AI labs design their own safety referee before any law arrives

The membership list has not been confirmed. Neither has the name, governance structure, funding model nor timeline. What is confirmed is that Anthropic, Google and OpenAI are discussing a joint body to set AI safety standards, and that the discussions are happening now, before any legislature has passed a binding law requiring it.^[1]

That sequencing is the detail buyers of AI tools should sit with. The labs are designing their own referee, and whether that referee carries any real power over them is a question none of the reported proposals have yet answered.

What is being proposed

Google DeepMind's Demis Hassabis put the clearest model on paper in a July essay. Hassabis called for a body modelled on a federally overseen public-private partnership or self-regulatory organisation, much like the Financial Industry Regulatory Authority (FINRA), with a board that includes independent leading technical experts and open-source representatives.^[2] The FINRA comparison is worth examining: FINRA regulates broker-dealers in the United States, funded by the industry it oversees, operating under government supervision but outside direct government control.

OpenAI's Sam Altman described his version a day later as "a US-led international forum that establishes accepted standards, provides expert and impartial analysis of capabilities and risks, and makes the technology available to nations and companies that participate and follow the rules."^[3] Both proposals share an architecture: voluntary participation, technical expert involvement, and a US anchor. Neither proposal, as publicly described, names an independent enforcement body with power to halt a model release.

The existing regulatory floor is thin. Executive Order 14409, signed by President Trump on 2 June 2026, directs a voluntary framework under which AI developers may provide the federal government up to 30 days of pre-release access to frontier models for cybersecurity and national security review.^[4] The proposed industry body would sit alongside that framework rather than replace it, adding another voluntary layer on top of an already voluntary one.

The week that made this urgent

The talks surfaced during a seven-day stretch that handed anyone watching the labs a concentrated dose of what self-regulation is meant to address.

On 9 September 2026, Jacob Coxon, who had worked as a pretraining researcher at both OpenAI and Anthropic, posted a public resignation thread. "Neither company is acting responsibly," Coxon said. "They are racing straight to self-improving superintelligence and gambling with our lives."^[5] It was a direct challenge to the safety culture both companies present publicly.

The following day, Anthropic published its September 2026 Threat Intelligence report. The report documented seven active harm categories, covering cases the company disrupted between December 2025 and August 2026: cyber operations, surveillance, influence, biological misuse, weapons development, distillation and scams.^[6] Nine months of documented, real-world misuse across seven categories is an operational record, not a theoretical risk paper.

The resignation and the threat report arrived on consecutive days, placing the safety standards conversation in a very different register from the policy essays published in July.

What buyers of AI tools should weigh

For marketing and media teams that have spent the past two years evaluating vendor safety claims, an industry standards body could eventually give those claims a common vocabulary and a shared baseline to test against. That would matter in a practical way.

An industry-written rulebook with no confirmed independent enforcement gives a buyer roughly the same assurance they already have: the word of the vendor. The FINRA analogy Hassabis reached for is instructive on that point. FINRA was created by an act of Congress, operates under Securities and Exchange Commission oversight, and carries real disciplinary power including fines and licence revocations. None of those features have been confirmed for the body under discussion.

Until they are, the proposed standards body remains a statement of intent from the same organisations whose safety practices a researcher publicly questioned on 9 September 2026. The next concrete marker to watch for is whether any government, US or otherwise, formally endorses or mandates participation. EO 14409's 30-day review window, signed on 2 June 2026, is the most concrete regulatory commitment currently in place.^[4]