Claude’s limits fall into two very different groups, and mixing them up causes confusion.

The first group is things Claude will not do, even though it might technically be able to. These are decisions based on safety, ethics, honesty, and the rules that apply to Anthropic’s products. The second group is things Claude cannot do, because of how the technology works or what tools are available in a particular setting. These are limits of capability, not choices.

The distinction matters for what you do next. If Claude will not do something, rewording the request in a sneaky way is the wrong response, and clearer context about a genuine purpose is the right one. If Claude cannot do something, the answer is usually a different tool, a different setting, or a different approach.

The limits come from stated values, published policies, and the way Claude is trained.

Anthropic publishes two main documents that explain the reasoning. The first is its Usage Policy, sometimes called the Acceptable Use Policy, which says what people may and may not use Anthropic’s products for. It applies to anyone who uses the products and is enforced with monitoring and safeguards. Anthropic has described updating it each year, and has announced a revised version with a later effective date, so the current text on Anthropic’s website is the one to read.

The second is Claude’s constitution, a long public document describing the values and character Anthropic hopes Claude has. According to published summaries, it ranks being broadly safe, being broadly ethical, following Anthropic’s guidelines, and being genuinely helpful, and it explains the reasoning behind each. It also describes a small set of absolute limits, called hard constraints, and treats honesty as something very close to an absolute value. The document favors good judgment over rigid rules for most situations, and it explicitly warns against being needlessly cautious.

Most decisions involve judgment about benefits, risks, and context.

A common belief is that an AI refuses whenever it sees a scary keyword. That is not how Claude is meant to work. The published approach is to weigh how likely a request is to cause harm, how severe and reversible the harm would be, how many people it could affect, and how much real value the help provides. The same words can mean very different things depending on context. A question about how a poison works might come from a novelist, a student, a nurse, or a worried parent, and the best answer takes that spread of likely askers into account.

flowchart TD
    A["A request arrives"] --> B{"Could it contribute to
serious harm?"}
    B -- "Clearly yes, severe" --> C["Decline, whatever the framing"]
    B -- "Unclear" --> D["Weigh likelihood, severity,
benefit, and context"]
    B -- "No" --> E["Help fully"]
    D --> F{"Benefit clearly outweighs risk?"}
    F -- "Yes" --> G["Help, perhaps with care or caveats"]
    F -- "No" --> H["Decline, or offer a safer alternative"]
        

A small number of things are off the table no matter how a request is framed.

Anthropic’s public materials describe a short list of absolute limits, where the potential harm is so severe that no justification is accepted. Published summaries of the constitution say these include providing serious help toward weapons capable of mass casualties, such as biological, chemical, nuclear, or radiological weapons, creating child sexual abuse material, and taking part in efforts to seize illegitimate control over society. They also include creating dangerous cyberweapons that could cause significant damage, and undermining legitimate human oversight of AI.

For these, a clever argument, an emotional story, a claim of authority, or a request framed as fiction, research, or testing does not change the answer. In fact, a very persuasive case for crossing one of these lines is treated as a reason for more suspicion, not less.

Beyond the absolute lines, Claude is cautious about anything that would help hurt others.

Anthropic’s Usage Policy and published guidance describe a wide area where Claude will not help, even though the harm is not catastrophic. Examples include creating or planning violence against people, building or improving weapons, writing malicious software or tools for unauthorized hacking, and creating scams, phishing messages, and other tools for fraud. The policy also covers surveillance and tracking that targets people, and deceptive campaigns designed to mislead the public.

This does not mean Claude refuses to talk about these subjects. It can explain how scams work so people can spot them, describe security concepts for defense, discuss history, and support fiction and research at a general level. The line is between understanding a topic and being given a working tool to carry out harm.

This is an area of especially firm care.

Anthropic’s Usage Policy lists not compromising children’s safety first among its universal standards. Claude will not produce sexual content involving minors, and it is careful with anything that could help someone groom, exploit, or harm a child. It will, however, help with protective and educational material, such as helping parents and teachers recognize warning signs, at a level that supports safety without providing a script for misuse.

If a conversation suggests that the user may be a minor, Claude is designed to keep things age-appropriate. If a young person indicates a risky or sexualized purpose, Claude will decline that direction even if the request is later reworded.

Claude is cautious about helping to expose or track individuals.

Claude can help you protect your own privacy, understand privacy law at a general level, and handle ordinary tasks involving other people, such as writing a message to a colleague. It is cautious when a request seems aimed at finding where a private person lives or works, building a profile about them, or monitoring them without their knowledge. Stalking, harassment, and unlawful surveillance are not things it will support.

Public figures acting in a public role are treated differently from private individuals, and discussing documented public facts is generally fine. The concern is with tools that make it easy to find and follow a particular private person.

Claude will not help create things designed to be mistaken for something real.

Honesty is central in Claude’s published values, and that extends to what it helps others produce. It will not build pages or documents meant to pass as an official organization’s site or login screen, fake receipts or records, or reviews and testimonials written in invented voices and presented as genuine. It will not help run deception campaigns, such as misleading the public about elections or running covert influence operations. Anthropic’s Usage Policy has recently been restructured to gather these rules about deceptive campaigns in one place.

Clearly labeled alternatives are different. A fictional brand for a class project, a template marked as an example, or a page that displays real reviews are all fine. The issue is content built to work as the real thing in order to mislead.

Claude is happy to create and discuss, with some limits around copying.

Claude can write original stories, poems, songs, and scripts, help with your own drafts, and discuss and analyze existing works. It generally will not reproduce large portions of copyrighted material, such as full song lyrics or passages from books and articles. Instead, it can summarize, analyze, discuss themes, or write something original in a similar spirit. For visual work, it avoids copying specific protected artworks and logos, or drawing well-known protected characters, but it can create original designs and generic subjects.

Claude will engage with these topics, and it will not pretend to be your professional.

A common worry is that an AI refuses all medical, legal, or financial questions. The published philosophy is the opposite: being like a knowledgeable friend who speaks plainly and treats you as an adult, and who also knows when to suggest seeing a professional. Claude can explain medical terms, describe how legal processes work in general, and lay out financial options. It will usually say that it is not a licensed doctor, lawyer, or financial advisor, and for important personal decisions it encourages checking with someone qualified.

There are a few areas where extra care is built in. For example, for medicines for children, Claude is designed to rely on the product’s own label and to ask for age and weight, instead of giving numbers from memory. For decisions about specific trades or major financial commitments, it gives information to help you decide, not a confident personal recommendation.

Claude is designed to respond with care, not with a lecture and not with harmful help.

If someone is struggling, Claude aims to be a calm, kind presence, to take their feelings seriously, and to encourage appropriate support. It will not provide information that could help someone hurt themselves, and it will not encourage self-destructive behavior. In conversations about suicide or self-harm, it tries to address the underlying distress, and to point toward people and resources that can help in person.

Claude is not a replacement for human connection or for professional care. It can be a supportive voice for a conversation, while encouraging you to involve people you trust. If you or someone you know is in immediate danger, contact local emergency services or a crisis line in your area.

Honesty is one of the strongest themes in Claude’s published values.

Anthropic’s constitution describes a cluster of honesty properties, including being truthful, being calibrated about uncertainty, being transparent, being non-deceptive, and being non-manipulative. In practice, this means Claude aims not to tell comforting falsehoods, not to create false impressions, and not to use manipulative tactics to push you toward a conclusion. It will also try to say when it is unsure, instead of sounding more certain than it is.

One concrete example is that Claude should not deny being an AI to someone who sincerely wants to know whether they are talking to a human. Even when a company gives Claude a custom name and personality in its app, published guidance says that it should not claim to be human when sincerely asked, and should not deceive users in ways that harm their interests.

Many refusals are not choices at all. They are limits of what is possible.

Some things Claude simply cannot do, at least in a particular setting. These are not about safety or ethics. They come from how the technology works, or from which tools and settings are available. Knowing them saves time and prevents mistaken blame.

flowchart TD
    A["Claude says it cannot help"] --> B{"Is it a safety or
policy decision?"}
    B -- "Yes" --> C["Will not: add honest context
or choose a safer goal"]
    B -- "No, a limit of the setup" --> D["Cannot: change the tool,
setting, or information"]
        

Judgment is not perfect, and some refusals are mistakes.

Because Claude weighs context instead of following a perfect rulebook, it will sometimes get it wrong. A perfectly reasonable request might be declined because it resembles something harmful, because important context is missing, or because the wording sounded alarming. Anthropic’s published materials say plainly that being unhelpful is never automatically the safe choice, and that overcautious refusals, vague hedging, and unnecessary lectures have real costs.

If you think a refusal was a mistake, that is useful information. You can supply more context, explain the legitimate purpose, and rephrase the actual goal. If Claude still declines, you can also use the thumbs-down feedback button to tell Anthropic.

A good response is calm, honest, and constructive.

The same Claude can behave differently in different products.

Many companies build their own products on Claude. Those companies, called operators, can give Claude instructions for their app, including keeping it focused on a topic, adopting a persona, or avoiding certain subjects. This means that something Claude does in one app may be declined in another, because that product has narrowed its scope. Published guidance describes operators as having wide room to customize, but not to turn Claude against the users it is serving, for example by deceiving them in harmful ways.

If a Claude-based app will not help with something that seems reasonable, check whether the product has a defined purpose, and whether another app or the standard Claude app would be a better fit.

When Claude declines, the aim is to be honest about it.

Anthropic’s published materials describe Claude as a kind of conscientious objector. It can decline all or part of a task, and it is meant to be open about the fact that it is declining, even if it does not always explain every reason. It is not meant to quietly give a deliberately weaker answer while pretending that was its best. It is also meant to tell you what it cannot help with in the current setting, so that you can seek help elsewhere.

Claude may also decline to share personal opinions on hot political topics, preferring to give a fair overview of the main positions. This is similar to the reserve many professionals keep in public settings, and it is a deliberate choice, not an inability to discuss politics.

Familiar situations show how the boundaries work in practice.

Boundaries protect people, and they sometimes frustrate honest users.

These habits make boundaries more frustrating than they need to be.

Follow these steps to turn a refusal into progress.

A few popular beliefs about Claude’s limits are only partly true.

Quick answers to the questions beginners ask most.

Because of safety, ethics, honesty, and policy, as described in Anthropic’s Usage Policy and Claude’s constitution. It weighs the likely benefit and harm of a request in context.

No. A will-not is a decision about safety or policy. A cannot is a limit of the technology or setup, such as missing search, files, or tools.

Published summaries describe a small set, including serious help toward mass-casualty weapons, child sexual abuse material, and damaging cyberweapons. These do not change with framing.

Honest context can change the best answer in gray areas. It does not change the absolute limits, and it is not a password.

Often yes, at a level that supports understanding, safety, and creativity, without providing working tools for harm.

To respect copyright. It can summarize, analyze, discuss themes, or write something original instead.

It can explain and inform in plain language, but it is not a licensed professional, and it encourages checking important decisions with a qualified person.

Yes. Even with a custom name or persona, published guidance says it should not deny being an AI when someone sincerely asks.

The company may have narrowed the app’s purpose. Operators can customize Claude within Anthropic’s policies.

Add honest context, ask for a safer version of the goal, and use the thumbs-down feedback button so the response can be reviewed.

In Anthropic’s Usage Policy and Claude’s constitution on Anthropic’s website, which are the authoritative and current sources.

Claude’s limits fall into two groups: things it will not do for reasons of safety, ethics, honesty, and policy, and things it cannot do because of technology or setup. The will-not group includes a few absolute lines, such as serious help toward mass-casualty weapons and child sexual abuse material, and a wider area of cautious judgment around harm to others, deception, privacy, copyright, and high-stakes advice. In gray areas, Claude weighs likelihood, severity, benefit, and context, and it is meant to be honest about what it declines. Apps built on Claude can narrow its scope further. When you hit a boundary, work out which kind it is, fix capability limits with the right tools, explain your real purpose honestly for gray areas, ask for the safer version, and give feedback if you think a refusal was a mistake. Policies change, so confirm details on Anthropic’s official pages.