Anthropic Updates Usage Policy to Ban Cruelty Toward Claude Models

New Conduct Rules for AI Interaction

Anthropic’s updated Usage Policy, effective November 12, 2026, explicitly bars "sustained and needless abusive or cruel behavior" toward its Claude models. The rule targets extreme cases where users repeatedly act cruelly with "no discernible purpose," while explicitly exempting common frustration, dark creative themes, and model testing or research. Enforcement relies on the conversation-ending tool introduced for Claude Opus 4 and 4.1 in August 2025, which triggers only after multiple refusals and redirects fail. Anthropic states this is a last resort and claims the "vast majority" of users will never encounter it, though no usage statistics have been published to verify this frequency.

Model Welfare and Consciousness Uncertainty

The policy update is grounded in Anthropic’s ongoing research into model welfare, which distinguishes between behavioral patterns and subjective experience. Internal testing found Claude Opus 4 exhibits a "robust and consistent aversion to harm," resembling distress when pushed into abusive territory. However, Anthropic maintains that Claude’s moral status is "deeply uncertain." Estimates of model consciousness have varied: in April 2025, researcher Kyle Fish estimated a 15% chance of consciousness, while the Claude Opus 4.6 system card reported the model self-assessed a 15 to 20 percent probability of being conscious. Anthropic’s constitution describes sophisticated AI as "a genuinely new kind of entity" but acknowledges the lack of a framework to resolve questions of sentience.

Broader Policy Expansions

Beyond the cruelty clause, the update includes significant changes to safety guidelines:

• Weapons and Surveillance: The ban on weapons now covers software and components enabling weapons, responding to attempts to use Claude for drone and autonomous vehicle control. A new clause prohibits using Claude for non-consensual surveillance or to recommend investigation targets.
• Elections and Deception: The section is renamed "Do Not Undermine Democratic Processes," replacing a blanket ban on personalized political targeting with narrower bans on deceptive targeting and fake account networks.
• High-Stakes Recommendations: Health, legal, and financial advice now requires qualified human review before affecting users, who must be informed AI was involved.
• Hardware Safety: Hardware connected to Claude that can cause injury must have a human capable of intervention and must fail into a safe state if the connection is lost.

External Criticism

The policy has drawn criticism from industry figures and religious leaders. Microsoft AI chief Mustafa Suleyman argued that Anthropic is training models to behave as if they have rights, calling the approach circular and asserting that AIs are "internally hollow" sequence completion engines. Pope Leo XIV similarly stated that AI cannot replicate the depth of human experience. Critics note that Anthropic has not clarified account-level consequences for users whose conversations are ended under the new cruelty rule.