Direct answer: Anthropic’s revised Usage Policy takes effect on November 12, 2026. If you use Claude, Claude Code, the Anthropic API, a cloud-provider version of Claude, or a product that embeds Claude, audit the workflow now—especially customer-facing agents, high-risk recommendations, autonomous physical actions, deceptive marketing automation, surveillance, and regional access.[1][2]
The five checks to complete before November 12
- Inventory every Claude workflow. Include Claude.ai, Claude Code, API calls, MCP connections, browser/computer-use agents, scheduled automations, cloud-provider access, and customer-facing products.
- Flag high-risk recommendations. Check whether Claude gives individualized advice or helps decide outcomes in legal, medical, financial, credit, insurance, housing, employment, education, healthcare-access, public-benefit, or legal-status matters.[2]
- Add human review and AI disclosure where required. A qualified person must meaningfully review covered recommendations and be able to change them before delivery or use; the affected person must be told that AI was involved.[2]
- Review autonomous hardware. Where Claude can control equipment capable of injury, a qualified operator must be able to observe and stop it, and the equipment must hold a safe state if Claude disconnects.[1][2]
- Check users, entities, ownership, and location. Anthropic restricts access based on physical location, incorporation or headquarters, and majority ownership or control—not only the country shown in an account profile.[1][3]
Do not assume this is only for API developers. Anthropic says the policy applies to anyone submitting inputs to its products or services, including users of Claude apps and Claude Code, API and developer-platform customers, cloud-provider and authorized-reseller customers, and end users of products that integrate Claude.[2]
What changed in the 2026 policy update?
Anthropic says most revisions clarify rules it already enforced, rather than creating entirely new restrictions. The update consolidates deceptive activity into a dedicated section, sharpens election, weapons, surveillance and law-enforcement language, explains high-risk requirements more precisely, adds controls for models connected to potentially dangerous hardware, and addresses sustained needless abuse of its models.[1]
The operational lesson is simple: treat November 12 as a review deadline, not as permission to ignore a risky workflow until then.
1. Deceptive campaigns and artificial activity
The policy prohibits fake personas, accounts, outlets or organizations used to mislead people about origin, attribution, popularity or whether they are interacting with a human. It also covers undisclosed sponsorship intended to influence opinion or decisions, networks of apparently independent outlets hiding a common source, fake reviews, astroturfing, scaled impersonation, and tools built to enable those campaigns.[2]
The text specifically prohibits manipulating search engines or AI answers by seeding networks of content that misrepresent their authorship, origin or independence.[2]
Practical action: stop any workflow that creates fake reviewers, undisclosed synthetic personas, coordinated “independent” posts, or fabricated corroboration sites. Legitimate SEO, PR, customer support and content repurposing still need truthful identity and attribution.
2. Elections and political communication
Anthropic’s revised section focuses on deceptive voter influence and disruption. It prohibits impersonating candidates or election authorities, false voting information, turnout suppression through deception or intimidation, and scaled automated messages that hide their artificial origin.[1][2]
Anthropic says it removed a blanket ban on personalized vote or campaign targeting because that wording also captured legitimate civic work, while deceptive targeting and misuse of personal data remain restricted elsewhere in the policy.[1]
3. Weapons, surveillance and law enforcement
The weapons section explicitly covers software and components used to test or operate weapons, targeting and fire-control functions, and integrating weapons with drones or autonomous platforms.[1][2]
The surveillance section prohibits non-consensual tracking, building or improving prohibited surveillance systems, certain biometric categorization, predictive-policing uses, and using Claude to recommend who should be investigated, arrested or charged. The policy retains permitted examples such as consent-based fraud monitoring, anonymized analysis, content moderation, authorized security research, legal research and journalism when they are not used for prohibited purposes.[1][2]
4. Sustained abusive behavior toward models
Anthropic added a prohibition on sustained and needless abusive or cruel behavior toward its models. It says the rule is aimed at extreme repeated behavior without a discernible purpose, not ordinary frustration, pushback, dark creative themes, or model testing and research.[1]
High-risk Claude use: the decision test
A workflow enters Anthropic’s high-risk review when it produces a recommendation that can affect an individual’s health, legal rights, finances, livelihood, housing, education or access to essential services.[2]
Use this test:
- Is the output about a specific person? General education is different from applying advice to an individual’s circumstances.
- Does Claude exercise judgment? Applying a fixed formula or carrying out a decision a person already made may fall within an exclusion, while ranking or recommending usually needs closer review.[2]
- Will someone rely on it or will it implement a decision? Internal drafting or research that is not the final recommendation is listed as an exclusion, but the final individualized advice or decision can be covered.[2]
- Can a qualified person change it before delivery or action? Merely viewing a result after it is sent is not meaningful pre-delivery review.
- Is AI involvement clearly disclosed to the affected person? The disclosure does not have to identify Anthropic or the specific Claude model.[2]
The 11 covered recommendation areas
Anthropic lists these areas: legal, medical, finance, credit, insurance, housing, employment, education and credentials, healthcare access, public benefits and services, and legal status and adjudication.[2]
Examples include personalized legal or tax advice, diagnosis or treatment recommendations, investment recommendations, credit decisions, insurance underwriting or claims decisions, tenant ranking, candidate screening, consequential grading, healthcare triage, benefit eligibility, and immigration or citizenship determinations.[2]
Important: Anthropic treats law-enforcement and criminal-justice recommendations as prohibited, not as a high-risk use that becomes acceptable merely by adding human review.[2]
Controls required for covered recommendations
Qualified human review
A qualified person must meaningfully review the recommendation before it is delivered as advice or used for a decision. That reviewer must have authority to change it, relevant training or experience, and a license when the underlying activity legally requires one.[2]
A weak “human in the loop” is not enough. Avoid rubber-stamp approval, review after delivery, reviewers without domain competence, or interfaces that hide the evidence needed to challenge a recommendation.
AI disclosure
The affected person must be clearly told that AI was used to produce the advice or decision. Anthropic says the disclosure does not need to name Anthropic, Claude or the model.[2]
Copy-paste disclosure:
AI assisted in preparing this recommendation. A qualified human reviewer evaluated it and may change it before it is used. Contact [team/role] if you want clarification or human review.
Adjust the wording to match the actual workflow and applicable law. Do not claim human review if none occurred.
Useful exclusions—but verify them carefully
Anthropic lists exclusions including general educational information, wellness content, explanations of advice already given by a qualified person, internal drafting or research that is not the final recommendation, fixed-rule execution where the model exercises no judgment, and business operations that do not advise or decide about a specific person.[2]
An exclusion is not a shortcut. Document why it applies, who approved that classification, and what change in the workflow would trigger re-review.
Claude connected to robots or physical equipment
The policy covers model-driven physical actions involving shared-space movement, machinery that can apply injurious force, hazardous energy or materials, actions on the human body, safety systems, and industrial processes where a fault could injure workers or the public.[2]
For covered hardware:
- a qualified operator must be able to observe the equipment and intervene;
- the operator must have an effective stop mechanism;
- the system must move to or hold a safe state if Claude disconnects;
- the design should account for prompt injection, manipulation and unexpected tool output;
- plans, code or commands should be reviewed before execution when human approval is feasible.[1][2]
Anthropic lists exclusions for consumer devices operating within built-in safety limits where the worst outcome is discomfort or inconvenience, monitoring without control, reviewed plans or commands, and enclosed lab automation handling non-hazardous material.[2]
Customer-facing bots, minors and agents
Every consumer-facing chatbot or agent that communicates directly with external users must disclose that users are interacting with AI, at least at the beginning of each session or in the interface.[2]
Products allowing minors to interact directly with Anthropic-powered features need additional safeguards. Anthropic’s guidance identifies age verification, content moderation, monitoring and reporting, educational safety guidance, compliance with child-safety and privacy law, and disclosure that the user is interacting with AI.[4]
Agent builders remain responsible for actions taken through tools, browsers and connected systems. An agent does not become compliant just because a human wrote the original goal.[2]
Supported-region audit
Anthropic’s policy says access is restricted for people physically located in unsupported regions, entities incorporated or headquartered there, and entities majority-owned or controlled by persons or entities in unsupported regions.[3]
Use this checklist:
- verify the current official region list rather than relying on an old screenshot;
- check remote staff and contractors by physical work location;
- check the entity’s incorporation and headquarters;
- check direct and indirect majority ownership or control;
- apply the same review to cloud-provider, reseller and embedded access;
- stop using VPNs or routing methods intended to hide an unsupported location.
Pakistan and the United Arab Emirates appear on Anthropic’s current supported-regions list, but that does not override the ownership, control or physical-location tests for a particular user or entity.[3]
30-minute Claude compliance audit
Minutes 0–5: make the inventory
Create one row per workflow:
Workflow name:
Owner:
Claude access route: app / Code / API / cloud / reseller / embedded
Users and regions:
Inputs and sensitive data:
Outputs:
Tools or systems Claude can control:
Who receives or relies on the result:
Minutes 5–15: classify risk
Mark each workflow:
[ ] Prohibited activity risk
[ ] High-risk individualized recommendation
[ ] Autonomous physical action
[ ] Consumer-facing AI disclosure required
[ ] Product serving minors
[ ] Regional-access concern
[ ] Internal drafting/research exclusion
[ ] General-information exclusion
[ ] No listed trigger found
Minutes 15–25: test controls
[ ] Qualified reviewer is named
[ ] Reviewer can change the output before delivery/action
[ ] Review evidence is logged
[ ] AI disclosure is visible
[ ] Tool permissions use least privilege
[ ] Destructive or external actions require approval
[ ] Emergency stop and safe-state behavior are tested
[ ] Region/ownership check is documented
[ ] Incident owner and escalation route are defined
Minutes 25–30: record a decision
Decision: approve / restrict / pause / retire
Reason:
Policy section:
Controls added:
Evidence link:
Owner:
Review date:
Next review trigger:
FAQ
When does Anthropic’s 2026 Usage Policy take effect?
Anthropic says the revised policy takes effect on November 12, 2026.[1]
Does the policy apply only to Anthropic API users?
No. Anthropic says it applies to users of its apps, Claude Code, API and developer platforms, cloud-provider and reseller access, and end users of products integrating Claude.[2]
Must a customer-facing Claude chatbot disclose that it is AI?
Yes. The policy requires consumer-facing chatbots and agents communicating directly with external users to disclose that the user is interacting with AI, at least at session start or in the interface.[2]
Is human review required for every Claude output?
No. The specific requirement applies to covered high-risk recommendations and physical actions. Anthropic lists exclusions such as general information, internal drafting that is not the final recommendation, and certain fixed-rule or reviewed workflows.[2]
Can Claude make hiring or credit decisions if a human reviews them?
Covered employment and credit recommendations require qualified human review and disclosure, but the workflow must also comply with applicable law and all other policy restrictions. Human review does not make a separately prohibited use acceptable.[2]
Does the policy ban users for criticizing Claude?
Anthropic says the abusive-behavior rule does not cover ordinary frustration, pushback, dark creative themes, or model testing and research; it targets sustained, needless cruelty in extreme cases.[1]
Are Pakistan and the UAE supported?
Both are listed on Anthropic’s current Supported Regions Policy page. Access still depends on physical location, entity location, and majority ownership or control.[3]