Anthropic, the company that develops the Claude models, published a new version of its usage policy (“Usage Policy”) on October 8, 2026. The revised text will take effect on November 12. The company points out that it updates this document every year, in line with the evolving capabilities of its models and feedback from its customers. According to its announcement, most of the changes aim to clarify existing rules; a few additions, however, are genuinely new.

Mostly clarifications

Anthropic explains that Claude now performs longer and more autonomous tasks, and that the new version adds examples showing how the rules apply to these capabilities. The text draws on abuses observed over the past year in three areas: influence operations, weapons development and surveillance.

  • Deceptive activities. The prohibitions, previously spread across the sections on elections, fraud, privacy and disinformation, are grouped together in a new section titled “Do Not Engage in Deceptive Campaigns or Artificial Activity”. It targets any deceptive activity, political or commercial: concealing the author of a message, amplifying content through fake accounts, or building the tools used for influence campaigns.
  • Weapons. The section specifies that the prohibition also covers the software and components that make a weapon work, such as guidance software, as well as the weaponization of drones and other autonomous vehicles. According to Anthropic, this reflects how the rule was already being applied.
  • Surveillance and criminal justice. Tracking people without their consent is prohibited, whether in real time or through the analysis of data already collected. Claude cannot be used to decide or recommend who should be investigated, arrested or charged. Consensual monitoring (such as fraud detection), content moderation, journalism and legal research remain explicitly permitted. Here too, the company says its practice is not changing.
  • High-risk uses. When a use may affect a person’s health, rights, finances, livelihood or access to essential services, Anthropic still requires a qualified “human in the loop”, empowered to review and modify Claude’s recommendations, as well as disclosure to the person concerned that AI is being used. These requirements are unchanged; the section now lists which recommendations are covered and which are not.

What is genuinely new

Three points change the substance of the rules. First, requirements apply to models connected to hardware that performs autonomous physical actions capable of causing injury: a qualified operator must be able to observe the equipment and stop it, and the equipment must be able to remain in a safe state if Claude is disconnected.

Second, the section on elections, renamed “Do Not Undermine Democratic Processes”, focuses on deceiving voters and disrupting elections. Anthropic removes the general prohibition on personalized voter targeting, which in its view blocked legitimate civic uses; targeting based on deception or the misuse of personal data remains prohibited by other sections.

Finally, a prohibition targets persistent abusive or cruel behavior toward the models with no apparent purpose. It only concerns extreme cases, not ordinary frustration, disagreement, dark creative themes or model testing.

Supported regions: a clarification on enforcement

Anthropic has also clarified its policy on supported regions. Use of Claude is prohibited for people physically located in an unsupported region, for entities incorporated or headquartered there, and for entities majority-owned or controlled by people or entities from those regions. The announcement does not list the regions concerned; it refers to the dedicated page.

A provider’s terms, not a law

This text should be read for what it is: a contractual commitment set unilaterally by a private provider toward its customers. It determines what Anthropic allows to be done with its models and what it can sanction, but it does not replace applicable law. A Swiss company that uses Claude must therefore comply with both these terms and its own legal obligations, for which it alone remains responsible. Complying with the usage policy does not guarantee legal compliance, and vice versa. For the teams concerned, particularly in healthcare or finance, the November 12 deadline is an opportunity to reread the primary text and check whether their use cases fall within the list of high-risk uses.

Sources