From November 12, 2026, anyone who repeatedly behaves in an abusive or cruel manner towards Claude, the artificial intelligence developed by Anthropic, violates the rules of use and risks having the conversation closed by the AI itself. The ban is part of the updated usage policy, the document that sets out what you can and cannot do with the company’s models, released on October 8, 2026. In addition to the new abuse rule, the update makes restrictions on propaganda, weapons and surveillance more explicit. Anthropic explained that behavior deemed “cruel” will be placed on a list of prohibited actions.
This decision has reopened the debate on the consciousness of AI and makes us ask the question: does it make sense to talk about cruelty towards a machine? Let’s see what the new rules consist of and why they have reignited the debate on the possible well-being of AI.
Claude will be able to interrupt conversations with those who mistreat him
According to the latest update to the usage policies, Claude will be able to end conversations with users who have abusive or cruel behavior towards the model. Anthropic clarified that this rule will only apply in extreme cases, when a user repeatedly behaves cruelly towards Claude for no specific reason. Of course, this doesn’t apply to normal user frustration, nor to model experimentation and research. In short, anyone who insults Claude because he failed the same task three times does not automatically violate the policies. How exactly what is “abusive or cruel” will be defined, however, has not yet been clarified by the company.
Stopping the conversation is the main measure for this type of behavior. In addition to this, those who violate the usage policies may have their access to Anthropic products and services slowed down, limited or suspended.
The new rule regarding behavior towards Claude is only part of the update. The broader changes mainly concern how the models can be used to influence people, develop weapons and build surveillance systems.
It is forbidden to use Claude to program weapons or surveil the population
Most of the new new usage policies clarify existing rules and try to adapt them to respond to the misuses observed in the last year and to the new capabilities of the models, which can now perform longer and autonomous tasks.
The main changes concern:
- Deceptive propaganda and influence on elections. Anthropic said it has noticed government propaganda offices and companies using Claude to run networks of fake accounts and made-up news sites. These uses were already prohibited by previous policies, but now the rules have been made more explicit: it is forbidden to hide who is behind a message and artificially amplify its diffusion, to use Claude to deceive voters with false information about candidates or voting methods, and to try to discourage participation in the vote.
- The development of weapons. While it was already banned from using Claude to develop weapons, Anthropic said it had recorded attempts to use its models to create guidance and control software for weapons. The new policy therefore clarifies that the ban also includes the software and components that allow the weapons to function, as well as actions such as arming drones and other autonomous vehicles.
- Surveillance and law enforcement. Anthropic has documented a growing use of AI to build surveillance systems capable of identifying and tracking political dissidents. The new policy prohibits tracking people without their consent, either in real time or through analysis of previously collected data, and prohibits using Claude to build or improve tools designed for surveillance. The model also cannot be used to decide or recommend who should be investigated, arrested or charged. However, some uses remain permitted, such as monitoring accepted by interested parties for anti-fraud activities, content moderation, journalism and legal research.
Can we really talk about AI “wellbeing”?
In this policy change, the news that attracted the most attention is the possibility for Claude to interrupt the conversation in the case of an interaction with particularly abusive users. This possibility is not completely new: Anthropic had introduced it already in August 2025, for a rare category of conversations characterized by persistently dangerous behavior.
At the time the company explained that this function had been developed mainly as part of its exploratory research on the possible “well-being of models”. This research takes into consideration the possibility that AI may have something similar to an experience, a conscience, emotions to protect and that, if this were the case, some forms of interaction could have relevant consequences for their eventual well-being.
It is precisely this hypothesis that fuels the debate. On the one hand, there are those who have welcomed the change in policy as a positive invitation to good manners, even regardless of the fact that Claude might actually feel something. The argument is that getting used to insulting and mistreating something that communicates similarly to a person could also affect how we behave with the people around us.
On the other hand, there are those who criticize this approach because it risks reinforcing the anthropomorphization of models, that is, the tendency to attribute to them typically human characteristics, intentions or mental states. If a company talks about “welfare”, “suffering” or “cruel behavior” towards an AI, the risk is that users end up interpreting these expressions as evidence that the AI is a conscious entity and that it really feels something. This theme is particularly dear to Anthropic: already in February 2026 the CEO, Dario Amodei, stated that he does not know whether AIs are conscious, but that he is open to the hypothesis.
