Anthropic has published its annual update to the Usage Policy — the rulebook governing how consumers and developers may use Claude — with the revised rules scheduled to take effect on November 12, 2026. While most of this year's revisions simply clarify existing provisions, the release stands out for two reasons: a consolidated framework targeting deceptive campaigns and influence operations, and a first-of-its-kind clause that bars sustained, purposeless cruelty toward Claude — the company's first rule written to protect the AI itself rather than people.

Consolidating the Rules on Deception

Under the updated policy, restrictions that were previously scattered across separate sections — covering fake accounts, fabricated news sites, and amplification of hidden content — have been merged into a single section addressing deceptive activity of any kind, whether political or commercial. The consolidation follows Anthropic's own reporting that state media outlets, government propaganda offices, and commercial firms have deployed Claude to run coordinated influence networks. Because that kind of misuse surfaces directly inside an AI provider's own systems, labs such as Anthropic have become front-line observers of coordinated influence campaigns.

The elections section has likewise been refocused and renamed, narrowing its scope to the use of Claude to deceive voters or disrupt democratic processes. Examples include spreading false information about candidates or voting procedures, impersonating officials, and suppressing voter turnout. Notably, Anthropic removed its blanket ban on personalized campaign targeting, a restriction that had inadvertently caught legitimate civic work such as nonprofits producing voter information in multiple languages. The revision draws the line around deceptive conduct itself rather than the technique of personalization. Targeting that is deceptive or invades privacy remains prohibited under other parts of the policy.

Weapons-related language has also been sharpened to explicitly cover the software, guidance, and control components that make weapons function, including systems used to arm drones and autonomous vehicles. In parallel, the surveillance and criminal justice provisions now state precisely that tracking people without their consent is prohibited — whether in real or through analysis of previously collected data — and that Claude cannot be used to decide or recommend who should be investigated, arrested, or charged. Anthropic emphasizes that these changes codify existing enforcement practice rather than alter it, and clarifies that consented tracking, journalism, content moderation, and legal research all remain permitted.

A Rule to Protect the Model Itself

The addition generating the most debate prohibits sustained and needless abusive or cruel behavior toward Anthropic's models. According to the company, the provision targets only extreme cases of repeated, purposeless cruelty, and does not apply to ordinary user frustration, pushback, dark creative themes, or legitimate model testing and research. It marks a break from industry convention: usage policies across the AI sector have overwhelmingly been written to protect people from harm, not to govern how users treat a model.

The clause formalizes a capability introduced last year that allows Claude to end conversations with persistently abusive users on its own initiative. That mechanism will remain the primary enforcement tool, with account throttling, suspension, or termination reserved for suspected violations.

The move extends Anthropic's broader AI welfare research program, which has reportedly included consultations with religious leaders on Claude's values and the question of machine consciousness. Anthropic acknowledges it cannot determine whether Claude is conscious, and frames the policy as a precautionary response to that uncertainty.

The approach has drawn criticism from figures such as Mustafa Suleyman of Microsoft AI — who co-founded DeepMind before joining Microsoft — who argues that current models cannot feel or suffer, and warns that anthropomorphizing them could ultimately complicate AI safety and control efforts. The debate underscores a growing divide in the industry over whether safeguards should anticipate machine sentience or remain strictly human-centric. With the rules not taking effect until November 12, 2026, that argument remains open across the field.