# What is Constitutional AI?

Constitutional AI is a training method in which a model's behavior is shaped by an explicit written set of principles, called a constitution, instead of relying solely on case-by-case human judgments. The model critiques and revises its own responses against those principles, and the revised outputs become training data. The approach makes the values guiding behavior inspectable and editable.

The method has two phases. In the supervised phase, the model generates a response, critiques it against a stated principle, and rewrites it, producing a corrected example. In the reinforcement phase, a judge model uses the same principles to rank candidate responses, and those rankings train a reward signal that further tunes behavior.

The distinguishing feature is that the governing criteria are written down. Where preference training from labels leaves values implicit in thousands of individual annotator decisions, a constitution states them as text that can be reviewed, debated, versioned, and revised. This changes value alignment from an opaque data artifact into a document with a change history.

Practical benefits include consistency and scale. The same principle is applied uniformly across every case, and generating self-critiques costs far less than collecting human judgments for each one. It also reduces the volume of harmful content human annotators must read, which is a genuine occupational concern in safety labeling work.

The approach does not eliminate the hard problems. Principles must still be written by people, may conflict with one another, and can be interpreted by the model in unintended ways. The model's ability to apply an abstract principle correctly is itself a capability that varies, so a well-written constitution does not guarantee faithful adherence.

## Key points

- Behavior shaped by an explicit written set of principles
- Model critiques and revises its own outputs against them
- Values become an inspectable, versionable document
- Reduces human exposure to harmful labeling content
- Principle conflicts and misinterpretation remain unsolved

## In practice

A principle states that responses should decline to help with weapons manufacturing while still answering chemistry questions with educational intent. The model drafts an answer to a borderline question, then critiques its own draft against that principle and produces a revision that explains the underlying chemistry without procedural detail. The revised pair becomes a training example. Reviewing the principle text later is far easier than reverse-engineering the same rule from ten thousand labels.

## Related terms

- [Reinforcement Learning From AI Feedback](/en/glossary/reinforcement-learning-from-ai-feedback)
- [Reinforcement Learning from Human Feedback](/en/glossary/reinforcement-learning-from-human-feedback)
- [Instruction Tuning](/en/glossary/instruction-tuning)
- [System Prompt](/en/glossary/system-prompt)
- [Fine-tuning](/en/glossary/fine-tuning)

[Back to the AI Glossary](/en/glossary)