Amodei Asked the Industry to Coordinate. It Had Been Coordinating Since July

On Monday, The Information reported that working groups from Anthropic, OpenAI and Google have been meeting regularly since at least July to discuss an industry-run standards body for frontier AI — a structure for testing and auditing models before they ship. The Washington Post reported the same talks. Both accounts rest on people familiar with the discussions; no company has confirmed anything on the record, and the reporting should be read as reporting rather than as an announcement.

If it holds, it changes what happened over the weekend. Dario Amodei's "We Must Pace the Frontier," Sam Altman's endorsement and the shelved OpenAI IPO were not the beginning of a coordination effort. They were the moment an existing one went public.

What is reported, and what is not

The described body would set shared safety benchmarks, run pre-release evaluations of frontier models, and give the results somewhere to live outside the company that produced them. The working groups are reported to have met as recently as last week. What is not reported is any agreed charter, any funding, any named participants, or any commitment by any lab to abide by a finding it dislikes.

That gap matters, because the three companies do not appear to want the same institution.

Three designs, three answers to one question

Hassabis put the FINRA comparison on the record on 14 July, calling for an independent standards body to test frontier models. Amodei's essay is more cautious about the same idea:

"An agreement by both sides to test their models before release for acute risks... could be done through a global standards body. I actually think creating such a body is likely feasible, but giving it real teeth will be a challenge."

Altman, according to The Information's account of an internal all-hands, has told staff the labs will have to build such a body themselves, without backing from the US government.

These are not three flavours of the same proposal. They are three different answers to the only question that determines whether a standards body does anything: when the tests come back bad, who has the power to say no. Amodei's answer is a regulator. Hassabis's is a self-regulatory organisation with a federal backstop. Altman's, for now, is nobody.

The antitrust problem is in the text

The obstacle here is not secrecy. Amodei names it himself:

"For antitrust reasons, it's helpful for the US government to mediate or at least enable these discussions — they don't need to participate, but do need to issue a narrow waiver for certain kinds of safety conversations."

That is a request for permission for competitors to agree on the rate and manner of their own output. It may well be the right thing to do. It is also, structurally, the thing competition law exists to prevent, which is why the plan cannot be executed without Washington — and Washington is currently pointing the other way.

A draft executive order that would establish a self-regulatory organisation for AI is reported to be circulating inside the administration and stalled by internal opposition. The president's own position was on display over the weekend in Ireland, where he downplayed the need for further checks on AI development: "But whoever wins with AI wins." Cohere's Aidan Gomez, writing on 13 September, called the broader coordination push "a sheep in wolf's clothing, a cartel by any other name."

There is already a forum

The Frontier Model Forum was launched in July 2023 by Anthropic, Google, Microsoft and OpenAI, with a mandate to develop safety standards and share information with governments. It now lists Amazon, Anthropic, Google, Meta, Microsoft and OpenAI as members.

The reported talks involve three of those six. That omission is the most informative detail in the story. Microsoft spent Monday publishing its own code of conduct for its first-party MAI models and opening it to a six-week public consultation — a unilateral document, not a joint one. Meta has resisted the regulatory turn entirely. And The Information reports that OpenAI and Anthropic have separately pressed the administration for consistent federal review standards that would apply to competitors including Meta and xAI.

A safety body narrower than the safety body that already exists is not primarily an attempt to coordinate the industry. It is an attempt by the labs that consider themselves frontier to define the frontier, and to set a bar that the companies outside the room would have to clear.

What to watch

Four things would turn this from reporting into fact: a company confirming the talks on the record; the executive order moving or dying; a membership list longer than three; and, eventually, a model release actually held back because an evaluation said so. Until then, the honest description is that the largest AI companies in the United States have spent two months discussing an institution none of them has agreed to be bound by.

Sources: The Washington Post: Anthropic, OpenAI and Google discussed creating new AI safety body · PYMNTS: Google, OpenAI and Anthropic float idea of AI standards body · Quartz: Anthropic, Google and OpenAI discussed a shared AI standards body · Dario Amodei, "We Must Pace the Frontier" · CNBC: Google DeepMind chief Demis Hassabis calls for U.S. to spearhead AI standards body · TechCrunch: DeepMind CEO calls for an independent standards body to regulate frontier AI · PBS NewsHour: Trump downplays the need to check AI development · Aidan Gomez, Cohere: Who gets to define the rules for AI · Microsoft AI: Humanist AI in practice — a public consultation on our Code of Conduct for MAI Models · Frontier Model Forum · Seoul Economic Daily: Rivals unite — Anthropic, OpenAI, Google push AI standards body