Constitutional AI
Constitutional AI is an alignment approach that guides model behavior by evaluating or shaping it according to an explicit set of principles.
Introduction to Generative AI, Second Edition
Constitutional AI is a training method that guides language models toward helpful, harmless, and honest behavior using a set of principles rather than relying solely on human preference labels.
- How it is used
- Guides safety post-training by replacing some human preference labeling with principle-based AI feedback.
- How deep it goes
- Explains principles, fine-tuning, model critique, and feedback, while presenting the architecture as simplified rather than implementation-level.
- Examples
- Uses human-rights principles, privacy terms, self-harm conversations, and model-generated critiques as concrete examples.
- What you can do after
- Evaluate how principles and model critiques can shape safer language-model responses without relying solely on human feedback.
Read how Introduction to Generative AI, Second Edition explains Constitutional AI →
Reinforcement Learning from Human Feedback
Constitutional AI is a model-training approach that uses a set of principles to guide and evaluate model behavior.
- How it is used
- Uses written principles to generate instruction revisions and preference data for RLHF-style post-training.
- How deep it goes
- Defines iterative critique and preference construction, but stops short of detailed optimization equations or implementation code.
- Examples
- Uses violence-avoidance and truthfulness critiques, iterative completions, and pairwise preference comparisons.
- What you can do after
- Build synthetic instruction and preference datasets, then apply them in fine-tuning and RLAIF-style training.
Read how Reinforcement Learning from Human Feedback explains Constitutional AI →
Related concepts
- constitution-as-code documentuses · A constitution-as-code document is a written set of principles and behavioral constraints used to guide or evaluate an AI system’s behavior.
- evaluatoruses · An evaluator is a model or component that assesses outputs against specified criteria and produces judgments, scores, comparisons, or feedback.
- learning signalproduces · A learning signal is information or feedback that indicates what a model should learn or how it should adjust its behavior or parameters.
- synthetic datauses · Synthetic data is data generated artificially rather than collected directly from the target process, often to approximate its properties or meet specified requirements.
- preference alignmentkind of · Preference alignment is adapting a model's behavior to produce outputs that better match stated preferences, human values, or task-specific criteria.
These guides are generated from the text of Manning books by Manning's knowledge graph and link to the sections they draw on. The full books are available on liveBook and at manning.com.