14 SEP 2026 — Anthropic, OpenAI and Google have been meeting since July to discuss an industry standards body for artificial intelligence, according to reporting by The Information on 13 September. The proposal on the table would audit and test frontier models, with shared protocols for evaluation before release.
The model the idea keeps returning to is FINRA, the body that polices America's securities brokers. That comparison is the right place to start, because FINRA has something this proposal so far does not: a regulator standing behind it.
What is reportedly being discussed
Representatives of the three laboratories have held working-group meetings regularly since July. The talks reportedly centre on auditing and testing frontier models through independent evaluations, pre-release safety reviews and standardised risk assessments. The standards would be voluntary.
The intellectual lineage is public. In a July essay, Google DeepMind's chief executive, Demis Hassabis, proposed a self-regulatory body for AI modelled on the Financial Industry Regulatory Authority. Sam Altman has since argued that the major laboratories will have to create a standards body themselves, and has said he believes they should do so without the backing of the United States government. Dario Amodei set out his own case the same week for slowing the pace at which model capabilities improve.
Government involvement is where the reporting gets murky. A draft executive order on the subject was reportedly prepared at the White House and ran into internal disagreement, with David Sacks, co-chair of the President's Council of Advisors on Science and Technology, opposing a government-led body. The talks between the laboratories appear to be moving in the space that dispute left open.
FINRA is not a club
The FINRA comparison flatters the proposal. FINRA is a self-regulatory organisation in the sense that the industry funds and staffs it, but it exists under federal securities law, operates under Securities and Exchange Commission oversight, and has teeth that come from that arrangement. Brokers must register with it. It fines them, suspends them and expels them, and an expelled firm cannot carry on the business.
Take away the statute and the oversight, and what remains is a trade association that publishes guidance. Those have their uses. They are not regulators, and nobody treats their standards as binding on a member who decides the commercial case for ignoring them is strong enough.
Test design is the easier half, since laboratories already publish system cards, run pre-release evaluations and describe capability thresholds. The harder question for the talks is what happens to a member whose model fails the shared test and ships anyway. If the answer is a published finding and nothing else, the body adds a logo to practices that already exist.
"A great idea from the cartel"
Not everyone invited to applaud has done so. Meta and Cohere have been reported as sceptical, and Cohere's chief executive, Aidan Gomez, dismissed the plan as "a great idea from the cartel".
The jibe lands on a real problem. Standards written by the three largest frontier developers will naturally encode what those three can afford. Evaluation regimes cost money and compute, and a testing protocol calibrated to the largest models can become a compliance cost that a smaller laboratory or an open-weights developer simply cannot carry. Incumbents setting the entry requirements for their own market is the textbook case regulators are supposed to be suspicious of, and it is not an accusation the three companies can wave away by pointing to good intentions.
Against that, frontier capabilities are concentrated in very few organisations, and those organisations employ the people able to design meaningful tests. Waiting for a government body that may never arrive leaves nothing in place at all, and a flawed industry standard may beat no standard.
Why it matters beyond the United States
Nothing here is a government process, and none of the governments that have written AI rules of their own is a party. Singapore's approach has leaned on voluntary testing frameworks and published evaluation toolkits rather than statute, which makes an industry-led testing standard look more familiar here than it does in Brussels. It also means an American industry body could end up setting de facto testing norms for models that are deployed across Asia by organisations that had no say in writing them.
That would not be new. Industry standards bodies set most of the technical rules the internet runs on. But those bodies are open to anyone willing to do the work, and nothing reported so far suggests this one would be.
The questions still open
Who gets in will define the thing. A body limited to three laboratories is a coordination arrangement, while one that admits open-weights developers, cloud providers and independent evaluators on equal terms would be a standards organisation.
Enforcement decides everything else. Look for any mechanism by which a member loses something for shipping a model the shared tests flagged.
The government question has not gone away either. A statutory backstop would make the body something like FINRA, and without one the comparison should be retired.
View author profile → · Editorial policy
