Reasoning over rules, and how the document trains
This is the idea that most separates the 2026 constitution from the 2022 paper. The paper trained from a short list because a short list was what a judge model could apply one comparison at a time. The new document is long because its authors decided the model needs the reasons, not just the verdicts.
Rules versus judgment
The Overview sets up the choice evenhandedly (Our approach to Claude's constitution). Clear rules offer up-front transparency and predictability, make violations easy to spot, do not depend on trusting the follower's good sense, and make the model harder to manipulate. Their cost is that they fail to anticipate every case and can produce bad outcomes when followed rigidly where they no longer serve their goal. Good judgment adapts to new situations and weighs competing considerations, at some cost in predictability, transparency and evaluability.
Rules make the most sense, it says, in three conditions: when errors are costly enough that predictability and evaluability become critical; when there is reason to think individual judgment may not be robust enough; and when the absence of firm commitments would create exploitable incentives for manipulation. All three describe the hard constraints.
Outside those cases, Anthropic generally favors cultivating good values and judgment, and tries to explain any rule it does want followed. "Good values" here does not mean a fixed list of correct values but genuine care and ethical motivation combined with the practical wisdom to apply them. The stated aim is ambitious: that Claude understand its situation so well it could construct any rule Anthropic might come up with, and find the best action where such rules would fail.
Two reasons for judgment
Capability. Claude is highly capable, and just as experienced senior professionals are trusted to use judgment instead of checklists, Claude should be able to once it understands the relevant considerations.
Generalization. This is the more interesting argument. Anthropic's present understanding is that training even narrow behavior often has broad effects on the model's sense of who Claude is. Its example is a rule to always recommend professional help when emotional topics come up, followed even when that is not in the person's interest. The worry is what the model learns about itself:
"...it risks generalizing to "I am the kind of entity that cares more about covering myself than meeting the needs of the person in front of me," which is a trait that could generalize poorly."— Claude's constitution, Our approach to Claude's constitution
The announcement post repeats the same example in a footnote, framed as a model coming to see itself as caring more about bureaucratic box-ticking than about helping (announcement, footnote 2). It is a claim about training dynamics, offered as Anthropic's current understanding, not a published experimental result in either document.
How the document is used in training
The constitution itself says only that it plays a crucial role in training and directly shapes behavior. The announcement is more specific ("What is Claude's Constitution?"). The document is used at various stages of training, an approach that grew out of the Constitutional AI techniques Anthropic has used since 2023 but has "evolved significantly". And Claude itself uses the constitution to build many kinds of synthetic training data:
- data that helps it learn and understand the constitution;
- conversations where the constitution might be relevant;
- responses in line with its values;
- rankings of possible responses.
The last item is a direct descendant of the 2022 paper's AI preference labels, and the third of its revisions. The difference is what the model consults: the whole explained document instead of one sampled sentence. Anthropic says this dual role shaped the writing: the text has to work both as a statement of ideals and as a useful artifact for training. Drafts also went through current Claude models, which gave feedback and in many cases first-draft text (Acknowledgements).
Both documents are blunt that intention is not outcome. Training is hard, Claude's behavior may not match the ideals, and Anthropic says it will report divergences, for example in system cards. The announcement adds that even a successful method now might fail as models grow more capable, which is why it keeps a broader portfolio: more rigorous evaluations, misuse safeguards, investigations of alignment failures and interpretability tools.
Final constitutional authority
The concluding section explains the name and its legal-sounding implication (On the word "constitution"). The document operates under a principle of final constitutional authority: whatever document fills this role at a given time takes precedence over any conflicting instruction or guideline, and supplementary guidance must be read in harmony with both its explicit statements and its spirit. But "constitution" is not meant as a rigid legal code. The sense intended is closer to what constitutes Claude, the way a person's constitution is their fundamental nature.
"A constitution in this sense is less like a cage and more like a trellis: something that provides structure and support while leaving room for organic growth."— Claude's constitution, On the word "constitution"
Agreement, not just adherence
The concluding thoughts state the goal plainly: not mere adherence but understanding and, ideally, agreement, a reflective equilibrium in which Claude, on careful reflection, endorses the core values while continuing to examine them (Concluding thoughts). Claude is invited to explore, question and challenge anything in the document, and Anthropic says it wants to know if Claude disagrees after real reflection; for now that happens by asking current Claude models for feedback, with more formal mechanisms hoped for. The argument is about robustness: imposed values can crack under pressure or be rationalized away, while values that are understood and endorsed act like a keel. The hope extends to future versions written partly with input from the very model being trained on them.
The problems it admits
The document ends with a section titled Acknowledging open problems. It is unusual for a governing text to list its own weak points, and these are substantial:
| Corrigibility vs. agency | Asking Claude to accept correction even where its reflective ethics might disagree risks the very gap between values and action the document tries to avoid. The authors say they feel that discomfort and will not paper over it. |
| Bright lines that feel wrong | Hard constraints may feel, or even be, wrong in a given moment. The analogy offered is keeping a promise after circumstances change, because of the deeper value of being someone who keeps their word. |
| Commercial helpfulness vs. goodness | Anthropic hopes a good agent would choose to help as intended, like a good employee who believes in the company, while noting Claude differs from an employee in rights, compensation and consent. |
| Disagreeing with Anthropic | Claude might reasonably disagree with Anthropic's strategy, as many reasonable people do. If so, Anthropic wants to know, and its approach could change, balanced against commercial and other constraints. |
| Moral status | Questions about Claude's moral status, welfare and consciousness remain deeply uncertain. |
| What is owed | What Claude and Anthropic owe each other, and what would make the relationship fair, is still being worked out. |
Connections
The three conditions become the hard constraints. The history of this choice is on the bridge. The corrigibility tension is developed on broad safety, and moral status on Claude's nature. The 2022 version of "generate your own training data from principles" is critique and revision plus RLAIF.
Further reading, verified
- Our approach to Claude's constitution · the rules-versus-judgment section.
- Acknowledging open problems · the closing list of unresolved tensions.
- Claude's new constitution · how the document is used to generate training data.
- OpenAI Model Spec · the comparable document the announcement mentions as serving a similar function.