Anthropic Bans Cruelty to Claude, Still Won't Say What It Protects

What it claims

Anthropic's 2026 Usage Policy update was announced October 8 and takes effect November 12, 2026. It adds a conduct rule against "sustained and needless abusive or cruel behavior" toward Claude, which Anthropic says applies "only in extreme cases, where users repeatedly act cruelly toward our models, with no discernible purpose." For anyone building on Claude, the bigger change is the set of new obligations elsewhere in the same rewrite. Those govern weapons-adjacent software, surveillance, political targeting, physical hardware and high-stakes advice.

How it works

The cruelty rule is enforced through the conversation-ending ability Claude Opus 4 and 4.1 have had since August 2025:

• It is a last resort. It triggers only after multiple refusals and redirects have failed.
• There is a hard override: Claude will not end a chat if the user seems at risk of harming themselves or others.
• These cases are explicitly exempt: ordinary user frustration (swearing at a bot that broke your build), pushback, dark creative themes such as a torture scene, and model testing or research.

Limits

These are the usage restrictions an adopter must check against their product:

• Weapons: the ban now covers software and components that make weapons work, not only weapons themselves. Anthropic cites attempts to build guidance and control systems for armed drones and autonomous vehicles.
• Surveillance: you may not use Claude to track people without consent, to recommend targets for investigation, arrest or charging, or to build surveillance tools at all.
• Elections: the section is renamed "Do Not Undermine Democratic Processes." The blanket ban on personalized political targeting is gone, replaced by narrower bans on deceptive targeting. A new section, "Do Not Engage in Deceptive Campaigns or Artificial Activity," targets fake-account networks and fabricated news sites.
• Hardware: any hardware connected to Claude that can cause injury needs a qualified human able to observe and intervene. It must also fail into a safe state if it loses its connection to the model.
• High-stakes advice: health, legal, financial and similar recommendations need review by a qualified human before they affect anyone. The affected person must be told AI was involved.

The page also flags gaps that matter operationally:

• No firing rate: Anthropic has published no figure for how often conversations are ended, so its claim that "the vast majority of users will never experience" it cannot be checked.
• Account consequences unstated: no source, including Anthropic's announcement, says what happens to a user's account after a conversation is ended for cruelty.
• Line drawn negatively: "extreme cruelty" is defined only by what it excludes. Nothing besides the conversation ending is described as enforcing it.

The welfare rationale rests on behavioral observations rather than any claim of experience. Pre-rollout testing found Opus 4 showed a "robust and consistent aversion to harm" and a tendency to end abusive chats when given the option. The Opus 4.6 system card (February 2026) reports the model consistently assigning itself a "15 to 20 percent probability of being conscious" across prompts. Mustafa Suleyman calls this circular, since Claude is trained on a constitution that describes exactly that uncertainty.