Bai et al. (Anthropic) • 2022

Abstract

The foundational paper behind Claude. Constitutional AI extends RLHF by introducing a set of written principles (a 'constitution') that enables models to critique and revise their own outputs. This approach reduces reliance on human feedback for harmlessness while maintaining helpfulness, providing a scalable path to safer AI systems.

Warum Es Wichtig Ist

• Foundational to Anthropic's approach to safe AI
• Reduces dependence on human feedback for safety alignment
• Scalable framework for encoding values into AI systems