WHAT WAS CONFIRMED
● Who: OpenAI, Anthropic and Google DeepMind, per OpenAI policy chief Chris Lehane.
● What: a self-regulatory standards body modelled on FINRA, testing powerful systems before release.
● Origin: a proposal Demis Hassabis first made in July.
● The response: Cohere's chief executive called it a cartel.
This is what the antitrust question was about
Six days ago we reported that OpenAI had asked members of Congress whether an industry-wide slowdown in developing the most capable systems could run afoul of antitrust law.
At the time that read as a sensible legal question following Amodei's pacing essay. It now reads as due diligence on something already being built.
THE OBJECTION IS THE OBVIOUS ONE, AND IT IS NOT WRONG
Three of the largest firms in a market agreeing on what gets tested and when it may ship is the shape of a cartel, whatever the motive. That is the structural criticism and it does not require anyone to act in bad faith.
It also matters who is not in the room. Cohere is not one of the three. Neither is Mistral, Z.ai, Moonshot, xAI or any open-weight lab. A body that sets release standards without them sets standards for them.
Why FINRA specifically
The Financial Industry Regulatory Authority is a self-regulatory organisation — the industry runs it, but it operates under government oversight and its rules are enforceable. That is a real model rather than a fig leaf, and it exists because securities regulation moved faster with industry expertise inside it than outside.
| What FINRA has |
Does the AI version? |
| Statutory authority | No. Congress has passed nothing |
| Government oversight | Not established |
| Membership across the industry | Three firms so far |
| Enforcement powers | Undefined |
| Funding independent of members | Hassabis has said funding is unresolved |
Without the first two rows, what remains is three competitors coordinating voluntarily — which is precisely the arrangement OpenAI asked Congress about, and precisely what Cohere objected to.
The mechanism problem, resolved unhelpfully
There were three ways an AI slowdown could happen. Voluntary coordination resembles a cartel. Unilateral restraint loses. Regulation works but Congress has passed nothing comprehensive in three years.
The labs have now chosen option one and a competitor has named it. That does not make them wrong to try — someone has to move first, and waiting for legislation has produced nothing. But it does mean the criticism was predictable, and the antitrust exposure is real rather than theoretical.
The three mechanisms, and why two of them cannot work →
What to watch
- Whether smaller labs are invited. A standards body of three is a very different thing from one with fifteen.
- Whether Congress responds. Statutory backing turns this from coordination into regulation, which is the only version that survives the objection.
- Who funds it. Hassabis has said funding is unresolved. A body funded by the firms it tests has an obvious problem.
- What "test before release" means in practice. Pre-release evaluation is either meaningful or a stamp, and the difference is publication rights.
Sources
FAQ
What is being proposed?
A self-regulatory standards body modelled on FINRA, developed by OpenAI, Anthropic and Google DeepMind, to test powerful AI systems before release. Demis Hassabis first proposed it in July.
Why did Cohere call it a cartel?
Three of the largest firms in a market agreeing on what gets tested and when it may ship resembles a cartel structurally, whatever the motive — particularly when competitors are not included.
Is it the same as FINRA?
Not yet. FINRA has statutory authority, government oversight and industry-wide membership. This has three firms, no legislation behind it, and unresolved funding.
Does this affect me?
Not today. If it gains authority it would affect when models ship and what testing they undergo first.
What would fix the objection?
Statutory backing, independent funding, and membership that includes smaller and open-weight labs. Without those, it remains voluntary coordination between competitors.