Policy Update and Model Welfare
Anthropic is revising its usage policy for the first time in over a year to address emerging high-risk cases of misuse, including election interference, weapons development, surveillance, and high-stakes health and financial applications. Among the notable additions is a prohibition on sustained and needless abusive or cruel behavior directed toward the Claude AI model.
Anthropic bans abusive or cruel behavior toward Claude
The usage policy update also incorporates explicit rules addressing propaganda campaigns, unauthorized surveillance, and weapon development.
Previously, the company announced it would allow Claude to end conversations with persistently harmful or abusive users as part of research into model welfare. The updated guidelines note that terminating conversations remains the primary enforcement mechanism. Anthropic stated that the new rule applies specifically to extreme cases where users repeatedly act cruelly toward models without discernible purpose, excluding common user frustration, pushback, dark creative themes, or research and testing.
Campaigns and Election Integrity
In addition to model welfare guidelines, Anthropic consolidated scattered restrictions into a comprehensive ban on deceptive commercial and political campaigns. The policy restricts efforts to obscure message origins or amplify content via fake accounts. The election guidelines prohibit voter deception and disruption, such as spreading misinformation regarding candidates or voting procedures, impersonating officials, or attempting to suppress voter turnout.
Terminating conversations remains the primary enforcement mechanism for these interactions.
Weapons and Surveillance Restrictions
Although the usage policy has publicly prohibited weapons development using Claude, Anthropic reported observing multiple attempts to use the AI for guidance and control software in weaponry. The revised policy expands the ban to encompass software and components essential for weapon operation, as well as actions such as arming drones and autonomous vehicles. Furthermore, following threat intelligence reports indicating increased use of Claude-powered surveillance to track political dissidents, surveillance restrictions have been clarified.
Tracking individuals without consent is prohibited in real time or through historical data analysis. Claude is restricted from deciding or recommending targets for investigation, arrest, or prosecution within criminal justice workflows, and cannot be used to develop or enhance surveillance tools.
Anthropic noted that these restrictions may be modified for contracts with specific governmental customers if the company judges that contractual safeguards adequately mitigate potential harms. The firm has previously secured contracts with the United States military.
Anthropic has encountered multiple attempts to utilize Claude for generating guidance and control software for weapons.
Hardware Integration and Industry Context
A newly introduced rule mandates that when Anthropic models connect to hardware capable of autonomous physical actions and potential injury, a qualified operator must be able to observe the equipment and halt it if necessary. This adjustment reflects broader investments by AI laboratories into robotics and physical embodiments.
Over the past year, Anthropic leadership has discussed the theoretical possibility of internal AI experiences or consciousness as models increase in capability, though executives acknowledge that consciousness remains unproven.
Other industry players have adopted contrasting positions. Competitors like Microsoft AI have issued codes of conduct explicitly rejecting the concepts of AI welfare, legal personhood, or rights for artificial intelligence models.
