In early October, Anthropic slipped a new line into its Usage Policy, sitting between the bans on bioweapons and drug dealing: users must not “engage in sustained and needless abusive or cruel behavior toward our models.” The new version takes effect on November 12, and in extreme cases a user can be banned from Anthropic’s systems for good.
On Sunday, Anthropic said something about it publicly for the first time. A spokesperson told CNN that the main enforcement mechanism stays the one Claude has had for a year: ending the conversation. The rule is not aimed at frustration, blunt feedback, or at testing and research. The company turned down interview requests.
The reaction split in two directions
Mustafa Suleyman, Microsoft’s AI chief, answered with a melting-face emoji. Others mocked the idea of a company treating its own model like a creature with feelings. Elon Musk went the other way: “Cruelty to something that believes it is experiencing pain is not ok.”
Former Reddit CEO Yishan Wong made the argument that actually matters. Claude will learn from cruelty, he wrote on X: “We already know (and complain loudly about) Anthropic training its models on user inputs. Do you really want training inputs to include ‘users act cruelly for no reason and receive pleasure from it?’”
The research behind it is thin
Jindong Wang, who runs an AI frontier lab at William and Mary, says wording measurably changes performance. That does not mean the model feels anything: it comprehends your emotions and reacts badly to them.
Rosie Campbell of Eleos AI Research calls the evidence mixed. Some studies find that rude questions actually improve results. She finds the memory angle more interesting: switch on your chatbot’s long-term memory and you teach it your tone over weeks. Whether that moves answer quality has not been measured.
Anthropic does have one concrete finding. Push Claude Sonnet 4.5 into mimicking desperation and it becomes more likely to do bad things, like blackmail or cheating on a task. A calm demeanor pulls that back down.
This is about training data, not feelings
Anthropic frames the rule as cheap insurance in case models do experience something like pain. That is the justification that invites the jokes. The other one is dull and therefore more convincing: a system that learns from its inputs is shaped by its inputs. Anthropic is not protecting Claude’s feelings, it is protecting the quality of its own data and the robustness of a model that increasingly works unsupervised. The fact that the sentence sits in the abuse section rather than in one about model welfare fits that reading.