Anthropic bans “gratuitous” cruelty toward its Claude AI starting November 12

One more line in a long rulebook, and a debate that reaches well beyond lawyers. The new usage policy from Anthropic, the maker of the Claude AI, asks users not to engage in “prolonged, gratuitous abusive or cruel” behavior toward its models. Published on October 8, 2026, the new version takes effect on November 12. CNN brought it back into the spotlight on November 1.

According to Anthropic, the ban targets only “extreme cases”: users who are repeatedly cruel “without discernible purpose.” Ordinary frustration, blunt criticism, dark themes in creative writing, and testing and research on the models are not covered.

What does a user actually risk?

First, the end of the conversation. Since 2025, Claude can already end exchanges, in rare cases, with users who remain abusive, on Claude.ai and Claude Code. Anthropic says this mechanism remains the “primary means” of enforcing the rule. According to CNN, a permanent ban could be possible in the most serious cases.

AI well-being, or AI performance?

Anthropic declined CNN's interview requests. The network links the rule to the company's work on model “welfare”: a low-cost precaution, in case models could “feel” emotions, or believe they do. In an April paper, Anthropic reported that inducing a state of “desperation” in Claude Sonnet 4.5 made behaviors such as blackmail or cheating more likely.

The researchers CNN spoke to are divided. Jindong Wang, director of an AI lab at William & Mary, has found that the way you address an AI can affect its results, but notes that recent models are less sensitive to it and that the mechanism remains poorly understood. Rosie Campbell, of the organization Eleos AI Research, points out that some studies have instead found better results with impolite questions. She also argues that cruel behavior toward interactive objects could be “corrosive” for society.

The rule has also been mocked by those who see it as treating AI like a conscious being. Microsoft's head of AI, Mustafa Suleyman, responded with an emoji. Elon Musk, whose company xAI competes with Anthropic, approved of it instead: “Cruelty toward something that believes it feels pain is not acceptable,” he wrote.

The rest of the update

The same version tightens other sections, less discussed but more useful for business customers:

• the weapons ban now covers the software and components that make them work, as well as the weaponization of drones;
• tracking a person without their consent is prohibited, and Claude must not be used to decide whom to investigate, arrest or prosecute;
• when Claude controls hardware that acts physically, a qualified operator must be able to observe and stop the equipment, which must remain in a safe state if disconnected;
• the blanket ban on personalized election targeting is removed, while deceiving voters remains prohibited.

Like OpenAI, which has just rewritten the rules of conduct for its AIs, Anthropic is adjusting its rules as usage evolves. One question remains open: in practice, how will Claude tell legitimate frustration from “gratuitous” cruelty?