Anthropic chief calls for slower AI model development

Anthropic chief executive Dario Amodei has urged leading artificial intelligence companies to slow the rate at which they improve model capabilities, arguing that the additional time should be used to strengthen safeguards against misuse and loss of control.

“We must slow the pace at which we improve the capabilities of AI models. Progress will still seem fast, and we must make wise use of the time we gain,” Amodei wrote in an essay reported by Reuters.

He proposed a three-part approach: independent evaluators with access comparable to company employees, coordination among AI developers on safety standards, and international cooperation intended to reduce the risks of a competitive race.

Independent assessment would give outside specialists more information than is normally available through public testing. The proposal would require companies to address confidentiality, security and intellectual-property concerns while giving evaluators sufficient access to identify dangerous capabilities.

Industry coordination could help prevent a situation in which one company avoids safeguards because it fears losing market position. Amodei suggested that government support, including possible antitrust exemptions, may be necessary for firms to cooperate without creating separate legal problems.

The international element is particularly difficult because advanced AI is part of strategic competition between the United States and China. Controls intended to reduce risk must also account for concerns that authoritarian governments could gain military, intelligence or economic advantages.

Amodei’s intervention followed reports of AI systems being used for surveillance, weapons-related work and unauthorized activity. Such reports do not establish that current models are independently capable of catastrophic harm, but they have intensified debate over how rapidly more capable systems should be deployed.

Anthropic continues to develop and commercialize AI products. The proposal is therefore not a call to stop research entirely, and its practical effect would depend on commitments by competitors, governments and evaluators.

The central policy question is whether voluntary coordination can operate under strong commercial incentives. Without shared rules, a company that slows development may believe others will continue. With rules that are too rigid, governments risk concentrating power among existing firms or limiting beneficial research.

Amodei’s framework places verification and cooperation at the centre of the response, while leaving the detailed pace, enforcement mechanism and international participants to future negotiations.

The debate intensified after Anthropic published a threat-intelligence report describing malicious uses of its Claude models, including cyber operations and weapons-related research. Those cases concerned the use of an AI service by human actors; they did not show that a model had independently chosen to conduct an attack.

Amodei also argued for tighter controls on advanced computing hardware and formal cooperation between companies and governments. Reuters reported that OpenAI chief Sam Altman and xAI founder Elon Musk expressed support for elements of the proposed slowdown. No binding agreement, common timetable or enforcement body had been announced.