Anthropic announced a new version of the Claude usage policy that takes effect on November 12, 2026, in an update focused largely on clarifying how the existing rules apply as models become more capable of carrying out longer and more autonomous tasks. The company says the policy is also based on patterns of misuse it observed over the past year in influence operations, weapons development, and surveillance.
New section on deceptive campaigns
Anthropic has grouped restrictions related to fake accounts, fabricated news, and concealing the party behind messages into a separate section titled “Do Not Engage in Deceptive Campaigns or Synthetic Activity.” The section prohibits using Claude to create networks of fake accounts or posts with the aim of amplifying content or concealing its source. It also covers building the tools and infrastructure needed to operate influence campaigns, whether political or commercial.
Defining the scope of election-related restrictions
The company renamed the elections section “Do Not Undermine Democratic Processes” and more clearly limited it to uses that deceive voters or disrupt elections, such as spreading false information about candidates or voting procedures, impersonating candidates or election officials, and attempting to suppress voter turnout.
At the same time, Anthropic removed the blanket ban on targeting voters and campaigns with tailored messaging. It explains that some legitimate civic uses, such as preparing voter information in different languages or sending election officials notifications to correct ballots, are no longer covered by that general prohibition. However, targeting based on deception or misuse of personal data remains prohibited under other sections.
Weapons, surveillance, and autonomous systems
The policy clarifies that the ban on weapons development includes guidance and control software and components that make a weapon function, in addition to weaponizing drones and autonomous vehicles. Anthropic says this change reflects how it applies the previous rules, rather than a shift in its enforcement practices.
The new wording also prohibits tracking people without their consent, whether in real time or through analysis of data collected previously, and prohibits using Claude to make a decision or provide a recommendation about investigating, arresting, or charging a person as part of law enforcement or criminal justice proceedings. The restrictions also cover building or improving surveillance tools. Uses approved by individuals, such as fraud monitoring, along with content moderation, journalism, and legal research, remain permitted uses according to the company’s announcement.
What changes in practice?
Anthropic has not changed the requirements for high-impact uses affecting health, legal rights, financial matters, livelihoods, or access to essential services, but it has rewritten them to directly define the covered cases. A qualified human with the authority to review and modify Claude’s recommendations when necessary must be available, and the person affected must be informed that artificial intelligence is being used.
The policy adds requirements for systems that connect Anthropic models to devices capable of autonomously carrying out physical actions that could cause injury. A qualified operator must be able to monitor and stop the equipment, and the equipment must enter a safe state if the connection to Claude is lost.
Additional restrictions and administrative clarifications
Anthropic added a prohibition on continuous and needlessly harsh behavior toward its models, limiting it to extreme cases. This does not include ordinary frustration, disagreement with a model’s answers, dark creative subjects, or testing and research.
It also clarified the supported-regions policy: Claude may not be used by people physically located in an unsupported region, by entities incorporated or headquartered there, or by entities owned or majority-controlled by people or entities in those regions.
Editorial reading: The core value of this update is not the announcement of new restrictions in every area, but the conversion of scattered rules or previous enforcement practices into more specific text as the use of models expands in political influence, physical systems, and sensitive decisions. The practical question for customers remains how to demonstrate the existence of human oversight capable of intervention, individual consent, and a safe state when the model is disconnected, particularly in applications that go beyond producing text to taking actions in the real world.