Anthropic Updated Its Policy and Prohibited Sustained Cruel Treatment of Models
XORit Editorial · Sources, AI assistance and corrections

Anthropic has updated the usage policy for its models. The company published the new version on October 8, 2026, and it takes effect on November 12. The document is revised every year in response to the changing capabilities of the models and customer feedback, and most of the edits merely clarify rules that are already in effect.
One of the edits stands out noticeably. The policy now includes a separate prohibition on sustained and needless abusive and cruel treatment of models.
This concerns extreme cases where a person is rude to the model again and again with no goal and no benefit. Ordinary irritation, a substantive argument, dark creative plots, testing and research do not fall under the prohibition. Claude already can end rare conversations with persistently rude users on Claude.ai and in Claude Code, and the company calls this the main enforcement mechanism for the new rule.
It is amusing that, formally, an inanimate model now has a rule protecting it from rudeness.
The remaining changes mostly clarify norms that are already in effect. Rules about misleading activity, from political to commercial, were gathered into a separate section, and the section on elections was narrowed to prohibiting deceiving voters and obstructing voting; the general ban on personalized campaigning was removed, since it affected legitimate civic work. Weapons prohibitions were made explicit for software and components that make weapons work, as well as for actions such as arming drones.
The section on surveillance and criminal justice was rewritten more precisely: tracking people without consent is prohibited, the model cannot decide whom to investigate, detain or charge, and consensual monitoring, content moderation, journalism and legal research remain permitted. For cases affecting health, rights, finances or access to services, the requirements remain for a human who can review and, if necessary, change the model's recommendations, and for notifying the affected person.
Image: Domdomegg CC BY 4.0 · License
Translation editor: Салтанов А.
Source: www.anthropic.com ↗