Anthropic, OpenAI, and Google Discuss an Industry-Led AI Standards Body

The meetings were closed, but one blueprint from the same month is public. Demis Hassabis, CEO of Google DeepMind, personally unveiled his concept for a Frontier AI standards body on July 14. His model was the self-regulatory organization of the U.S. securities industry: funded by the industry but operating independently under federal oversight, standing as a self-regulatory body rather than a government agency.

The backbone of the proposal, which METAL reviewed, is pre-release review. It's a three-stage design: companies would submit models to the body 30 days before launch, starting voluntarily and then hardening into a requirement for entering the U.S. market once the approach proves itself. In the final stage, the body would develop its own test questions that it never discloses externally, to stop companies from tuning their models to benchmarks they already know.

The proposal even spells out what would be tested. Cybersecurity vulnerabilities, potential biological threats, and autonomous agentic behavior were named as national-security-relevant evaluation areas. Hassabis also proposed testing whether models try to circumvent safeguards or show deceptive tendencies, watermarking AI-generated output, and requiring interpretable reasoning. Which models clear this bar would be decided by capability thresholds the body manages, not by which company built them.

Hassabis himself drew the line on the body's scope. He wrote that the standards body would be "responsible for developing evaluation protocols and partnering with appropriate federal agencies and U.S. national laboratories to conduct testing in national-security-relevant domains." That makes it a technical-standards-setting organization rather than a regulator, leaving final judgment authority with the government. He also stipulated that the board should include independent technical experts and representatives from the open-source community.

He also answered the question of why now. Hassabis diagnosed the moment this way: "We are currently trapped in an extremely intense, multi-layered commercial and geopolitical race." It's a scene of racers trapped in the race coming together to create a referee — except the racers pay the referee's salary. Hassabis didn't dodge that point either, writing that funding would need to be substantial to attract technical talent and support large-scale testing.

Over the weekend, this idea took a step out of the closed meeting room. Microsoft CEO Satya Nadella said on September 13 that he welcomes deliberate pacing to get alignment right, and said he would publish a code of conduct for Microsoft's own MAI models the next day for public comment. He wrote that "any pursuit of superintelligence must be rooted in the core principle that AI we build should not be pursued unless it helps humanity and remains under human control." He also included a line saying governance should not be controlled by a small number of actors and should have broad representation spanning the ecosystem, governments, and academia.

Where the three companies' discussion sits in the broader regulatory debate becomes clear only when placed alongside other events this week. METAL reported that Anthropic CEO Dario Amodei published an essay calling for slowing the pace of frontier development and bringing outside evaluators inside the company. That's exactly the deliberate pacing Nadella welcomed. Public statements and closed-door working meetings are pointing in the same direction.

How far voluntary compliance can hold has already been tested once. METAL reported that Anthropic did not submit Claude Mythos 5.1, released on September 1, to the UK AI Safety Institute's pre-release evaluation. As long as handing a model to an evaluator stays voluntary, the option not to hand it over always exists alongside it. That's why Hassabis set the final stage of his three-stage design as mandatory.

There's also the unresolved problem that slowing down together may not be legally safe. METAL reported that OpenAI asked Congress to rule on whether jointly slowing development with competitors amounts to collusion. A body that sets safety standards jointly sits right at the center of that question, because the moment competitors agree on the same standard, that standard becomes a barrier to entry for any other company trying to enter the market.

That's exactly the risk experts have flagged. Cascade Insights co-founder Sean Campbell warned that such a body could drift into regulatory capture, and Suzu Labs CEO Michael Bell questioned whether creating yet another bureaucratic organization is really the best answer. Michelle Lopez Maldonado of the Information Technology and Innovation Foundation recommended building up the capacity of existing standards bodies instead of creating a new one. The suspicion that arises whenever the players build the referee never really changes.

What's clear right now is the sequence. Congress has said it won't step in, the White House has said there's no reason to slow down, and companies have been filling that vacuum among themselves since July. Hassabis's blueprint is public, Nadella's code of conduct has gone out for comment, and the three companies' working-group meetings haven't produced results yet. The rules of the AI industry are being written in meeting rooms right now, not in legislatures.