Glossary
What is constitutional AI
In short
Constitutional AI is a method from Anthropic in which a model is trained for safe behavior from a written list of principles: it criticizes and corrects its own answers, and people only write the rules.
Language models · constitutional AI
MMOOZZGG.com editorsChecked: 2 min read
For a child
In a classroom, you can hang rules on the wall: do not hurt each other, tell the truth, help. Then everyone checks what they said and corrects themselves if they broke a rule. This way the teacher does not need to watch every step. Some neural networks are taught in a similar way: they get a list of rules and correct themselves.
In plain words
Usually, to teach a neural network not to give harmful answers, people rate its answers by the thousands. Here it works differently. Developers write a short set of principles, a “constitution”. The model answers a request, then rereads the answer, looks for what breaks the principles and rewrites it. It is further trained on the corrected answers. Then a copy of the model compares pairs of answers and decides which is closer to the rules, and these ratings also go into training. People do not have to read piles of dangerous texts.
For business
For business, this approach is valuable for its transparency: the rules by which the model refuses or answers are written in words and published, so you can read them and compare them with your policy. Anthropic released the text of its models’ constitution under the CC0 license. The method itself works during training at the developer, and the user does not configure it. But the idea carries over to a company: your own principles on what an assistant may tell customers and what it must keep quiet about go into the system instruction and the answer review rules. The risk is that the wording does not cover every situation, so edge cases are tested by hand. Such rules fit well into an AI use policy.
For a specialist
Constitutional AI (CAI) was described by Bai and co-authors (2022). Human oversight comes down to a list of principles. The first phase is supervised: the model generates answers to provocative requests, criticizes them by a randomly chosen principle, rewrites them, and is fine-tuned on the corrected versions. In the second phase, the model compares pairs of answers, a preference model is trained on these comparisons, and it serves as the reward in reinforcement learning. The authors called this reinforcement learning from AI feedback (RLAIF). Both phases can use step-by-step reasoning (chain-of-thought), which makes the model’s decisions more transparent. The first published Claude constitution (May 2023) drew on the Universal Declaration of Human Rights, Apple’s terms of service, DeepMind’s Sparrow rules and other sources. The new version came out on 21 January 2026 and sets an order of priorities: safety, ethics, following Anthropic’s rules, helpfulness.
Example
A company writes a short set of rules for its assistant: give no legal opinions, do not discuss other companies, call an employee when in doubt. A separate model checks draft answers against these rules before they go to the customer.
Related terms
Questions and answers
What is Claude’s constitution?
How does constitutional AI differ from RLHF?
Sources
Not sure whether your company needs this?
AI solutions map