Anthropic expands Claude policy on elections, surveillance and abuse
Anthropic has revised its usage policy for Claude, adding explicit prohibitions on election interference, weapons software, surveillance and prolonged verbal abuse of the model. The update also sets out restrictions on deceptive campaigns, including the use of Claude to operate fake accounts or create fabricated news outlets.
One of the most notable additions bars users from pursuing persistently harmful or abusive interactions with Claude. Anthropic said the rule is intended for extreme situations in which users repeatedly act cruelly toward its models without a discernible purpose.
A narrow rule on abusive interactions
Anthropic said the new provision does not cover ordinary user frustration, pushback, dark creative themes, model testing or research. Claude has been trained since an August update to end conversations involving persistently harmful or abusive user interactions; the revised policy now expressly prohibits users from continuing to pursue such exchanges.
The policy change follows Anthropic’s engagement with prominent religious scholars, including efforts in some cases to persuade them that Claude could be conscious or possess a soul. The updated language turns the company’s approach to harmful treatment of its model into a stated condition of use.
Election and deception restrictions
Anthropic has placed rules on voter deception and election disruption in a section titled Do Not Undermine Democratic Processes. The policy also prohibits broadly deceptive activity, extending beyond elections to campaigns that use fabricated identities or news operations.
The changes add policy boundaries around high-risk uses of Claude at a time when the company is also focused on protecting its model capabilities. In that context, Anthropic’s Claude capability extraction concerns illustrates Anthropic’s concern about the extraction of Claude’s capabilities and the controls surrounding its deployment.
What organisations should do
Businesses using Claude should review internal prompts, automation flows and approval processes against the revised restrictions. Political communications, public-facing content, identity-related activity and work involving surveillance or weapons software require particular scrutiny, while teams should document legitimate testing and research activity so that it remains clearly distinguishable from prohibited conduct.
The practical implication is that organisations need governance that reflects both the technical use of Claude and the policy limits attached to that use before deploying the model in sensitive workflows.