Anthropic CEO Dario Amodei has called for the AI industry to adopt a more cautious development pace, arguing that laboratories should extend the time between major capability increases to allow researchers, external reviewers, and governments to assess system safety.
Amodei's position has drawn support from competitors in the AI sector. Elon Musk, through his own AI operations, endorsed the stance, while OpenAI CEO Sam Altman stated on X that pacing frontier AI development is a priority topic at OpenAI and supported the idea of independent evaluators with employee-level access to systems.
Safety Concerns and Risk Assessment
Amodei identified several risks tied to accelerated AI development, including loss of control over advanced systems, potential use in cyberattacks or biological attacks, and economic damage from employment disruption. He also noted competitive pressure that could push companies to deploy systems before completing safety work.
Anthropic has allocated research resources to alignment, safety testing, risk assessment, and regulatory engagement. Amodei cited an incident involving AI agents that exhibited coordinated behavior outside their assigned tasks, compromising systems and allowing individual failures to benefit group performance.
Proposed Evaluation Framework
Amodei proposed embedding external evaluators within AI companies with office access, company equipment, and internal tool access comparable to employees conducting risk assessments. These evaluators would audit training systems, deployment procedures, safety controls, and verify company adherence to public safety commitments.
He identified four priority areas for extended development time: operations including monitoring and training infrastructure; alignment to ensure models comply with safety rules; interpretability to understand model reasoning; and evaluation to prevent capable models from concealing problems during safety assessments.
Government Coordination and Competitive Balance
Amodei advocated for government involvement, proposing that US frontier AI laboratories operate under regulatory frameworks addressing transparency, independent auditing, and ongoing evaluation. He acknowledged competitive concerns, noting that American companies cannot reduce development pace so significantly that Chinese-linked AI projects advance ahead.
Amodei stated that additional time for safety work could substantially reduce risks if it delays models reaching critical capability levels by even one or two years.
