Anthropic Calls for Co-ordinated Pause in AI Development
- •Anthropic urges a co-ordinated industry pause on advanced AI development to mitigate recursive self-improvement risks.
- •The startup reports that AI task capability has been doubling roughly every four months, outpacing societal readiness.
- •Anthropic plans to convene policymakers and rival labs to establish risk management standards for frontier AI models.
Anthropic is calling for a co-ordinated and verifiable pause in the development of advanced artificial intelligence, warning that rapid technological progress could soon result in recursive self-improvement. The company states that the ability of AI systems to perform tasks on their own has been doubling roughly every four months. In a blog post published on Thursday, Anthropic co-founder Jack Clark and Anthropic Institute lead Marina Favaro emphasized that if systems become capable of building their own successors, humanity must prioritize securing and monitoring these technologies before recursive self-improvement—the point where AI improves without human intervention—becomes inevitable.
The company argued that while a unilateral pause by a single lab is possible, it would only shift the industry front-runner rather than addressing the lack of a broader, deliberative process. Anthropic cautioned that poorly co-ordinated slowdowns could be counterproductive if less cautious actors continue to advance, thereby potentially reducing overall safety. To facilitate a meaningful pause, the startup is calling for agreements among multiple well-resourced labs, alongside established rules for triggering and lifting such a hiatus under external oversight.
In response to these concerns, the Anthropic Institute plans to convene policymakers, researchers, rival firms, and civil society groups in the coming months to discuss managing risks associated with self-improving systems. As of the report, major industry players including OpenAI, xAI, Alphabet, Meta Platforms, and Mistral have not responded to requests for comment. This call for caution follows Anthropic's own recent growth, including a confidential filing for an initial public offering on Monday and a valuation of $965 billion. Despite its focus on safety, the company noted in February that it would no longer withhold potentially dangerous capabilities if competitors were close to matching its models. Meanwhile, the US government has moved toward a voluntary testing framework, with an executive order from President Donald Trump requesting that labs submit their most capable models for cybersecurity testing prior to public release. Anthropic currently faces a potential national security blacklist, effective later in 2026, stemming from its refusal to allow the US military to utilize its models for domestic surveillance or autonomous weapons.