Anthropic warns AI-built AI could outrun safety and tilt markets

Anthropic said in a June 4 evaluation that its Claude system had generated more than 80% of the code its engineers contributed to company projects by May 2026, and that in August Claude led 26% of AI research and development tasks, up from under 1% in February. The company said none of this work was fully autonomous and that no system has achieved recursive self-improvement, meaning an AI creates and trains its successor without human input. Anthropic warned that faster AI progress could outpace safety checks and regulatory oversight, and CEO Dario Amodei called for slower capability development and independent evaluators. Stanford’s 2026 AI Index also reported private AI investment rose 127.5% in 2025, with US investment 23 times China’s.