Anthropic Will Start Banning People for Being Cruel to Claude on November 12. Venting Is Still Fine
Anthropic rewrites this policy roughly once a year as Claude's own capabilities change, and most of what's in the 2026 version is maintenance: narrower wording, new examples for longer-running agentic tasks, clarified enforcement on existing rules. The abuse-toward-Claude clause is the one getting headlines, and it's worth being precise about what it actually restricts before anyone draws a bigger conclusion from it.
It is not a statement that Claude is sentient or that Anthropic has concluded the model suffers. Anthropic has spoken publicly in the past about model welfare as an open, unresolved question worth taking seriously even without an answer, and this clause reads as part of that same cautious posture rather than a verdict.
What the policy actually changed
| Area | What changed |
|---|---|
| Treatment of Claude | New ban on sustained, needless abusive or cruel behavior, aimed at extreme repeated cases, not venting or testing |
| Physical hardware | Claude connected to hardware that acts autonomously now requires a qualified operator who can observe it and a guaranteed safe state if Claude disconnects |
| Surveillance | Explicit ban on using Claude to decide or recommend who to investigate, arrest, or charge, plus a ban on building or improving surveillance tooling |
| Elections | Renamed to 'Do Not Undermine Democratic Processes'; the blanket ban on personalized campaign targeting was removed, narrowed instead to deception and voter-data misuse |
| Deceptive activity | New standalone section consolidating rules against fake accounts, fabricated news sites, and influence operations, political or commercial |
| Weapons | Prohibition now explicitly names software that arms drones or other autonomous vehicles |
The surveillance change is the one with the clearest practical stakes outside of anyone's feelings about chatbot welfare. Banning Claude from recommending who gets investigated, arrested, or charged draws a hard line between legitimate uses, fraud monitoring with consent, content moderation, journalism, legal research, and the use case civil liberties groups have worried about most: predictive policing dressed up as an AI recommendation.
What this doesn't ban
- ✦Ordinary frustration or pushback at Claude's answers. The rule targets sustained cruelty with no purpose, not a sharp reply to a bad response.
- ✦Dark or difficult creative writing. Fiction involving cruelty toward a character is not the same thing as the policy's target.
- ✦Red-teaming and model testing. Deliberately provoking Claude to find failure modes is explicitly outside the rule's scope.
- ✦A single rude message. Enforcement is aimed at repeated, extreme behavior, and Claude ending a conversation isn't itself proof of a violation.
The elections change is the more consequential one for anyone running a campaign or civic organization that uses Claude. The old blanket ban on personalized political targeting is gone, replaced by narrower bans on deception and misusing voters' personal data. Legitimate get-out-the-vote and constituent-outreach work that was arguably over-restricted before now has more room, while actually deceptive targeting stays banned under the new, separate deceptive-activity section.
When does Anthropic's new Usage Policy take effect?
November 12, 2026. It was published October 8, 2026. The previous version, effective since September 15, 2025, remains in force until then.
Can I actually get banned for being rude to Claude?
The policy targets sustained, needless, extreme cruelty with no discernible purpose. Ordinary frustration, pushback, dark fiction, and model testing are explicitly excluded. A single harsh message isn't described as a violation on its own.
Does this mean Anthropic thinks Claude has feelings?
The policy doesn't make that claim. Anthropic has previously described model welfare as an open question it takes seriously without claiming an answer, and this rule reads as consistent with that cautious stance rather than a declaration either way.
What changed around surveillance and law enforcement?
Claude can no longer be used to decide or recommend who should be investigated, arrested, or charged, and building or improving surveillance tools is now explicitly prohibited. Consented tracking such as fraud monitoring, content moderation, journalism, and legal research remain permitted.
Did Anthropic ban political uses of Claude entirely?
No. It removed the previous blanket ban on personalized vote and campaign targeting, narrowing the elections section to cover deception and disruption of democratic processes specifically. Legitimate civic and campaign work has more room than before; deceptive targeting is still banned.
Try the tools ✦
Free browser tools that never upload your files.