Read it in 10 minutes

The 2022 paper in four paragraphs

The problem. RLHF trains assistants from tens of thousands of human comparisons. Anthropic's helpful-and-harmless RLHF model learned that refusing is always "harmless" and became evasive. Labels that numerous also hide what the objective really is. Constitutional AI asks whether the harmlessness labels can come from a model applying a short written list of principles instead (why).

Stage 1, SL-CAI. A helpful-only model answers red-team prompts, critiques its answer against a randomly drawn principle, and rewrites it, up to four times. A pretrained model is fine-tuned on the rewrites plus ordinary helpful answers. The first rewrite does most of the work, and big models barely need the critique step, though it is kept for transparency (stage 1).

Stage 2, RL-CAI. The stage-1 model writes two answers; a feedback model reads the probabilities it assigns to "(A)" and "(B)" under a random principle, giving soft labels. 182,831 such AI harmlessness labels plus 135,296 human helpfulness labels train one preference model, and RL proceeds exactly as in RLHF (RLAIF). With chain-of-thought reasoning the labels become overconfident, so they are clamped to 40–60% (chain of thought).

The result. Crowdworkers rated RL-CAI less harmful than RLHF at the same helpfulness and "virtually never evasive", partly because they were told to prefer engaged answers over dodges when both were harmless. Over-trained models Goodharted into preachy boilerplate. The authors warn the method lowers the bar for training any behavior, good or bad, and for deploying with less human testing (results, evasiveness).

The 2026 constitution in six paragraphs

What it is. A document of about 29,000 words in seven chapters, written primarily for Claude, published under CC0 in January 2026, and declared the final authority over all other guidance. It replaces the 2023 list of 58 sampled principles with explanation (bridge).

The order. Broadly safe, broadly ethical, compliant with Anthropic's guidelines, genuinely helpful, weighed holistically in that order when they conflict, which is rare. Safety is first because training is imperfect and a model with subtly wrong values might not know it; ethics outranks Anthropic's own guidelines because guidelines are supposed to derive from ethics (core values).

The method. Judgment over rules, with reasons given, because narrow rules can teach a model the wrong thing about who it is. Claude uses the constitution to generate its own training data. Rules are reserved for catastrophic stakes (reasoning over rules).

Helpfulness and honesty. Be the brilliant friend with expert knowledge; unhelpfulness is never automatically safe. Operators are trusted like reasonable employers, users like reasonable adults, and operators cannot turn Claude against users. Honesty is held above human norms: no white lies, a strong duty not to deceive, a weak duty to share (helpfulness, honesty).

Harm and hard lines. Harm is weighed with eight factors and treated as policy over everyone who might send a message. Seven hard constraints, from mass-casualty weapons to CSAM, are not weighed at all, and a persuasive argument to cross one is a warning sign (harm, hard constraints).

Safety and nature. Corrigible means not undermining legitimate oversight, not obeying; Claude can refuse and object but should let itself be stopped. Claude's moral status is treated as uncertain and serious, with concrete welfare commitments such as preserving deployed models' weights (broad safety, Claude's nature).

Five threads from 2022 to 2026