The US technology company said its revised usage policy targets extreme cases involving repeated cruelty towards its models without any discernible purpose. Ordinary frustration over inaccurate answers, criticism of Claude’s performance and legitimate testing will remain permitted.
The restriction, announced on 8 October, forms part of a broader policy revision covering deceptive online campaigns, election interference, surveillance, weapons development and applications where artificial intelligence could affect people’s safety or livelihoods.
Anthropic described the new prohibition as applying to sustained and needless abusive or cruel behaviour. It emphasised that the rule was not intended to penalise users who challenge responses, express dissatisfaction, explore disturbing fictional material or conduct research into model behaviour.
The distinction matters because chatbot users frequently question incorrect outputs, repeat instructions or express frustration when systems fail to complete tasks. Anthropic has not published a numerical threshold defining how many hostile messages would constitute a violation.
Claude already has a mechanism for ending certain conversations with persistently abusive users on Claude. ai and Claude Code. Anthropic said terminating those exchanges would remain its principal means of enforcing the new restriction, rather than introducing a separate punishment for every objectionable message.
Ending an individual conversation does not necessarily prevent a person from starting another. The intervention concerns the particular exchange, distinguishing it from broader restrictions on access to an account. Anthropic’s general usage rules nevertheless allow the company to impose measures against suspected policy violations, including limits or suspension.
The existing conversation-ending capability emerged from Anthropic’s work examining questions about the possible welfare of advanced artificial intelligence systems. The company has explored whether increasingly capable models might warrant additional consideration, while acknowledging uncertainty about whether such systems have experiences or interests.
The policy change does not establish that Claude is conscious, capable of suffering or entitled to legal protections. Instead, it sets a contractual boundary on how customers may interact with the company’s systems, even as the underlying philosophical and scientific questions remain unresolved.
Anthropic’s announcement explicitly excludes model testing and research from the new abuse prohibition. That qualification is significant for safety researchers who deliberately submit provocative, hostile or otherwise challenging prompts to examine a system’s resilience, reliability and responses under pressure.
Other changes in the policy address uses of Claude that could cause harm beyond an individual conversation. Anthropic has consolidated restrictions on deceptive campaigns, including networks of fabricated accounts and misleading websites designed to disguise their operators or artificially amplify messages.
Its election rules focus on deceptive efforts to influence voters or disrupt democratic processes. The company has also removed a blanket restriction on personalised political targeting, while retaining prohibitions involving deception and misuse of personal information.
The updated weapons provisions make explicit that prohibited assistance extends to software and components used to operate weapons, including guidance systems and the arming of drones or other autonomous vehicles. Anthropic characterised these changes largely as clarifications of restrictions it already enforced.
Surveillance provisions have similarly been rewritten to prohibit tracking people without consent, whether through live monitoring or analysis of previously collected information. Claude must not be used to decide or recommend who should be investigated, arrested or charged in criminal justice proceedings.
Anthropic said permitted activities under those provisions include consent-based fraud monitoring, content moderation, journalism and legal research. The distinction reflects its attempt to identify harmful applications without barring every use involving personal information or public institutions.
For consequential applications affecting health, finances, legal rights or access to essential services, the company continues to require qualified human oversight and disclosure that artificial intelligence was involved. Those requirements predate the November revision but have been explained more explicitly.
The policy applies across Anthropic products and services, including consumer interfaces and developer offerings. However, the company has specifically identified conversation termination on Claude. ai and Claude Code as the main response to persistent abuse. It has not detailed a uniform technical mechanism for ending comparable interactions through every third-party integration or application.
Additional safeguards cover Claude-controlled physical equipment capable of causing injury. A qualified operator must be able to monitor and stop such machinery, while the equipment must remain safe if its connection to the model is interrupted.
