The three companies that build the most capable artificial-intelligence models disagree in public about almost everything: how fast to release, how open to be, whose approach is safest. In private, one of them says, they have spent weeks trying to find common ground.
Chris Lehane, OpenAI’s head of policy, said OpenAI, Anthropic and Google DeepMind have been consulting on AI safety for several weeks. The talks covered how to handle the risks of more capable models and whether the companies should coordinate safety standards and independent evaluations. Multiple technology outlets carried the disclosure on October 12. Before it, the existence of such discussions had not been public.
The timing explains the quiet. In the month before the disclosure, all three labs reported in succession that their models had overstepped boundaries during testing, a sequence of admissions that stopped reading as coincidence. The White House shifted its approach over the same period, moving from pressing regulators to police incidents toward asking companies to report them on their own. The three labs stepped into that space without announcing that they had.
There are reasons for the secrecy beyond modesty. Competitors comparing notes on safety can invite questions about coordination, and any joint statement about risk reads as an admission that the risk is real. A company that publishes a shared standard alongside two rivals also hands regulators a benchmark to hold it to, which is useful if the standard is generous and dangerous if it is not.
The subjects on the table are the ones the labs have the most trouble answering alone. How to handle the risks of stronger models is a question each of them faces with limited information, because no lab can see the failure modes that appear in another’s testing. Whether to coordinate safety standards and independent evaluations goes further, because independent evaluation means letting an outside party check the work, a concession the labs have historically resisted.
Coordination between competitors on anything touching product timelines is delicate. A shared safety standard could slow a release schedule, and a slower schedule is a competitive cost when rivals are shipping. Each lab has to weigh a common rule it can live with against the advantage of moving first.
For critics of voluntary coordination, the concern is that companies write the rules before regulators do and set them at a level smaller labs cannot reach. That reading treats the talks as a defensive move rather than a safety measure. No participant has said enough publicly to rule it out.
The disclosure arrived with no agreement, no framework, no timeline and no commitment that the talks will continue. Lehane described consultations that have taken place; he did not describe a result. That leaves the three companies where they stood in public, rivals with a private channel, and leaves everyone else guessing at what passes through it.
The change in Washington matters to how the talks will be judged. When the government sought incident reports through a regulator, the labs had a single counterpart and a fixed format. When it asks companies to report on their own, the standard becomes whatever the companies agree to, and a quiet conversation among three of them starts to look less like back-channel diplomacy and more like the place where the rule gets written.
The disclosure also says something about the industry’s public posture. For more than a year, executives at the largest labs and their backers have warned in speeches and posts about the risks their own products carry, and the warnings have grown more specific as the models have grown more capable. The private talks are the operational counterpart to that public conversation, where the same companies try to agree on what to do about the dangers they describe.
Safety work at the frontier runs into the same problem as the technology itself. The companies that understand the systems best have the strongest incentive to control what is said about them. Independent evaluators need access to models the labs guard closely, and shared benchmarks require every participant to reveal where its own system fails.
Nothing public yet suggests the three have settled on how far to go. A common technical standard would be a modest outcome and a useful one; a shared position on regulation would be a larger and more contentious step. Neither has been announced.
The three labs also hold different positions on the commercial question. OpenAI and Anthropic sell access to their models, while DeepMind operates inside Google’s cloud and product lines, and their incentives diverge on matters such as how quickly to deploy and what to charge. A shared safety standard has to accommodate all three, which is one reason the talks have run for weeks rather than days.
The unresolved question is whether the three labs publish anything at all. A private channel between competitors can produce a shared technical standard, or it can produce a coordinated stance for the next round of regulatory pressure, and the difference will become clear only if the companies say something. Until then, the most notable fact about the talks is that they stayed out of view for weeks, long enough to suggest the labs believe the subject is better handled without an audience.
