Constitutional AI
Definition pending
We've cataloged "Constitutional AI" but haven't written a full definition yet. Definitions are hand-curated rather than auto-generated, so it takes time to cover every term.
Want this one prioritized? Email us and we'll bump it.
Practical example
Constitutional AI means the model self-critiques its outputs against a written "constitution" of principles before showing them to the user. Operationally, this adds one extra inference pass per response — doubling your GPU cost. The trade: higher safety at 2x latency and cost. Only enable it on user-facing endpoints where safety failures have compliance implications.
Workflow example
Implement Constitutional AI as a post-processing layer: (1) model generates candidate response, (2) second inference pass critiques against your safety constitution (list of 10-20 rules), (3) if critique finds violations, model revises response, (4) revised response goes to user. Monitor extra_latency_ms and critique_trigger_rate. If trigger rate <1%, your constitution is too loose; if >20%, it's too strict and you're burning GPU.