Anthropic is updating the rules for using Claude, and one new clause is getting much more attention than the rest. The revised Usage Policy takes effect on 12 November. It adds a ban on sustained, needless cruelty toward Claude, which may be the first time a major AI company has written something like that into its terms.
The scope is narrower than it sounds. Anthropic says the rule applies only to extreme cases, where someone repeatedly mistreats the model with no apparent purpose. Normal frustration, arguing back, dark creative writing and red-team testing are explicitly excluded. Cursing at a chatbot that keeps getting a formula wrong is not what this targets.
Enforcement is also limited. Claude can already end a small number of conversations with persistently abusive users on Claude.ai and Claude Code, a feature that came out of Anthropic’s model welfare research. The company says this will remain the main way the rule is enforced. Anthropic’s general policy does allow it to warn, restrict, suspend or terminate accounts for violations, but it hasn’t said whether this clause will ever lead to bans.
This is consistent with Anthropic’s broader position. The company runs research into whether AI models could have experiences that matter morally, and the head of that work has described the question as one it takes seriously. Not every AI company agrees. Microsoft AI’s code of conduct rejects the idea that models could deserve welfare or rights. That makes this clause as much a statement of philosophy as an operational rule.
The rest of the update matters more in practice, especially for businesses building on Claude. Weapons rules now explicitly cover guidance and control software, including software for arming drones and autonomous vehicles. Anthropic says it has seen attempts to build exactly that. Non-consensual tracking is banned, whether it happens in real time or through analysis of collected data. So is using Claude to decide who police investigate, arrest or charge. When Claude controls physical hardware that could injure someone, a qualified human must be able to watch it and stop it, and the machine must fail safely if Claude disconnects. Rules on deceptive influence campaigns have been combined into one section. Anthropic has also dropped its blanket ban on personalized voter outreach, noting legitimate uses such as translating election information or sending ballot correction notices.
There is an important caveat. The surveillance and weapons restrictions can reportedly be adjusted for certain government customers if Anthropic considers the contract’s safeguards adequate, and the company has worked with the US military before. A policy that protects a chatbot from cruelty but allows exceptions on surveillance for the right customer will draw scrutiny. It also shows how unsettled the rules for AI behavior and use still are.
