OpenAI, Anthropic and Google DeepMind are designing their own AI regulator, and courting the adviser who promised there'd never be one

On Sept. 15, at a Washington briefing, OpenAI chief global affairs officer Chris Lehane confirmed what the three biggest US frontier labs had reportedly been doing quietly for weeks: designing a private body to review each other's models before they ship. "It's better to try to work together to prioritize safety," Lehane said -- adding, pointedly, that OpenAI believes no antitrust waiver is needed for OpenAI, Anthropic and Google DeepMind to coordinate this way. Neither Anthropic nor Google has confirmed the talks to reporters directly; Lehane's is the only on-the-record voice so far.

Reporting has settled on a working name -- the Frontier AI Standards Agency -- for a body with a specific job description: industry-funded, federally overseen, and modeled on FINRA, the Financial Industry Regulatory Authority that polices Wall Street brokers. Member labs would submit frontier models for third-party review up to 30 days before release, accept incident-reporting rules, publish model cards, fund safety research, and agree on who counts as a qualified auditor. The pitch is a body that can move at the speed AI ships, something no act of Congress has managed in three years of trying.

None of this is brand-new. It is five weeks old with a name attached. Google DeepMind chief Demis Hassabis posted the original FINRA-for-AI proposal in mid-July: funding from the labs themselves, a board that includes independent technical experts and open-source-community representatives, capability-assessment protocols, and a path from voluntary to mandatory submission. Reaction split immediately and has stayed split. Microsoft's Satya Nadella endorsed it outright; OpenAI's Sam Altman called it "thoughtful"; even Elon Musk allowed it was "a thoughtful framework overall and certainly a good starting point."

The skeptics who spoke up in July were arguing from inside the safety community, not against it. Turing Award winner Yoshua Bengio warned the voluntary-to-mandatory sequencing was backwards: "we absolutely need a clear and precise roadmap to transition from a voluntary to a mandatory model," not a promise to get there eventually. Governance researcher Nader Henein went further: "self-regulation is not viable," full stop, citing both vendor capacity to actually staff serious reviews and the built-in conflict of a lab paying for its own audit.

The man they want to run it said no FDA, ever

The leadership search is where the story gets pointed. Per Sept. 24 reporting, the labs have approached Sriram Krishnan -- the Trump administration's senior AI policy adviser until he departed at the end of June -- to serve as the new body's chief executive. Other names reportedly in the mix: former Biden-era technology official Arati Prabhakar, ex-diplomat Condoleezza Rice, and venture capitalist David Friedberg. Krishnan is the one whose record actually says something about what kind of body this would be.

“There will not be an FDA for AI.” — Sriram Krishnan, in a July Financial Times interview, weeks before the labs he'd overseen as a White House adviser approached him to run the body they're building to review AI models

That line is from a Financial Times interview Krishnan gave in July, shortly after leaving the White House, explaining why the Trump administration would not create a formal AI licensing regime. Requiring a company to clear a team of lawyers before shipping a model, he said, would put "sand in the gears" of the AI revolution. Handing the CEO job to the industry's most prominent opponent of centralized AI oversight is either, as one outlet's own analysis put it, "the cleanest way to head off statutory rules" -- or a legitimacy problem the labs are choosing not to see.

(FINRA itself is not a government agency. It's a private, non-profit body that Wall Street brokers are legally required to join, funded by membership fees and fines, with the SEC as its overseer of last resort -- the model the labs are explicitly copying.) The timing is also pointed. The approach to Krishnan landed two days after France convened a rare UN Security Council session on AI safety, where Sam Altman and Dario Amodei renewed a call for international, government-led AI-pacing coordination -- the same pledge four subscribers are currently suing four labs over. A private, US-only standards body is a narrower and faster-moving answer than anything a UN session produces, and it arrived while that slower answer was still being pitched on the record.

Why the skeptics don't agree on much except the diagnosis

David Sacks, the White House's own AI adviser, has called the labs' self-regulatory instincts "regulatory capture" or an election-season distraction more than once this year -- most pointedly on Sept. 13, when he accused Sam Altman and Dario Amodei of using a separate pacing pledge to argue for exactly the kind of antitrust carve-out Lehane now says isn't needed. Sacks's objection comes from the right: let the market discipline unreliable AI systems, not a body the labs fund themselves.

Cohere chief executive Aidan Gomez -- who runs a frontier lab that is not one of the three building this body -- has made a related but distinct historical argument to Forkast: self-regulatory organizations have a track record of licensing rapid growth rather than braking it. He points to the SEC's 1975 designation of Nationally Recognized Statistical Rating Organizations, which entrenched the same three credit-rating agencies for decades, and the EU's 1985 Motor Vehicle Block Exemption, which let carmakers write their own distribution rules for twenty years. Both, in his telling, "lowered political friction while entrenching existing players" rather than disciplining them.

A Yale Journal on Regulation analysis by legal scholar going by the byline Frazier makes a narrower but harder-edged claim: this specific design may not survive a lawsuit. The private-nondelegation doctrine, running back to the 1935 Supreme Court case *Schechter Poultry*, holds that "Congress cannot escape constitutional restraints by delegating government authority to private parties to accomplish indirectly what Congress cannot do directly." Later cases require a private regulator to "function subordinately to" a government overseer with real "authority over and surveillance of their work" -- and Frazier's read of nine decades of securities-regulation history is that satisfying that requirement is exactly what would slow the body down to normal regulatory speed, defeating the entire point of building it. His conclusion: "Congress can delegate, but it cannot delegate and walk away."

The Council on Foreign Relations takes the proposal more seriously as a starting point -- its authors broadly support a FINRA-style body -- while warning that specific design choices will decide whether it earns trust or has to be rebuilt after losing it. Their sharpest flags: an issuer-pays funding model was exactly what undercut credit-rating agencies' credibility before the 2008 financial crisis, and a body holding pre-release access to unreleased frontier models across every founding lab would become, in their words, "one of the world's highest-value espionage targets" -- a risk FINRA itself never had to manage. Their fix is structural: separate who sets safety specifications, who funds infrastructure, and who actually evaluates models into three different institutions, rather than folding all three jobs into one industry-funded agency.

None of this means the agency won't get built. FINRA itself proves an industry-funded self-regulator can last -- it has policed Wall Street for fifty years, badly in 2008 and adequately most other times. What none of this week's reporting settles is which FINRA the labs are actually building: the one that polices its own members credibly, or the one whose 2008 near-failure is the CFR piece's own cautionary tale. That answer depends on funding structure, on whether Krishnan or someone with a different record ends up running it, and on whether Congress's own FRONTIER Act -- which Lehane says OpenAI already supports in part -- reaches a floor vote first.

For a policy beat that spent most of 2026 watching state legislatures and Congress try and mostly fail to write binding AI rules, this is the first serious sign that the labs themselves have concluded Washington won't finish the job in time -- and have decided to write their own rulebook rather than wait. Whether that rulebook restrains them is now a design question, not a hypothetical one, and design questions have deadlines: a funding structure, a charter, and a CEO's answer are all things a specific future event can settle, not just argue about.