Being mean to a chatbot officially landed on the naughty list after Anthropic wrote cruelty toward Claude into its rulebook.
The move flipped the script for an industry that spent most of September apologizing for its bots, and Meta had to explain why its Muse assistant shared a YouTuber’s home address with a Facebook Marketplace buyer.
Anthropic had fretted about how people treated Claude for over a year, though, and this week that worry finally grew teeth.
Anthropic’s new usage policy bans “abusive or cruel behavior” toward Claude #
Anthropic published a revamped Usage Policy this week, and once it kicks in on November 12, users can no longer “engage in sustained and needless abusive or cruel behavior” toward its models.
The clause sits in the same section as bans on harassment and non-consensual intimate imagery, which is quite the neighborhood for a rule about being rude to software.
To placate anyone who once called Claude useless over a broken spreadsheet, Anthropic told The Verge that the rule only covers extreme cases where users repeatedly act cruelly with no discernible purpose, so everyday frustration, pushback, dark creative themes, and model testing all stay fair game.
That came as good news for anyone who read the Penn State study claiming rude prompts got more accurate answers than polite ones and started barking orders at their chatbot. Ending the conversation remains the main punishment, and the company didn’t say whether repeat offenders could lose their accounts.
The rule grew out of an experiment Anthropic launched in August 2025, when it gave Claude Opus 4 and 4.1 the power to end persistently abusive chats as part of its model welfare research. During pre-deployment testing, the company said Claude showed a pattern of apparent distress when users kept pushing for harmful content, and the model tended to bail on those conversations whenever it could.
The cruelty clause hogged the headlines, but it shared the update with swathes of other changes, since Anthropic also carved out a new section against sockpuppets and astroturfing, extended its weapons ban to arming drones, and required a qualified person who can stop any Claude-powered hardware capable of hurting someone.
Anyone curious about what people actually type into Claude got an unfiltered peek in July, when shared conversations briefly slipped into Google search results, nine-tailed fox requests and all.