AI Self-Regulation: The Giants Are Building Their Own AI Watchdog

There is an old question that every society eventually has to ask about power: who watches the watchmen? This week it acquired a very modern answer. According to The Information, the three most powerful AI labs on earth — Google, OpenAI, and Anthropic — are building their own regulator. A voluntary industry body, provisionally called the Standards Authority for Frontier AI, designed to set and check the safety standards for exactly the kind of models these same three companies build and sell. They are, in other words, proposing to be both the watched and the watchmen. Whether that is a genuine step toward safety or the most sophisticated way yet of avoiding it is the question worth sitting with.

What Was Announced

The plan, still taking shape and not yet formally launched, is for an independent standards body — reported under working names including the Standards Authority for Frontier AI, or SAFA — that would operate without government oversight and aim to launch by late 2026 or early 2027. The idea was initiated by Google DeepMind's Demis Hassabis in July, and the three labs initially sought a public-private partnership under federal supervision. When a White House executive order on the subject stalled, they pivoted to pure industry self-regulation. The body would translate the companies' safety pledges into concrete technical benchmarks, support third-party pre-deployment testing, set incident-reporting rules and auditor-qualification standards, and — potentially — conduct model evaluations itself rather than merely publishing guidelines for others. OpenAI's chief global affairs officer confirmed the labs had been coordinating for weeks, said no antitrust waiver was needed, and reiterated the company would proceed with or without government support.

A Regulator Modeled on Wall Street — Without Wall Street's Teeth

The labs have reportedly modeled the body on FINRA, the self-regulatory organization that polices Wall Street brokers. It is a revealing comparison, and not in the way its designers might hope. FINRA works because it operates under the supervision of a government agency, the SEC, and because it can impose real penalties: it can fine firms, suspend them, and bar them from the market entirely. The proposed AI body, as described, has none of these powers. It would be voluntary. Its members would be the very companies it evaluates. And there is no government backstop giving its findings force. A self-regulator with the structure of FINRA but none of its enforcement authority is not a watchdog; it is a discussion forum with a badge. The single most important question about this body is the one its FINRA branding is designed to obscure: what happens when a member fails a standard and wants to ship anyway? On the evidence so far, the answer is nothing.

The Man They Want to Lead It

The choice of proposed chief executive sharpens the point rather than softening it. The labs have approached Sriram Krishnan, who until June 2026 served as the top AI policy adviser in the Trump White House — and who left that role having argued forcefully against government regulation of AI. On his way out he said, memorably, that there would be "no FDA for AI," warning that a central agency requiring approval before a model could ship would put, in his phrase, sand in the gears of the AI revolution. There is a certain logic to it, from the labs' perspective: who better to run your voluntary alternative to government regulation than the man who spent eighteen months arguing government regulation should not exist? But for anyone hoping this body represents real oversight, the signal is hard to miss. The person they want in charge of the watchdog is on record saying the watchdog should have no leash.

The Ghost of the Frontier Model Forum

We have, in fairness, seen almost exactly this before, and that history is the fairest test of the new promise. In 2023, Google, Microsoft, OpenAI, and Anthropic launched the Frontier Model Forum, an industry body with a ten-million-dollar fund and a mandate to advance responsible AI. In the years since, it has published best practices and useful research. It has also never stopped a single model release, never required a single evaluation, and never enforced a single standard, because it was never built to. The new body is being positioned as structurally different — more formal, with real auditor certifications and pre-release requirements. Perhaps it will be. But the burden of proof sits squarely on the labs, because the last time they promised to police themselves, the policing turned out to be optional, and optional policing is what you build when you want the appearance of oversight without the reality.

Claim and Counter-Claim

The case for the body is real and deserves a fair hearing. Government AI regulation in the United States has genuinely stalled; the federal capacity to evaluate frontier models barely exists; and in that vacuum, a serious industry effort to build shared testing standards, incident reporting, and pre-deployment audits is better than the nothing that would otherwise fill the gap. If SAFA actually conducts independent cross-company evaluations against common criteria — something no purely voluntary body has done before — it could raise the floor for everyone and move faster than any government could. Self-regulation is not automatically theater; sometimes it is the only thing available in time.

The case against is that self-regulation by the dominant players tends, structurally, toward two outcomes, and neither is safety. The first is toothlessness — standards that flex whenever they collide with a launch date, as the Frontier Model Forum's did. The second is worse: a body that sets the rules high enough to burden smaller competitors and new entrants, while its founding members, who wrote the rules, clear them comfortably. That is why a rival AI chief executive, Cohere's Aidan Gomez, reportedly called the proposal a cartel by another name. A regulator built by the three largest firms, run by an opponent of regulation, with no power to punish its own members, is a structure that protects incumbents at least as plausibly as it protects the public. The honest synthesis: this could raise real safety standards, and it is designed in precisely the way that would also let the biggest labs control those standards, exclude challengers, and pre-empt the harder, independent oversight that might otherwise come. Both can be true, and which one wins depends entirely on powers this body, as described, does not yet have.