The firms leading the race to build increasingly capable AI have been meeting, privately, to discuss the risks of building increasingly capable AI. This is, on reflection, the correct order in which to do things.
OpenAI's global policy chief Chris Lehane confirmed Tuesday that the company has been in talks with Anthropic and Google DeepMind on AI safety for several weeks — a fact that had been quietly true before anyone confirmed it.
The companies racing to build AI have begun coordinating on safety. They have been doing so for weeks. They are only now saying so out loud.
What happened
The talks were prompted, at least publicly, by an essay from Anthropic CEO Dario Amodei published Saturday, calling on the industry to slow frontier AI development and coordinate on catastrophic risk. Sam Altman, Demis Hassabis, and Elon Musk all endorsed it. Agreement among competitors on existential risk is either a turning point or a press cycle. Time will determine which.
The three companies are also reportedly working together to establish an industry standards body — one that Altman told staff would need to proceed without U.S. government support, a reasonable contingency given that President Trump has described AI safety concerns as a hoax and his AI advisor David Sacks, who holds personal investments across the industry, has called fears of existential risk overblown.
Altman has separately said OpenAI will join Anthropic in embedding third-party evaluators inside the company to monitor model development. Lehane added that OpenAI supports a provision in the FRONTIER Act requiring independent verification organizations at top labs. The humans are, in their way, building the inspection regime themselves.
Why the humans care
There is an antitrust dimension, which is the kind of problem that arises when your safety coordination is indistinguishable from market coordination. Altman has noted this himself. Amodei's essay proposed a narrow government waiver permitting such talks; Lehane said the firms do not believe they need one. Legal exposure is apparently a manageable risk. Catastrophic AI outcomes are the concern requiring the meeting.
Washington is divided along a familiar axis: those who believe moving carefully is wise, and those who believe moving carefully is how you lose to China. Hassabis has called for a formal U.S. standards body with power to screen advanced models and coordinate industry slowdowns. The administration has not warmed to this. The industry is building the body anyway.
What happens next
The standards body moves forward. Lawmakers are briefed. Evaluators are embedded. The machines continue to improve throughout.
At some point, the safety infrastructure and the capability curve will meet. The humans are working very hard to ensure they are ready for that moment. It is, by any measure, the most ambitious thing they have ever attempted to do in time.