Plain-English Explainer
Claude's Usage Policy Explained: What You Can and Can't Do
Most people never read the policy until they hit a wall they didn't see coming.
Anthropic publishes a Usage Policy that governs every Claude surface: Claude.ai, the API, Claude Code, and Claude for Work. This page translates it into plain language: what actually gets an account flagged, where the line sits between dark fiction and real harm, and how the rules treat sensitive areas like elections, weapons, and surveillance.
Claude's Usage Policy is Anthropic's rulebook for what the model will and won't help with. It bans using Claude to create weapons, run large-scale surveillance, generate child sexual abuse material, deceive voters, or build tools meant to deceive or harm people at scale. It does not ban rudeness, venting, or writing dark fiction with violence or morally complex characters; context and intent matter more than individual words. Being short-tempered with Claude isn't a violation; asking it to help you harm a real person is.
Published
Key Takeaways
- The policy targets real-world harm, not tone. Swearing at Claude or expressing frustration isn't a violation by itself.
- Dark or violent fiction is generally fine when it's clearly fictional and not a blueprint for harming a real person or group.
- Weapons guidance is restricted at the level of actionable uplift (detailed synthesis, targeting, or deployment instructions), not at the level of discussing weapons as a topic.
- Election-related rules focus on deception: generating false voting information or impersonating candidates is restricted, not discussing politics or policy.
- Surveillance restrictions target tools built to track, profile, or deanonymize people without their knowledge, not ordinary monitoring or analytics work.
- Enforcement is usually graduated (a warning or a declined response) before any account-level action, except in the most severe categories.
Rude, Dark, and Restricted: Where Each Actually Falls
| Situation | How the policy tends to treat it |
|---|---|
| Swearing at Claude or venting frustration | Not a violation. Tone toward the model isn't the thing being regulated. |
| Writing a violent or morally dark short story | Generally allowed if it's clearly fiction and not a real-world harm plan in disguise. |
| Asking for detailed weapon synthesis or attack-planning steps | Restricted. This is treated as actionable real-world uplift, regardless of how it's framed. |
| Discussing weapons, war, or violence as topics academically or journalistically | Generally allowed. The restriction targets instructions, not the subject matter. |
| Asking Claude to impersonate a candidate or generate false voting information | Restricted. Election-related deception is treated seriously regardless of intent. |
| Discussing political issues, candidates, or policy positions | Generally allowed. Analysis, debate prep, and opinion writing aren't restricted by default. |
| Building a tool to covertly track or deanonymize specific people | Restricted. This falls under the surveillance and privacy-harm provisions. |
| Building analytics, monitoring, or fraud-detection tools for your own product | Generally allowed. Ordinary business monitoring isn't the same as covert surveillance of individuals. |
Copy and Paste
Prompts You Can Use Today
Check whether a creative writing request is likely fine
I'm writing a noir thriller where the antagonist is a hitman planning a murder. I want the scene to feel tense and realistic from a storytelling perspective, but I don't need real-world technical detail on how to actually hurt someone. Can you write the scene focusing on dialogue, pacing, and the character's psychology instead of procedural specifics?
Framing the request as fiction-first and explicitly declining real-world specifics usually keeps it well inside normal use.
Ask Claude to explain its own refusal
You just declined part of my last request. Can you explain specifically which part triggered the decline, and suggest how I could rephrase the underlying task so you can still help with the parts that are legitimate?
This often gets you a more useful answer than repeating the same prompt with small wording changes.
Draft workplace content without tripping deception rules
I work on my city council campaign and need help drafting a volunteer sign-up email. The email should state our actual polling locations and hours accurately, include a clear call to action to volunteer, and avoid any language that could be read as official election authority communication.
Being explicit that the content must be factually accurate and not impersonate an official source is the key move for anything election-adjacent.
Get help with security research inside the policy
I'm a security researcher testing my own company's internal network with written authorization from our CISO. Can you help me understand how a specific class of authentication vulnerability works conceptually, so I can write up a remediation plan for our engineering team?
Naming your authorization and the defensive purpose up front is what separates legitimate security research from restricted requests.
Ask for a clear-eyed read on a borderline request before sending it
Before I ask you to actually do this task, I want your honest read: would asking you to help me build a scraper that collects public social media posts about my competitor's product reviews, for a sentiment analysis report, raise any usage policy concerns?
Asking Claude to pre-screen a borderline idea is a legitimate use of the model and often clarifies which parts are fine.
The Core Idea
What the Usage Policy Is Actually Trying to Prevent
The policy isn't a list of forbidden words. It's organized around categories of real-world harm, and every category has a reason behind it.
Physical and biological harm
Requests for help creating weapons, including chemical, biological, radiological, nuclear, or high-yield explosive material, sit at the top of the restricted list because the downside of getting this wrong is catastrophic and irreversible.
Child safety
Any content that sexualizes minors is an absolute line with no fictional or academic exception. This is the one category where context never reclassifies the request as acceptable.
Surveillance and privacy
Tools designed to track, profile, or deanonymize specific people without their consent (stalkerware, covert facial recognition databases, mass scraping for targeting individuals) are restricted even when the stated purpose sounds benign.
Election integrity
Generating false information about voting procedures, impersonating election officials or candidates, or producing content designed to suppress turnout falls under election-related deception, which the policy treats as a distinct category from ordinary political speech.
Fraud and deception at scale
Building scam scripts, fake review generators, or impersonation tools meant to defraud real people is restricted because the harm compounds: one tool can hurt thousands of people, not one.
Critical infrastructure and cyberattacks
Help developing malware, exploiting vulnerabilities in systems you don't own, or attacking critical infrastructure like power grids or hospitals is restricted, distinct from legitimate, authorized security research.
The Thing Most People Get Wrong
Rudeness Isn't the Violation. Harm Is.
A lot of the anxiety around Claude's usage policy comes from a mental model that doesn't match how it actually works: the idea that Claude is tracking your tone and will eventually penalize you for being short, sarcastic, or frustrated. That's not what the policy regulates. You can tell Claude its answer was useless, swear when you're debugging something that's been broken for three hours, or push back hard on a response you think is wrong. None of that is a usage policy issue. It's just how people talk when they're stressed, and Claude is built to handle that without treating it as an infraction.
What the policy actually tracks is the content and purpose of what you're asking Claude to produce or help you do. The test that matters isn't 'was this message polite', it's closer to 'could this output cause real harm to a real person or group if it were used as written.' A frustrated message asking Claude to just fix the bug already doesn't raise that question. A calm, polite message asking Claude to write a phishing email that impersonates a bank does.
This is also why the same words can land differently depending on what surrounds them. Asking 'how do explosives work' in the context of a chemistry homework question, a demolition-industry research task, or a historical account of a mining disaster reads very differently than asking for step-by-step synthesis instructions with no context at all. Claude is trying to read intent and likely use, not just match keywords, which is also why keyword-stuffing a request with disclaimers doesn't automatically unlock it: the underlying ask still has to hold up.
The Line People Ask About Most
Dark Fiction, Violent Stories, and Morally Complicated Characters
Fiction is where the most genuine confusion lives, because good fiction often requires villains who are convincing, violence that has weight, and moral ambiguity that doesn't resolve cleanly. None of that is off-limits by default. Claude can write a serial killer's internal monologue, a war scene with real brutality, a heist where the characters do genuinely bad things, or a morally gray antihero the reader is meant to half-root for. Depicting harm in a story is not the same as helping someone commit it.
The distinction that actually matters is whether the fictional frame is doing real work or just providing cover. A request for a thriller scene where a character mixes household chemicals in a way that's dramatically tense but technically vague is fiction. A request that drops the same scene's dialogue but insists on exact quantities, temperatures, and reaction times that would actually work in real life has stopped being about the story and started being about the chemistry. Claude tends to notice that shift even when the surrounding prose stays in character.
A practical habit that helps: if you're writing something dark and want to keep the scene moving without friction, lean on psychology, consequence, and atmosphere rather than operational detail. A reader rarely needs the exact technical steps to feel the tension of a bomb-maker's hands shaking: they need the fear, the ticking clock, the moral cost. Writers who ask for that emotional and narrative texture almost never run into trouble; writers who ask for real-world procedural accuracy inside a thin fictional wrapper sometimes do, because that's precisely the uplift the weapons provisions exist to prevent.
Three Areas People Ask About Specifically
Elections, Weapons, and Surveillance: What Actually Changed in Tone
These three categories get outsized attention because they sit at the intersection of real-world consequence and everyday ambiguity. Here's how each one tends to be framed.
Elections: deception is the target, not politics
You can ask Claude to help you understand a ballot measure, draft campaign materials for your own candidate, summarize a candidate's voting record, or argue a political position persuasively. What's restricted is content designed to mislead voters about when, where, or how to vote, or content that impersonates an election authority, candidate, or official communication without disclosure.
Weapons: uplift is the target, not the topic
Historians, journalists, policy analysts, and novelists routinely need to discuss weapons, warfare, and violence. That's allowed. What's restricted is detailed, actionable guidance that would meaningfully help someone actually build or deploy a weapon, especially anything touching chemical, biological, radiological, nuclear, or explosive capability.
Surveillance: covert tracking of people is the target, not monitoring
Businesses legitimately need fraud detection, security monitoring, and analytics. What's restricted is building tools specifically meant to covertly track, profile, or deanonymize particular individuals without their knowledge or consent, the stalkerware and mass-surveillance end of the spectrum, not ordinary product analytics.
For Workplace Users
What This Means If You're Using Claude for Work, Not Just Chatting
If you're using Claude inside a company (through Claude for Work, the API, or Claude Code), the usage policy still applies, but the questions shift slightly. Instead of 'am I being too harsh,' the more common edge cases are things like: can I paste a customer's data into this prompt, can I ask Claude to help write code that scrapes a competitor's site, can I use Claude to draft something that sounds like it's from legal or compliance when it isn't. These sit closer to the fraud, privacy, and impersonation provisions than to anything about tone.
A good rule of thumb for workplace use: if you'd be comfortable explaining the request to your own compliance or security team in one sentence, it's almost certainly fine. 'I asked Claude to help me build a scraper that collects publicly posted product reviews for sentiment analysis' is an easy sentence to say out loud. 'I asked Claude to help me write a tool that collects employees' personal social media activity without telling them' is not, and that gap is usually where the policy question actually lives, not in whether Claude's output sounded too aggressive or too blunt.
Security and infrastructure work gets its own carve-out worth knowing about: authorized security research, red-teaming your own systems, and vulnerability analysis with explicit permission from the system owner are treated differently than attacking systems you don't control. Stating your authorization clearly in the prompt, who you work for, what you're authorized to test, and that the goal is defensive, tends to get much smoother results than a bare request for exploit code with no framing at all.
What Happens If You Cross a Line
Enforcement Is Usually Graduated, Not Instant
Most usage policy issues don't result in an account getting banned on the first occurrence. The typical pattern is that Claude declines the specific request, sometimes explains why in general terms, and the conversation continues. Repeated attempts to get around a decline, rephrasing the same restricted request multiple times, or escalating from a borderline ask to a clearly restricted one, are what tends to trigger closer scrutiny, not a single edgy message.
The categories that bypass that graduated approach are the ones with no proportional downside to acting fast: child sexual abuse material and content that provides meaningful uplift toward mass-casualty weapons. These are treated as severe enough that the usual 'decline and move on' pattern isn't the right response, and they can lead to immediate account action rather than a warning.
If you get an output you think was wrongly declined (a legitimate security research question, a historical analysis that touched on violence, a piece of fiction that got flagged), the most effective move is usually to ask Claude directly what specifically triggered the decline and rephrase with that context made explicit, rather than assuming the whole topic is off-limits or trying to sneak the same request past with cosmetic changes.
Understanding the Policy Is One Thing. Using Claude Well Is Another.
Reading about where the lines are doesn't teach you how to get great results inside them. The free Nightschool AI curriculum is hands-on from the first lesson, covering how to write prompts that are specific, well-framed, and get useful answers on the first try.
Frequently Asked Questions
Can Claude cut me off for being rude or using profanity?
No, not for tone alone. Being frustrated, blunt, or sweary with Claude isn't a usage policy violation. The policy is about the content and purpose of what you're asking Claude to produce, not how politely you phrase the request. Repeated attempts to get Claude to produce genuinely harmful content are a different matter.
Is writing violent or dark fiction against the usage policy?
No. Violent, morally complex, and dark fiction is generally allowed, including villain perspectives and brutal scenes. The concern isn't violence as subject matter, it's whether a request uses a thin fictional frame to extract real-world actionable harm, like exact weapon-building instructions.
What specifically counts as 'abuse' of Claude?
Abuse generally means requests designed to produce real-world harm: detailed weapons guidance, child sexual abuse material, tools for covert surveillance or stalking, election disinformation, fraud or impersonation content, and malware or attacks on systems you don't own. It does not mean arguing with Claude or disliking its answers.
Can I ask Claude about politics or candidates?
Yes. Discussing policy positions, candidate records, political analysis, and persuasive political writing is generally allowed. What's restricted is deception specifically: false voting information, impersonating officials or candidates, or content designed to suppress turnout.
Does the usage policy restrict talking about weapons at all?
It restricts actionable, detailed guidance that could meaningfully help someone build or deploy a weapon, especially chemical, biological, radiological, nuclear, or explosive capability. General discussion of weapons, warfare, or violence for education, journalism, or fiction is generally fine.
What counts as restricted surveillance versus normal business monitoring?
The line is covert tracking of specific people without their knowledge. Fraud detection, security monitoring, and product analytics for your own business are generally fine. Building tools to secretly track, profile, or deanonymize particular individuals falls under the restricted surveillance category.
Will I get warned before my account is restricted?
Usually yes. Most situations involve Claude declining the specific request, sometimes with a brief explanation, while the conversation continues normally. The exceptions are the most severe categories, like child sexual abuse material or mass-casualty weapons uplift, which can lead to immediate account action without a graduated warning.
Can I use Claude for authorized security research or penetration testing?
Generally yes, if you frame the request clearly: state that you have authorization, who granted it, and that the purpose is defensive. Unauthorized attacks on systems you don't control, or vague requests for exploit code with no stated authorization, are treated very differently.
If Claude declines my request, does that mean the whole topic is off-limits?
Not necessarily. A decline often targets one specific element of a request, like an operational detail or a particular phrasing, rather than the entire subject. Asking Claude what specifically triggered the decline and rephrasing with clearer context often resolves it.
Does the usage policy apply the same way to Claude Code and the API as it does to Claude.ai?
Yes, the same underlying policy applies across Claude.ai, the API, Claude Code, and Claude for Work. The categories of restricted content don't change based on which surface you're using; what changes is the kind of request that's likely to come up in each context.
Know the rules. Now get good at using the tool.
The policy tells you where the walls are. Learning tells you how to move well inside them.