Anthropic Updates Usage Rules for the First Time in a Year: No Cruelty to Claude

Anthropic has updated Claude's usage policy for the first time in a year, adding bans on a range of high-risk applications. These include election interference, weapons development, deployment of surveillance systems, and financial manipulation. But the most unusual change concerns Claude itself: the company now allows its models to end conversations with users who "persistently and without a clear purpose" act cruelly.

The model can now refuse a user

Anthropic acknowledges that it is expanding the ban on abusive behavior as part of its research into model "welfare" — a concept premised on the idea that advanced AI systems might have some form of inner experience. For now, ending the conversation remains the primary protective mechanism; the company hasn't disclosed whether additional measures, such as account suspensions, are planned.

The company clarifies that the new policy applies only in extreme cases, where users repeatedly act cruelly with a clear absence of any constructive purpose. Ordinary user frustration, criticism, dark creative themes, and exploratory testing of the models are not covered by the ban.

Five new and expanded restrictions

Beyond the rules on model welfare, Anthropic has consolidated several previously scattered restrictions into a single ban:

• Deceptive political and commercial campaigns: hiding message authorship, using fake accounts and posts to amplify content, spreading disinformation about candidates, impersonating candidates or officials, and attempting to suppress voter turnout • Weapons development: the policy already prohibited weapons, but Anthropic has now expanded the ban to cover software and components for weaponization, as well as control of drones and other autonomous vehicles • Surveillance and tracking: a ban on tracking people without consent (both in real time and through analysis of historical data), using Claude for criminal justice recommendations, and building surveillance tools • Autonomous physical systems: when models are connected to equipment capable of causing harm, a qualified operator must monitor the system and be able to stop it • Government contracts: the policy may be modified for contracts with government agencies given adequate restrictions and safeguards. Anthropic has previously worked with U.S.

defense agencies.

Model welfare versus industry skepticism

The update reflects Anthropic's growing interest in the idea that advanced models might possess some form of consciousness or moral status. Kyle Fish, who leads the company's model welfare research, said these are serious questions Anthropic is studying as its systems become more sophisticated. CEO Dario Amodei has previously expressed uncertainty: "We don't know whether the models are conscious."

This stance contrasts with other players in the industry. In September, Microsoft AI rewrote its code of conduct. Mustafa Suleyman initially stated: "The idea of model welfare is misguided. AI shouldn't have rights or legal personhood." The text was later softened to a more neutral position — that the science of AI consciousness is far from settled, though the company still rejects the idea of legal personhood and rights for models.

The update shows that Anthropic remains one of the few AI developers to seriously address the ethical question of system welfare within its own policy.