Anthropic published a revised Claude Usage Policy on October 8, 2026. The company says it takes effect on November 12. Most edits clarify existing rules, while the update gives clearer examples for deceptive campaigns, privacy, high-risk decisions and AI systems that can take actions beyond a chat. Teams should review the policy before the effective date and compare it with the real capabilities of their product.
What changes on November 12?
Anthropic says the new text reflects the way Claude is being used for longer and more independent work. It does not mean every edited line creates a new ban: the company says many changes explain rules it already enforced. There is no single migration step for all customers; the right review depends on purpose, data, users and the actions a system can take.
Deceptive campaigns and election integrity
The policy groups deceptive activity in a section called “Do Not Engage in Deceptive Campaigns or Artificial Activity.” It covers hiding who is behind a message, amplifying content through fake accounts or posts, and building infrastructure for coordinated influence campaigns.
The election section focuses on deception and disruption, such as misleading people about voting, impersonating candidates or election officials, and trying to suppress turnout. Anthropic also removed its blanket restriction on personalized vote and campaign targeting. It says legitimate civic work, such as preparing voter information in other languages, can be allowed when it does not involve deception or misuse of personal data.
Marketing and communications teams should separate ordinary content production from fake identities and artificial engagement. Keep account ownership clear, preserve approvals and make sure automated publishing cannot hide who is responsible for a message. Our summary of Anthropic's threat reporting explains why source identity and workflow controls matter.
Physical harm, surveillance and law-enforcement use
The revised policy makes clear that restrictions cover software and components that support prohibited physical harm, not only direct instructions. It also clarifies limits on surveillance and law-enforcement uses.
Anthropic says tracking people without consent is prohibited, whether it happens live or through analysis of previously collected data. Claude cannot be used to decide or recommend who should be investigated, arrested or charged, or to build or improve surveillance tools. The policy describes permitted areas such as consent-based fraud monitoring, content moderation, journalism and legal research.
For developers, review the goal, data source, permissions and real-world effect instead of relying only on a feature name or a user's description of a project.
High-risk decisions still need human oversight
Anthropic says its existing human-in-the-loop requirements remain in place for uses that can affect health, legal rights, finances, livelihood or access to essential services. A qualified person must be able to review and change Claude's recommendation, and the affected person must be told AI was used.
For systems connected to physical equipment that could cause injury, Anthropic adds clearer requirements for a qualified operator who can observe and stop the equipment. The equipment must also be able to reach a safe state if Claude disconnects.
A review button is not enough if nobody owns the decision. Assign a reviewer, define which actions need approval, record the decision and test the stop procedure before connecting Claude to a high-impact workflow or physical system.
What the policy says about abusive behavior
Anthropic added a prohibition on sustained and needless abusive or cruel behavior toward its models. It says the rule targets extreme cases and excludes ordinary frustration, pushback, dark creative themes, model testing and research. This is a narrow policy rule, not a claim that Claude is conscious or that every harsh prompt is prohibited.
Checklist for Claude teams
- Inventory Claude workflows, including scheduled jobs and agents that can act without a new prompt.
- Review workflows that publish messages, use personal data, track people or connect to physical systems.
- For high-impact decisions, confirm that a qualified person can review and change the result.
- For physical systems, test the operator stop control and safe state if the model or network fails.
- Check the Supported Regions Policy if your team, customers or processing span countries.
- Record the policy version reviewed and assign an owner to check future changes.
If your team customizes Claude Code, our Claude Code Mods guide explains why custom behavior and tool access should be reviewed along with written instructions.
What remains unclear
Anthropic has not published an enforcement outcome for every example. Because some wording clarifies rules the company says it already enforced, readers should not treat every edit as a new restriction. Check the live policy for the exact wording that applies to your use case.
Frequently asked questions
When does the revised policy take effect?
Anthropic says November 12, 2026. The announcement was published on October 8.
Does the update ban ordinary frustration with Claude?
No. Anthropic says the new rule targets sustained, needless cruelty in extreme cases and excludes normal frustration, testing and research.
Can Claude still help with voter information?
Yes, legitimate civic work can remain allowed when it does not involve deception, voter suppression or misuse of personal data.
Do high-risk decisions need a human reviewer?
Yes. A qualified person must be able to review and change recommendations in covered high-risk use cases, and the affected person must be told AI was used.
Sources
- Anthropic: Usage Policy (revised October 8, effective November 12, 2026)
- Anthropic: Supported Regions Policy