Zuckerberg pushes back on coordinated AI safety pauses: 'Every lab has the responsibility'

Mark Zuckerberg says AI labs do not need a coordinated pact to slow down on safety. Market pressure will do that for them, and Meta's recent behavior is his proof. Anthropic has gone on the public record asking for government regulation and a mechanism that would let leading labs pause when warning signs appear. Nvidia's Jensen Huang has argued for independent evaluations and personal liability for the companies shipping frontier AI models. The frontier-lab safety fight split three ways this week, and the live argument is over who gets to decide when to press the brakes, not whether to press them at all.

Zuckerberg laid out the case on X on Tuesday, using Meta's handling of its own unreleased Muse model as the example. Meta delayed Muse "for several months" to strengthen safety and security, Zuckerberg wrote, and he framed the call as exactly the kind of decision a competitive market will keep rewarding rather than punishing. "Every lab has the responsibility and incentive to move at the pace required to train its models safely," he argued, and he tied that to "alignment" (the field's term for making sure AI systems follow user intent rather than drift toward unwanted behavior) as a feature labs will compete on, not a tax they will pay.

The Anthropic counter-position, as Business Insider summarized, is a regulator with binding pause power: a government body that can compel the leading labs to slow down when warning signs emerge, rather than trust each lab to act alone. Huang's frame, in the same coverage, hands the call to outside evaluators and corporate officers, with personal liability for AI companies as the backstop. Neither position trusts the others to self-pause cleanly.

Meta's chief AI officer, Alexandr Wang, reinforced the market-driven posture within hours. He called "racing on recursive self-improvement" (letting a model iteratively rewrite or train its successor with minimal human oversight) "one of the riskiest pathways for potential loss of control to powerful models" and committed Meta to directing "the significant majority of our compute towards serving people rather than racing on RSI." The language tracks Zuckerberg's: governance posture, not compute caps, is what the two executives want on the public record.

All three positions start from the same premise: frontier AI development may need to slow down at moments. The split is over who has the authority to demand it. Zuckerberg wants the market. Anthropic wants the state. Huang wants a third-party evaluator with personal liability behind it. The strongest test of the market-self-pause thesis is the first frontier lab that visibly does not slow down on its own, when a regulator-less response cannot contain the fallout.