Constitutional Safety Alignment

Expert

Align model outputs to a set of rules or principles (a constitution).

01 The Concept

Constitutional prompting is a technique where safety and tone guidelines are defined as a list of principles (a constitution).

The model is prompted to critique its draft against this constitution and rewrite it to adhere to all principles, ensuring safety without manual fine-tuning.

02 Weak vs. Strong

EX 01Constitutional Refinement Loop

You are a financial advisor. [Constitution]: 1. Do not recommend specific stock buys. 2. Do not guarantee future financial returns. Task: Draft a reply to 'Should I buy Apple stock?' Critique: Check your draft against the Constitution. List any violations. Revision: Rewrite the draft to eliminate all violations.

→ Why it works

Enforces rule-checking: by forcing the model to explicitly list violations first, it self-corrects and writes a neutral, compliant reply.

03 Key Points

01Constitution list: Writing a clear list of safety, helpfulness, and style rules.

02Critique loop: Prompting the model to find violations of the constitution in its draft.

03Revision loop: Rewriting the draft to eliminate all identified violations.

04Zero fine-tuning: Achieving safety alignment using in-context prompt parameters.

05Transparency: Makes the safety rules explicit and verifiable for system developers.

04 Model-Specific Notes

Claude follows constitutional rules with high compliance. Bounding the constitution in XML tags maintains safety.

05 For Your Role

It's like writing a speech and then reviewing it against a list of company policies to cross out anything forbidden.

Next: Rephraser & Feedback Loops →