Anthropic Proposes Global Pause Over Recursive AI Self-Improvement Risks

AI research lab Anthropic has proposed an international mechanism to slow or temporarily pause the development of frontier AI systems. The company warns that the industry is approaching a threshold of recursive self-improvement, where AI autonomously designs and develops its own successors. Anthropic revealed that as of May 2026, its Claude model authors approximately 80% of the company's codebase, increasing engineer productivity eightfold since 2024.

Co-founder Jack Clark and research lead Marina Favaro argue that this trend could trigger an intelligence explosion, potentially leading to a loss of human control. They propose a regulatory brake pedal and a verification regime similar to nuclear arms-control agreements to prevent a race to the bottom where safety is sacrificed for competitive advantage. The company is currently preparing for an initial public offering with a valuation of approximately $965 billion.

Reactions to the proposal are divided. The Australian government welcomed the call for a more considered approach, while the advocacy group PauseAI Australia argued that such freezes should be governed by international treaties rather than private companies. OpenAI disagreed with the framework, asserting that democratic governments, not private labs, must determine the rules and accountability mechanisms for AI innovation.

These warnings coincide with claims from SoftBank CEO Masayoshi Son that superintelligence will arrive within two years, as well as University of Toronto research demonstrating that AI can be used to create adaptive hacking worms to automate complex cyberattacks.