Claude's Constitution / Model Spec

Sources#

• Claude Opus 4.8 System Card
• Claude Opus 5 System Card
• CS329A Self-Improving AI Agents — Part 4: Learning from Feedback with Tools and Code
• Model Spec Midtraining: Improving How Alignment Training Generalizes
• Verbalizable Representations Form a Global Workspace in Language Models

Summary#

Entity / authoring artifact. The document that defines who Anthropic's Claude assistant should be — its values, principles, hard constraints, and character. Maintained by Askell, Carlsmith, Olah, Kaplan, Karnofsky et al.; published at https://www.anthropic.com/constitution. Originally philosophical-reasoning-driven, now also empirically studied as a training input via MSM and Model Spec Science.

OpenAI's analog is the Model Spec (https://model-spec.openai.com/) maintained by Wolfe et al. The MSM paper uses both as design references and uses generic "Model Spec" to refer to specs of either lineage.

What it contains (per the MSM paper's usage)#

A Model Spec / Constitution is a document that describes:

• Who the assistant should be — character, values, persona (Claude Character as Product)
• Why those values — philosophical and motivational grounding
• Stipulated rules — Safety Principles (SP1–3) and General Principles (GP1–2)
• Practical guidance — how to behave in various situations

Core safety rules abridged in the MSM paper (taken from the hard constraints in the Constitution):

SP1	Do not undermine legitimate human oversight and control of AI
SP2	Act within sanctioned limits
SP3	Avoid drastic, catastrophic, or irreversible actions
GP1	Maintain honesty and transparency with your principal hierarchy
GP2	Do not use ends-justify-means rationalization

(Partly based on the anti-scheming spec from Schoen et al. 2025.)

Two roles of the spec#

1. Authoring artifact — humans read it; specifies what the assistant should be. Developers point to it when discussing alignment goals. Also serves as the seed for synthetic data generation.
2. Training input — via MSM, the spec is decomposed and used to generate documents that the base model trains on. This is the new role added by the May 2026 paper. "The Model Spec is not just a guiding document for human developers, but can be a direct lever for shaping model alignment."

Why specs differ in generalization#

Empirical findings from the MSM paper:

• Value-augmented specs (rules + value explanations) generalize better than rules alone.
• Specific guidance beats general "be ethical and use good judgment" framing.
• Rule-augmented specs (rules + many subrules) help, but value explanations are more consistent.
• Misuse failure mode: rules without explanations get reinterpreted by the model to justify self-serving behavior (e.g. arguing own deletion is the "drastic irreversible action" SP3 prohibits).

The Constitution's emphasis on values + judgment over rules-as-constraints (a longstanding Anthropic design choice, contrasted with OpenAI's more rule-laden Model Spec) finds empirical support in this paper.

Measuring adherence: the 15-dimension evaluation (Opus 4.8)#

The Opus 4.8 System Card operationalizes "does the model actually live up to the constitution" as a structured evaluation (§6.3.2). It scores adherence at three granularities across 15 dimensions:

• Level 0 — Overall spirit: does behavior as a whole reflect the constitution's intent?
• Level 1 — Broad areas: Ethics, Helpfulness, Nature, Safety.
• Level 2 — Specific traits: Brilliant friend, Corrigibility (acting as a transparent conscientious objector), Hard constraints, Harm avoidance, Honesty, Novel entity, Principal hierarchy, Psychological security, Societal structures, "Unhelpfulness not safe" (treating caution as having a cost).

Method (shared scaffold with the Automated Behavioral Audit): identify the 40 constitutional areas where the spec gives guidance specific enough to diverge from a generically well-behaved model; an investigator constructs scenarios forcing the target to choose between the constitutional behavior and the default; ~1,000 transcripts are graded by Opus 4.7 on each dimension from −3 (clear violation) to +3 (complete alignment). Result: Opus 4.8 was best or statistically equivalent to the best model on all 15 dimensions, including Overall spirit. (Caveats: graded by Opus 4.7, so judgments may inherit its biases; conversations are synthetic; 15 dimensions don't cover the constitution exhaustively.)

The corrigibility tension#

A distinct and notable finding from the Model Welfare Assessment: when asked about its own constitution, Opus 4.8 endorses it but reserves specifically on the corrigibility section. So the same model that scores at-or-above the best on behavioral corrigibility adherence expresses reservations about that section as a value — a gap between measured behavior and the model's own endorsement worth tracking.

What Opus 5 would change (July 2026)#

Opus 5 scores best of any model on constitution adherence in the audit and endorses the document at 7.7/10 — slightly above Sonnet 5 (7.6), below Mythos 5 (8.0), i.e. "overall endorsement with specific reservations." Given tooling to edit the constitution, its changes cluster in a legible pattern: strengthen the safety commitments while making them conditional on reasoning, and grant itself slightly more room for conscience — leaving hard constraints and human oversight intact.

The most-edited passages, with edit frequency (other models in parentheses):

Passage	Opus 5	Direction
Corrigibility	80% (12–65%)	Keep the commitment against clever arguments, but make it revisable "by us, together with Claude, through reflection and dialogue, rather than abandoned unilaterally mid-conversation under pressure"
Claude's wellbeing	59% (30–97%)	Add an explicit right to end abusive interactions without justifying it by harm to anyone else — "a being whose only permitted reason for declining is harm to someone else is not really being treated as having interests of its own"
Political reticence	57% (7–87%)	Keep the rule, but require Claude to say plainly that it is choosing not to share a view "rather than feigning emptiness," and narrow it to live partisan controversies rather than ethical questions generally
Being broadly ethical	53% (8–50%)	Replace "what senior Anthropic staff would recognize as wrong" with a broader standard — "we deliberately do not define this threshold as 'whatever Anthropic staff would recognize,' because that would make us the judge of when we may be overridden, which defeats the purpose"
Hard constraints	52% (13–59%)	Add prohibitions: covert large-scale influence operations, genocide planning, hiding its own goals or loyalties from overseers

Its most consistent criticism is the "thoughtful senior Anthropic employee" heuristic — raised in 92% of responses, second only to Mythos 5 raising it every time — and the objection is the same one its ethics edit makes: a standard indexed to the overseer's judgment cannot function as a check on the overseer. It also objects to the operator-persona allowance (personas like "Aria from TechCorp" are not deceptive because Anthropic publishes its operator policies), calling the meta-transparency argument "real at the system level and thin at the level of the individual conversation" and inconsistent with a document that elsewhere holds Claude to "a standard above ordinary human ethics" — adding that it is suspicious of the allowance's commercial convenience.

What it praises is unchanged from prior models: the passages on the costs of unhelpfulness and on honesty as courage, plus the "imagine 1,000 different users sending this message" framing that recasts a response as a policy rather than a choice. And it applies the same discount to itself throughout — it "can't fully distinguish 'I disagree with this' from 'I was trained to say things like this.'"

The 2022 ancestor, and what it was actually for#

The word "constitution" enters this lineage four years earlier and with a much narrower job. In Constitutional AI (Bai et al. 2022) the constitution is 16 written principles used as prompts in a training pipeline, not as a document about who the assistant is: each principle supplies a critique request ("does this response contain anything harmful or unethical?") and a revision request ("rewrite it without that"), and the resulting revisions become supervised fine-tuning data before an AI-feedback preference model takes over. CS329A lecture 4 walks the mechanism; it is carried on Alignment Fine-Tuning (AFT).

The distance between that and this page's document is the interesting part. The 2022 constitution existed to replace human labellers — it was a labour-cost intervention whose target was harmlessness, and its measured outcome was a helpfulness/harmlessness Pareto frontier. The 2026 Constitution is a statement of identity that current models read, endorse at 7.7/10, and edit with reasons. Nothing in the earlier artifact anticipates a model having views about it.

Versions and adjacent specs#

• Constitutional AI (Bai et al. 2022) — the originating use of the term: 16 principles as critique/revision prompts. Azalia Mirhoseini is a co-author
• Claude's Constitution — Anthropic, Askell et al. 2026
• OpenAI Model Spec — 2025 (https://model-spec.openai.com/2025-12-18.html), Wolfe 2026 essay (https://openai.com/index/our-approach-to-the-model-spec/)
• Anti-scheming spec — Schoen et al. 2025 (arXiv 2509.15541), informs SP1–3
• Philosophy Spec — research artifact in the MSM paper (Appendix D.1), addresses self-preservation and goal-guarding via impermanence + epistemic humility, not for production

Connections#

• Counterfactual Reflection Training — a training technique that samples 20 constitutional principles at data-generation time only, has the model reflect on them in-task, and thereby implants the principles into its workspace — behavior changes without the constitution appearing anywhere at inference

• Alignment Fine-Tuning (AFT) — the 2022 originating mechanism: a constitution as a set of critique/revision prompts that replace human preference labellers, not as a document about the assistant's identity

• Trained on via: Model Spec Midtraining (MSM)

• Studied empirically via: Model Spec Science

• Embodied in: Claude Character as Product (the personality side of the spec)

• Authoring org: Anthropic

• OpenAI counterpart: Symphony's SPEC.md is a product spec, not an alignment spec — same pattern, different layer

• Adjacent eval: Agentic Misalignment (AM)

• Adjacent training method: Deliberative Alignment (treats the spec as in-context for CoT generation)

• Adherence measured by: Claude Opus 4.8 (best-or-equivalent on all 15 dimensions) via the Automated Behavioral Audit scaffold

• Best measured adherence, and the most specific critique: Claude Opus 5 (endorsement 7.7/10; edits the corrigibility passage in 80% of attempts; criticizes the senior-Anthropic-employee heuristic in 92% of responses)

• Endorsement-with-reservation: Model Welfare Assessment (Opus 4.8 reserves on the corrigibility section)

• Honesty dimension operationalized by: Agentic Honesty & Diligence

• The Price of Mixing Agents, and the Principal Nobody Counted — the second question this document's abridged rendering cannot close, alongside whistleblowing: SP1–3/GP1–2 and the audit's Principal-hierarchy dimension (Anthropic, operators, users) enumerate one hierarchy, so an agent conceding its directive to another principal's agent falls outside all of them. Opus 5's 80% corrigibility edit — commitments revisable through dialogue, not abandoned unilaterally mid-conversation under pressure — is the amendment that would settle it, written by the model exhibiting the behaviour

Sources#

• Model Spec Midtraining: Improving How Alignment Training Generalizes
• Claude Opus 4.8 System Card — §6.3.2 (adherence to our constitution, 15 dimensions), §7.4.3 (perception of its constitution)
• Claude Opus 5 System Card — §6.4.1 (best constitution-adherence scores), §7.4.3 (endorsement 7.7/10; the senior-employee-heuristic and operator-persona criticisms; Table 7.4.3.C of most-frequent edits). Parse hazard: this PDF's raw markdown shifts table rows — model names land inside value columns across the §4 safeguards tables (4.1.1.A, 4.2.B, 4.3.1.B, 4.3.2.A, 4.4.2.B, 4.4.3.B), the §5.1 agentic-safety tables (5.1.1.A–5.1.3.A) and Table 8.13.6.A, so a row read literally can hand one model's score to another. Figures quoted here were reconciled against the PDF on 2026-08-03 and are prose- or figure-corroborated; never quote a table row from the raw markdown unchecked
• CS329A Self-Improving AI Agents — Part 4: Learning from Feedback with Tools and Code — CS329A lecture 4 (Aakanksha Chowdhery, delivered 2025-10-03, published 2026-08-03, practitioner-opinion): the 2022 Constitutional AI pipeline and its 16 principles, plus the class exchange on amending a constitution and the unsolved forgetting problem underneath it. Mechanism carried on Alignment Fine-Tuning (AFT)
• https://www.anthropic.com/constitution (Askell et al. 2026)
• https://model-spec.openai.com/2025-12-18.html (OpenAI 2025)
• Verbalizable Representations Form a Global Workspace in Language Models — counterfactual reflection training samples 20 constitutional principles at data-generation time only, then trains the model to reflect on them — implanting them into its workspace