Training a language model to behave helpfully and avoid harmful outputs traditionally relies heavily on reinforcement learning from human feedback (RLHF). Human raters review pairs of model outputs and indicate which is better, and the model is trained to produce outputs more like the ones humans preferred. Anthropic's Constitutional AI method, published in 2022, proposed a meaningful modification to this process: instead of relying on human judgment for every single comparison, have the model itself critique and revise its own outputs against an explicit, written set of principles (a "constitution") reducing (though not eliminating) the amount of human labeling required, and making the standards the model is trained against more explicit and inspectable.

The two-phase process

Constitutional AI works in roughly two stages. In the first, supervised phase, the model is prompted to generate a response, then prompted again to critique that response against specific principles from the written constitution (things like avoiding harmful, deceptive, or discriminatory content), and then to revise its own response based on that critique. This self-critique-and-revise cycle can run multiple times, and the resulting revised responses are used to fine-tune the model directly.

In the second phase, the model generates multiple responses to a prompt, and instead of a human rater choosing the better response, another instance of the model (again guided by the written constitution) evaluates which response better follows the stated principles. That AI-generated preference data is then used in a reinforcement learning process similar to standard RLHF, but with AI-generated rather than purely human-generated preference judgments, sometimes described as RLAIF (reinforcement learning from AI feedback) rather than RLHF.

Why this approach, and what problem it addresses

Pure human-feedback approaches face a real scaling constraint: producing enough high-quality human preference judgments to shape a model's behavior across the enormous range of situations a deployed model encounters is expensive and slow, and human raters themselves can have inconsistent or unstated standards that make training signal noisier than a written, explicit set of principles would. Constitutional AI's explicit written principles are also more inspectable than an implicit standard embedded only in scattered human preference judgments. You can actually read the constitution a model was trained against, which is a meaningfully different transparency posture than "trust that our anonymous human raters had good judgment."

The connection to red-teaming and alignment more broadly

Constitutional AI is one specific, published technique within the much broader alignment challenge we cover in our separate explainer. It's an attempt to more efficiently and more explicitly encode intended behavior into a model's training process. It doesn't replace red-teaming, covered separately; red-teaming remains necessary to find cases where a model's actual behavior diverges from the constitution's intent, which happens regardless of how well the training process worked, because no training process perfectly generalizes to every situation a deployed model will encounter.

What it doesn't solve

It's worth being precise about the limits here: Constitutional AI shifts and partially automates part of the alignment training process, but it doesn't resolve the deeper, harder problem of specifying a fully complete and correct set of principles in the first place, a poorly written or incomplete constitution would still produce a poorly aligned model, just via a different mechanism than poorly calibrated human raters would. It's a genuine methodological advance in how alignment training is done, not a solved-problem claim about alignment itself.

Why this matters for anyone evaluating AI safety claims

Constitutional AI is a useful example of what a concrete, published, technically specific alignment technique actually looks like, as opposed to a vague claim that a company "prioritizes safety." When evaluating any AI lab's safety claims, the presence of specific, published, technically detailed methodology (like this one) is a meaningfully stronger signal than general safety messaging without technical substance behind it.