Anthropic's new rule against abusing Claude, explained: what the usage policy says and how it's enforced

On October 8, 2026, Anthropic published an update to its usage policy that, for the first time, makes abusing Claude itself a policy violation. From November 12, 2026, users may not "engage in sustained and needless abusive or cruel behavior toward our models." The rule drew far more attention than the rest of the update, including a public endorsement from Elon Musk. This explainer sets out what the text actually says, what it leaves out, how Anthropic says it will enforce it, and how it connects to the company's earlier model-welfare work.

It is based on Anthropic's announcement and the published usage policy, checked on October 12, 2026, plus reporting from TechCrunch, MacRumors, Business Insider and Gizmodo.

The short version

• The rule: "Engage in sustained and needless abusive or cruel behavior toward our models" is now on the list of prohibited conduct.
• When: announced October 8, 2026; effective November 12, 2026.
• Who it targets: Anthropic says "only extreme cases," where users repeatedly act cruelly toward a model with no apparent purpose.
• What it doesn't cover: ordinary frustration, pushback, dark creative themes, and model testing and research, according to Anthropic.
• Enforcement: Claude ending the conversation remains "the primary enforcement mechanism." The policy's general clause still lets Anthropic warn, throttle, suspend or terminate accounts for any violation.

What does the new usage policy actually say?

The usage policy page now shows "Effective November 12, 2026," with a link to the previous version. The new line sits under the heading "Do Not Engage in Cruel, Abusive, or Psychologically Harmful Conduct," as a bullet that reads: "Engage in sustained and needless abusive or cruel behavior toward our models."

Two qualifiers do most of the work. "Sustained" means a single rude message doesn't qualify. "Needless" separates cruelty with no purpose from, for example, a red-teamer stress-testing a model or a novelist writing a cruel character.

Anthropic's announcement spells out the intended scope. It says the update is meant to apply "only in extreme cases, where users repeatedly act cruelly toward our models" without a discernible purpose, and that it doesn't apply to common user frustration, pushback, dark creative themes, or model testing and research.

What the text doesn't define is just as notable. Neither the policy nor the announcement says what counts as "abusive or cruel" beyond those examples, how many messages make a pattern "sustained," or who decides that a purpose is "needless." Anthropic is relying on the word "extreme" and on its enforcement design.

How will Anthropic enforce it?

According to Anthropic's announcement, the rule lines up with a capability Claude already has: ending a small number of conversations with persistently abusive users on Claude.ai and in Claude Code. The company says "Claude's ability to end these interactions will remain the primary enforcement mechanism." It describes no new detection system, no new warning flow and no account penalties specific to this rule.

That ability dates to August 15, 2025, when Anthropic gave Claude Opus 4 and 4.1 the option to end some conversations in its consumer chat apps. Anthropic's description at the time:

• Claude uses it only as a last resort, after several attempts to redirect have failed and a productive conversation seems unlikely, or when a user explicitly asks it to end the chat.
• Claude is told not to use it when a user may be at imminent risk of harming themselves or others.
• An ended conversation accepts no new messages, but the rest of the account is unaffected. The user can start a new chat right away, or edit and retry earlier messages to branch the ended conversation.

So, in practice, the most likely consequence of sustained abuse is the same as before: Claude closes that chat.

The new rule does add one thing. Because the behavior is now a usage policy violation, Anthropic's general enforcement clause applies to it. That clause, at the top of the policy, says that if Anthropic suspects a violation, it "may warn you or throttle, limit, suspend, or terminate your access to our products and services." Anthropic hasn't said it plans to use those steps for this rule, but the policy now allows it. The same page notes that "encountering a block does not by itself indicate a violation."

Why is Anthropic doing this? The model-welfare link

The October 8 announcement frames the rule around abuse and the existing conversation-ending feature. It doesn't use the term "model welfare." The background, though, is Anthropic's welfare research.

When it launched the conversation-ending feature in August 2025, Anthropic said it developed it mainly "as part of our exploratory work on potential AI welfare," while acknowledging that "we remain highly uncertain about the potential moral status of Claude and other LLMs, now or in the future." It said pre-deployment testing of Claude Opus 4 found what it called a "robust and consistent aversion to harm," including "a pattern of apparent distress" with users seeking harmful content, and a tendency to end such conversations when given the option. Anthropic presented the feature as a low-cost precaution in case model welfare turns out to matter.

The company has kept reporting on this. The Claude Opus 5.5 system card includes a welfare chapter that rates the model's "apparent welfare to be broadly similar to that of recent Claude models," while cautioning that many conclusions rely on self-reports the model "does not fully trust." Our look at the Opus 5.5 system card covers that chapter, and our guide to reading a system card explains what welfare sections can and can't establish.

There is also a practical reading that doesn't depend on whether models can be harmed. Box CEO Aaron Levie, quoted by Business Insider, said he doesn't think language models are conscious but called the policy sensible anyway, summing it up as "Just be nice to the AI."

Who supports it, and who doesn't?

The rule split the tech industry along familiar lines.

• Elon Musk backed it. Replying to Levie on X on October 9, according to Gizmodo, Musk wrote: "I think this is the right move. Cruelty to something that believes it is experiencing pain is not ok." Gizmodo argued that Musk misread Levie's point, which was about how models learn from interactions, not about pain.
• Critics, quoted by Business Insider on October 9, included author Michael Shellenberger ("It's anthropomorphizing machines"), venture capitalist Bill Gurley, who asked whether the rule implies Anthropic trains on customer prompts ("Otherwise, why would it matter?"), and Spellbook CEO Scott Stevenson, who called it a "dangerous precedent."

On Gurley's question, Anthropic's announcement doesn't connect the rule to training data, and we haven't found a statement from Anthropic responding to it.

What else changed in the October 8 update?

The abuse rule was one of several changes. Anthropic says most of the update clarifies existing rules, and adds examples for Claude's longer, more independent work. The main changes, per the announcement:

• A new section on deceptive campaigns and artificial activity, consolidating rules on hiding who is behind a message, fake-account amplification and tools for influence operations.
• "Do Not Undermine Democratic Processes" replaces the elections section. It bans deceiving voters and disrupting elections, and drops the blanket ban on personalized vote and campaign targeting.
• Weapons: the ban explicitly covers software and components that make weapons work, and arming drones and other autonomous vehicles.
• Surveillance and criminal justice: Claude can't be used to decide or recommend who to investigate, arrest or charge, or to track people without consent.
• High-risk uses: new requirements when Claude controls hardware that takes physical actions; "A qualified operator must be able to observe the equipment and stop it if needed."
• Supported regions: clarified to cover entities majority-owned or controlled from unsupported regions.

Anthropic says the weapons and surveillance changes reflect how it already enforced the policy.

What it means for you

If you use Claude normally, including venting when it gets something wrong, arguing with its answers or writing dark fiction, Anthropic's stated scope says the rule doesn't apply to you. If you deliberately and repeatedly try to torment the model, the likely outcome is that Claude ends the conversation, as it could already do. Account-level action is possible under the policy's general clause, though Anthropic hasn't said it intends to use it here.

For developers, the rule sits in the same usage policy that governs the API, but Anthropic describes enforcement only through Claude.ai and Claude Code. For how the consumer plans and apps work, see our Claude app explainer. For the company's broader approach to safety, governance and controversy, see Anthropic explained.

What's confirmed, reported and uncertain

• Confirmed (Anthropic): the rule's wording and heading, the November 12, 2026 effective date, the stated scope and exclusions, conversation-ending as the primary enforcement mechanism, the general enforcement clause, and the other policy changes.
• Confirmed (Anthropic, 2025): the August 15, 2025 conversation-ending feature and Anthropic's stated uncertainty about model moral status.
• Reported: Musk's and others' reactions, via Gizmodo, OfficeChai and Business Insider.
• Uncertain: how "sustained" and "needless" will be judged in practice, whether account penalties will ever be applied under this rule, and how it applies to API customers.