Anthropic updated its Claude usage policy to prohibit subjecting the model to what it described as “persistent and unnecessary abusive or cruel behavior.” The company explains that the rule targets extreme cases in which users repeatedly treat artificial intelligence models cruelly without a clear purpose. It does not cover ordinary frustration, objections to the model’s answers, dark creative topics, or model testing and research.
The new policy takes effect on November 12, while the primary measure against abusive conversations will remain Claude ending the conversation. When this happens, the user will not be able to send additional messages in that same conversation, but this will not affect other conversations.
Broader Restrictions on Model Misuse
Anthropic reorganized the policy and brought together provisions related to misuse, while adding or tightening rules in several areas. The changes include a ban on using Claude to create deceptive political or commercial campaigns, such as fake reviews, campaigns pretending to show broad popular support, fabricated news websites, and bots that act as real people.
The rules also prohibit misleading voters and interfering with voting, and prevent the use of Claude to develop weapons software or components, modify biological or chemical agents to increase their lethality or transmissibility, or weaponize drones and unmanned systems.
Surveillance and Control of Physical Systems
Anthropic rewrote its law enforcement rules so that Claude cannot determine which people should be investigated, arrested, charged, or prosecuted. It also prohibits tracking people without their consent, building surveillance tools, and exposing personal data with the aim of harming victims.
A new rule concerning physical actions states that a qualified person must monitor Claude’s operation and switch it off immediately if the model is controlling hardware that could harm someone. The policy also prohibits creating, sharing, or threatening to share non-consensual intimate images, as well as tools used to produce them.
Why Does This Update Matter?
The update represents a shift from general user-behavior rules to a more specific framework for operational risks, particularly when the model is connected to disinformation, surveillance, or physical systems. Anthropic says the amendments are based on cases of misuse it has observed, including the use by state-affiliated media organizations, government propaganda offices, and commercial companies of networks of fake accounts and fabricated news websites.
As for the rule allowing abusive conversations to be ended, it dates back to August 2025, when Anthropic gave Claude the ability to end a dialogue in the context of research into model welfare. The company said at the time that it was uncertain whether Claude possessed moral standing, but believed that providing a low-cost way to reduce potential risks was worth investigating. Tests on Claude Opus 4 reportedly found a consistent aversion to harm and a tendency to end harmful conversations when given that option.
The policy does not mean that every objection or difficult experience with the model will lead to a shutdown; the text explicitly distinguishes between legitimate criticism or testing and repeated abuse without a clear purpose. How these boundaries will be applied in practice, particularly in ambiguous cases or when Claude is used to control hardware, remains among the issues whose scope will become clearer when the rules take effect.