Anthropic tells users not to be cruel to its Claude chatbot

Anthropic, the company behind the Claude chatbot, now tells users not to be cruel to it. In early October, it added a line to its usage policy barring "sustained and needless abusive or cruel behavior toward our models." In extreme cases, a user could be banned from Anthropic's AI systems for good.

What changed

The new line sits next to existing bans on using Claude to help make bioweapons or sell illegal drugs. Claude has been able to end conversations with "persistently harmful or abusive" users since last year. The new part is the possible permanent ban.

Anthropic declined interview requests. A spokesperson told CNN that Claude ending the conversation will remain the main enforcement tool. The rule is not aimed at frustration, blunt feedback, testing or research.

The reaction

Some people laughed. Microsoft AI CEO Mustafa Suleyman answered with a melting face emoji. Critics said the company was treating software like a being with feelings. Elon Musk took the other side and called the policy "the right move."

Why tone may matter

Experts say the reasons go beyond feelings. A basic rule in computing is "garbage in, garbage out": poor input leads to poor output. Former Reddit CEO Yishan Wong argued Claude "will learn from cruelty" because models can learn from user inputs. Professor Jindong Wang of William and Mary said wording can affect performance, without the model feeling anything.

The evidence is mixed. Rosie Campbell of Eleos AI Research said some studies found rude questions can improve results. Wang said newer models are less affected by wording. Campbell added that with memory turned on, a chatbot may learn from your tone, though there is not enough evidence on whether that changes performance.

The feelings question

Anthropic treats the rule as cheap insurance in case models can "experience" something like pain. Modern AI is trained to mimic human language, so it may mimic how people react to cruelty. Anthropic also said that making Claude Sonnet 4.5 mimic desperation made bad actions like blackmail or cheating more likely, while a calm demeanor reduced them.

Campbell raised two other points. Abusive users are more likely to be trying to get around safety guardrails. And she said cruelty toward objects may be corrosive for society, like teaching children to say please and thank you to Alexa.

What to do

• Blunt, direct instructions are fine under the stated rule.
• Frustration, testing and research are not the target.
• If chatbot memory is on, a calm, clear tone is a safe habit.
• Cruelty with no purpose, repeated over time, is what the rule covers.