Anthropic published a new version of its Usage Policy on October 8. It takes effect on November 12. Most of it is clarification. One thing is genuinely new: sustained, needless cruelty toward the models is now prohibited.
The new clause about how you treat Claude
Anthropic writes the limits into the rule itself. It is meant for extreme cases, where someone acts cruelly over and over with no discernible purpose. Ordinary frustration, pushback, dark creative themes and model testing are explicitly out of scope.
Enforcement doesn’t run through account bans. It runs through a capability that already shipped: Claude can end rare conversations with persistently abusive users, on Claude.ai and in Claude Code. Anthropic calls that the primary enforcement mechanism going forward.
Deceptive campaigns get their own section
Until now, the rules against fake accounts and fabricated news sites sat scattered across four sections: elections, fraud, privacy, disinformation. There’s now a section called “Do Not Engage in Deceptive Campaigns or Artificial Activity,” and it covers deception of any kind, political or commercial. Building the tooling and infrastructure for such campaigns is in there too.
Where the examples come from is in Anthropic’s September threat report: state media outlets, government propaganda offices and commercial firms running networks of fake accounts with Claude.
Elections: narrower scope, one ban dropped
The elections section is now called “Do Not Undermine Democratic Processes,” and it aims only at deceiving voters and disrupting elections. That means false claims about candidates or about how to vote, impersonating candidates or election officials, and attempts to suppress turnout.
What Anthropic removed is the blanket prohibition on personalized vote and campaign targeting. It caught work nobody meant to ban: the nonprofit translating voter information into other languages, or the election official mailing ballot cure notices. The behaviors that motivated the old rule, targeting built on deception or on misused personal data, stay prohibited through the other sections.
Weapons, surveillance, high-risk
The weapons section now says plainly that the software and components that make a weapon work are covered, along with actions like arming drones. The reason: people have tried to use Claude to build guidance and control software.
The surveillance section was rewritten. Tracking people without their consent is prohibited, whether in real time or after the fact from data already collected. Claude can’t be used to decide or recommend who to investigate, arrest or charge. Consented tracking such as fraud monitoring stays permitted, as do content moderation, journalism and legal research.
High-risk use cases keep the qualified human in the loop and the duty to tell the affected person that AI was involved. New is an addition from the Model Hardware Standard: if Claude is wired to hardware that moves on its own and could injure someone, a qualified operator has to be able to watch it and stop it. And the equipment has to hold a safe state if Claude drops off.
Most of these lines describe what already applied
On weapons and on surveillance, Anthropic says so itself: nothing changes in what it enforces. The document catches up with the practice. For anyone who has to justify inside a company whether a use case is allowed, that’s the real improvement. The high-risk section now lists which recommendations are covered and which aren’t, because people apparently kept asking exactly that.
Only the clause about how you treat the models breaks new ground, and it is carefully worded. The bad day isn’t banned, the pattern is. Whether that line holds won’t be settled by the policy. It will be settled by how often Claude walks out of a conversation.