Principles the model applies to itself
Constitutional AI gives a model a written set of principles — a constitution — and has the model review and improve its own responses against them. It is an alignment method developed by Anthropic, and it underpins how the company builds its models.
How it differs from RLHF
In standard RLHF, human raters judge large numbers of answers and the model learns their preferences. That is expensive in human effort, transfers rater bias directly into the model, and leaves the criteria opaque.
Constitutional AI shifts much of that evaluation to the model itself. Against stated principles such as avoid content that harms people and be genuinely helpful, the model critiques its own answer and rewrites it. That self-improvement loop is built into training.
What it gains
- Explicit criteria: the principles a model behaves by can be read, checked, and argued with
- Scalability: tuning can proceed from written principles without hiring raters at scale
- Consistency: less variance between individual raters, so the policy holds together
Being able to publish and examine the rules a model was trained against was a meaningful step for transparency.
Context worth having
Constitutional AI is behind the behavior Claude is often described as having — careful and candid — and it symbolizes Anthropic’s emphasis on safety. Experiments have included drawing on public input when writing the principles, which connects to the wider question of who decides the rules AI follows. Taken with RLHF and alignment, it rounds out the picture of AI safety work.
Constitution is a metaphor: in practice it is a list of a few dozen behavioral principles. The name comes from their functioning as the highest-level rules for the model’s conduct.