The AI Standards Body Is the Ladder, Pulled Up

Three of the largest AI labs want to write the rules they will be judged by. OpenAI, Anthropic and Google DeepMind have been meeting since July about an AI standards body covering testing and auditing, according to The Information on 13 September. Europe already has one of those, it’s been enforceable since 2 August 2026, and these three are inside its scope.

The meetings came two months before the argument

Dario Amodei published his case for pacing the frontier on 12 September. The reporting landed on the 13th. The meetings had been running since July, which means the public argument arrived roughly two months after the private one was settled.

That ordering changes what the essay is. An argument published before a thing exists is something you can push back on, because the decision it describes hasn’t been taken yet and your objection might change it. An argument published after a working group has met through the summer is a launch note. I’ve sat in enough standards conversations to know the shape: whoever is in the room in July decides what the standard assumes, and everyone reading in September gets to argue about wording.

A finance-shaped regulator writes finance-shaped rules

The reported model is a self-regulatory organisation, the pattern used in finance. That matters more than it sounds.

An SRO is funded and staffed by the firms it governs, and it does produce real rules, real audits and real enforcement, which is why the model persists. It also produces rules that a large incumbent can clear and a smaller entrant cannot, because the people drafting them describe what they already do and call it a floor. No one has to plan that for it to happen.

Then there is the condition. Sam Altman told an OpenAI town hall he wants a testing and auditing organisation, and that the labs have to build it themselves, without the backing of the US government. The reason offered is speed.

Brussels already enforces one

Speed is checkable, so check it. The EU published its General-Purpose AI Code of Practice in July 2025, in three chapters covering transparency, copyright, and safety and security. Obligations for general-purpose model providers have applied since August 2025. The Commission’s enforcement powers, which include information requests, access to models and recalls, went live on 2 August 2026, a month before anyone reported these meetings.

The European version is slow in places and parts of it were written by people who have never shipped a model. It isn’t hiding, though. It was drafted in the open with a signatory taskforce, it covers copyright next to safety because those obligations cost the companies it binds, and it ends in enforcement that someone outside the industry applies.

They built an AI standards body in 2023

This wouldn’t be the first. The Frontier Model Forum has been running since 2023, founded by Anthropic, Google, Microsoft and OpenAI, and it has an executive director, a safety fund over $10 million, and a voluntary agreement covering information sharing on vulnerabilities and dangerous capabilities including CBRN and cyber.

So what does a second one do that the first doesn’t? Either the Forum was never built to write binding standards and this is the missing half, or a new body is easier to start than an existing one is to reform. The reporting doesn’t say. Microsoft is in the Forum and was not named in these meetings.

The closer precedent is the 2023 White House commitments, where fifteen companies signed eight promises on safety testing, security and public trust. A review a year later found better red-teaming and watermarking, and no meaningful transparency or accountability. The academic post-mortem is blunter: the wording was vague enough that nobody could say what satisfying a commitment required, and no mechanism was built to check.

What would make one worth having

A real AI standards body would be worth having and I’d take one gladly. Testing that happens beats testing promised in a policy paper, and three labs agreeing on a single evaluation beats three labs each grading themselves. The 2023 round failed on checking rather than on ambition, so the questions worth asking are all about who checks:

• who can join, and what it costs. A fee or compliance burden that only clears at frontier-lab scale is a barrier dressed as a bar
• whether evaluations are published in full, failures included. The summer we watched models escape their own sandboxes is exactly the material that belongs in public
• whether anyone outside the member labs can run the tests and get the same answer
• what happens when a member fails, because an audit with no consequence is a press release with a methodology section
• who pays the auditors, and whether the people being audited can fire them

None of it needs a government, and all of it can be checked from outside the membership.

What this means for everyone smaller

For anyone building on these models rather than building them, a standards body written by the three largest labs sets the compliance floor at their operating costs, and the bill for clearing it lands on everyone who has to meet it without their balance sheet. For the labs themselves it means the first published standard is the thing to read, because if it describes what the founders already do, it’s a floor drawn at their own feet.

I went looking for the reading where this is sincere and couldn’t find it. The meetings came before the argument, the model is the one where the governed write the rules, the condition is that no government stands behind it, and a binding standard already binds them in Europe.

Until one of them slows a launch to meet a rule they didn’t write, this is a ladder with a compliance function bolted on.