Anthropic's Amodei Pushes Three-Step Plan to Slow AI Development as Industry Split Widens

The AI industry's brief, fragile consensus on regulation cracked open again this week. Anthropic CEO Dario Amodei spent the weekend circulating a three-part proposal for slowing the pace of frontier AI development, and by Monday the fault lines that have divided labs, investors, and policymakers for two years were back in plain view. Amodei's plan calls for third-party evaluators stationed inside AI labs, coordination among domestic developers, and international agreements meant to keep the technology from advancing faster than oversight can follow.

The first prong is the most concrete. Rather than relying on labs to police themselves, Amodei wants outside evaluators embedded in development teams with enough access to test models before release. That is a meaningful shift from the voluntary commitments several major labs signed at the White House in 2023, which left testing and disclosure largely in companies' own hands. Independent access would give regulators and researchers a view into capability jumps that currently surface only after a model ships.

The second prong, domestic coordination, addresses a structural problem that has dogged every US regulatory effort: labs compete on speed, so any single company that slows down risks losing ground to rivals. Amodei's answer is to make restraint collective rather than unilateral, so no lab is penalized for holding back. The third prong pushes the same logic across borders, where differing national rules have so far allowed developers to shop for the lightest oversight.

The timing matters. The proposal lands as Washington, Brussels, and Beijing are all weighing or implementing AI rules, and as the companies themselves remain split on how far restrictions should go. Some labs and investors have argued that aggressive curbs would hand the lead to overseas competitors; others have warned that safety testing is lagging behind model capability. Amodei's plan tries to bridge that gap by making slowdown a coordinated move rather than a solo bet.

What remains unanswered is enforcement. Amodei's proposal describes who should do the evaluating and at what level coordination should happen, but not who holds labs to it or what happens when they refuse. Third-party evaluators need legal authority and access guarantees that currently do not exist in the US, and international agreements of this kind take years to negotiate. Until those gaps close, the plan reads more as a framework for debate than a rulebook, and the smackdown over AI regulation is not over.