Anthropic CEO Proposes Three-Part Plan to Slow AI

Anthropic CEO Dario Amodei has published a plan to slow the development of advanced artificial intelligence. The proposal comes amid intensified warnings from researchers about AI's potential dangers.

In a new blog post, Amodei argued that the pace of AI capability improvement must be reduced. He cited two primary reasons: the recent OpenAI-HuggingFace security breach and the observation that AI is advancing "drastically faster," particularly in its growing ability to build the next generation of AI systems. "We must slow the pace at which we improve the capabilities of AI models," Amodei wrote. "Progress will still seem fast, and we must make wise use of the time we gain."

Other tech leaders reacted positively. OpenAI CEO Sam Altman agreed that pacing the frontier is necessary, calling it a primary topic of recent internal discussions. SpaceX CEO Elon Musk simply posted, "Dario is right." The debate follows researcher Jacob Coxon's resignation from Anthropic, where he claimed leading AI companies are "gambling with our lives" and that some builders believe the technology "could kill us all by the end of the decade."

Embedded Third-Party Evaluators

The first pillar of Amodei's plan involves inviting independent "embedded evaluators" into AI companies. These evaluators, from organizations like METR, would verify that companies follow their safety and pacing commitments and ensure safety incidents are reported. Amodei compared them to regulators embedded within banks.

Anthropic is unilaterally committing to this step and calls on governments to require other frontier AI companies to do the same. This commitment means providing evaluators with company badges, desks, laptops, and access comparable to internal risk assessment teams, barring legal or contractual restrictions. Altman called this a "good idea" and stated OpenAI would follow suit.

Coordination Among Democratic Nations

Amodei's second proposal calls for leading AI companies within democratic countries to coordinate on common safety standards and limits on the rate of unchecked AI progress. He acknowledged that antitrust concerns have made such coordination seem unlikely.

"For antitrust reasons, it's helpful for the US government to mediate or at least enable these discussions," Amodei wrote. He suggested the government issue a narrow waiver for certain safety conversations. Amodei also addressed the common argument that slowing development risks Chinese AI dominance. He claimed the U.S. Could widen its lead over the next three to five years by refusing to sell powerful chips or semiconductor manufacturing equipment to Chinese firms and by cracking down on model distillation techniques.

Attempts at Global Coordination

The final strategy involves "global coordination," where the United States and its allies attempt to cooperate with authoritarian governments, including China. Amodei admitted there are "stark limits on what can be achieved" but suggested potential areas for agreement.

These could include prohibiting narrow and obviously dangerous uses of AI, such as employing the technology to produce biological weapons or allowing users to do so.

Criticism and Response

Amodei's warnings have drawn criticism from some quarters. Industry boosters have labeled him a "doomer" whose comments feed into a broader backlash against AI. Others are skeptical of apocalyptic warnings. Journalist Brian Merchant, for example, said he has yet to see a credible, step-by-step documentation of how AI could lead to human extinction. He suggested such proposals might only serve Anthropic and OpenAI, calling it "what regulatory capture looks like in action."

In response, Amodei argued the backlash is "fundamentally a crisis of trust" stemming from public skepticism of tech companies and government. He reiterated his belief in AI's potential benefits. "My desire to achieve these benefits is undimmed," Amodei wrote. "But the benefits will only be achieved if we build the technology in the right way."