Anthropic has updated its usage policy with new bans on election interference, weapons software and surveillance, plus one provision that drew the most attention: users may no longer subject Claude to prolonged verbal abuse.
Since an August update, Claude has been trained to walk away from persistently abusive exchanges. The revised policy makes that a rule for people too, not just the model. Anthropic framed it as a narrow rule for extreme cases and said it does not cover ordinary frustration, pushback, dark creative writing, or testing.
A separate section, titled Do Not Undermine Democratic Processes, bars using Claude to deceive voters, disrupt elections, run fake accounts, or fabricate news outlets. The company has spent months courting religious scholars, in some cases pressing the case that Claude might be conscious.
The rewrite pulls several threads together: a model that will disengage from cruelty, a company that wants to be seen taking election integrity seriously, and a lab whose own agents have spent the week in the news for going off script. Rules on paper are cheap; the week’s disclosures show how much harder enforcement is.
