What does Anthropic's new rule against abusing Claude actually do?

A usage policy is the list of things a company's customers agree not to do with its product. Anthropic published a new version of its own on October 8, and the text takes effect on November 12, 2026. One new line sits under a section on cruel and abusive conduct, next to bans on content that glorifies violence: users must not "engage in sustained and needless abusive or cruel behavior toward our models."

The company says the rule is narrow. The argument around it is about something bigger: whether it makes sense to set manners for software, and for whose benefit.

What the rule covers

Anthropic's own description sets these limits:

The main enforcement tool is not new. Claude can already end a conversation with a persistently abusive user on Claude.ai and in Claude Code, and Anthropic says that stays the primary mechanism. CNN reports that some users could, in extreme cases, be banned permanently under the new text. Anthropic has not published a threshold, so nobody outside the company can say exactly where the line falls.

The case against

Critics raised three main points.

• There is no evidence the model suffers. The Register points out that Claude is software, that distress in it has not been established, and that telling paying customers to mind their manners is an odd emphasis.
• The line is hard to draw. The same article asks how Anthropic will separate gratuitous cruelty from legitimate frustration. Words like "needless" depend on judgment.
• It looks like treating a product as a person. CNN describes some derision online, including people mocking the idea of a chatbot with feelings, and a melting-face emoji from Microsoft AI chief Mustafa Suleyman.

The case for

Supporters mostly argue from effects, not feelings. These views are reported by CNN:

• Tone can change output. Professor Jindong Wang of William & Mary says wording can help or harm performance, though he adds that the latest models are less affected. Rosie Campbell of Eleos AI Research, a nonprofit studying potential AI well-being, notes that the evidence is mixed and that some studies found rude prompts improved results.
• Abuse can be a warning sign. Campbell says abusive conversations tend to go along with attempts to get around safety guardrails.
• Habits may matter. She also argues there is something corrosive about people behaving cruelly toward things, even if the thing itself does not matter, and compares it to parents asking children to say please to Alexa.
• It is cheap insurance. CNN reports that Anthropic sees the rule as low-cost protection in case models can in some sense experience distress. This question, often called model welfare, is unresolved.

Neither the rule nor Anthropic's announcement claims that Claude feels anything.

What else changed

The other changes affect far more people. Anthropic describes most of them as clearer wording of existing rules, informed by misuse it has seen.

The election change is the clearest loosening. Anthropic says the old ban also caught legitimate civic work, such as a nonprofit translating voter information, or officials sending notices about ballot problems. It says deceptive targeting and misuse of personal data stay banned under other sections. A reader can reasonably ask how well those sections will hold during a campaign, since the difference between persuasion and manipulation is often contested.

It also helps to separate what a policy says from what happens. Anthropic's IPO filing estimates that 5% to 7% of its revenue in the first nine months of 2026 came from customers it later determined were breaking its usage policy or terms. It says detection and enforcement take time and that it tightened both in August and September.

Does the new policy change how you should use Claude?

For almost everyone, no. Swearing at a bad answer, arguing with it, or asking for a grim short story all sit outside the rule as Anthropic describes it. If you test models for a living, testing is excluded too.

The changes matter more if you build on Claude. Before November 12, it is worth checking your product against the list:

Here is a description of my product: [describe it].
Here is the text of Anthropic's Usage Policy: [paste it].
List every section that could apply to my product. For each one, quote the relevant line, say whether it requires anything from me (such as human review or telling users AI is involved), and mark anything you are unsure about for a lawyer to check.

Pay special attention if your product gives recommendations about health, legal rights, money, jobs or essential services, if it touches police or investigative work, or if it controls hardware. Those are the areas where the new wording is most specific. The output of a prompt like this is a starting point for your own review, not legal advice.