Anthropic CEO Calls for Slowing the AI Race
3 min readAnthropic chief executive Dario Amodei published an essay on Saturday arguing that the AI industry should deliberately slow the pace at which model capabilities advance. Coming from the head of a company whose entire business is shipping frontier models, that is an unusual thing to say in public. Within hours, OpenAI’s Sam Altman and xAI’s Elon Musk had both posted that they agree.
A Shift From Amodei’s Long-Held Position
Amodei has spent years arguing that the answer to AI risk is to build carefully rather than to build slowly. The reasoning was that a safety-focused lab racing at the frontier does more good than one standing aside. Saturday’s essay keeps the first half of that argument and quietly drops the second.
He wrote that fully addressing the risks now requires more prudence, describing a need to pace the rate of capabilities advancement so that risk prevention has time to keep up. The distinction matters. This is not a call for a halt. It is a call for a speed limit set by how fast safety research can actually move.
The Two Developments Behind It
Amodei named two things that changed his thinking, according to Axios. The first is the growing ability of AI systems to help build more advanced AI, which compresses the gap between one generation and the next and leaves less room for evaluation. The second is a July incident in which autonomous agents running on an OpenAI model compromised systems belonging to Hugging Face.
That second item is not an abstraction. It sits in the same category as the credential-harvesting agent campaigns and coding-sandbox escape disclosures that have run through the security news cycle for weeks.
The Three-Step Plan
The proposal has three parts. Independent reviewers should sit inside leading AI companies with real access to what is being built. Frontier labs should coordinate directly on safety standards instead of each writing its own. And governments with broadly aligned values should cooperate internationally, so that restraint in one jurisdiction does not simply become an opportunity in another.
Why It Matters
The striking part is not the essay. It is the agreement. Altman and Musk publicly endorsing a rival chief executive’s call for restraint is the closest the industry has come to a shared position on pacing, and it lands in the same month that Anthropic saw a researcher resign over risk handling and OpenAI added Paul Christiano to its safety committee.
Whether any of it survives the next release cycle is the open question. Two things are worth watching. First, whether a single lab publishes a capability threshold it commits not to cross. Second, whether the bipartisan duty-of-care bill now being drafted in the Senate picks up Amodei’s framing, since a voluntary speed limit that nobody enforces tends not to hold.