Anthropic updates usage policy: ban on verbal abuse of model alongside election and weapons restrictions

Anthropic published an update to its usage policy on Thursday. The update adds explicit prohibitions on election interference, weapons software, and surveillance, but the change that drew the most attention is a new clause banning sustained verbal abuse of the model.

According to the announcement, since an August update Claude has been tuned to end conversations in which a user displays "harmful or abusive interactions on a sustained basis." The written policy now forbids users from continuing such conversations in the first place.

In its official statement the company clarified that the clause targets only extreme cases — situations where users repeatedly act cruelly toward the model without any apparent legitimate purpose. The ban does not apply to ordinary user frustration, disagreement with answers, dark creative topics, or testing and research on the model. The wording is intended to narrow the scope to patterns that are repeated and lack a legitimate aim, not to block everyday interactions where users express dissatisfaction or probe boundaries.

Behind the update lies a separate Anthropic effort engaging prominent religious scholars, in some cases attempting to persuade them that Claude may be conscious or possess a soul. The connection between that theological outreach and the abuse policy was not detailed in the announcement, but the timing raises questions about how the company frames its relationship to its models, both externally and toward users.

Separate clauses in the updated policy prohibit users from running large-scale deceptive campaigns, including using Claude to operate fake accounts or fabricated media outlets. A dedicated section titled "Do not undermine democratic processes" groups rules against voter deception and election disruption alongside the bans on weapons software and surveillance applications. That structure places model-abuse restrictions next to more familiar risk categories, signaling a broader risk perception than in previous versions.