Anthropic CEO rare lengthy defense: AI is approaching a point where it must slow down, yet the company still needs to stay at the forefront.

Beating AI News Flash: Anthropic recently submitted a draft AI safety bill to U.S. lawmakers, drawing skepticism from U.S. AI safety advocates.

The draft sets a very high threshold for the government to emergency-halt dangerous models: the risk must involve biological weapons, cyberattacks, or AI loss of control, and could cause more than 50 deaths or serious injuries, or more than $1 billion in losses. Some provisions could also limit state governments' regulatory authority.

Nathan Calvin, strategy lead at AI safety organization Encode AI, subsequently dug up a series of Anthropic decisions over the past year. In February this year, Anthropic revised its originally strict safety policy. Previously, the company promised to pause training or release if safety measures could not keep pace with model capabilities; the new policy no longer requires the company to unilaterally stop R&D at higher-risk stages, on the grounds that competitors might seize the opportunity to overtake.

At the same time, the company has invested massive computing power, built a chip team, and let Claude take on an increasing number of AI R&D tasks. Nathan questioned whether Anthropic, while emphasizing AI safety, is loosening constraints on itself and even proposing provisions in the draft legislation that could weaken government oversight. Are these decisions accidental missteps, or choices the company made to maintain a competitive edge?

On October 21, Anthropic CEO Dario Amodei rarely personally published a long post on X to rebut. He acknowledged that the company has made some mistakes, but insisted that competing for the most advanced AI capabilities has never contradicted Anthropic's safety mission. The company has always pursued a "race to the top" strategy: only by keeping itself at the technological frontier can it be capable of pushing the entire industry to raise safety standards. He also listed nine safety decisions that had cost the company commercially, including restricting access to Mythos-class models for several months and strictly limiting biology-related requests—preferring to mistakenly block legitimate research and lose customers rather than allow the technology to be misused.

Dario also acknowledged that AI is approaching a stage where it could accelerate next-generation AI R&D and make safety measures hard to keep up. Anthropic's safety roadmap updated on October 19 also explicitly states that existing policies mainly target the deployment of dangerous models, and future policies may need to directly limit R&D speed, because speed itself can also become a risk. The company plans to propose relevant policy options by January 15, 2027, or explain why it has not.