Washington stalled, so the labs are drawing their own maps

It took about 24 hours for two frontier labs to quit the same status quo in opposite directions.

On Tuesday, Google DeepMind CEO Demis Hassabis published a proposal, reported first by Axios’ Mike Allen, Zachary Basu and Madison Mills, calling for a U.S.-led global body modeled on Wall Street’s FINRA to screen frontier AI models before they ship: industry-funded, majority-independent board, operational “before year-end.”

On Wednesday, a POLITICO story republished by Business Insider under Hallam Bullock’s byline detailed Anthropic’s answer to the same problem: back state AI-safety bills one legislature at a time, each tougher than the last, because, in the words of Cesar Fernandez, the company’s head of U.S. state and local government relations, a government response to the risks of advanced AI models “can’t wait for action in Washington.”

Same diagnosis, two prescriptions pointing opposite ways: one national referee, or fifty state floors. Read each against its author’s competitive position and a pattern surfaces. Each is the flavor of regulation its author is best built to survive. Our read: rules are coming either way. The live fight is over which lab’s preferred venue gets to write them.

A referee paid by the players

The mechanism in Hassabis’s plan is specific. “Initially, Frontier Labs would voluntarily share models with the Standards Body for review up to 30 days before release,” he wrote, per TechCrunch’s Russell Brandom. The body would test for cyber, biological and deception capabilities, outsourcing evaluations to specialized AI safety organizations, and once the process proved itself, review would become mandatory for U.S. market approval. Axios reports the scope would cover all frontier-class models regardless of origin or openness, with a board salted with Turing Award winners. Hassabis called current AI-driven cyber threats “warning shots.”

“The strength of this approach is it would be technically focused, while at the same time supporting innovation and incentivising responsible behaviour,” Hassabis argued.

“Technically focused” is the phrase doing the work. A single gate staffed by eval experts advantages the labs with the deepest evaluation infrastructure and the scale to absorb a 30-day pre-release freeze; a startup racing to ship cannot idle a flagship model for a month as comfortably as Google can. One federal gate is also far cheaper than fifty compliance regimes for one of the most globally distributed AI companies on earth. And a body whose writ covers open-weight and foreign models points the same gun at the rivals Google can’t out-lobby, only out-test.

FINRA’s other feature, the one that makes it beloved as a template: the regulated pay for the regulator.

Fifty floors, each one higher

Anthropic’s map runs through statehouses. Per the POLITICO story, the company endorsed California’s 2025 frontier-model law, alone among the leading labs. It backed an Illinois law, signed this month by Gov. JB Pritzker, requiring annual independent third-party audits of safety plans. And in late June it endorsed regulations under development for a Massachusetts economic development bond bill (the nation’s strongest state AI-safety proposal, by Anthropic’s own description) that would put independent evaluators on catastrophic risks, bioweapons included, with enforcement by the state attorney general. New York’s bill moved the other way: OpenAI lobbyists pressed Gov. Kathy Hochul to amend it to more closely resemble California’s rules. Fernandez said the company is “looking for legislation that meaningfully raises the bar on safety for the most capable AI systems.”

“Each one of those bills was stronger than the previous bill, and the bills all moved real safety obligations forward.”

That’s Fernandez describing the strategy, and it is a ratchet by design. Notice where the floor keeps rising: toward wherever Anthropic already stands. The company’s own policy framework notes that publishing safety practices is “already required by California and New York law” and boasts that its proposal “goes further” with regular risk reports and independent evaluators. Laws mandating published safety frameworks and third-party assessments cost roughly nothing for the lab that already publishes them. For everyone else, they’re a new bill.

This route also runs directly against the industry’s checkbook. Leading the Future, the super PAC reportedly seeded with $25 million each from Marc Andreessen and Ben Horowitz and backed by OpenAI president Greg Brockman, is spending toward a single national AI framework and against the state-by-state patchwork, as The Next Web’s Alina Maria Stan reported; the newer Innovation Council Action has roughly $100 million pledged toward the same end. The money follows a Senate that stripped federal preemption language 99-1, and more than 1,000 AI bills introduced in the states in 2025 alone. Anthropic’s $20 million, per the same reporting, went to educating the public on AI policy rather than campaigns.

The obvious objection

The counter-read is that these are complementary hedges, not opposite bets. Anthropic says plainly that it wants Congress to act; Hassabis’s body starts voluntary. Neither burns the other road.

True as far as it goes. But venue selection is revealed preference. Google could endorse the Massachusetts bill tomorrow; there’s no sign it has. Anthropic could call for one industry-funded federal gatekeeper; instead it set a preemption price — “We do not believe Congress should preempt state law unless it enacts a federal law that is at least as strong as the framework we are proposing today” — that a Senate which stripped preemption language 99-1 shows no sign of being able to pay. Complementary in theory. In practice, each lab is pouring the foundation of the arena where it holds home advantage, and “regulate us, but our way” is the through-line.

One more clock worth reading. FINRA polices brokers with brokers’ money, which is exactly what makes it a comfortable template for industry. But its lineage runs back to the National Association of Securities Dealers, registered with the SEC in 1939; the merger that produced the modern FINRA didn’t close until 2007. Hassabis has given his version until December.