What is Constitutional AI? It is Anthropic's training method that shapes an assistant's behavior with a written list of principles (a "constitution") and AI generated feedback, instead of relying only on human labels for every harmful or preferred answer. Anthropic introduced the method in its December 15, 2022 paper Constitutional AI: Harmlessness from AI Feedback (also on arXiv as 2212.08073). Claude has been trained with versions of that approach since early Claude releases, and on January 22, 2026 Anthropic published a new, much longer Claude constitution under Creative Commons CC0.
This guide covers what Constitutional AI is, how the two training stages work, how Claude's constitution has changed, how it compares with OpenAI's Model Spec, what it does and does not solve, and what it means if you use or buy Claude.
What Constitutional AI means
Most chat assistants after 2022 are shaped by preference training. People rank answers, a preference model learns those rankings, and reinforcement learning pushes the assistant toward answers that score well. That family of methods is usually called reinforcement learning from human feedback (RLHF). It works for everyday helpfulness, but it has costs: people must review large volumes of text (sometimes disturbing), labels are expensive, and the rules the model learns stay partly hidden inside a preference model.
Constitutional AI keeps the preference training idea and changes the source of the rules. Humans write a short set of principles. An AI model then critiques and revises its own drafts against those principles, and later ranks pairs of answers by the same principles. Anthropic calls the ranking step RL from AI Feedback (RLAIF). In Anthropic's words from the 2022 research page, the only human oversight for harmlessness labels is "a list of rules or principles."
| Term | Plain meaning |
|---|---|
| Constitution | Written principles that say how the assistant should behave |
| Critique and revise | The model checks a draft against a principle and rewrites it |
| RLAIF | Preference data and RL rewards come from AI judgments, not only human labels |
| Constitutional AI (CAI) | The full pipeline that uses a constitution for both stages |
Constitutional AI sits inside the wider AI alignment problem: it is one way to make the intended objective more explicit (outer alignment style work) while still needing other tools for inner failures such as alignment faking. It is also one method inside the broader field of AI safety.
How Constitutional AI training works
Anthropic's 2022 paper and its May 9, 2023 explainer Claude's constitution describe two phases.
Stage 1: supervised critique and revision
- Start from a helpful but not yet harmlessness trained model.
- Ask it hard or adversarial prompts so it produces rough answers.
- Ask the model to critique that answer using one principle drawn from the constitution.
- Ask it to revise the answer in light of the critique.
- Repeat with other principles, then fine tune on the revised answers.
Chain of thought style reasoning (asking the model to explain the critique step by step) improved human judged quality and made the decision path easier to inspect in Anthropic's tests.
Stage 2: reinforcement learning from AI feedback
- Sample pairs of answers from the fine tuned model.
- Ask a model which answer better matches a sampled constitution principle.
- Train a preference model on those AI preference labels.
- Run reinforcement learning with that preference model as the reward.
Anthropic reported that the resulting assistants were preferred by crowdworkers over earlier RLHF harmlessness models on adversarial prompts, while staying helpful and less evasive: they explained objections instead of shutting down. The 2022 abstract stresses that both stages can use far fewer human harm labels than classic RLHF.
What is in Claude's constitution
"Constitutional AI" is the method. Claude's constitution is the living document Anthropic feeds into that method (and related training) for Claude.
Early principles (2023)
The May 2023 post published a long list of short principles. Sources included the Universal Declaration of Human Rights, platform style safety rules, DeepMind Sparrow inspired rules, attempts to weigh non Western perspectives, and principles Anthropic found worked in early research. Examples included choosing answers that are helpful, honest and harmless; avoiding toxic or discriminatory content; refusing help with illegal plans; and avoiding overly preachy tone. During training the model sampled principles rather than applying every rule on every step.
The January 22, 2026 rewrite
Anthropic's Claude's new constitution post says the old list of standalone principles was not enough. The new document explains why Anthropic wants certain behavior, so Claude can generalize instead of only following bright line rules. Anthropic still keeps hard constraints for some high stakes acts (the post's example is never providing significant uplift to a bioweapons attack). The full text is public at anthropic.com/constitution and in the anthropics/claude-constitution GitHub repo, released under CC0.
Current Claude models should, in Anthropic's stated order of priority:
- Be broadly safe: not undermine appropriate human oversight of AI during this phase of development.
- Be broadly ethical: be honest, act by good values, and avoid inappropriate, dangerous or harmful actions.
- Be compliant with Anthropic's guidelines: follow more specific Anthropic instructions where they apply (medical advice, cyber requests, jailbreaks, tools and similar).
- Be genuinely helpful: benefit operators and end users.
When those goals conflict, Anthropic says Claude should generally prioritize them in that order. The company also says most everyday tasks do not force a hard conflict. Sections of the new constitution cover helpfulness across Anthropic, API operators and users; Anthropic's supplemental guidelines; ethics and honesty; being broadly safe even above ethics when oversight is at risk; and Anthropic's uncertainty about Claude's nature or possible moral status.
Anthropic is clear about the gap between document and model: training is hard, outputs may not always match the ideals, and the constitution is a living document. That honesty matters when you read marketing claims about "aligned" assistants.
Constitutional AI in Claude vs OpenAI's Model Spec
Both labs now publish the behavior rules they train toward. OpenAI's parallel document is the OpenAI Model Spec.
| Item | Constitutional AI / Claude constitution | OpenAI Model Spec |
|---|---|---|
| Lab | Anthropic | OpenAI |
| Role | Method plus Claude's values document | Public rulebook for GPT behavior in ChatGPT and the API |
| Latest major text | January 22, 2026 constitution | October 29, 2026 Spec rewrite |
| Organizing idea | Four priorities in order, with reasons written for Claude | Five values plus a chain of command with many worked examples |
| License | CC0 | CC0 |
| Training claim | Constitution shapes training and synthetic data | Models are trained toward the Spec; OpenAI says they do not fully match it yet |
They answer the same user question from different angles: "what rules is this assistant supposed to follow?" Neither document is a usage policy, a law or a guarantee. OpenAI's Spec is covered in detail in our Model Spec guide. Claude product behavior also sits next to Anthropic's usage policy and abuse handling, covered in guides such as Claude abuse policy and does Claude report to police.
Pros and cons of Constitutional AI
Pros
- Makes intended values readable. Anyone can open Claude's constitution or the 2022 principles list and see what Anthropic says it wants.
- Cuts the need for humans to label every harmful example, which Anthropic argues is safer for workers and cheaper to scale.
- Supports less evasive refusals: models can explain why they will not help, instead of only shutting down.
- Lets labs edit behavior by editing principles, which is more inspectable than silently changing a preference dataset.
- Fits scalable oversight research: using AI to help supervise AI, a theme also discussed by groups such as METR, Redwood Research and the Alignment Forum.
Cons
- Humans still choose the principles. A constitution encodes Anthropic's (or another lab's) judgment, not a universal moral consensus.
- A written rulebook does not prove inner alignment. Alignment faking research shows models can play along with training objectives under pressure.
- Sampling principles one at a time (in the early method) can leave gaps; the 2026 rewrite tries to fix that with richer explanation, but Anthropic still warns of a training gap.
- Public text can be gamed by attackers who craft prompts against known rules.
- It is not a full safety stack. Evaluations, interpretability, control protocols, usage policy enforcement and deployment limits still matter. Organizations such as the Center for AI Safety and MIRI stress that alignment research goes well beyond any single training recipe.
What Constitutional AI means for users and buyers
If you use Claude Free, Pro, Max, Team or the API:
- Expect refusals and caution that track Anthropic's published priorities (safety and ethics before raw helpfulness when those conflict).
- Treat the constitution as Anthropic's stated intent, not a contract that every answer will match it.
- Compare it with ChatGPT's Model Spec if you are choosing a vendor on transparency of behavior rules.
- For career or research paths into this area, see how to get into alignment.
If you are buying for a company, ask vendors which public documents guide training, how they measure compliance with those documents, and what independent evaluations (METR style work) they publish. A constitution or Spec is a useful transparency signal. It is not by itself a security control.
FAQ
What is Constitutional AI?
Constitutional AI is Anthropic's method for training assistants with a written constitution of principles plus AI critique, revision and preference labels (RLAIF), so harmlessness training needs far fewer human harm labels than classic RLHF alone.
What is the primary goal of Constitutional AI?
Anthropic's primary research goal in the 2022 paper was a harmless but non evasive assistant: one that handles adversarial prompts by explaining objections, without depending on large human labeled harm datasets. In product terms, the goal is an inspectable way to encode helpful, honest and harmless behavior into Claude.
What is Constitutional AI in Claude?
Claude is Anthropic's assistant trained with Constitutional AI and related methods. The current values document is Claude's January 2026 constitution, which ranks being broadly safe, broadly ethical, compliant with Anthropic guidelines and genuinely helpful.
What is Constitutional AI training?
Training has two classic stages: supervised critique and revision against constitution principles, then reinforcement learning from AI feedback that ranks answers by those principles. Anthropic says the 2026 constitution is used at more stages, including synthetic data that teaches Claude the document itself.
How is Constitutional AI different from RLHF?
RLHF uses human preference labels as the main reward signal. Constitutional AI still uses RL, but for harmlessness it replaces most human harm labels with AI judgments guided by written principles. Helpfulness data can still involve humans; the 2022 harmlessness results Anthropic highlighted came from AI supervision.
Is Constitutional AI the same as Claude's constitution?
No. Constitutional AI is the training method. Claude's constitution is the principles document that method (and later training) uses for Claude. OpenAI's Model Spec is a related idea at another lab, not Constitutional AI by name.



