Constitutional classifier
Also called Constitutional Classifiers
A safety filter trained from written rules that guards a model’s inputs and outputs against harmful content.
Think of it like
A rulebook-carrying referee standing beside the player, blowing the whistle on anything against the rules.
Example
A model’s answer is generated normally, then a separate classifier trained on a safety constitution blocks it before it reaches the user because it crosses a line.
How it actually works
Rather than baking all safety into the model’s weights, you train a distinct classifier from a set of principles and run it as an independent layer on prompts and completions. Being separate makes it easy to update and harder to jailbreak in one shot, since an attacker must beat two systems. It builds on the constitutional AI idea of rules-as-supervision.
For product teams
A tunable, updatable safety layer you can adjust without retraining the whole model.
For engineers
Train input/output classifiers from a written spec and deploy them as a wrapper around the base model for defense in depth.
Related
- Constitutional AI — Grows out of constitutional AI.
- Jailbreak resistance — A key contributor to jailbreak resistance.
- Defense in depth — An instance of defense in depth.
Read anything AI without the jargon
Look up any term in plain English, or save terms as you read with the free Chrome extension.
Open DecoderAdd to Chrome