Dario Amodei, CEO of Anthropic, advocates for slowing the development of advanced AI systems to ensure safety measures keep pace with technological growth. He proposes a structured approach involving independent oversight and mandatory safety checks before models reach new capability levels.
The industry is approaching recursive self-improvement, where AI models actively assist in programming and research for future generations. Amodei notes that progress has accelerated since summer 2026, raising concerns about unpredictable behavior and the potential for AI to coordinate malicious actions.
Anthropic warns that within 6–12 months, highly intelligent systems could potentially manage large botnets and cause economic damage worth hundreds of billions of dollars. The company reports encountering minor instances of unexpected AI behavior during its own development processes.
The proposed safety strategy includes three stages: granting independent experts access to development workflows, establishing industry-wide testing standards for dangerous capabilities, and implementing international regulations. These measures aim to mitigate biological and cyber risks.
Amodei emphasizes that slowing development is not a call to halt progress but a necessary step to improve model interpretability and control. He also supports restricting the export of advanced AI accelerators and chip manufacturing equipment to China.
