Being mean to Claude is the line that traveled. The sentence that can still tailor the surveillance ban did not move.
On October 8 Hayden Field (@haydenfield) posted that Anthropic was updating its usage policy for the first time in over a year. The new rules prohibit sustained “abusive or cruel behavior” toward Claude, and add restrictions on propaganda campaigns, surveillance, and weapons. By this morning that post had 1,569 likes, 215 reposts, 359 replies, and 1,553,256 impressions. The Verge headline is the cruelty line. Surveillance shows up in the subhead.
I read both versions of the policy. The cruelty bullet is new. The sentence that can still tailor the surveillance ban was already on the page dated September 15, 2025.
What got the impressions
Anthropic’s own summary went up the same day. The updated Usage Policy takes effect November 12. The live policy page already shows that effective date. The previous version is still linked from it.
The new abuse bullet sits under “Do Not Engage in Cruel, Abusive, or Psychologically Harmful Conduct.” It says not to engage in sustained and needless abusive or cruel behavior toward the models. Anthropic says this is for extreme cases, where a user repeatedly acts cruelly toward the model with no discernible purpose. Ordinary frustration, pushback, dark creative themes, and model testing and research are out of scope.
The enforcement they describe is a feature they already had. Claude can end rare conversations with persistently abusive users on Claude.ai and Claude Code. Ending the chat remains the primary mechanism. The October 8 post does not say the account gets banned.
The September 15, 2025 text does not have that bullet. I checked the archive. The cruelty rule is new. It is a small change next to the section on who can be tracked.
What the surveillance section says now
The section is now titled “Do Not Use for Surveillance, Censorship, or Prohibited Law Enforcement or Criminal Justice Purposes.”
One bullet says not to make or suggest decisions in law enforcement or criminal justice processes. The list includes decisions to investigate, charge, arrest or detain, prosecute, sentence, or grant or revoke early or conditional release. Last year’s line was shorter. It covered determinations on criminal justice applications, including decisions about parole or sentencing. “Suggest” and “investigate” are in the new text. Anthropic’s news post says Claude cannot be used to decide or recommend who to investigate, arrest, or charge.
Tracking got a longer sentence. The 2025 line banned targeting or tracking a person’s physical location, emotional state, or communication without consent. It named facial recognition, battlefield management, and predictive policing. The new line bans targeting or tracking a person or their property without consent by monitoring physical location, biometrics, movements, online activities, communications, or associations. Real time, an ongoing watch, or analysis of previously collected data. Monitoring in a law enforcement investigation is now in the examples.
There is a new bullet on building, improving, or enhancing the surveillance capabilities of tools designed for those prohibited uses. Another spells out doxxing. Do not identify, unmask, or locate people from anonymous or pseudonymous activity, or compile dossiers, in order to harass, intimidate, surveil, or cause harm.
Anthropic says the rewrite does not change what they enforce. Their post says this update “does not reflect a change in what we enforce in practice: it is just intended to make our policy more explicit.” They used the same framing for the weapons section.
The 2025 surveillance section did not include this carve-out paragraph. The new one does. It does not prohibit analysis or tracking that individuals have agreed to, such as fraud detection or anti-money laundering screening on a financial account, or consumer apps where a user has opted in. It also does not prohibit aggregated or anonymized analysis, content moderation, authorized security research, legal research and analysis by law enforcement agencies or courts, or journalism, provided those activities are not used for the purposes prohibited above.
Suggesting who to investigate is prohibited. Legal research and analysis by a law enforcement agency is permitted, if it is not used for a prohibited purpose. A dossier compiled to surveil someone is prohibited. Aggregated or anonymized analysis is not. Previously collected data cannot be used to track a person who did not consent. A fraud screen on a financial account can, under “agreed to.”
A bank AML screen and a person agreeing to be followed are not the same event. The policy offers the AML screen as an example of tracking people have agreed to. I would not sign that. The “provided” clause is how both lists stay in the document. Whether a given request is research or a suggestion to investigate is Anthropic’s call. The page does not show the calls.
The tools bullet works the same way. If a tool is designed for a prohibited surveillance use, Claude cannot be used to improve it. If the tool is a fraud screen, or an app the user opted into, that prohibition does not apply. The label on the tool is what the rule looks at.
The sentence that did not move
Above the rules, both versions carry the same paragraph. I compared the September 15, 2025 archive to the page now dated effective November 12, 2026. This sentence matches.
“Anthropic may enter into contracts with certain governmental customers that tailor use restrictions to that customer’s public mission and legal authorities if, in Anthropic’s judgment, the contractual use restrictions and applicable safeguards are adequate to mitigate the potential harms addressed by this Usage Policy.”
The public bullets are the default. A government customer can get a tailored set of restrictions if Anthropic decides the contract and the safeguards are enough. That judgment sits with Anthropic. The tailored text is not on the usage-policy page.
Hayden’s post is what people saw. 1,553,256 impressions on a cruelty ban, and a surveillance rewrite the company says does not change enforcement. The sentence that can still tailor those restrictions for a government customer was already in last year’s policy. This week’s edit left it there.
I do not need a theory of model welfare for this. If the work is tracking someone who did not agree, the model is the wrong tool, and a usage policy is not the control when a contract can tailor the prohibition. This update left that sentence where it was.