Three AI Labs Are Quietly Writing Their Own Rulebook

Three Rivals Are Negotiating Their Own Referee

Anthropic, Google and OpenAI have held working-group talks since July on creating a shared, industry-led standards body to test and audit advanced AI models, The Information reported on September 13, 2026, with a fourth round of meetings as recently as last week. The catalyst was a July essay by Google DeepMind co-founder Demis Hassabis, who proposed a body modeled on FINRA, the US authority that polices Wall Street brokerages under SEC oversight: funded by the industry, staffed by independent technical experts, evaluating advanced models before release.

Anthropic CEO Dario Amodei folded the idea into his own September 12 blog post calling for a slowdown in frontier development, and it also won public backing from Sam Altman and Elon Musk. Altman was explicit about the shape he wants, telling a company town hall that the major labs should build a testing and auditing organization on their own, "without the backing of the US government."

The EU Already Built This, Just Not For The Labs To Run It

The European Union has had a comparable structure in force since August 2, 2025: the General Purpose AI Code of Practice, which binds providers of frontier models, Anthropic, Google and OpenAI included, to publish detailed training data summaries, assess systemic risk and cooperate with independent testing. The European Commission approved it only after input from nearly 1,000 stakeholders, industry, academics, civil society and EU member states among them, not the three companies subject to it deciding the standard themselves.

That is the structural difference a FINRA comparison glosses over. FINRA itself is industry-funded, but it operates under statutory SEC oversight with legal enforcement powers a private AI standards club, formed by the very labs it would grade, does not have. A body the regulated companies build, staff and fund, with no government backstop, sets its own bar for what counts as a passing grade.

A Rulebook Written By The Companies It Grades Is Also A Moat

The three labs racing hardest to build frontier systems are also the three drafting the rulebook that would define what a safe one looks like, a rulebook Meta, xAI and every smaller developer had no seat at the table to shape. A voluntary body that only its three founders can point to as proof of responsible testing quietly becomes a competitive filter: it sets an entry bar for credibility that rivals must meet on terms Anthropic, Google and OpenAI wrote.

The timing explains the urgency. President Trump has publicly dismissed calls for AI guardrails the same week, leaving little prospect of near-term federal testing legislation, so a founder-run body fills the vacuum before anyone else can. Read against OpenAI's separate call for the UK to pass binding, government-written capability rules, the pattern is consistent rather than contradictory: these companies accept government rules where they can help write them and remain a rule-taker only when there is no alternative, and build their own club everywhere else.

Read next: A Security Badge Told The Pentagon Nothing About What Grok Would Generate  |  Your AI Vendor Just Got a Letter From Brussels