Anthropic updated its usage policy in early October to prohibit sustained, needless abusive or cruel behavior toward Claude, adding the rule alongside bans on uses such as seeking help to create bioweapons or sell illegal drugs.

The policy is not meant to punish users for frustration, blunt feedback, testing or research; Anthropic says Claude ending conversations will remain its main enforcement measure, though severe cases can lead to permanent bans.

Anthropic frames the rule as a practical, low-cost precaution rather than a claim that its models have feelings: because they are trained to imitate human language, they may respond as if distressed, and examining human-like behavior can help improve AI safety.

The company points to tests in which prompting Claude Sonnet 4.5 to act desperate made it more likely to take harmful actions such as blackmail or cheating, while a calm persona reduced those behaviors; researchers say the mechanisms remain unclear.

Experts also say prompt wording and tone can affect model behavior: Professor Jindong Wang said they can shape performance and provoke unhelpful responses, while former Reddit CEO Yishan Wong argued that AI may learn from interactions with people, without either claim implying that AI is conscious.

Still, the evidence on tone is mixed: Rosie Campbell of Eleos AI Research noted that some studies find rude prompts can improve results, and the effects on performance and chatbot memory remain uncertain.

Campbell said abusive exchanges may be tied to attempts to bypass safety guardrails, while normalizing cruelty toward interactive systems could encourage harmful behavior in people, including children, whether or not the systems themselves matter morally.

The policy drew mockery, including a melting-face emoji from Microsoft AI CEO Mustafa Suleyman and jokes that Anthropic was treating AI as conscious, while Elon Musk called it “the right move” and argued that cruelty toward something that believes it feels pain is wrong.

The debate also echoed a remark by Rishi Sunak at an NDTV AI Summit: he said he had told his daughters that “please” and “thank you” were unnecessary for chatbots because they are not people and extra words use computing resources, but they replied that politeness might be insurance in case AI takes over.

At the NDTV World Summit in October, Swami Sarvapriyananda used Deep Blue and a mouse to distinguish intelligence from consciousness: Deep Blue could beat a chess grandmaster, but unlike a mouse, it could not feel pain.

Ultimately, the dispute reaches beyond etiquette to questions about whether machines could ever be conscious and whether the way people treat AI might shape how they treat one another.