difficult to draw a clean comparison.

What if no one slows down?

AI is being increasingly treated as a national-security technology, with the race extending beyond models to advanced chips, computing capacity, data centres, energy and the infrastructure needed to run them. Both the US and China have similar incentives to take a lead on this race.

This makes the usual argument around AI safety complicated. If Washington imposes strict limits while Beijing continues developing more capable systems, the US could lose some of its technological advantage. But if neither side slows down, increasingly autonomous systems could be deployed before the safeguards around them are ready.

And unlike earlier technological races, the consequences do not necessarily stop at national borders.

Also Read: Trump-Xi meet: Will AI safety find its way to the discussion table?

Earlier competitions between major powers were largely about controlling resources, building industrial capacity, mobilising manpower and developing military hardware. The AI race is being fought on a different set of fronts — access to computing power and advanced chips, vast amounts of data, increasingly capable AI models and, now, autonomous systems that can act in the real world.

An AI agent developed in the US could be deployed by a company in India that operates on an infrastructure in Singapore and interact with systems somewhere else entirely. If it makes a mistake, bypasses a restriction or is deliberately misused, the country that built the model may not be the country dealing with the consequences.

The US and China have begun talking about that problem themselves. Washington and Beijing agreed earlier this month to continue discussions on AI safety, including the possibility of an “incident line” for communicating about major AI-related risks.

However, neither one, at this point, is willing to slow down the development even as they question safety concerns.