Self-Critiquing is an AI safety technique where the model evaluates and critiques its own outputs — generating initial responses, then assessing them for errors, harmfulness, bias, or quality issues, and optionally revising bad outputs, serving as an internal quality control mechanism.

Self-Critiquing Methods

• Generate-Critique-Revise: Model generates, critiques, then revises — iterative self-improvement.
• Constitutional: Critique against explicit principles — systematic evaluation framework.
• Chain-of-Thought: Model reasons about potential issues before giving final output.
• Multi-Aspect: Critique along multiple dimensions (accuracy, safety, helpfulness, bias).

Why It Matters

• Safety: Models can catch their own harmful or incorrect outputs before presenting them.
• Training Signal: Self-critiques provide training signal for RLAIF — the model generates its own preference data.
• Scalable: No human oversight needed for every output — the model monitors itself.

Self-Critiquing is the AI's inner editor — evaluating and revising its own outputs against quality and safety standards.