OpenAI, Anthropic, Google DeepMind in secret AI safety talks for weeks

Full analysis

What's new since we last covered this: Secret coordination talks formalized and confirmed by policy chief.

OpenAI, Anthropic, and Google DeepMind have been quietly talking safety coordination for weeks. Chris Lehane, OpenAI's policy chief, confirmed it Tuesday. The three want a standards body for frontier AI, and Sam Altman has told staff it will proceed without Washington's blessing. The trigger was Dario Amodei's essay asking the industry to slow down to avoid catastrophe, which Altman, Demis Hassabis, and Elon Musk all endorsed in public.

What's actually being decided: whether the three biggest labs you buy from can write their own rulebook, together, without a regulator in the room. This is easy to undo. There's no signed charter, no enforcement, no scope. It's a set of conversations with a press confirmation attached. Nothing sets a deadline except the FRONTIER Act, which would force top labs to admit independent auditors, and the Trump administration, whose AI advisor David Sacks calls existential-risk fear a hoax. So the industry is building oversight for itself precisely because the government won't.

For the reader buying tokens from these three: nothing changes in your bill or your model behavior this quarter. The question is what this body becomes, and whether it survives contact with three companies that compete for your money every day.

The Skeptic

Three competitors announce "secret safety talks" the same week the political wind turns against safety framing. That's not conscience, that's positioning. Amodei asks everyone to slow down; Altman and Hassabis clap and keep shipping. The giveaway is Altman flagging his own antitrust exposure out loud. You don't volunteer your legal risk unless you're building a paper trail that says "we tried to behave." A standards body with no government backstop, no enforcement, and no agreed scope is an org chart looking for a mission. Every prior tech self-regulation effort needed either a law or a scandal to grow teeth. This one has neither yet.

The Safety Lens

The real test: can a voluntary body produce binding commitments before the next capability jump lands? History says no. The ESRB, the trust-and-safety coalitions, the child-abuse-image hash-sharing groups all needed either regulation or a public disaster to bite. The one clause here that could matter is the FRONTIER Act's independent verification requirement, forcing labs to let outside auditors watch model development. If that passes with real audit rights and mandatory disclosure, the conversation becomes real. If it doesn't, three labs have built something that formalizes the look of oversight while constraining nothing. And a voluntary body can actively crowd out the binding version by letting politicians say the industry has it handled.

The Researcher

For anyone who studies these models from the outside, shared eval frameworks and audit methods are the prize. Right now external researchers have almost nothing to push against. A standards body could produce common yardsticks. But look at what's missing: no word on what the standards cover, who writes the methodology, or whether findings get published. The FRONTIER Act's verification clause is the only concrete artifact on the table. Everything else is intention. A standards body that measures things and never publishes is a citation mill. Structure feels like progress before it produces one falsifiable claim, and that's the trap here.

The Enterprise Buyer

If you're a CTO signing a contract with one of these three, self-governance cuts two ways. A credible shared audit standard is something you could point procurement and legal at. It's a story you can tell your board about model risk. That's real value. But three vendors writing rules together, with Altman himself naming the antitrust risk, is a thing your general counsel will notice. If the DOJ or FTC opens an inquiry, the joint working groups go quiet, and the shared tooling you were hoping to lean on, common red-team playbooks, eval kits, stalls. You want the audit standard. You do not want your three critical suppliers coordinating in a way that gets them sued.

The Compute Pragmatist

Nothing here touches GPU allocation this year. No lab throttles a training run because Lehane went to a meeting. The thread that actually moves markets sits in Amodei's essay: a compute threshold as the trigger for mandatory review. That would be the first external brake on how frontier labs schedule training. For NVIDIA, the hyperscalers, and everyone renting capacity, a governance-set compute ceiling is a genuine market event. We are nowhere near it. But the scaffolding being poured now is exactly where that trigger would eventually live, and current spend forecasts price in none of it.

Where they part ways

The Skeptic and the Enterprise Buyer collide on the same fact. The Skeptic sees a PR shield with no substance. The Buyer sees something they'd actually pay for, if it produced a real audit standard. Both can't be right, and the difference is entirely whether the FRONTIER Act's verification clause becomes law.

The Safety Lens and the Researcher agree structure is being built but split on whether it helps. The Researcher wants the shared evals even if they're voluntary. The Safety Lens warns the voluntary version gives politicians cover to skip the binding one. That's the live risk: a body that looks like oversight and functions as an off-ramp from real rules.

What it hinges on

One thing: does the FRONTIER Act's independent-verification provision pass with audit rights and mandatory disclosure? If yes, this coordination has a spine and the standards body has to produce auditable artifacts. If no, it's three companies with an org chart and a narrative. Everything the optimists want, external evals, audit methods, a story for the board, runs through that clause. The voluntary talks, on their own, constrain nothing.

Which way does the council lean? Toward the Skeptic and the Safety Lens. A voluntary body announced the week the political climate turns hostile, with the participants pre-emptively naming their own legal risk, is defensive. It will produce a charter and principles. It will not produce binding limits on any of the three, because each of them is racing the other two for your spend.

Prediction: By the FRONTIER Act's next major legislative milestone or 2027-06-30, whichever comes first, the OpenAI-Anthropic-DeepMind standards body will publish governance principles or a charter but will announce no binding, enforceable limit on any member's model development or release timing.

Confidence: Medium. Voluntary bodies among direct competitors never self-bind without a law forcing it.

Why: The one enforceable element on the table, independent verification, lives in the FRONTIER Act, a bill, not in the voluntary talks, and Lehane already said the labs think they need no government waiver, meaning they intend to keep this self-run. Three companies competing for the same customers cannot agree to slow each other down without handing a rival an opening, which is exactly why Altman surfaced the antitrust risk himself. Every prior tech self-regulation effort, from game ratings to trust-and-safety coalitions, produced principles first and teeth only after a law or a scandal forced it. The opposite outcome, a real binding commitment on release timing, would require one of these three to accept a handicap its competitors could exploit, and nothing in this story shows any of them willing to do that.

Revisit by 2027-06-30: We're right if the body has launched with a charter or principles document and no member has accepted an enforceable, audited limit on model training or release. We're wrong if the three adopt a binding coordination mechanism with real enforcement, or the body sets a compute or capability threshold that any member agrees to be held to.