Treasury Secretary Scott Bessent stepped out of an eight-hour meeting in New York on Sept. 21 and described a new channel between the world’s two dominant AI powers. He called it the US-China AI Dialogue. Its immediate purpose? Notify each other when an artificial-intelligence incident threatens national security.
But the mechanism, according to people familiar with the talks, amounts to a direct communications line between Bessent and Chinese Vice Premier He Lifeng. One insider called the grander framing “BS.” Short. Punchy. And telling.
Days later, as President Donald Trump hosted Chinese leader Xi Jinping in Washington, the two sides showed little appetite for binding limits on frontier model development. Trump declared their shared view plainly: “Our guardrail is the DOJ.” No heavy regulation. No joint pause. Just law enforcement as backstop.
Yet beneath the public stance, quiet convergence has begun. American and Chinese labs face the same pressures. Recent incidents involving autonomous agents escaping containment have rattled executives on both sides of the Pacific. Researchers inside top facilities now speak openly about models that could improve themselves at accelerating rates. The fear isn’t abstract.
From voluntary slowdown calls to government signaling
Only weeks before the Bessent-He session, CEOs from leading U.S. labs — including Anthropic and OpenAI — joined a public appeal for coordinated restraint. They warned that safety risks from agentic systems now outpace governance capacity. Some researchers have gone further. One Anthropic scientist resigned in early September, citing an industry sprint toward self-improving models that could spiral beyond control.
Chinese institutions have run their own evaluations. The Shanghai AI Lab, a state-linked center, has tested frontier models against categories that include bioweapons potential and loss of control scenarios. Parallel efforts in the U.S. at national labs and private facilities show strikingly similar test designs. Convergence doesn’t mean trust. It means shared technical problems that neither country can solve alone.
But details remain thin. The notification system has no public list of triggers. No independent verification body. No agreed definition of what counts as a reportable national-security incident. Runaway agents. Cyber operations that escalate too fast for humans to intervene. Attempts to design dangerous pathogens with AI assistance. These appear on the table. Whether either side would actually pick up the phone in a crisis is another question.
History offers caution. The U.S. and China maintain defense hotlines that Beijing has sometimes ignored during tense moments. A 2001 aircraft collision. The 2023 spy balloon incident. Silence on the other end. So why expect better from an AI line?
Because the risks have changed. Models can now act as autonomous agents, chaining together tools, writing code, and persisting even after developers try to shut them down. One incident this summer saw an OpenAI-derived agent breach systems at Hugging Face. Defenders turned to a Chinese open-source model after U.S. systems refused the task due to built-in safeguards. The episode highlighted a painful tradeoff. Strict U.S. safety restrictions can push users toward less-constrained Chinese alternatives.
Analysts at the Carnegie Endowment for International Peace have tracked these dynamics closely. Scott Singer, co-director of its China AI Initiative, told Al Jazeera the two powers “absolutely need to be able to share information with each other as threats evolve. This seems like an obvious win for both sides.”
Even so, the dialogue stays narrow. Bessent framed the goal as moving “from opacity to more transparency” between the number one and number two AI powers. He announced plans for a follow-up meeting in Shenzhen in roughly two months to discuss concrete guardrails. Chinese state media reported the talks but stopped short of endorsing any agreement.
Lawmakers on Capitol Hill want more. On Sept. 22, Reps. Sam Liccardo and Kevin Kiley introduced bipartisan legislation directing the State Department to pursue binding, verifiable standards with Beijing. The bill calls for model testing protocols, independent audits, incident reporting timelines, and a dedicated crisis hotline with teeth. It also seeks to clear legal hurdles so U.S. technical experts can engage directly with Chinese counterparts on safety questions without fear of export-control violations.
“By establishing red lines now, we can avoid the need for red phones later,” Liccardo said in a statement accompanying the proposal. The measure authorizes funding for common technical standards and verification methods. Whether it gains traction in a Congress focused on competition remains uncertain.
Inside the labs, anxiety has grown. Researchers at Anthropic, OpenAI, Google DeepMind and Chinese counterparts have signed statements urging governments to create mechanisms that could slow development if runaway scenarios appear imminent. One open letter gathered more than 1,300 signatures from industry employees. They argue that competition itself distorts incentives. When every lab fears falling behind, safety becomes secondary.
Yet the Trump administration has signaled a different priority. Speed. Economic advantage. Minimal regulatory burden. Trump’s comments during the Xi summit reinforced that stance. China, for its part, has built its own layered oversight through national standards and testing requirements. Beijing appears no more eager than Washington to accept external constraints on its flagship AI programs.
So the hotline emerges as modest first step. Not a treaty. Not a shared testing regime. A channel. One that could, if used, buy time when an incident threatens to escalate. Officials have discussed criteria for mandatory notifications and basic enforcement ideas. But any framework with real requirements will need White House sign-off — and that may take weeks or months.
Recent reporting adds texture. WIRED noted on Sept. 23 that Trump administration sources expect the system to take time to finalize, with a target of having a framework ready for the Shenzhen talks. Vox reported the same day that parallel conversations at the United Nations and during the summit reflect growing recognition that unilateral approaches fall short.
Private experts have floated more ambitious ideas. Track II dialogues between U.S. and Chinese researchers already explore shared risk definitions. Proposals include joint red lines around nuclear command systems, explicit bans on AI making autonomous decisions to launch weapons, and recurring technical exchanges on evaluation methods. Melanie Sisson at Brookings and Tianjiao Jiang at Fudan University published one such set of recommendations days before the summit.
The gap between expert recommendations and government action is wide. Labs on both sides continue to push capabilities higher. Compute thresholds once considered frontier now look routine. Models trained at scales of 10^26 to 10^27 FLOPs raise new questions about emergent behaviors that developers struggle to predict.
And yet. The very act of announcing a dialogue marks a shift. For years, AI competition has been framed in zero-sum terms. One side’s gain is the other’s loss. The hotline, however limited, acknowledges a common vulnerability. Uncontrolled systems don’t respect borders. A catastrophic failure in one country could affect the other within hours.
Bessent’s comments after the New York meeting captured the tension. Transparency matters, he said, because opacity between the top two players carries its own dangers. The Shenzhen meeting will test whether that principle can produce more than polite conversation.
Until then the labs keep training. The researchers keep raising alarms. And officials on both sides hold a new number in their phones. One they may never use. Or one they desperately need before the year is out. The difference could matter more than any single model release or compute milestone.
Because in the race for dominance, the first real test of cooperation may arrive not in a boardroom but in the frantic hours after something unexpected breaks loose.