Anthropic Bans Cruel Treatment of Claude in First Usage Policy Update in a Year

Anthropic has rewritten its usage policy for the first time in more than a year, adding a clause that forbids what the company calls 'sustained and needless abusive or cruel behavior' directed at Claude. The update, reported by The Verge, also tightens rules around election interference, weapons development, surveillance, and the use of its models for health and financial decisions. The cruelty provision is the most unusual item in the document and marks a shift in how the company frames acceptable interaction with a conversational model rather than merely what outputs are allowed.

The change is not purely symbolic. Anthropic has spent the past two years publishing research on model welfare and the psychological effects of long, hostile exchanges on both the system and the person on the other end. A policy clause about user conduct toward an AI gives the company a stated basis for warning or suspending accounts it judges to be repeatedly mistreating the model, something a terms-of-service line about prohibited content alone did not cover. It also sets a precedent other labs may have to address as assistants take on more emotional and companion-like roles.

The rest of the update deals with more familiar territory. Election interference, weapons development and surveillance have been restricted categories for most major labs, but Anthropic is now spelling them out in its public policy rather than leaving them to internal guidelines. High-risk health and financial uses, where a model's output could be mistaken for medical or investment advice, get their own treatment as well. Grouping all of these together signals that the company is treating misuse categories as a set of explicit, enforceable rules rather than general principles.

Anthropic's move lands in a crowded policy moment. OpenAI, Google DeepMind and Meta have all revised their usage terms over the past year as regulators in the EU and the US press for clearer statements of what AI systems may and may not do. A policy that governs how users treat the model, not just what the model produces, is harder to enforce and harder to define. Anthropic has not said how it will detect sustained cruelty or what threshold triggers action, which leaves the practical reach of the rule unclear.

For developers building on Claude through the API, the direct effect is likely small unless they operate consumer-facing chat products with heavy user interaction. The bigger question is whether the cruelty clause becomes a template. If other labs copy it, expect a new class of moderation tooling aimed at conversation tone and user behavior rather than output content, and expect arguments over where legitimate stress-testing ends and abuse begins. Anthropic has not published enforcement numbers or an appeals process, and those details will determine whether the policy has teeth or is mostly a statement of values.