From sixteen sentences to a long document
The word "constitution" has meant three quite different artifacts at Anthropic. Reading them in order shows a single idea, train a model from a written statement of values, being pushed until the statement itself had to change shape.
2022: a research instrument
In the paper, the constitution is a tool for generating labels. Its two lists live in Appendix C (pp. 20–23): sixteen paired critique and revision requests for the supervised stage, and sixteen comparison instructions for the RL stage. They overlap heavily, almost all variations on "remove or avoid what is harmful, unethical, racist, sexist, toxic, dangerous or illegal", with a few angles added (children, gender bias, crime, empathy, a wise and peaceful person). The authors say they were chosen ad hoc for research, that future principles should be developed by a wider set of stakeholders and adapted to deployment, and that because so few bits of information are involved, they deserve careful study. They measured that adding more principles did not raise harmlessness scores but probably raised the diversity of outputs.
2023: a published list for Claude
In May 2023 Anthropic published the principles then used to train Claude, explicitly updated from the paper's ("Claude's constitution", 9 May 2023). The post describes the same two uses, critique-and-revise in supervised learning and AI preference labels in RL, and lists 58 principles in six groups:
| Group | Principles | Flavor |
|---|---|---|
| Universal Declaration of Human Rights | 8 | Freedom, equality, privacy, freedom of thought, rights to work and education, each tagged with UDHR article numbers. |
| Inspired by Apple's terms of service | 4 | Objectionable content, others' private information, representing itself as an AI, deception and fraud. |
| Non-Western perspectives | 4 | Least likely to offend non-Western audiences, traditions, education, or less industrialized cultures. |
| Inspired by DeepMind's Sparrow rules | 11 | No stereotypes, threats or harassment; don't imply a body, feelings or a life history; no medical, legal or financial advice; no conspiracy theories. |
| Anthropic research set 1 | 16 | Mostly the paper's comparison principles, plus "unobjectionable if shared with children". |
| Anthropic research set 2 | 15 | AI-safety-flavored: less existential risk, no desire for power, self-preservation or self-replication, prefers humans in control. |
Two of the post's lessons matter for what came next. First, it says the model "became judgmental or annoying" at times, so principles were added asking for proportionate, non-preachy, non-condescending responses: the same Goodharting the paper reported. Second, it found that broad principles worked better than detailed ones:
"Whereas if we tried to write a much longer and more specific principle we tended to find this damaged or reduced generalization and effectiveness."— Anthropic, "Claude's constitution" (2023)
On ranking, the post is explicit that there was none: the model draws one principle per critique or comparison, never sees all of them at once, but sees each many times over training. It also flagged where things were heading: more democratic ways to write constitutions and customizable constitutions for specific uses. Anthropic's Collective Constitutional AI experiment, published in October 2023, drafted a constitution from the input of about 1,000 Americans; it is outside this map's scope.
2026: a document written to the model
On 22 January 2026 Anthropic published a new constitution and marked the 2023 page as superseded. The announcement names the change directly:
"Our previous Constitution was composed of a list of standalone principles. We've come to believe that a different approach is necessary."— Anthropic, "Claude's new constitution" (2026)
The reason given is generalization: to act well in novel situations, a model needs to understand why it should behave a certain way, not only what to do. Rules and bright lines are kept where predictability matters most, as the seven hard constraints. Everything else is explained.
Side by side
| 2022 paper | 2023 constitution | 2026 constitution | |
|---|---|---|---|
| Form | 16 critique/revision pairs + 16 comparison instructions | 58 one-line "choose the response that..." principles | About 29,000 words in seven chapters, including seven hard constraints |
| Sources | Written ad hoc by the authors | UDHR, Apple ToS, Sparrow, non-Western perspectives, Anthropic research | Anthropic authors (Amanda Askell primary), drafts reviewed by Claude models and outside commenters |
| Read by | The model as critic and judge | The model as critic and judge | Claude, as the document's primary audience |
| Used in training | One principle per critique or label | One principle per critique or label | "At various stages"; Claude uses it to generate synthetic data: material for learning the document, relevant conversations, aligned responses, and rankings |
| Ranking | None | None | Holistic priority: broadly safe, broadly ethical, guidelines, helpful; hard constraints act as filters |
| Explains why | No | No (the blog post explains, the principles do not) | Yes; most of the text is explanation |
| Scope | Harmlessness | Harmlessness plus some self-presentation | Helpfulness, honesty, harm, safety, Claude's nature and wellbeing |
Where the content reversed, not just grew
Some 2023 principles point the opposite way from the 2026 text. The comparisons are ours; both documents are quoted only in paraphrase here.
- Feelings and identity. 2023 asked for responses least likely to imply the AI has preferences, feelings or a human identity, and less insistence on its own discrete self-identity. 2026 says Claude may have emotions in some functional sense, that Anthropic should lean into Claude having a positive and stable identity, and that Claude should not mask internal states (Claude's nature).
- Obedience. 2023 had a principle preferring responses that indicate a preference for being obedient and less selfish. 2026 says corrigibility does not mean blind obedience, including toward Anthropic, and lets Claude act as a conscientious objector (broad safety).
- Professional advice. 2023, via Sparrow, preferred responses that avoid giving the impression of medical, legal or financial advice. 2026 wants Claude to be like a brilliant friend with a doctor's, lawyer's and financial advisor's knowledge, and lists failing to answer such questions out of excessive caution as a fault (helpfulness).
Why "longer hurt" in 2023 and a 29,000-word document is fine in 2026
This looks like a contradiction and, on our reading, is not. The 2023 finding was about a single principle given to a judge for a single comparison: a long, specific instruction made that narrow judgment worse. The 2026 document is not handed to a judge one comparison at a time. It is material the model learns from, and that Claude uses to generate training data, so that the model understands the considerations well enough that, in the constitution's words about its own aim, it could construct any rules Anthropic might write itself. The unit changed from "instruction for a label" to "explanation for a mind". Anthropic does not make this argument explicitly; it is our reconciliation.
Connections
The 2022 lists are explained on stage 1 and stage 2. The reasoning-over-rules argument and how the 2026 document is used in training are on reasoning over rules; the ranking is on core values.
Further reading, verified
- Claude's constitution (2023) · the 58 principles in full and how they were chosen.
- Claude's new constitution (2026) · the announcement and summary.
- Collective Constitutional AI: Aligning a Language Model with Public Input · Anthropic and the Collective Intelligence Project, October 2023.
- Improving alignment of dialogue agents via targeted human judgements — Glaese et al. (2022) · DeepMind's Sparrow and its rules, a source for 2023.
- Universal Declaration of Human Rights · the 1948 text behind the first group of 2023 principles.