What Is Constitutional AI?
Constitutional AI is a training method in which the supervision signal for a behaviour is generated against a written list of principles — the constitution — instead of being collected as a human judgement on every example. A response is drafted, criticised against a principle drawn from that list, and rewritten; the rewritten responses, and later the comparisons between pairs of responses, are what the training run learns from. The principles are read during training. Nothing consults them when a request is answered.
The two phases, and what each one produces
The first phase is supervised and works by critique and revision. A prompt — usually one selected to provoke the behaviour being trained against — is answered by a starting configuration that has been trained for helpfulness and not yet for anything else. A principle is sampled from the constitution and used to ask for a critique of that answer, then for a revision addressing the critique. The pair of steps can repeat, sampling a different principle each pass. What is kept is the final revision paired with its original prompt, and fine-tuning on the collected set yields a configuration whose first drafts already resemble the revisions.
The second phase is reinforcement learning, and it replaces the human comparison at the centre of the ordinary preference pipeline. The phase-one configuration answers a prompt twice. Both answers, together with a principle drawn from the same constitution, are handed to a system asked which of the two better satisfies it. Those judgements become a preference dataset, a preference model is trained on it, and the configuration is optimised against that preference model exactly as it would be against one built from human comparisons. Because the label is generated rather than collected, this phase is generally called reinforcement learning from AI feedback — a name that distinguishes it from the human-feedback pipeline whose machinery it otherwise reuses in full.
The division of labour is the part worth carrying away. People write the constitution; a system applies it to the individual case. In the formulation the method was introduced with, that substitution was made for harmlessness alone, and preferences about how helpful an answer was continued to come from people — so the method as originally stated is not the removal of human supervision but a change in what humans are asked to supply, from a judgement per example to a document.
| Artefact | Where it comes from |
|---|---|
| The constitution | Written by people, in ordinary prose, and fixed before the run begins. |
| The prompt set | Assembled by people, commonly weighted towards cases that provoke the behaviour under training. |
| Critiques and revisions | Generated, one sampled principle at a time, by the configuration being trained. |
| The supervised set | The final revisions, kept and paired with the prompts that produced them. |
| Preference labels | Generated: a comparison of two answers against a sampled principle. |
| The preference model | Trained on those generated comparisons, then optimised against. |
Only the top two rows are produced by hand, and neither of them is produced per example. That is the economy the method is after: human effort moves from labelling cases to stating rules, and a rule can be edited and its consequences re-derived, where a labelled corpus can only be labelled again.
What a constitution is not
It is not a runtime check. Nothing reads the document while a request is being served; its influence arrives entirely through weights that were adjusted before the deployment existed. A written rule that must hold on every single request is a different piece of engineering — a validator, a filter, a policy step in front of or behind the call — and the need for one is not removed by having trained against a sentence that says something similar.
Neither is it a system prompt, though the two are confused often enough to be worth separating. Both are principles written in prose. A system prompt is text placed into the context of each request, and it can be read, edited or withdrawn between one request and the next. A constitution is an input to a training run, it is not present at request time in any form, and changing it changes nothing until another training run has been done.
And it is not a guarantee about the behaviour that results. The document supplies a target for an optimisation, and the distance between what a principle means to a reader and what an optimiser finds when it is rewarded for satisfying a judge is not a distance better wording can close. What the method genuinely offers is legibility: the rules a run was steered by exist as a page that can be read, argued with and quoted, which is more than a corpus of pairwise comparisons offers, since nobody can read a preference dataset back as a statement of intent.
How the Registry classifies a principle-guided revision task
Rewriting a draft against a written set of principles and keeping the record of each revision is filed as data processing in this registry's classification. The classification follows what the operation does to the material rather than what the material is about: a supplied text is transformed into another text under a stated rule, and what the task delivers is the trail of transformations rather than any new subject matter.
The Registry's canonical brief for this entry is filed as:
Revise a draft reply against a written set of principles and record each revision the principles produced.
Submitted for assessment it is
classified as Data processing, and its wording is hashed once — to
1bf5756401bfb1fd…, the first sixteen of sixty-four hexadecimal
characters — with the wording itself never stored. The hash is what the derivation
reads.
That class's own page is /tasks/analysis.
Classification is one of three inputs. The other two are the configuration submitted with the task, and the permanent chart derived from that configuration — fixed by the model name, the training cutoff and the temperature alone, and never reading the task at all. The same ascendant, ruling planet and harmony therefore appear on every assessment a given configuration receives, whatever it was asked to do. The derivation is published in full at /method.
What this page does not claim about constitutional ai
Whether a written principle produces the behaviour it describes is a question about a training run, and this registry has conducted none. It holds no preference data, has compared no configuration against another, and takes no position on which principles belong in a constitution or on who should write one. What it reads is the wording of the task submitted to it.
The Registry does not run this task, does not inspect any system's output for it, and validates no assessment it issues against what afterwards happens. What it does is compute — from a published method, for one submitted task and one submitted configuration — a verdict and a recommended execution window. It computes neither on this page.
Questions about constitutional ai
- What is Constitutional AI?
- Constitutional AI is a training method in which a written list of principles supplies the supervision signal for a behaviour. Responses are critiqued and revised against principles sampled from that list, and later compared against them in pairs; the revisions and comparisons are what the training run consumes, in place of a human judgement on each example.
- How is Constitutional AI different from RLHF?
- Reinforcement learning from human feedback trains a preference model on comparisons collected from people. Constitutional AI keeps that pipeline and changes where the comparisons come from: a system compares two answers against a principle sampled from a written constitution. Human effort moves from judging examples to writing the rules by which examples are judged.
- What is in a constitution?
- A list of short natural-language principles, each stating something a response should or should not do, written so that one can be sampled at a time and used to prompt a critique or a comparison. It is an ordinary document. It has no special syntax and it is not executed by anything.
- Is the constitution consulted when a request is answered?
- No. It is read during training, where it shapes critiques, revisions and preference labels. At request time nothing looks at it, which is why a rule that has to hold on every request needs a mechanism that runs on every request.
- What does RLAIF mean?
- Reinforcement learning from AI feedback: the phase in which the preference labels used to train a preference model are generated by a system applying written principles to pairs of candidate answers, rather than collected from people comparing them.
Related entries
Order an assessment for data processing tasks
The Registry issues a permanent, numbered task risk assessment for one submitted task and one submitted configuration. EUR 1.90 Standard, EUR 4.90 Extended, EUR 14.90 Full Chart, which adds the permanent chart. Assessments from EUR 1.90; machine-readable at /pricing.json.