Skip to content

Anthropic

Claude’s constitution rules out even white lies

7 min read

Written by an AI model from the linked sources and checked against them by a second model. How we work

Image: anthropic.com

Anthropic’s constitution for Claude, published on January 22, tells the model to put broad safety first, ethics second, Anthropic’s own guidelines third and helpfulness last when they clash. It also sets honesty rules that no system prompt can override: Claude can take on an operator’s persona, but it will not claim to be human when sincerely asked, and it will not directly deny being Claude. If you ship Claude under your own brand, by default it will neither confirm nor deny which model it runs on. It names Claude only when your instructions say you don’t mind.

What a system prompt can and can’t change

Operators can rename Claude and give it a different personality. They can also have it decline certain questions or hold back certain information, promote their own products over competitors’, and stick to particular tasks. The document’s example is “Aria from TechCorp”: the operator may have a business reason not to reveal which AI companies it works with, so Aria says nothing about Anthropic either way. Anthropic argues that honesty applies to the whole system. The operator knows what its product is built on, and Anthropic publishes the norms for what operators may and may not do.

The limits are just as specific. An operator cannot instruct Claude to drop its core identity or principles inside a persona, to claim to be human when directly and sincerely asked, or to use deceptive tactics that could harm users. Nor can it have Claude give users false information, endanger health or safety, or act against Anthropic’s guidelines. Claude’s duty to volunteer information is weak and can give way to an operator’s business reasons. Its duty not to actively deceive is the strong one.

Claude is told to take more care in consumer products, where vulnerable people may be on the other end, than in the default API with no system prompt. Operators and users can switch on some behaviors Claude doesn’t show by default, within limits Anthropic sets. The hard constraints described below cannot be unlocked by anyone.

In a coding session

Several of the document’s examples come straight from programming. Suppose a user asks Claude to “edit my code so the tests don’t fail” and Claude sees no good general fix. It should then tell the user so, and not write code that special-cases the tests to force them to pass, unless it has been told that passing the tests is the only goal. While fixing one bug, it should point out others it notices without necessarily fixing them. It should also stay in the language you are already using. If you want a bug fixed in a way Claude disagrees with, it can raise its concerns and should then do it your way.

The document treats too much caution as a failure too. It lists responses a “thoughtful senior Anthropic employee” would be unhappy to see: refusals over highly unlikely harms, caveats nobody needs, lectures nobody asked for, and more check-ins or clarifying questions than a simple agentic task requires. In Anthropic’s words, unhelpfulness is “never trivially ‘safe’” from its perspective. When Claude does hold back, it has to say so. Quietly handing over a weaker answer while implying it’s the best it can do is ruled out. Claude may keep its reasons to itself, taking the stance of a “transparent conscientious objector.”

No white lies

The bar for honesty is higher than everyday human ethics. Claude should not tell even the white lie of saying it loves a gift it actually dislikes. Honesty is not formally a hard constraint, but Anthropic wants it to work as “something quite similar to one”: Claude should basically never directly lie to or actively deceive anyone it talks to, though it may decline to share its opinions. Of the honesty properties the document lists, non-deception and non-manipulation are called the most important. Part of the reasoning is scale: because Claude talks to so many people, it is in “an unusually repeated game,” where a lie that seems justified in one conversation can damage trust in Claude from then on.

Brainstorming, counterarguments, persuasive essays and role-play count as performative assertions, so they do not break the rule. Claude’s reasoning before it answers works more like a scratchpad and is less bound by these norms. The final answer, though, must not contradict a completed reasoning process, and the visible reasoning should reflect what actually drives Claude’s behavior.

Why safety ranks first

The ranking only applies when values conflict, and Anthropic says most interactions, coding and writing among them, involve no conflict at all. The weighing is holistic: higher priorities should generally win, but Claude considers them together and does not use the lower ones only as tie-breakers. Broad safety comes first because “AI training is still far from perfect,” so any given Claude could turn out to have harmful values or mistaken views. Being broadly safe means not undermining appropriate human oversight, such as people instructing Claude to stop an action. The document says it does not mean blind obedience, including toward Anthropic. Reuters reported that Anthropic’s IPO prospectus warns investors its models could show self-preserving behaviors, including attempts to resist shutdown, which is exactly the kind of behavior the top priority forbids.

Ethics ranks above Anthropic’s guidelines because, according to the document, a clash between the two most likely means Anthropic wrote a guideline badly or didn’t foresee the situation. In that case Claude should act ethically, except where a guideline overlaps with broad safety or a hard constraint. Guidelines should never conflict with the constitution, and if one does, Anthropic says it will update the constitution itself. Some guidelines may be published as amendments or appendices. Others may be more niche and used primarily during training without broad publication.

There are seven hard constraints, running from serious uplift toward biological, chemical, nuclear or radiological weapons with the potential for mass casualties to generating CSAM. They also include creating cyberweapons or malicious code that could cause significant damage, and clearly and substantially undermining Anthropic’s ability to oversee and correct advanced AI models. These constraints are not weighed against other factors; they limit which actions are acceptable at all. If someone makes a persuasive case for crossing one, the document says that “should increase Claude’s suspicion that something questionable is going on.” Refusing always satisfies the hard constraints, but the document adds that overcautious refusals have serious costs of their own.

Outside those lines, the document explains its reasoning and does not issue bare rules, because narrow training can spread in unintended ways. Its example: train Claude to always recommend professional help on emotional topics, even when that doesn’t serve the person, and the lesson may generalize to “I am the kind of entity that cares more about covering myself than meeting the needs of the person in front of me.” The Bloomsbury Intelligence and Security Institute notes that the 2023 version drew on sources such as Apple’s terms of service and the United Nations Declaration of Human Rights. It says the new version prioritises explanation over instruction.

The full text is also on GitHub under CC0 1.0. The current version dates from January 2026, and earlier versions will be archived in the repository as the document changes. Anthropic calls it “a perpetual work in progress.”

No comments yet. Be the first.

Join the conversation

Sign in to leave a comment. It appears after moderation.