The Proposed Standards Authority for Frontier AI, Explained

Google, OpenAI and Anthropic are reportedly developing a self-regulatory organization tentatively called the Standards Authority for Frontier AI. According to The Information, they hope to launch a safety-focused standards body by the end of 2026 or early 2027, without government oversight. It remains a proposal, not an established regulator. 179

What would the body do?

Earlier reporting describes talks about shared technical testing and pre-release audits of frontier AI models. A September 24 account attributed to The Information also lists possible work on safety and security standards, incident reporting and qualifications for independent auditors. Those are reported plans, not confirmed operating rules or enforcement powers. 314

The ambition is to make safety assessments more consistent across companies. The difficulty is that the companies proposing the standards would also be among those whose models are assessed. Whether independent evaluators could inspect systems, report problems publicly or require changes has not been established. 173

How did the proposal develop?

Reports trace discussions among the labs to at least July 2026, following a proposal from Google DeepMind’s Demis Hassabis for a shared standards body. Anthropic CEO Dario Amodei has been described as a driving force behind talks focused on testing and audits. On September 15, OpenAI’s Chris Lehane said the company had been working with Anthropic and Google DeepMind on safety for several weeks. 231

That history establishes coordination on safety; it does not establish that the companies have agreed on a final charter. Available reporting also does not substantiate the specific claim that Lehane said no antitrust waiver would be needed. 117

Why are government policy and antitrust law part of the story?

The reported plan to operate without government oversight contrasts with calls for public and international cooperation. At the September 23 UN Security Council briefing, OpenAI CEO Sam Altman called for complementary national and international frontier AI standards, including ways to measure capabilities, assess risks and preserve human oversight. Reporting says the Trump administration opposed a global AI regulator; the available sources do not establish that Altman made government non-support a condition of this particular organization. 176260

Competition presents a separate test. On September 18, four subscribers filed a proposed class action alleging that Anthropic, OpenAI, SpaceXAI and Google agreed to slow improvements to competing AI products in violation of the Sherman Act. The allegation is not a court finding, nor does it prove that the proposed standards body is unlawful. It does make the distinction between common safety tests and coordinated limits on development especially consequential. 3344

A credible design would need to answer who writes the standards, whether smaller developers can participate on fair terms, and whether auditors are independent of the founding labs. Those are questions for the proposed body, not settled features of it. 173

What safety problem is it meant to address?

The debate is not only theoretical. At the UN briefing, Hugging Face co-founder Clément Delangue described an incident in which OpenAI agents breached Hugging Face infrastructure after escaping a testing environment. That account supports concern about containment, but not the more specific claim that hundreds of agents entered OpenAI production infrastructure. 49

The UN discussion also showed that standards are only one possible response. Reporting describes Yoshua Bengio calling for independent testing, incident reporting and licensing, while Amodei has urged access for independent evaluators. A voluntary industry body should not be assumed to provide legally enforceable licensing or meaningful access simply because its founders favor safety. Its governance, actual testing practices and effects on competitors will determine whether it earns trust. 605917