Anthropic CEO outlines plan to slow AI development

Anthropic CEO Dario Amodei has called for a slower pace of frontier AI development, arguing that recent advances have increased both the capabilities and the risks of the systems being built. He proposed independent evaluators inside leading AI companies, coordination on safety standards, and limited international agreements on dangerous uses.

The proposal marks a shift from broad warnings about AI risk toward specific mechanisms for constraining development. OpenAI CEO Sam Altman said OpenAI agreed that companies need to “pace the frontier” and indicated that it would adopt embedded evaluators, although he did not provide a detailed implementation plan. SpaceX CEO Elon Musk also publicly endorsed Amodei’s position.

Why Amodei is calling for a slower frontier

In a new blog post, Amodei said two developments persuaded him that AI companies should take a more cautious approach: the reported OpenAI-HuggingFace hack and what he described as a sharp acceleration in AI capabilities. He placed particular emphasis on systems becoming better at helping build the next generation of AI.

Amodei’s argument is not for abandoning AI development. Rather, he said companies should reduce the speed at which model capabilities improve while using the additional time to strengthen safety practices. He maintained that AI could substantially improve people’s lives, but argued that those benefits depend on building and deploying the technology with greater care.

The timing also reflects growing pressure inside and outside the major AI labs. TechCrunch reported that Anthropic researcher Jacob Coxon resigned over concerns that leading companies were taking unacceptable risks. Amodei’s post did not explicitly address that resignation, and the available evidence does not establish that it directly prompted his proposal. It does show, however, that the debate over AI safety has become more visible even within frontier-model organizations.

Embedded evaluators would get internal access

Amodei’s most immediate proposal is to place third-party safety evaluators inside frontier AI companies. He pointed to organizations such as METR, which could verify whether firms are honoring their pacing and safety commitments and help ensure that significant incidents are reported.

The proposed arrangement would give evaluators company identification, workspaces, laptops, and access broadly comparable to that available to internal risk-assessment teams. Amodei said exceptions could apply where laws or contracts require them. He compared the model with regulators working inside financial institutions rather than relying solely on after-the-fact disclosures.

Anthropic is “unilaterally committing” to this approach, according to Amodei’s post. Altman described the idea as good and said OpenAI would do the same, with more details to come. Those commitments remain executive statements rather than independently verified operating programs. The evidence available so far does not specify the evaluators’ authority, reporting structure, funding, or ability to publish findings.

That distinction matters for AI builders and enterprise buyers. An evaluator with access but no power to halt a release may provide visibility without changing decisions. Conversely, a genuinely independent review function could affect model launch schedules, access policies, incident response, and the evidence customers receive before deploying AI agents in sensitive workflows.

Coordination faces competition and antitrust barriers

Amodei also proposed that major AI companies in democratic countries coordinate on common safety standards and limits on unchecked progress. He acknowledged that direct coordination among competitors could raise antitrust concerns, particularly if companies appear to be jointly restricting products or slowing competition.

His solution was for the U.S. government to mediate or enable narrowly defined safety discussions and provide a limited antitrust waiver. That proposal would require policymakers to distinguish technical safety cooperation from anti-competitive conduct. The source material does not indicate that such a waiver exists or that governments have agreed to create one.

The geopolitical challenge is equally significant. Critics of slowing AI development often argue that restrictions could allow China to gain ground. Amodei countered that export controls on powerful chips and semiconductor manufacturing equipment, combined with tougher action against model distillation, could widen the United States’ lead over the next three to five years.

Those are strategic judgments rather than established outcomes. They depend on enforcement, the pace of domestic innovation, China’s ability to substitute restricted hardware and software, and the willingness of allied governments to coordinate. Amodei’s proposal therefore links AI safety policy to industrial policy and national-security decisions, rather than treating model development as an issue confined to company laboratories.

The narrowest path may be global safety rules

Amodei’s final proposal was international coordination among the United States, its allies, and authoritarian governments where possible. He acknowledged that cooperation with China would have clear limits, but suggested that governments might still agree on a small number of obviously dangerous applications.

He cited AI-assisted biological weapons production as an example of an area where a narrow prohibition could be more achievable than a broad global agreement on AI development. Such rules would not resolve the wider debate over model capabilities, but could create minimum boundaries around high-consequence uses.

The approach also responds to a broader trust problem. Amodei argued that public backlash against AI reflects skepticism toward technology companies, the technology sector, and governments. Industry critics, including journalist Brian Merchant as cited by TechCrunch, have questioned whether apocalyptic scenarios are sufficiently supported by step-by-step evidence and warned that sweeping safety proposals could strengthen the market position of Anthropic and OpenAI.

That criticism highlights a central credibility problem: the companies asking for more oversight would also help define the standards, access rules, and pace limits. Independent governance would need to prevent safety coordination from becoming a mechanism for regulatory capture or for protecting incumbent labs from new competitors.

What to watch next

The clearest near-term signal will be whether OpenAI publishes a concrete plan for embedded evaluators, including their independence, access rights, escalation powers, and incident-reporting obligations. Anthropic’s own implementation will be similarly important; a public commitment alone will not show whether evaluators can influence deployment decisions.

Policymakers’ response will provide a second test. Watch for proposals covering narrow antitrust protection for safety discussions, formal incident-reporting requirements, and rules governing access to frontier-model development. Any government action will reveal whether Amodei’s ideas can move beyond executive statements.

AI builders and enterprise customers should also monitor how these commitments affect release cadence, model access, safety documentation, and restrictions on high-risk capabilities. If pacing becomes operational, companies may see slower model updates but stronger evidence about reliability and misuse risks. If it remains voluntary, the competitive pressure to release increasingly capable systems may continue unchanged.

Creati.ai perspective

Amodei’s announcement is significant because it turns the phrase “pace the frontier” into a proposed operating model: outside evaluators inside labs, common safety discussions, and limited international rules. But the proposal is still at the commitment stage. Its value will depend on whether evaluators are genuinely independent and whether safety findings can delay or alter launches.

For the market, the tension is not simply between fast and slow AI development. It is between speed that remains difficult to audit and speed accompanied by credible oversight. Builders and buyers should treat the latest statements from Anthropic and OpenAI as signals of changing governance expectations, not yet as proof that frontier development has materially slowed.