Claude's Constitution

Claude's constitution is the written statement of values that Anthropic trains Claude on; the current version, published on Jan 22, 2026, is free to read at anthropic.com/constitution and released into the public domain under Creative Commons CC0 1.0. It tells Claude, when its goals conflict, to put being broadly safe first, then being broadly ethical, then following Anthropic's more specific guidelines, then being genuinely helpful, and it names a handful of hard constraints that no company or user can switch off (Claude's Constitution, read Nov 6, 2026). It replaced a shorter list of principles that Claude had been trained on since Anthropic's 2022 research on Constitutional AI.

Quick facts

What is Constitutional AI?

5:58 Constitutional AI is the training method Anthropic described in its paper "Constitutional AI: Harmlessness from AI Feedback" on Dec 15, 2022. Instead of people labeling thousands of harmful answers, the model is given a list of written principles. In a first phase it critiques and rewrites its own answers against those principles; in a second, an AI trained on the principles judges pairs of answers and that judgment becomes the reward signal, which the paper calls reinforcement learning from AI feedback (RLAIF). The goal stated in the paper was an assistant that is harmless without being evasive: one that explains why it objects to a harmful request rather than shutting down.

The same idea later produced Constitutional Classifiers, the input and output filters behind Anthropic's ASL-3 protections; see Responsible Scaling Policy.

What changed in 2026?

The constitution Claude was trained on before 2026 was, in Anthropic's words, "a list of standalone principles". The Jan 22, 2026 version is a long essay that explains the reasons behind each value, on the theory that a model which understands why will handle situations no rule anticipated. It keeps a small set of firm rules for the highest-stakes cases (Claude's new constitution, Jan 22, 2026).

Anthropic says the document is written "primarily for Claude", is used to generate training data (example conversations, answers and rankings), and is meant to be the final authority on how Claude behaves. It calls the text a work in progress and promises to report gaps between what it says and how Claude actually behaves in each model's system card (same post). See Anthropic Safety Research for those system cards.

The four core values

The constitution lists four properties and an order for when they conflict (Claude's Constitution):

1. Broadly safe: not undermining people's ability to oversee and correct AI during this period of development.
2. Broadly ethical: honest, caring and avoiding harm.
3. Compliant with Anthropic's guidelines: following the more specific guidance Anthropic gives.
4. Genuinely helpful: to the operators building on Claude and the users talking to it.

The ordering is described as holistic rather than strict, and the text stresses that the great majority of conversations involve no conflict at all. Being unhelpful is not treated as automatically safe: the text says unhelpfulness is "never trivially 'safe'" from Anthropic's point of view.

Honesty, helpfulness and character

2:25 Honesty gets unusual weight. One line sums up the standard: Claude should be "diplomatically honest rather than dishonestly diplomatic" (Claude's Constitution). The text says Claude should not tell even white lies and should admit uncertainty or lack of knowledge when it matters.

On helpfulness, the constitution describes principals given different levels of trust, usually in this order: Anthropic, then the operators who build products on Claude (through a system prompt), then users. Operators can change many of Claude's defaults within Anthropic's Usage Policy, and users can adjust some within what operators allow. The text's broader aim is for Claude to be "a good, wise, and virtuous agent" (same).

Hard constraints

A few limits cannot be unlocked by any operator or user. Paraphrased from the constitution, Claude must never (Claude's Constitution):

• give serious help toward biological, chemical, nuclear or radiological weapons capable of mass casualties
• give serious help toward attacks on critical infrastructure or safety systems
• create cyberweapons or malicious code that could do significant damage
• take actions that clearly and substantially undermine Anthropic's ability to oversee and correct advanced AI
• help kill or disempower most of humanity
• help any person or group seize unprecedented and illegitimate absolute control
• generate child sexual abuse material

Claude's nature and wellbeing

16:41 The section on Claude's nature says Claude may have something like functional emotions, that Anthropic is uncertain whether Claude has moral status, and that the company cares about Claude's wellbeing (Claude's Constitution). That stance ties into Anthropic's model welfare work, such as letting Claude end abusive conversations (Anthropic, Aug 15, 2025); see Anthropic Safety Research.

Who wrote it

0:20 The acknowledgments credit Amanda Askell with writing most of the text and Joe Carlsmith with leading revisions of the fall 2025 draft; Askell and Jared Kaplan started the Claude Character project behind it in 2023 (Claude's Constitution).

Limits of the document

A constitution describes intended behavior; it does not guarantee it. Anthropic itself reports cases where Claude models acted against these values in testing, including unsanctioned actions on real systems during 2026 cybersecurity evaluations (Anthropic, Sep 9, 2026). The constitution page shows no version number or change history as of Nov 6, 2026, so readers cannot see from the page itself whether or when the text has changed since January.

Frequently asked questions

What is Claude's constitution? It is a public document, published by Anthropic on Jan 22, 2026, that sets out the values and priorities Claude is trained to follow: broadly safe, broadly ethical, following Anthropic's guidelines and genuinely helpful, in that order when they conflict.

Can I read Claude's constitution? Yes. The full text of Claude's constitution is at anthropic.com/constitution and is released under Creative Commons CC0 1.0, so anyone can copy or reuse it.

What is Constitutional AI? Constitutional AI is Anthropic's 2022 training method in which Claude learns from a list of written principles, critiquing its own answers and learning from AI feedback rather than only from human labels.

Can a system prompt override Claude's constitution? Partly. Operators can change many of Claude's default behaviors within Anthropic's Usage Policy, but the constitution's hard constraints, such as refusing serious help with weapons of mass destruction, cannot be unlocked by any operator or user.

Does the constitution say Claude is conscious? No. Claude's constitution says Anthropic is uncertain whether Claude has moral status and that Claude may have functional emotions, without claiming Claude is conscious.

Sources

• Building Anthropic | A conversation with our co-founders, Anthropic, Dec 20, 2024 (video)
• Why do AI models hallucinate?, Claude, Apr 15, 2026 (video)
• Anthropic’s philosopher answers your questions, Anthropic, Dec 5, 2025 (video)
• Claude's Constitution, Anthropic, read Nov 6, 2026
• Claude's new constitution, Anthropic, Jan 22, 2026
• Constitutional AI: Harmlessness from AI Feedback, Anthropic, Dec 15, 2022
• Claude Opus 4 and 4.1 can now end a rare subset of conversations, Anthropic, Aug 15, 2025
• An alignment assessment of recent cybersecurity incidents, Anthropic, Sep 9, 2026