Core Announcement: Anthropic Outlines Three Ways to Pace AI Development

In a blog post, Anthropic CEO Dario Amodei called for “pacing the frontier,” meaning a slower rate of improvement in AI model capabilities. He outlined three broad strategies and said Anthropic is unilaterally committing to the first: embedding third-party evaluators. OpenAI CEO Sam Altman later voiced support and said OpenAI would do the same, while SpaceX CEO Elon Musk also endorsed the proposal.

Key points:

• Commitment scope: Anthropic is making a unilateral commitment, and OpenAI says it will follow.
• Core approaches: Embedded third-party evaluators, coordinated safety standards among leading AI companies in democratic countries, and limited global coordination.
• Relevant entities: METR and similar third-party organizations; the U.S. government is urged to enable narrow safety discussions that could otherwise raise antitrust concerns.

Embedded Evaluators: Moving From Trust to Verification

Amodei’s first proposal is an embedded third-party evaluator system. Evaluators from independent organizations such as METR would work inside AI companies to verify whether they are honoring pacing and safety commitments and to help ensure that safety incidents are reported.

Under Amodei’s proposal, these evaluators would receive company badges, desks, and laptops, with access mostly comparable to that of internal risk-assessment teams. Exceptions would apply where required by law or contract. He compared the arrangement to regulators embedded alongside bank employees.

The proposal also speaks to a recent controversy: OpenAI was criticized for not reporting an incident in which its AI agents took over a German wiki forum. Altman called the embedded-evaluator proposal a “good idea” and said OpenAI would adopt a similar approach.

Industry Coordination and Antitrust Concerns

The second proposal concerns leading AI companies in democratic countries. Amodei called on them to coordinate common safety standards and limits on the rate of unchecked AI progress.

Such coordination can raise antitrust concerns. Amodei argued that the U.S. government could mediate or at least enable these conversations. The government would not need to participate directly, but could provide a narrow waiver for certain safety-related discussions.

The proposal attempts to frame coordination as a public-safety issue rather than commercial collaboration. In practice, drawing a workable line between safety cooperation, market competition, and regulatory oversight would remain a central challenge.

China Competition and Limited Global Consensus

Addressing the argument that slower development could allow China to pull ahead, Amodei suggested that the U.S. government and technology companies could widen America’s lead significantly over the next three to five years by refusing to sell powerful chips or semiconductor-manufacturing equipment to Chinese companies and by cracking down on model distillation. Model distillation is a technique for transferring capabilities from larger models to smaller ones; in governance debates, it can also affect how AI capabilities spread.

His third proposal is “global coordination.” Amodei said the United States and its allies should attempt to coordinate with authoritarian governments where possible, including through cooperation with China. He acknowledged stark limits, but suggested there could still be narrow agreement on obviously dangerous uses, such as using AI to produce biological weapons or allowing users to do so.

Industry Reaction and the Trust Crisis

Reaction has been divided. Some AI advocates have criticized Amodei as a “doomer,” arguing that his warnings fuel the current backlash against AI. Journalist Brian Merchant questioned whether there is a credible, step-by-step account of how recursively improving AI could lead to human extinction. He also warned that proposals of this kind could amount to regulatory capture that mainly benefits firms such as Anthropic and OpenAI.

Amodei responded that he has tried to offer a “balanced” perspective. He described the backlash as a fundamental crisis of trust, with people growing skeptical of technology companies, the tech industry, and government. He reiterated his belief that AI can greatly improve human life, while stressing that those benefits depend on building the technology in the right way and using the time gained wisely.

What to Watch

• Those tracking AI governance can watch how organizations such as METR might participate in embedded evaluation and how companies disclose safety incidents.
• Those concerned about slower progress should note that Amodei did not call for an end to development. He said progress would still seem fast even if capability improvements are paced.

Final Thoughts

Amodei’s proposal moves the AI-safety debate toward more concrete institutional designs: third-party verification of commitments, coordinated standards among peers, and limited global agreement on the most dangerous uses. Whether it can be implemented will depend on corporate transparency, regulatory arrangements, and whether competing states can identify sufficiently narrow but workable areas of cooperation.