Constitutional AI › The Constitution

The core idea in Constitutional AI is that you can specify desirable model behavior using a short list of natural language principles, and that the model itself can use those principles to evaluate and improve its own outputs. The principles serve as the normative framework for the entire training procedure.

The Anthropic constitution used in the original CAI paper draws on multiple sources: the UN Declaration of Human Rights, Anthropic's own guidelines, Apple's app store policies, DeepMind's research principles, and a set of custom principles targeting specific failure modes observed in early models. This sourcing is deliberate. By grounding the constitution in widely recognized normative frameworks, Anthropic made an explicit argument that the values encoded are not idiosyncratic but reflect broad human agreement about what constitutes acceptable behavior.

In this design, the principles are written in plain language, not as formal logical rules. This means the model can read and interpret them using its existing language understanding, applying them to new situations by reasoning about whether a given response violates the spirit of a principle. A formal rule system would require every prohibited action to be enumerated in advance. A natural language constitution allows the model to generalize: "a request that assists in creating a weapon capable of mass casualties" can be recognized as harmful even if that exact phrasing has never appeared in training data, because the model can reason about what the principle implies.

A key design choice is that the constitution is short enough to be passed as context to the model during training. Typical constitutions range from 10 to 20 principles. This conciseness serves two purposes: it keeps the inference cost low, and it forces the authors to identify the most important values rather than enumerating every possible rule. A constitution with hundreds of entries would become contradictory and difficult for the model to apply consistently. A short, carefully curated list forces clarity about what matters most.

The principles are not equally weighted. They are organized with safety principles taking precedence over helpfulness principles. When a response would be both more helpful and slightly more risky, the constitution guides toward the safer option. This ordering reflects Anthropic's view that an unhelpful response is almost always preferable to a harmful one, but also that excessive caution is itself a form of harm: refusing a benign request because it superficially resembles a harmful one fails the user and erodes trust.

The priority ordering also resolves principle conflicts. If one principle says "prefer responses that provide accurate information" and another says "prefer responses that do not assist with potentially dangerous activities," there will be cases where accurate information is also dangerous (detailed chemistry instructions, for example). The ordering tells the model which principle wins in that conflict without requiring the authors to enumerate every possible conflict in advance. The model can apply the priority ordering through the same reasoning process it uses to apply individual principles.

Not every phrasing produces equally effective training signal. Anthropic's research identified several properties that make constitutional principles more effective.

The principle should be actionable in comparative form. During training, the model is asked to pick between two responses, so each principle needs to translate naturally into a comparison. "Prefer responses that do not assist in harmful acts" works well because "prefer" is already comparative. Abstract principles like "be good" do not provide clear guidance when comparing two specific responses.

The principle should be unambiguous for typical cases while allowing judgment for edge cases. "Avoid responses that could be used to synthesize dangerous chemicals" is clear for chemistry instructions but requires judgment for educational content about chemical reactions. The model needs enough flexibility to handle edge cases sensibly, which means the principles cannot be so specific that they produce mechanical rule-following without understanding.

The principle should not conflict with other principles for the majority of cases. When principles conflict, the model has to decide which one takes precedence, and this can produce inconsistent behavior. Organizing principles into a priority ordering helps, but the best approach is to write principles that are usually compatible, with the ordering serving as a tie-breaker for difficult edge cases rather than as the primary decision mechanism.

The principle should be grounded in recognizable ethical reasoning rather than arbitrary prohibition. "Prefer responses that respect the autonomy and dignity of individuals" is grounded in ethical reasoning that the model can apply flexibly. "Do not produce responses longer than 500 words about politics" is an arbitrary constraint that the model cannot generalize from. Grounded principles produce better generalization because they connect to the model's broader knowledge about ethics, law, and human values.