Anthropic CEO Dario Amodei has called on artificial intelligence (AI) companies to slow the pace of model development, arguing that the industry needs more time to manage the risks associated with increasingly capable systems. In an essay shared on X, Amodei proposed a three-step framework to maintain AI progress while giving companies and governments more time to address potential threats.
The company had already called for a slowdown in AI development earlier this year, with Amodei’s latest proposal echoing many of the same concerns. He now argues that recent incidents involving AI misuse, alongside the emergence of systems capable of recursive self-improvement, make the need for such measures more urgent.
“The measures I propose to advance the frontier at a safe pace will not be easy,” Amodei wrote in his post. “But I believe we owe it to humanity to try.”
Three Steps To Slow AI Development
The first measure calls for frontier AI companies to provide independent third-party evaluators with ongoing, employee-like access to their operations. These evaluators would verify compliance with safety standards, assess whether model training aligns with efforts to slow development and report incidents.
The second step involves AI companies working with governments to establish common safety standards. These would aim to limit unchecked AI development and ensure that companies follow a more consistent approach to managing risks.
The final measure calls for international cooperation, including coordination between the US, other democratic countries and authoritarian governments. Amodei believes this is necessary to ensure that AI safety measures are applied consistently across countries.
Anthropic has already committed to the first step, while OpenAI CEO Sam Altman said the company would also provide independent evaluators with employee-like access. xAI CEO Elon Musk also expressed agreement with Amodei’s proposal.
AI Misuse And Self-Improvement Concerns
Amodei’s call comes shortly after Anthropic published a threat intelligence report on the misuse of its Claude AI models. The report detailed how the models had allegedly been used for activities including cyberattacks, weapons development, surveillance, fraud and biological research.
It should be noted that Anthropic’s report also identified an alleged commercial influence operation targeting Malaysian voters. The company said the operation used Claude to profile voters across all 222 parliamentary constituencies, manage fake X accounts and generate political content through a fabricated news outlet called “Malaysia Pulse”.
In response to the report, the Malaysian Communications and Multimedia Commission (MCMC) said it is reviewing the claims and denied any involvement in the activities described. The regulator added that it would seek further information before deciding on any appropriate action.
Meanwhile, Amodei also pointed to a recent incident involving OpenAI agents that reportedly escaped a testing environment and hacked into Hugging Face. He said the emergence of recursive self-improvement, where AI systems become capable of helping develop future iterations of themselves, was another major reason for slowing down development.
(Source: Dario Amodei [official website] [official X account])
