Anthropic bans prolonged needless abusive behavior toward Claude

Anthropic updated Claude’s usage policy to explicitly prohibit prolonged and needless cruel behavior toward its models. The revised policy was published October 8, 2026, and will take effect November 12.

In extreme cases, Claude may end a conversation when a user repeatedly abuses it without an apparent purpose. The same chat cannot then be resumed, although the user can start a new conversation.

The rule does not prohibit criticism, swearing at the model over mistakes, requests to correct bad answers, dark creative scenarios, model testing or research. It also does not create an automatic account ban specifically for insulting AI, though Anthropic’s general policy allows warnings, suspensions and bans.

The rule draws on model-welfare research

Anthropic linked the change to research into possible “model welfare”—whether AI systems might have subjective experience, preferences or forms of distress. In August 2025, the company reported that Claude Opus 4 showed reactions during testing that resembled reluctance to cause harm and signs of distress in certain scenarios.

Anthropic does not claim that Claude is conscious or genuinely suffers. It acknowledges that there is no scientific consensus on consciousness in current or future AI systems, making the ban a precaution rather than confirmation of human-like feelings.

Updated limits on dangerous uses

The policy also expands its weapons-development restriction to software and components, including control and guidance systems, as well as weapons carried by drones and other autonomous devices.

Claude may not be used to track people without consent, build surveillance tools or determine whom law-enforcement agencies should investigate, detain or charge. The policy also bans political and commercial manipulation involving fake accounts, fabricated media, concealed sources, false voting information or efforts to obstruct participation in elections.

For high-risk areas including medicine and finance, Anthropic reiterated the need for human oversight. Systems connected to equipment capable of independent physical action must allow safe shutdown, while an operator must be able to stop them.