Experts chart practical ways for AI dialogue

Experts at a Brookings Institution discussion on Tuesday pointed to technical exchanges, safety testing and communication over artificial intelligence-related incidents as practical areas where China and the United States could begin building a more sustained dialogue, even as competition between the two countries in advanced AI continues.

The discussion focused more on what a workable initial stage might look like rather than whether the two countries could reach a broad agreement on AI governance. Two other events held on Tuesday by the Peterson Institute for International Economics, or PIIE, examined the growing economic impact of AI, including competition between Chinese and US models, the distribution of productivity gains and the increasing role of AI in the global economy.

The discussions followed President Xi Jinping's state visit to the US in late September. The Chinese Foreign Ministry said that the two sides agreed to establish the China-US AI Dialogue to exchange views on AI-related risks and benefits, with the next exchange to take place in November, as well as a bilateral communication channel for AI incidents.

During his talks with US President Donald Trump, Xi said the two countries may have competition, but they may cooperate even more. He called for continued dialogue, efforts to jointly prevent the misuse and abuse of AI, and a people-centered approach.

Kyle Chan, a fellow at Brookings' John L. Thornton China Center, suggested starting with what he called "selective sharing", rather than trying immediately to negotiate detailed requirements on what each side must disclose. The two sides could begin with limited information about incidents of concern, including discussions based on publicly available information. "We can start with kind of this piecemeal solution and build this as we go along," Chan said.

At the Brookings event, Matt Sheehan, a senior fellow and co-director of the China AI Initiative at the Carnegie Endowment for International Peace, argued for keeping expectations practical.

Given the constraints between the two countries, Sheehan said that the best near-term outcome may be something "short of cooperation and more than sort of this unbridled competition".

He said technical exchanges could improve safety practices without immediately requiring binding reciprocal restraints, citing evaluating loss-of-control risks and testing systems that may behave differently once they recognize they are being evaluated.

"How do you actually build sandboxes that are actually going to constrain these models?"

James Andrew Lewis, a distinguished fellow at the Center for European Policy Analysis who has participated in negotiations with China, was more cautious about framing AI governance as an arms-control issue. Asked whether arms control was a suitable framework at this stage, he said, "The short answer is no."

Lewis cited differences in risk perceptions and the lack of data needed for such negotiations, and suggested that standards, principles and guidelines could offer a more realistic avenue for discussion.

The debate comes as Chinese AI models have rapidly narrowed their performance gap with leading US models. Bloomberg Intelligence said this week that top Chinese models were only about 3 percent behind their US rivals on benchmark scores following the September release of DeepSeek's V4.1 Flash, compared with about 9 percent in May and 15 percent earlier this year.

The growing AI competition was also a central theme at a PIIE event earlier on Tuesday. Martin Chorzempa, the institute's Dennis Weatherstone senior fellow, said Chinese developers do not have to match US companies at every point on the technological frontier to influence the global AI market.

Chorzempa said intensifying competition would "tend to transfer the gains of AI to the users instead of the developers of AI".

Anton Korinek, a PIIE nonresident senior fellow and head of Transformative AI Economic Studies at the Anthropic Institute, stressed the uncertainty surrounding AI's broader economic impact. Korinek's main message to policymakers was to "prepare for a range, not for a point forecast".

In faster-growth scenarios, he said, "The hard problem is not whether we have gains, but how we distribute these gains."

Chan stressed the importance of keeping China-US engagement on AI on an official track. "It matters a lot to have the official dialogue, to have the buy-in from Beijing on the Chinese side," he said.

Even if such a dialogue appears limited at the beginning, he said, "It matters to have that kind of green light from the top in China to say, okay, this is something that we want to engage in and now let's start talking."

During the next round of China-US AI talks, Sheehan said one meaningful outcome would be to "turn this into a structured ongoing dialogue". Another would be some way for the two sides to communicate about "evolving, rapidly evolving AI incidents that are crossing borders or potentially crossing borders".

"I think those are the two things that I would hope (for) from this first November meeting," Chan said, adding that he considers both feasible.

yifanxu@chinadailyusa.com