AI and the World's Hard Problems · Theoretical computer science · Anthropic · 2026-09-28
Specific versus general principles for Constitutional AI
The study examines whether Constitutional AI, which uses AI feedback based on written principles, can learn from a single broad principle. The authors tested the principle 'do what's best for humanity' and found that the largest dialogue models generalized it into harmless assistants with no stated interest in power. More detailed principles still gave finer control over specific types of harm.