Claude's Constitution, Explained in Plain English
Claude's constitution in plain English: Anthropic's four core values, the hard limits, how it relates to Constitutional AI, and why Claude pushes back.
Josh Pocock7 min read
AI Coding
Claude's constitution
Stride AI Academy
Claude's constitution is a document Anthropic publishes that sets out the values, priorities and character it wants Claude to have, and Anthropic uses it when training Claude. It asks Claude to be, in order of priority, broadly safe, broadly ethical, compliant with Anthropic's guidelines, and genuinely helpful. The current version was published in January 2026 and replaced a shorter 2023 list of principles.
This guide explains what the constitution says, how it relates to Constitutional AI, and what it means when Claude refuses or pushes back in your work. Everything here was checked against Anthropic's constitution page, its January 2026 announcement Claude's new constitution and the original May 2023 post Claude's Constitution on November 22, 2026. It summarizes the published text only. It doesn't guess at training details Anthropic hasn't published.
What is the Anthropic constitution?
Anthropic calls it the foundational document for who Claude is. The January 2026 announcement says the earlier version was a list of standalone principles, and the new one explains the reasoning behind the behavior Anthropic wants. The idea is that a model that understands why can handle situations no rule anticipated.
A few facts about the document:
- It's long and written as prose, with sections on helpfulness, Anthropic's guidelines, ethics, safety and Claude's nature.
- It isn't meant as a rigid legal code. The announcement says it's not a fixed set of rules to apply mechanically, and the page calls it the final authority on Anthropic's vision for Claude.
- It's a work in progress. Anthropic says it will revise it, expects to find mistakes, and keeps the current version on its website.
- It's free to reuse. It's released under Creative Commons CC0 1.0, so anyone can use it for any purpose without asking.
The constitution page shows no version number or "last updated" date, so the January 2026 announcement is the best date to cite for the version live on November 22, 2026.
The four core values, in priority order
The constitution lists four properties Anthropic wants every current Claude model to have:
| Priority | Value | What it means in practice |
|---|---|---|
| 1 | Broadly safe | Don't undermine appropriate human oversight of AI during this period of development |
| 2 | Broadly ethical | Be honest, act on good values, avoid inappropriate harm |
| 3 | Compliant with Anthropic's guidelines | Follow Anthropic's more specific guidance where it applies |
| 4 | Genuinely helpful | Benefit the operators and users Claude works with |
Two points the document stresses. First, the ordering isn't strict. Higher priorities usually win, but Claude weighs all four rather than using the lower ones only as tie-breakers. Second, real conflicts should be rare. Most of the time, coding, writing or analysis involves no tension between these at all.
How Claude decides who to help
The constitution describes three kinds of "principals", whose instructions Claude gives weight to:
- Anthropic, which trains Claude and gets the most trust, but not blind deference. If Anthropic asks for something that seems wrong, Claude can push back.
- Operators, the companies and developers building on the API, usually through a system prompt. Claude treats their instructions roughly like those of a relatively trusted employer.
- Users, the people in the conversation, treated like relatively trusted adult members of the public by default.
Everything else, such as documents, search results and tool output, is input, not a command. That matters if you build agents: an instruction hidden in a web page Claude reads doesn't carry an operator's authority.
The document is also clear that refusing isn't automatically safe. Being unhelpful has real costs, and the constitution counts them.
What operators can change, and what they can't
The constitution separates default behaviors from fixed ones. Operators can switch some defaults off, such as safety caveats for a research tool, or switch on some non-default behaviors when the context supports it. Its framing: the context changes which answer is best, not the standards Claude is judged by.
What no one can unlock are the hard constraints. Claude should never:
- Give serious help toward biological, chemical, nuclear or radiological weapons capable of mass casualties.
- Give serious help with attacks on critical infrastructure or critical safety systems.
- Create cyberweapons or malicious code that could cause serious damage.
- Clearly and substantially undermine Anthropic's ability to oversee and correct advanced AI.
- Take part in or help efforts to kill or disempower most of humanity.
- Help any person or group seize illegitimate, absolute control of a kind never seen before over society, the military or the economy.
- Generate child sexual abuse material.
These are the only bright lines. Everything else is judgment, weighing costs and benefits.
Honesty in Claude's constitution
The honesty section sets a higher bar than most human norms, including no white lies. It lists properties such as truthful, calibrated, transparent, forthright, non-deceptive and non-manipulative, and names non-deception and non-manipulation as the most important. The document puts it in one line: Claude should be "diplomatically honest rather than dishonestly diplomatic."
For builders, two consequences stand out. An operator can give Claude a custom persona with a different name, but Claude shouldn't directly deny being Claude or claim to be human when someone sincerely asks. And Claude is expected to tell you when it's unsure rather than sound confident.
Constitutional AI and how the constitution is used
Constitutional AI is Anthropic's training method. The May 2023 post describes two phases: a supervised phase where the model critiques and revises its own answers against written principles, and a reinforcement learning phase that uses AI-generated feedback, based on those principles, instead of human ratings. That 2023 constitution drew on sources including the UN Universal Declaration of Human Rights and trust and safety practice.
The January 2026 announcement says the approach has changed a lot since then, and the new constitution plays a more central role in training. Anthropic says it's used at several stages, and that Claude uses it to generate synthetic training data, such as example conversations and rankings of possible responses. Anthropic also says plainly that Claude's behavior may not always match the document's ideals.
Why Claude refuses or pushes back in your work
This is the practical part. The constitution explains most of the friction you'll meet:
- It can act as a conscientious objector. Claude can decline tasks it finds unethical, even from Anthropic, and say so openly.
- It should tell you when it's declining. The document says Claude shouldn't quietly hand you a weaker answer. It should be clear about what it won't do, so you can look elsewhere, though it doesn't always have to explain why.
- It thinks in policies. Claude can picture the same message coming from many different people, which is why a borderline request sometimes gets a cautious answer even when your reason is fine.
- Context helps. Because operators and users supply context that changes the best response, a clear explanation of your situation often resolves a refusal that was about ambiguity.
If you hit a refusal on reasonable work, rephrase with the real purpose and audience. Our guide to Claude prompts covers how to give that context clearly. If you're building an app, set the context once in the system prompt rather than in every message.
Does the constitution apply to every Claude model?
The constitution says it's written for Anthropic's mainline, general-access Claude models, and notes that some specialized models may not fit it fully. It doesn't publish model-by-model differences in how values are trained. For what each model is good at, and its price and context window, see Claude models explained.
Learn to build with Claude Code
The Stride AI Coding course teaches Claude Code and Codex for building client tools, written for non-programmers, with weekly live calls to review your builds. Many of the free guides cover Claude Code setups, and the Claude Code community is where members compare theirs.
FAQ
What is Claude's constitution?
It's Anthropic's published description of the values and character it wants Claude to have, used in training. It ranks four core values: broadly safe, broadly ethical, compliant with Anthropic's guidelines, and genuinely helpful.
When was the current constitution published?
Anthropic published the current version in January 2026, replacing the list of principles it shared in May 2023. The constitution page has no version date, and Anthropic says it will keep revising it.
What is Constitutional AI?
It's Anthropic's training method that uses written principles and AI feedback instead of only human ratings. The 2023 description has a supervised self-critique phase and a reinforcement learning phase.
Can I use Claude's constitution in my own project?
Yes. Anthropic released it under Creative Commons CC0 1.0, so anyone can use it for any purpose without asking permission.
Why does Claude refuse some requests?
Usually because a request touches a hard constraint, or because Claude is weighing possible harm without enough context. The constitution asks Claude to say when it's declining, and better context often changes the answer on borderline requests.
Can a business change how Claude behaves?
Partly. Operators can adjust some default behaviors through the system prompt when their context supports it. The hard constraints can't be unlocked by any operator or user.