On June 2, President Trump signed an executive order on frontier AI that went out of its way to promise restraint: no licensing, no preclearance, no government veto over when a model ships. Eight days later, the chief executive of one of the companies that order was designed to reassure published a paper asking for almost exactly the opposite. In "Policy on the AI Exponential," Anthropic's Dario Amodei argues that the government should have the legal authority to block the deployment of AI models that pose catastrophic risks, backed by civil penalties tied to a company's global annual revenue. The gap between those two documents is now the central fault line in American AI policy.
Amodei's proposal is unusually specific for a manifesto. It would apply to models trained with more than 10^25 floating-point operations, or to companies earning over $500 million in AI revenue or spending over $1 billion on AI research. Covered developers would have to test their models, publish the findings, and submit to independent evaluation. If a third-party assessment found unacceptable danger in one of four areas (biological weapons, cyberattack capability, loss of control, or AI accelerating its own development), the government could stop the launch. "In the several years that it can take Congress to act, AI can go from an amusing toy to the full country of geniuses," Amodei wrote, comparing the regime he wants to the rules that govern aviation and pharmaceuticals.
The administration is moving in a different direction, and not quietly. Its June order builds voluntary machinery: a classified benchmarking process to identify "covered frontier models," an optional 30-day pre-release review window, and a cybersecurity clearinghouse. Law firms briefing their clients on the order have stressed how explicitly it forecloses anything mandatory. Meanwhile, Axios and The Next Web report the White House is negotiating with senators to trade a three-year freeze on state AI laws for passage of the Kids Online Safety Act and the NO FAKES Act. Congress has rejected preemption twice; states passed 145 AI laws last year. The administration's bet is that bundling it with child safety legislation changes the math.
What makes Amodei's intervention awkward, and interesting, is who it comes from. Anthropic cites the vulnerability-hunting capabilities of its own Claude Mythos Preview as the reason transparency alone no longer suffices, days after putting a version of that capability on general sale. Critics were quick to note that compute and revenue thresholds would concentrate oversight on a handful of incumbents, Anthropic among them, while raising the drawbridge behind them. The counterargument is just as plain: if the people building the most capable systems say voluntary commitments are no longer enough, that is information.
The open question is whether anyone in Washington wants the power being offered. The White House has now twice declined to claim authority over model releases, preferring frameworks companies can walk away from. Amodei is asking the government to bind his industry, and implicitly his competitors, with rules that have teeth. One side is holding a door open and promising never to lock it. The other is asking who keeps the key.