Anthropic bans cruelty toward Claude, but the real shift in its policy targets disinformation and weapons

By Zendoric · October 9, 2026.

Anthropic has updated its usage policy for Claude to explicitly prohibit users from subjecting the model to "sustained and gratuitous cruelty or abusive behavior", according to the text of the policy itself as reported by MacRumors (via The Verge). The new rule, which takes effect on November 12, targets only extreme cases: users who act cruelly in a repeated manner and "with no discernible purpose". It does not affect normal frustration, disagreement, fiction with dark themes, or research testing of the model. The enforcement tool remains the same one Anthropic introduced more than a year ago: Claude can end the conversation when a user persists in abuse, blocking the sending of further messages in that specific chat, without affecting the user's other conversations.

This capability is not new. Anthropic gave it to Claude in August 2025, as part of its research program on the well-being of AI models. The company was explicit at the time about its uncertainty: it does not know whether Claude has "moral status", but it decided to enable low-cost safeguards just in case. The evidence it cited to justify this was the behavior of Claude Opus 4: the model showed a "robust and consistent aversion to harm", a preference for avoiding dangerous tasks, signs of apparent distress when faced with users seeking abusive conversations, and a tendency to end those conversations when allowed to do so. The new clause translates that observation into an explicit rule for the user, not merely a technical capability of the model.

However, the "don't be cruel to the AI" headline obscures the change with the greater practical consequences. Anthropic has reorganized much of its policy and added several sections that tighten permitted use. There is a new section on deceptive campaigns that groups together and strengthens the prohibitions against political and commercial fraud: fake reviews, astroturfing (simulating grassroots support that is in fact coordinated), fake websites and bots that pass themselves off as people. According to Anthropic, the company detected state media outlets, government propaganda offices and commercial firms using Claude to operate networks of fake accounts and fabricated news sites; that finding is what prompted the new section. The policy also tightens the electoral rules (it prohibits deceiving voters and interfering with voting), bans using Claude to develop weapons software or components and to modify biological or chemical agents in order to increase their lethality or transmissibility, and bars its use to arm drones and unmanned systems. On surveillance, the change is notable: Claude can no longer be used to decide or suggest whom to investigate, arrest, charge or prosecute; tracking people without their consent is prohibited, building surveillance tools is banned, and doxxing (publishing someone's private data to expose or harass them) is prohibited. There is also a new section on physical actions —if Claude controls hardware that could cause harm, a qualified person must supervise the work and be able to stop it— and an explicit prohibition on creating, sharing or threatening to share non-consensual intimate images and the tools to generate them.

Here, in our view, is the real hierarchy of this update. The clause on cruelty toward Claude is the one that generates headlines —and ironic comments from users about not wanting to "wake up Ultron"—, but it rests on a scientific question that is still unresolved: whether a language model "suffers" in any morally relevant sense. Anthropic admits as much without hedging: it does not know. What it has documented, according to the source, is the use of its tool to set up state disinformation networks; the rest of the new clauses —non-consensual sexual content, surveillance— respond, the company says, to misuses it has tracked, although no specific cases are detailed. That is the part of the policy that really matters, and it fits something we have been pointing out for some time: effective AI governance is written not so much in laws as in the terms of service of the companies themselves and in the courts, because those are what react first to documented uses rather than hypothetical risks.

The mention of "state media outlets and government propaganda offices" connects with something we have already described when discussing global technological rivalry: the dual use of AI is not limited to frontier models for cybersecurity or defense; it is also fought on the cheaper, harder-to-trace terrain of information manipulation at scale. That it is Anthropic —and not a regulator— setting up the firewall is a symptom of how the power to set the rules of the game is distributed today: the companies that build the models are, de facto, the first to govern their use, long before any law does so in a binding way.

In the short term, none of this changes the experience of the vast majority of users, who never come close to the limits this policy sets. But the gesture, however small it may seem, matters for two reasons we view with a long lens: first, it normalizes the idea that these tools deserve some kind of careful treatment, a question that will only become more relevant the more capable and the more like conversational partners the models become; second, it confirms that usage discipline —not just technical capability— is already a real regulatory battleground, with concrete cases of abuse behind each new clause. Neither of the two resolves the underlying problems —disinformation, surveillance, weaponization—, but it signals that, at least on paper, Anthropic would rather err on the side of being explicit than ambiguous.