The groundwork was laid last August, when Anthropic gave Claude the ability to walk away from users who were “persistently harmful or abusive.” The company described that as part of its research into model welfare, and said ending the conversation—not banning the user—would remain its “primary enforcement mechanism.”¹

By September, Anthropic said it had also documented more familiar dangers: state-linked media, propaganda offices and commercial operators using AI to run fake accounts and fabricated news sites. Its new policy groups those previously scattered restrictions under a single ban on deceptive campaigns, covering both political and commercial attempts to hide a message’s origin or manufacture amplification.²

The update, announced Thursday and taking effect November 12, sharpens several other lines. Claude may not be used to deceive voters, impersonate election officials or suppress turnout; it also cannot help build weapons guidance or control software, including systems for arming drones and other autonomous vehicles. Anthropic further says Claude cannot be used to identify and track people without consent, or recommend whom law enforcement should investigate, arrest or charge.²

For health, finance, legal rights and other high-stakes decisions, the company reiterates that a qualified human must be able to review and change Claude’s recommendations, with affected people told AI was involved. Hardware connected to the model must have an operator able to stop it and a safe state if Claude disconnects.²

But the clause drawing the most attention bars “sustained and needless abusive or cruel behavior” toward Claude. Anthropic says it applies only where users repeatedly act cruelly “with no discernible purpose,” not to frustration, pushback, dark fiction or model research.² Human-focused reporting has framed the change as a striking escalation from allowing Claude to exit abusive chats to explicitly prohibiting users from pursuing them.³