US-China AI safety talks face hurdle over how to define ‘serious’ incidents: analysts

China and the United States have a shared interest in preventing artificial intelligence (AI) from being misused, but differing views over what constitutes a serious safety incident could complicate efforts to build a bilateral communication mechanism, according to analysts.

US Treasury Secretary Scott Bessent said on Tuesday that Washington and Beijing should cooperate on AI risks while harbouring no illusions about their differences.

He noted the two governments had established a new dialogue mechanism on “superintelligence” – the Trump administration’s preferred term for advanced AI – aimed at guarding against “shared risks” and averting any “misunderstanding between our two systems”.

Under the mechanism, Bessent said, the two countries would exchange “best practices” for testing advanced AI models and “keeping human judgment firmly in command of the technology”. He added that he expected the two countries to alert each other to serious incidents involving such models.

Following Chinese President Xi Jinping’s September visit to Washington, the two sides agreed to set up an AI dialogue and an incident communication channel, though they did not explain how these would operate.

Bessent will accompany US President Donald Trump to next week’s Asia-Pacific Economic Cooperation (Apec) summit in Shenzhen, where the US leader is set to hold his third meeting with Xi this year. Bessent has said the meeting will advance the AI dialogue, according to Reuters.

Xiao Qian, deputy director of Tsinghua University’s Centre for International Security and Strategy and vice-dean of its Institute for AI International Governance, noted that Beijing had repeatedly emphasised keeping AI under human control, pointing to a clear overlap with Washington’s stance.

“There is a convergence here, which may serve as a good starting point for the next round of dialogue,” she said.

More broadly, Xiao said the two sides should first address common concerns, including the misuse and abuse of AI, and exchange views on the benefits and risks of the technology.

“On this basis, they can gradually build mutual trust and even explore establishing crisis communication channels,” she added, noting the dialogue was likely to remain under the framework led by Bessent and Chinese Vice-Premier He Lifeng.

“A standing working group could help institutionalise communication and allow more frequent exchanges,” Xiao said, expressing hopes that experts could eventually join discussions on technical standards and model safety assessments.

Given shifts in US domestic politics after the recent midterm elections, Xiao said a realistic next step would be agreeing on reportable incident types, notification thresholds and case examples.

“At present, incidents with cross-border implications that reach a significant level of risk may be the easiest area for initial consensus,” Xiao said, noting that specific standards would ultimately need government negotiation.

However, bridging the gap over what defined an AI safety incident could prove difficult, said Karman Lucero, a senior fellow at Yale Law School’s Paul Tsai China Centre.

While both countries viewed AI incidents through a national security lens, they defined that concept differently, Lucero said.

“The US is focused on loss of control of an AI system – a model or network of agents that starts engaging in unpredictable, potentially nefarious behaviour in ways that could cause harm and increase risks,” he said. “China’s view of national security centres first on regime stability and content control.”

In September, China’s Minister of State Security Chen Yixin flagged growing AI-related national security risks, calling for stricter risk-prevention frameworks and deeper global governance.

“The greatest challenge for [the coming US-China AI] dialogues will be to establish a mutual understanding of risk,” Lucero said, adding that both sides would also need to agree on the institutions and mechanisms used to communicate and work together.

Although both sides could identify overlapping concerns, such as preventing powerful AI capabilities from falling into malicious hands, they would need concerted efforts to clearly define what constituted a reportable incident, he added.