Anthropic has published its 2026 usage policy, which takes effect on November 12. It is the first rewrite in more than a year, and the provision drawing the most attention is a new ban: users may not subject Claude to persistent and gratuitous abuse or cruelty.
Yet the clause takes up very little room in the overall update. According to the company, most of the changes clarify existing rules and add examples for a time when Claude takes on longer, more autonomous work.
How the policy defines "abuse" of a model
Anthropic confines the ban to extreme cases: a user repeatedly showing cruelty toward the model with no discernible purpose. The company carved out what is not affected:
"It does not apply to common versions of user frustration, pushback, dark creative themes, or model testing and research."
In other words, ordinary user annoyance, arguing back, creative writing with dark themes, and model testing and research all fall outside the rule.
Enforcement leans on Claude's own ability to end a conversation. That feature arrived in August 2025 for Opus 4 and 4.1 and is now live in both Claude.ai and Claude Code; the company says it will remain the main enforcement tool. The policy text does not say whether accounts could be banned over this. Some media reports say the safety team retains the power to warn, rate-limit and deactivate accounts, but that cannot currently be verified from the official text.
The other changes
- Deceptive activity gets its own section: rules previously scattered across elections, fraud, privacy and misinformation are merged. Whether the aim is political or commercial, the policy bars concealing where information comes from, using fake accounts or fake posts to amplify content, and building tools for influence operations.
- The election clause is narrower: it now prohibits only deceiving voters and disrupting elections. The blanket ban on personalized voter targeting is gone, because it caught legitimate work such as nonprofits offering voters multilingual information.
- The weapons clause is spelled out: it covers guidance and control software and components for weapons, and adding weapons to autonomous vehicles such as drones. The company says this reflects how the rules were already enforced and is not a new ban.
- The surveillance clause is rewritten: it bars tracking people without their consent, using Claude to decide or recommend whom law enforcement should investigate, arrest or prosecute, and building or improving surveillance tools. Consent-based fraud detection, content moderation, and journalistic and legal research are explicitly allowed.
- High-risk scenarios get a longer list: uses that affect a person's health, legal rights, finances or livelihood still require review by a qualified person and disclosure to those affected that AI was used. A new requirement applies when Claude is connected to hardware that can take physical actions on its own: a qualified operator must be able to observe and stop the device, and the device must remain safe once Claude disconnects.
What it means for users in China
The part most directly relevant to users in China is the enforcement note in the "supported regions" section. The new text lists three categories of prohibited parties: individuals located in unsupported regions; entities registered or headquartered in those regions; and entities whose majority ownership or control sits with individuals or entities in those regions.
Mainland China, Hong Kong and Macau are all absent from Claude's supported list. Drawing the line by ownership share is something Anthropic already put forward in its September 2025 sales restrictions; this time it is written into the body of the usage policy. When a number of Hong Kong users had their accounts deactivated on October 1, Anthropic did not respond publicly. Once the new policy takes effect, such actions will have clearer wording to rest on.
For teams that registered overseas but are controlled by Chinese capital and are expanding abroad, the change is that compliance risk moves from the sales terms into the usage policy everyone must follow. Anthropic has not said how ownership structures would be checked under the new terms.
Companies in China working on embodied AI should also take note of the physical-hardware provision. Many robotics projects use overseas large models for high-level planning, and under the new rules the stop mechanism and the safe state after disconnection have to be considered at the product design stage.
Sources: Anthropic's official policy update notes, The Verge, CocoLoop, MacRumors; the text of Anthropic's usage policy was checked for the effective date and the wording of each clause.