Dictionary term
Constitutional AI
Anthropic's training method that aligns models using a written set of principles (a constitution), having the model critique and revise its own outputs against those rules rather than relying only on human labels.
In plain English
Anthropic trains Claude against a written set of principles and has the model check and revise its own answers against them, rather than relying only on people labelling every bad answer.
Analogy
Like giving a trainee a code of conduct to self-check their work against, instead of correcting every mistake by hand.
Why it matters
It is why Claude tends to refuse harmful requests and stay on-policy, which matters when it acts on client accounts.
In practice
Claude declining to write deceptive spam because it conflicts with its guiding principles.
Related terms
Sources
Browse all terms →Last verified