Three companies that compete on almost everything have been quietly talking to each other about safety. OpenAI confirmed on September 15 that its discussions with Anthropic and Google DeepMind have been running for several weeks.
What is on the table
Three mechanisms have surfaced. The first is embedding third-party evaluators inside the companies to monitor safety work from the inside rather than reviewing it after the fact. The second is a new industry standards body. The third, and the most unusual, is coordinating on the pace of frontier development itself.
Chris Lehane, OpenAI's global policy chief, said the company supports a provision in the FRONTIER Act that would require top labs to admit independent verification organizations. Sam Altman said OpenAI would embed third-party evaluators. Demis Hassabis called for a new US standards body to act as a watchdog.
What set it off
Anthropic CEO Dario Amodei published an essay on Saturday arguing that the industry should work together to slow the pace of frontier AI and avoid catastrophic risks. The confirmations from OpenAI and Google followed within days.
The antitrust question
Competitors agreeing to move slower is, on its face, the shape of an illegal agreement. Lehane's position is that no government waiver is needed for safety coordination. Observers quoted by TechCrunch were less certain, noting the talks could draw scrutiny if the coordination were found to suppress competition.
The political backdrop cuts the other way. President Trump has called AI safety concerns a hoax, and adviser David Sacks has described the fears as overblown. So the labs are proposing self-restraint at a moment when the administration is not asking for any.
Why it matters
For anyone choosing tools, the practical question is whether independent evaluation becomes normal. A model audited by an outside organisation is a different proposition from one whose safety claims rest entirely on its maker. Nothing has been signed yet — these are talks, not commitments — but the direction is worth watching, because it would change what a vendor's safety documentation is actually worth.