Advanced · Safety

Constitutional AI

Visual diagram · (in preparation) · Math · (in preparation) · Worked example · 3 difficulty levels.

TL;DR. An approach to aligning AI models, particularly LLMs, by providing them with a 'constitution' of principles to guide their behavior without human feedback.

Technical Definition

An approach to aligning AI models, particularly LLMs, by providing them with a 'constitution' of principles to guide their behavior without human feedback.

How it works

Constitutional AI seeks to automate the alignment process by allowing an AI to critique and revise its own harmful outputs based on a set of guiding principles, without direct human labeling. This reduces the need for extensive human oversight in fine-tuning, offering a scalable method for instilling ethical behavior and safety guardrails directly into the model's self-correction mechanism.

Related Concepts

• RLHF (Reinforcement Learning from Human Feedback) — A technique that aligns LLM outputs with human preferences by training a reward model from human comparisons.
• AI Safety & Alignment — The field ensuring AI systems behave as intended, remain under human control, and avoid unintended harm.
• Model Alignment — The process of training AI models, especially LLMs, to behave in accordance with human values, intentions, and ethical principles.