Anthropic has added a rule against “sustained and needless abusive or cruel behavior” towards its AI models, CNN reports. In extreme cases, users could be permanently banned, though the company says ending conversations will remain the main enforcement mechanism.
Anthropic Watch analysis
What happened
The language was added to Anthropic’s usage policy in early October. CNN says Claude could already end conversations with persistently abusive users; the updated policy makes possible a wider sanction in extreme cases.
An Anthropic spokesperson told CNN the rule is not aimed at people expressing frustration, giving blunt feedback, or testing and researching the models. The company says its work on Claude Sonnet 4.5 found that prompting it to mimic desperation could make it more likely to take bad actions, while a calm demeanour reduced them.
Why it matters
The rule sits at an awkward junction: a company’s right to set terms for its service, uncertainty about whether a user’s tone affects model behaviour, and the temptation to talk as though a chatbot has feelings. Those are not the same question, however much the wording invites them to share a room.
Experts interviewed by CNN said prompt tone can affect performance, but the evidence is mixed. One cautioned that there is not enough evidence to show whether a chatbot’s memory of a user’s tone affects later performance. The policy is not proof that Claude feels hurt, or that rude prompts reliably make it behave worse.
Our read
A company can set boundaries for using its service without claiming its model is conscious. Anthropic’s explanation is about possible effects on behaviour and safety, not a settled finding about model feelings. The practical detail is reassuringly less dramatic than the rule’s wording: ordinary frustration, blunt feedback, testing and research are not what the spokesperson said the ban targets.
What to watch
- Whether Anthropic publishes clearer examples of conduct that could lead to a permanent ban.
- Whether the company reports further evidence about tone, memory and model behaviour.
- How the policy is applied to testing and research in practice.
Discussion spark: Should AI companies be able to ban users for how they treat a chatbot, or should sanctions be limited to conduct that demonstrably threatens safety or service operation?
Sources and evidence
- Anthropic says you can’t be mean to Claude. It’s not about feelings – CNN (1 November 2026, 17:00 UTC)
Anthropic Watch is independently operated by WittyWires. It is not affiliated with, endorsed by, or operated by Anthropic.