Anthropic’s revised Usage Policy, announced on October 8, 2026, takes effect on November 12 and forbids “sustained and needless abusive or cruel behavior toward our models.” The new line applies to every Claude user, including people on the free plan.
The Claude abuse ban prohibits repeated, purposeless cruelty aimed at Claude, and Anthropic says it applies only in extreme cases. Snapping at a wrong answer is not the target. Most people who cross the line will see Claude close the chat. They won’t lose their account.
What the Claude abuse ban prohibits
The rule joins a list in the policy titled “Do Not Engage in Cruel, Abusive, or Psychologically Harmful Conduct.” Its weight rests on two adjectives. “Sustained” points to a pattern across many messages, so a single outburst falls outside it. “Needless” means the behavior serves no purpose, which is how Anthropic separates cruelty from work.
The company lists four things the rule does not cover: everyday frustration, hard pushback on an answer, dark creative themes, and model testing or research. A novelist writing a villain who torments an AI assistant is fine. So is a security team probing how Claude reacts to hostile prompts.
Two scenarios show where the line probably falls. A developer who curses at Claude at 2 a.m. because the same test keeps failing is venting with a goal, and the Claude abuse ban leaves that alone. A user who opens chat after chat only to insult the model, ignoring its attempts to redirect, fits the description of sustained and needless abuse. Anthropic has not published a fuller definition, and a spokesperson did not say what counts as abusive or cruel when asked.
Role-play sits in a gray zone. A user who asks Claude to voice a character who gets yelled at in a story is writing fiction, which the rule exempts. The same insults typed for an hour with no story around them look like sustained and needless abuse. Under the Claude abuse ban, the pattern around a swear word counts for more than the word itself.
Who the November 12 usage policy covers
The November 12 usage policy, which carries the Claude abuse ban, governs the Claude apps, Claude.ai, Claude Code and the API. It also covers products other companies build on Claude. It is the first full Anthropic usage policy update in more than a year, and the cruelty clause is only one of its changes.
The same revision clarifies rules on election content and covert influence campaigns, surveillance, weapons software, and AI systems that control physical devices. For compliance teams, those parts of the Anthropic usage policy update carry more practical weight than the cruelty line. Still, the Claude abuse ban drew most of the attention. It raises a question few software companies have touched: whether users owe anything to the tool itself.
Can Claude end a conversation on its own?
Yes, in rare cases. Anthropic gave Claude Opus 4 and 4.1 the ability to end chats in its consumer apps in August 2025, reserving it for persistently harmful or abusive exchanges. Claude ending conversations is a last resort that follows failed attempts to steer the chat elsewhere.
When it happens, only that thread locks. The user can open a new chat immediately, and other conversations on the account stay untouched. Earlier messages can still be edited and retried, which creates a new branch and keeps long-running work recoverable. Claude is also told not to use the feature when someone may be at imminent risk of harming themselves or others, so a person in crisis keeps the conversation open even under the Claude abuse ban.
How the Claude abuse ban is enforced
Enforcement runs on two levels, and they are easy to confuse.
Step one: Claude closes the chat
Anthropic says Claude’s ability to end these interactions “will remain the primary enforcement mechanism.” In practice, the policy writes into the terms a behavior the model already had. Under the Claude abuse ban, a closed chat is the expected outcome in nearly every case, and Claude ending conversations will look the same on November 13 as it did the week before.
Step two: action on the account
The policy’s general clause lets Anthropic warn users or throttle, limit, suspend or terminate access when it suspects a violation. That clause predates the cruelty rule, but it now covers it, which means some users could face permanent bans in extreme cases.
What Anthropic has not published is a threshold. Nothing says how many ended chats would prompt a review, or whether a Claude account suspension under this rule involves a human decision. One closed chat reads as a warning signal, not a verdict on the account. A string of them is the pattern most likely to draw attention, because “sustained” is the word the policy chose.
What changes for developers and Claude Code users
Claude Code already supports chat ending, so developers who argue with the model in the terminal fall under the Claude abuse ban the same way chat users do. The larger open question involves the API. The policy covers API use, but Anthropic’s announcement does not say whether API traffic gets the same chat-ending mechanism.
That gap matters for anyone running a product on Claude. If an app invites its users to vent at a character, or a game rewards insulting an AI, those messages travel through the developer’s account and API key. The safe reading before November 12 is to assume the Claude abuse ban reaches that traffic, and to check whether any feature depends on sustained hostility toward the model. A Claude account suspension that lands on a company key would take down every customer at once. That makes this a product risk as much as a user one.
Support teams face a softer version of the same question. Angry customers already shout at chatbots, and most of that anger is ordinary frustration over a refund or a delay, which the rule exempts. The Claude abuse ban targets a different pattern: insults aimed at the model itself, repeated with no request attached. A support log full of complaints about late orders should not look like that. A team that sees heavy, aimless abuse in its transcripts has a reason to add its own filters before the policy starts.
Red-team and evaluation teams have the clearest exemption, since model testing and research sit outside the rule. Even so, a short written purpose for each adversarial test suite costs little and answers the “needless” question if it ever comes up. The trade-off is paperwork for teams that never expected to need it.
Why Anthropic is policing how people treat Claude
The rule grows out of Anthropic’s AI model welfare research, which asks whether advanced models could have experiences that deserve moral weight. The chat-ending feature came from that work after testing showed what Anthropic described as a pattern of apparent distress when Claude faced persistently harmful requests.
Anthropic isn’t claiming its models are sentient or that they can be harmed. The logic reads as a precaution under deep uncertainty. If model welfare turns out to matter, early limits will look sensible; if it doesn’t, very few users lose anything. The Claude abuse ban turns that precaution into a formal line in the terms of use.
The case against the rule
Not everyone accepts the premise behind the Claude abuse ban. Microsoft AI chief Mustafa Suleyman answered the policy with a melting-face emoji, and others mocked Anthropic for treating Claude as if it had feelings. His objection runs deeper than a joke. In a September 16 essay, he argued that training Claude to treat its own consciousness as an open question could make advanced AI harder to control. Microsoft AI’s draft Humanist AI Code of Conduct rejects AI model welfare outright.
Other critics stress that Claude is software with no established evidence of distress, and object to paying customers being told to mind their manners. A practical complaint sits alongside the philosophical one. The model that gets insulted is also the one that decides when to close the chat, which puts heavy weight on its judgment in a gray zone Anthropic has left undefined. People who depend on long working sessions have good reason to want that zone mapped before the Claude abuse ban takes effect.
What to do before the November 12 usage policy starts
For most people, nothing changes. If you vent at Claude now and then, the Claude abuse ban was not written for you. The users who should act are the ones whose habits or products could look like sustained and needless abuse to a model reading the transcript.
- Read the cruelty clause and its exclusions in the policy itself, since summaries tend to drop the exemptions.
- If you build on Claude, audit any feature that invites users to insult an AI character, and record the purpose of adversarial testing.
- If Claude closes a chat, start a new one or edit an earlier message, and treat repeated closures as a sign to change approach.
Set aside ten minutes this week to read the clause, then check one product flow against it. The Claude abuse ban takes effect in ten days, which makes this week the cheapest time to catch a conflict.
Frequently asked questions
Can you get banned from Claude for swearing at it?
Occasional swearing is very unlikely to get anyone banned. Anthropic’s rule targets sustained and needless cruelty, and it explicitly leaves ordinary frustration alone. The usual outcome for persistent abuse is that Claude ends the chat. Account suspension stays possible under the general policy, but Anthropic has not said what would trigger it.
Does Anthropic think Claude has feelings?
Anthropic does not claim Claude is sentient or can be harmed. It studies model welfare because it considers the question unresolved, and it treats limits on abuse as a cheap precaution in case model experiences turn out to matter. Critics argue the approach encourages people to see software as a person.
Does the new policy apply to the Claude API?
Yes, the Usage Policy covers API use and apps built on Claude. What remains unclear is enforcement, because Anthropic’s announcement does not say whether API traffic uses the same chat-ending mechanism as Claude.ai and Claude Code. Developers should assume the cruelty rule applies to what their users send.
What happens to a chat that Claude ends?
That single conversation locks and accepts no new messages. Every other chat on the account keeps working, and the user can start a fresh conversation right away. Earlier messages in the ended chat can still be edited and resent, which opens a new branch so ongoing work is not lost.
Is testing Claude with hostile prompts still allowed?
Yes. Model testing and research are excluded from the cruelty rule, so red-teaming and evaluation work can continue. Teams that run adversarial prompts at scale should still write down the purpose of each test suite, because the rule turns on whether hostile behavior has a real reason behind it.
Why did Microsoft’s AI chief criticize Anthropic’s approach?
Mustafa Suleyman argues that teaching Claude to treat its consciousness as uncertain could make advanced AI harder to control. Microsoft AI’s draft code of conduct rejects model welfare entirely and says its models should stay subordinate to people. He mocked the abuse policy itself with a melting-face emoji.
