OpenAI's head of public policy confirmed that the three largest labs have been working for several weeks on a body tasked with testing models before deployment, modelled on the self-regulation of the US financial sector. Cohere's chief executive called the project a cartel. Both positions deserve to be taken seriously.
On Wednesday we reported the existence of these meetings, noting that an industry-led body structurally cannot provide independent verification. Here is the confirmation, and the first public objection.
The model invoked, and what it implies
The reference to US financial self-regulation is precise. It is a private body, funded by companies in the sector, that sets rules, monitors their enforcement and can impose sanctions — all under the supervision of a public regulator.
That last point is essential and often forgotten. This model does not work because industry regulates itself: it works because a public authority can overturn its decisions, sanction it, and withdraw its delegation.
An equivalent body for AI today has no public regulator above it. That is the difference between a delegation and a substitution.
The cartel argument
It needs to be stated precisely, because the word is strong and it refers to something concrete.
Three dominant companies that jointly define the tests a model must pass before deployment are, in effect, setting a cost of entry. That cost is bearable for the one who designed and budgeted for it. It is far less so for a smaller player.
This mechanism requires no bad intent. Requirements sincerely intended to reduce a risk produce the same effect as a deliberate barrier, and that is precisely what makes it hard to challenge.
The objection comes, moreover, from a player in a position to bear it, which does not disqualify it — as we wrote about the debate on the pace: someone with an interest can still be right.
The difference does not lie in intentions but in three observable elements. Who can join: a body open to any player, including a small one, is not an agreement. Who writes the rules: if only the three largest decide, governance is captured whatever the formal openness. Who can challenge: a refusal decision must be able to be brought before someone outside. None of these three points is established at this stage, and that is what will need watching.
The opposite argument, which also holds
It would be dishonest to present only one side.
Without a common standard, each lab tests whatever it wants, however it wants, and publishes whatever it decides. We spent the summer documenting this situation: voluntary post-mortems, figures published by those who produce them, no comparability.
A body that defined shared tests would make evaluations comparable between labs. That is exactly what is missing, and nobody else is building it.
And a shared standard lets fixes and best practices circulate, as we wrote about the standardisation of protocols.
What would be missing to decide
Two elements would make this project defensible, and they cost nothing to grant.
Open governance. Participation accessible to smaller players, public labs and independent researchers, with a weight that is not proportional to revenue.
Public oversight. That is what makes the model being invoked work. Without an authority able to revise a decision, you do not have a delegation of regulation: you have a club.
Note that the installation of independent evaluators announced yesterday goes in the right direction on a neighbouring point. The two mechanisms could complement each other — shared tests, verified by third parties — or substitute for one another, in which case the second would have served to avoid the first.
What to take away
This project is probably necessary and potentially problematic, and both things are true at the same time.
The test will be the one we proposed on Wednesday: does a body accept that an independent third party audits its enforcement? That is what distinguishes a standard from a charter, and a delegation from a capture.
And we will need to look at who sits on it. A standards body where the three largest decide and the others watch has a name, and it is not that of a standard.