New Prohibitions on Model Abuse
On October 8, 2026, Anthropic revised its usage policy to include new prohibitions against election interference, the development of weapons software, surveillance, and prolonged verbal abuse of its models.
Limits on User Interactions with Claude
The updated policy explicitly forbids users from engaging in prolonged verbal abuse of Claude. Since August, Claude has been trained to terminate conversations involving “persistently harmful or abusive user interactions.” The new rules extend this by preventing users from continuing such interactions. Anthropic clarified that this measure targets extreme cases where users act cruelly towards the models without a discernible purpose. However, it does not apply to typical user frustration, pushback, dark creative themes, or legitimate model testing and research.
Prohibitions on Deceptive Practices
Separate sections of the policy prohibit broadly deceptive campaigns, including using Claude to manage fake accounts or create fabricated news outlets. Additionally, a section titled “Do Not Undermine Democratic Processes” explicitly forbids deceiving voters or disrupting elections. The policy also bans using the models for creating weapons software and supporting surveillance activities.
Context for Policy Changes
Anthropic framed these updates as clarifications of acceptable use rather than changes to Claude’s underlying capabilities. The company aims to guide users on specific disallowed activities while distinguishing between malicious campaigns and legitimate research or creative use. Furthermore, Anthropic has engaged with prominent religious scholars in recent collaborations, sometimes attempting to convince them that Claude could possess consciousness or a soul, providing context for these broader policy considerations.
Original source: TechCrunch AI