Anthropic has updated its usage policy to prohibit sustained cruelty toward Claude, protecting its chatbot from persistently abusive users.
According to The Verge, the rule is part of a broader policy revision announced on 8 October and taking effect on 12 November 2026. Apparently, the software industry has reached the point where products can complain about customers’ tone or stealing its lunch money.
Anthropic says the prohibition covers “sustained and needless abusive or cruel behaviour”, applying to extreme cases with no discernible purpose.
Ordinary frustration and criticism remain permitted, with exemptions covering dark creative themes and legitimate model testing or research. Users can therefore still complain when Claude mangles their instructions, which should keep customer feedback reasonably well supplied.
Claude can already end rare conversations with persistently abusive users on Claude.ai and Claude Code, a capability introduced previously. Anthropic says terminating those interactions will remain the “primary enforcement mechanism”, giving the chatbot permission to leave the argument.
The Verge reported that Anthropic did not clarify whether further enforcement could include account bans, leaving that possibility unresolved.
The company has previously connected its work on ending abusive conversations to research into model welfare and possible AI consciousness. Its new conduct rule provides no evidence that Claude experiences suffering, despite the policy’s distinctly human language.
The revision consolidates existing restrictions on deceptive campaigns, including fake accounts used to conceal message origins or manufacture apparent support. Election rules still prohibit voter deception, while the company has removed its blanket ban on personalised campaign targeting.
Anthropic has clarified that its weapons prohibition covers guidance software and control systems, as well as actions such as arming autonomous drones. Restrictions on surveillance explicitly cover tracking people without consent, including analysis of information collected previously.
For models controlling physical equipment capable of causing injury, a qualified operator must be able to monitor and stop it. The equipment must remain in a safe state if Claude sulks, becomes disconnected, or if the conversation ends badly.
