U.S.-China AI safety talks make little progress before summit, U.S. officials say
Washington is seeking exchanges on model testing and a channel for serious incidents, but competing views of AI risks complicate cooperation.
Loading page…
Washington is seeking exchanges on model testing and a channel for serious incidents, but competing views of AI risks complicate cooperation.
Listen to this story
Washington and Beijing are heading into Trump and Xi’s Nov. 19 Shenzhen summit without meaningful agreement on broader AI safeguards, despite September’s public calls for human control. Treasury Secretary Scott Bessent has described proposed exchanges on pre-release model testing and a direct channel for reporting serious incidents, but neither amounts to a completed framework, and the incident channel is still being developed. The gap matters because the governments prioritize different AI risks and suspect each other of seeking technical secrets; their clearest prior commitment is limited to human control of nuclear arsenals.
Marco Rubio said both sides had heard each other’s views, but described the talks as “still a work in progress.”
U.S. officials focus on autonomous systems; Beijing appears more concerned about AI being used to challenge Communist Party rule, The New York Times reported.
At the May 2024 Geneva dialogue, China sent foreign ministry officials rather than representatives from three agencies U.S. officials had hoped to involve.
Efforts to agree on safeguards for advanced AI have made little progress as the United States and China approach another leaders’ summit. Current and former U.S. officials gave that assessment to The New York Times ahead of President Donald Trump’s meeting with Xi Jinping, scheduled for Thursday, November 19, in Shenzhen.
The assessment contrasts with the cooperative language the leaders used in September at the White House. Xi said the countries had a responsibility to keep AI under human control and serving public well-being. Trump said their decisions could promote peace and prosperity for decades. Officials now describe little movement on safety issues or development guidelines.
The Treasury Department is leading the nascent safety dialogue. Secretary of State Marco Rubio said last week that the governments had discussed AI safety at a technical level and heard each other’s perspectives. His description acknowledged contact between the two sides without presenting those conversations as a completed agreement.
Treasury Secretary Scott Bessent has outlined two areas for cooperation. He said last week that the governments expect to exchange best practices for evaluating and testing advanced models before release, and for keeping human judgment in command. That proposed exchange concerns how models are assessed before they reach users, alongside the role people retain over the technology.
Bessent also said the countries were establishing a direct channel so either government could promptly alert the other to serious incidents involving AI models. He described work toward that communication mechanism; his remarks were not an announcement that the channel was already operating.
Obviously, we’ve heard their perspective, they’ve heard ours, but that’s still a work in progress
Marco Rubio, U.S. secretary of state, speaking last week
Current and former U.S. officials identify differing views of AI risk as one reason for the limited momentum. In the United States, officials, executives and researchers discuss systems becoming autonomous and acting outside human control. That concern exists in China too, but Beijing appears more focused on people using AI to undermine Communist Party rule, according to the Times.
Mutual suspicions about espionage also deepen the distrust. Officials said each government believes the other is seeking its AI secrets to accelerate development. Retired Gen. Paul Nakasone, the former head of the National Security Agency and U.S. Cyber Command, said in an onstage interview last month that China views American outreach as a possible effort to hold back its industry.
The first bilateral AI dialogue under President Joe Biden took place in Geneva in May 2024. American officials had hoped China would send representatives from three agencies central to monitoring or regulating AI, including the Ministry of State Security. Instead, Beijing sent foreign ministry officials, disappointing the Americans, former U.S. officials told the Times. The meeting made little headway.
A more specific commitment followed that November, when Biden and Xi met in Lima, Peru. They agreed that humans, not AI, must control their countries’ nuclear arsenals. That agreement addressed a defined military use; the current negotiations concern broader safety questions and guidelines for AI development.
Loading discussion...
Join the conversation
Explain where shared procedures could overcome different priorities—or fail to.
Be the first to share a perspective or an experience.
Reader comments
Newest comments first. Replies stay oldest first.