Anthropic Calls for Global Pause in Frontier AI Development, Warns of Recursive Self-Improvement Risks

Recursive self-improvement, or RSI, refers to AI helping design and train more capable successor models, creating a cycle of accelerating capabilities. Anthropic argues that model failures could become more frequent and harder to understand if competition among frontier labs outpaces safety governance and society’s ability to adapt. It says the world must first establish verifiable safeguards that allow developers to slow or pause in coordination.

Anthropic published “When AI Builds Itself” on June 4, 2026, saying Claude had written more than 80% of the code merged at the company as of May that year. Engineers were merging eight times as much code per quarter as they did during 2021–2025, while the success rate on the most difficult and ambiguously specified coding tasks had risen to 76% within six months. The company said human review had become a bottleneck and that it would push governments and frontier AI developers to coordinate mechanisms for temporarily slowing development.