Anthropic, the wildly successful AI company that has positioned itself as the most safety-conscious among top research labs, is dropping the central pledge of its flagship safety policy, company officials told TIME.
In 2023, Anthropic committed to never training an AI system unless it could guarantee in advance that its safety measures were adequate. For years, its leaders touted this promise—the central pillar of their Responsible Scaling Policy (RSP)—as evidence that they were a responsible company capable of withstanding market incentives to rush the development of potentially dangerous technology.
However, in recent months, the company decided to radically overhaul the RSP. This decision includes scrapping the promise not to release or train AI models if Anthropic cannot guarantee proper risk mitigations in advance.
“We felt that it wouldn’t actually help anyone for us to stop training AI models,” Jared Kaplan, Anthropic’s Chief Science Officer and co-founder, said in an exclusive interview with TIME. “With the rapid advance of AI and competitors blazing ahead, we didn’t feel it made sense for us to make unilateral commitments.”
The new version of the policy includes commitments to be more transparent about the safety risks of AI, such as making additional disclosures about how Anthropic’s own models fare in safety testing. The company commits to matching or surpassing the safety efforts of competitors and promises to “delay” development only if leaders consider Anthropic to be the clear leader of the AI race and perceive the risks of catastrophe to be significant.
This change comes as Anthropic—previously considered to be trailing OpenAI—rides a wave of technological and commercial success. Its Claude models, particularly the Claude Code software tool, have won devoted fans. In February, Anthropic raised $30 billion in new investments, valuing the company at $380 billion, with annualized revenue reportedly growing at a rate of 10x per year.
Kaplan denied that the decision was a capitulation to market incentives as the race for superintelligence accelerates, framing it instead as a pragmatic response to emerging political and scientific realities. “I don’t think we’re making any kind of U-turn,” Kaplan said.
When Anthropic introduced the RSP in 2023, it hoped to encourage rivals to adopt similar measures. However, binding national regulations never materialized. Instead, the Trump Administration has endorsed a “let-it-rip” attitude toward AI development. While a global governance framework seemed possible in 2023, three years later, it is clear that door has closed. Meanwhile, competition for AI supremacy—between both companies and nations—has only intensified.
Furthermore, the science of AI evaluations has proven more complicated than expected. In 2025, Anthropic announced it could not rule out the possibility of its models facilitating bio-terrorist attacks. However, they also lacked strong scientific evidence that the models did pose that specific danger. What the company previously imagined as a “bright red line” was becoming a fuzzy gradient.
In February, the board unanimously approved the new RSP, which states: “If one AI developer paused development while others moved forward without strong mitigations, that could result in a world that is less safe. The developers with the weakest protections would set the pace, and responsible developers would lose their ability to do safety research.”
Anthropic argues the retooled RSP maintains the spirit of the old one by committing to regularly release “Frontier Safety Roadmaps” and “Risk Reports” every three to six months. These documents will detail future safety goals and explain how capabilities, threat models, and active mitigations fit together.
Chris Painter, Director of Policy at METR, noted that while the change is understandable, it is a bearish signal for the world’s ability to navigate AI catastrophes. He stated that Anthropic “believes it needs to shift into triage mode… because methods to assess and mitigate risk are not keeping up with the pace of capabilities.”
