Anthropic Usage Policy Now Bars Cruelty Toward Claude

Anthropic published a new Anthropic Usage Policy on 8 October, its first refresh of the document in more than a year. It takes effect on 12 November, and most of the changes tighten rules on election interference, weapons software and surveillance.

One of them stands apart. The policy now prohibits what it calls sustained and needless abusive or cruel behavior toward Claude itself.

What the Anthropic Usage Policy actually prohibits

The wording, reported in detail by The Guardian and The Verge, is narrower than the headlines suggest. Anthropic says it applies only in extreme cases, where users repeatedly act cruelly toward its models with no discernible purpose, and lists what it does not cover: ordinary frustration, pushback, dark creative themes, and model testing. Arguing with Claude remains allowed. Making a habit of it does not.

Enforcement is left vague. Ending the conversation is still the primary mechanism, a capability the company described in August under its model welfare research. Anthropic has not said whether suspending the accounts of repeat offenders is on the table.

The larger changes concern deception

Strip out the cruelty provision and the rest is a consolidation exercise. Anthropic gathered rules scattered across its fraud, privacy and disinformation sections into one new section called Do Not Engage in Deceptive Campaigns or Artificial Activity, which covers fake account networks and influence-operation tooling, as TechCrunch reported. The elections section was rewritten as Do Not Undermine Democratic Processes.

One restriction was removed rather than added: a blanket ban on personalized vote and campaign targeting, which Anthropic says had been catching legitimate civic work such as nonprofits translating voter information. The full document is published at anthropic.com/legal/aup.

Weapons, surveillance and hardware

The weapons prohibition now explicitly covers the software that makes weapons work, including arming drones and other autonomous vehicles. Surveillance rules tightened: tracking people without consent is banned, Claude cannot recommend who to investigate, arrest or charge, and consented tracking, content moderation, journalism and legal research all remain permitted.

A new clause covers hardware. When Claude is connected to equipment that takes autonomous physical actions and could cause injury, a qualified operator must be able to observe and stop it. The policy does not specify whether that operator has to be a person.

The pushback

Not everyone accepts the framing. Reason argued that banning cruelty toward a chatbot still treats it as closer to a person than it is. Others ask who benefits when users are polite, given that courteous conversations make for cleaner training data. The more practical criticism is about attention: the cruelty rule is the least consequential item in the document. The provisions governing influence operations, weapons software and state surveillance describe what Anthropic says it has already been fighting.

One clause is easy to miss. Restrictions may be modified for contracts with certain governmental customers when, in Anthropic’s judgment, the safeguards are adequate. Anthropic has previously contracted with the US military, and its models are already deployed inside enterprise security platforms. The company has also taken similarly firm positions before, having called for the AI industry to slow down.

What to watch

The policy takes effect on 12 November, and its importance depends entirely on enforcement. A rule against sustained cruelty is only meaningful if conversations actually end. The duller provisions, how often the government carve-out is invoked and whether the removed vote-targeting ban stays removed, will matter more than whether a chatbot hangs up on someone being rude to it.