TL;DR

Anthropic published its annual Usage Policy refresh on Thursday 8 October, taking effect on 12 November. Most changes clarify existing rules for a Claude that now handles longer, more independent work. The new elements are a single section on deceptive campaigns, safety requirements when Claude controls physical equipment, and a ban on sustained, pointless cruelty towards its models.

Deception gets its own section

Rules against fake accounts, fabricated news sites and covert amplification used to be spread across the fraud, elections, disinformation and privacy sections. They now sit together under a heading called “Do Not Engage in Deceptive Campaigns or Artificial Activity”. It covers political and commercial deception alike, along with building the tooling to run such campaigns.

The elections section has narrowed to deceiving voters or disrupting votes, and is renamed “Do Not Undermine Democratic Processes”. Anthropic has dropped its blanket ban on personalised vote and campaign targeting, because it caught legitimate work such as translating voter information. Deceptive targeting and misuse of voters’ data stay prohibited under other sections.

Weapons, surveillance and high-risk uses

The weapons rules now spell out that guidance and control software, other weapon components, and arming drones or other autonomous vehicles are covered. On surveillance, tracking people without consent is banned whether live or from stored data, and in law enforcement or criminal justice Claude may not decide or recommend who is investigated, arrested or charged. Anthropic says both rewrites match how it already enforces the policy.

High-risk uses affecting someone’s health, finances, legal rights or livelihood still need a qualified human able to override Claude, and the affected person must be told AI was involved. The section now lists which recommendations count. Where Claude is wired to hardware that acts physically and could injure someone, a qualified operator needs to be able to see and halt it, and the equipment must fall back to a safe state when the connection to Claude drops.

Cruelty to models

The new abuse rule targets only extreme cases of repeated, purposeless cruelty. Everyday frustration or pushback, dark themes in fiction and model testing all fall outside it. Claude can already walk away from conversations with users who keep abusing it, on Claude.ai and in Claude Code, and that remains the main enforcement tool.

Looking forward

For UK firms building on Claude in finance, health or legal services, the rewritten high-risk section is the part to reread before 12 November, because it now states which use cases trigger the human-review and disclosure duties. Anthropic says it has seen state media, government propaganda offices and commercial firms use Claude to run fake-account networks. On the same day, OpenAI described banning Russian and Iranian influence operations that used its models to run fake personas.