US–China AI safety talks are “on”—except the White House says they aren’t

The superpowers are finally putting AI safety on its own agenda—if the sourcing holds.

Reuters, citing people briefed on planning, says Washington and Beijing are preparing mid-September talks devoted solely to AI safety, with Treasury Secretary Scott Bessent leading the U.S. side. The U.S. wants joint monitoring of AI-driven cyberattacks and some form of lab self-policing with information sharing; China sees the dialogue as a deliverable for a Sept. 24 leaders’ meeting. The Japan Times carries the same thread—and the White House rebuttal that no mid-September AI meeting is currently planned.

The calendar is crowded. Frontier launches, agent breakouts, and critical-infrastructure defense are the same week’s news. Putting Treasury, not a science adviser, in the lead chair marks the topic as macro-meets-security, not a friendly lab workshop.

Take: What you can read today is the seam in the messaging, not a communiqué draft. If talks happen, symbolism will outrun enforceable standards—“labs policing themselves” sounds thin after the Hugging Face intrusion. If the White House denial sticks, AI safety may still be bargained off the summit menu. For industry, harder constraints remain domestic: AG probes, copyright suits, and supply-chain stamps move faster than an unconfirmed Beijing dialogue.