Anthropic has revised the usage rules for its Claude AI models, adding a prohibition on sustained, purposeless cruelty toward the software alongside broader changes governing weapons, surveillance, political influence and safety-critical equipment. The revised policy is scheduled to take effect on November 12, 2026.
The company describes the cruelty provision as a narrow rule for extreme and repeated conduct with no discernible purpose. It says ordinary frustration, disagreement, dark creative material, and model testing or research are outside the restriction. Claude's existing ability to end some abusive conversations is expected to remain the main response. Anthropic introduced that mechanism for certain model versions in August 2025 as part of work it calls model-welfare research.
The distinction matters because there is no established evidence that a chatbot experiences distress. The policy regulates user behavior toward the system without resolving questions about machine consciousness. The supplied reporting also does not explain whether an ended conversation can lead to account sanctions or disclose how often Claude has invoked the conversation-ending feature.
Other provisions address more immediate safety risks. Anthropic has clarified that its weapons restrictions cover software and components, not only complete weapons. The clarification follows attempts to use Claude for guidance and control work involving armed drones and autonomous vehicles, according to the report.
The surveillance rules prohibit using Claude to decide whom police should investigate, arrest or charge. They also bar building surveillance tools and tracking people without consent, whether tracking occurs live or after the fact. Anthropic characterizes several of these revisions as clearer statements of existing restrictions.
Political-use rules have also changed. The company removed a broad prohibition on personalized political targeting after concluding it could capture legitimate activities, including nonprofit organizations translating voter information. Deceptive targeting and misuse of personal data remain barred. A separate restriction covers deceptive campaigns and artificial activity, after Anthropic said it observed state media, government propaganda operations and commercial groups using Claude to operate fake accounts and fabricated news sites.
For Claude-connected physical equipment capable of causing injury, the policy requires a qualified human operator who can monitor and stop the system. The equipment must also move to or remain in a safe state if its AI connection fails. These provisions treat a language model's output as one component in a supervised system rather than an autonomous safety authority.
The cruelty clause is likely to attract attention because it applies social language to software. In operational terms, however, it is one part of a wider policy update focused on how increasingly capable models may be connected to weapons, public decisions, influence operations and physical machinery. Enforcement details and real-world incident rates remain important unanswered questions.
