Constitutional AI: Harmlessness from AI Feedback
webCredibility Rating
High quality. Established institution or organization with editorial oversight and accountability.
Rating inherited from publication venue: Anthropic
Foundational Anthropic paper introducing Constitutional AI and RLAIF, directly influential on Claude's training methodology and a major contribution to scalable alignment research.
Metadata
Summary
Anthropic introduces Constitutional AI (CAI), a method for training AI systems to be harmless using a set of principles (a 'constitution') and AI-generated feedback rather than relying solely on human labelers. The approach uses a two-stage process: supervised learning from AI-critiqued revisions, followed by reinforcement learning from AI feedback (RLAIF). This reduces dependence on human feedback for identifying harmful outputs while maintaining helpfulness.
Key Points
- •Introduces Constitutional AI: a self-improvement loop where the AI critiques and revises its own outputs guided by a written set of principles.
- •Uses RLAIF (Reinforcement Learning from AI Feedback) as a scalable alternative to RLHF for harmlessness training.
- •Reduces need for human-labeled harmful content, making safety training more scalable and transparent.
- •The 'constitution' makes AI values explicit and auditable, improving interpretability of alignment objectives.
- •Trained models show strong harmlessness without significant sacrifice in helpfulness, validated through human evaluations.
Cited by 5 pages
| Page | Type | Quality |
|---|---|---|
| Large Language Models | Concept | 62.0 |
| Racing Dynamics Impact Model | Analysis | 61.0 |
| Paul Christiano | Person | 39.0 |
| AI Value Lock-in | Risk | 64.0 |
| AI Proliferation | Risk | 60.0 |
Cached Content Preview
Alignment Research Constitutional AI: Harmlessness from AI Feedback Dec 15, 2022 Read Paper Abstract As AI systems become more capable, we would like to enlist their help to supervise other AIs. We experiment with methods for training a harmless AI assistant through self-improvement, without any human labels identifying harmful outputs. The only human oversight is provided through a list of rules or principles, and so we refer to the method as 'Constitutional AI'. The process involves both a supervised learning and a reinforcement learning phase. In the supervised phase we sample from an initial model, then generate self-critiques and revisions, and then finetune the original model on revised responses. In the RL phase, we sample from the finetuned model, use a model to evaluate which of the two samples is better, and then train a preference model from this dataset of AI preferences. We then train with RL using the preference model as the reward signal, i.e. we use 'RL from AI Feedback' (RLAIF). As a result we are able to train a harmless but non-evasive AI assistant that engages with harmful queries by explaining its objections to them. Both the SL and RL methods can leverage chain-of-thought style reasoning to improve the human-judged performance and transparency of AI decision making. These methods make it possible to control AI behavior more precisely and with far fewer human labels. Policy Memo Constitutional AI Policy Memo Related content Emotion concepts and their function in a large language model Read more How Australia Uses Claude: Findings from the Anthropic Economic Index Read more Anthropic Economic Index report: Learning curves Anthropic's fifth Economic Index report studies Claude usage in February 2026, building on the economic primitives framework introduced in our previous report. Read more Constitutional AI: Harmlessness from AI Feedback \ Anthropic
1000c5dea784ef64 | Stable ID: sid_zNjyCZdCMB