Anthropic is drawing a line against cruelty toward Claude even as researchers continue to debate whether AI models can experience suffering.
The company’s updated Usage Policy prohibits “sustained and needless abusive or cruel behavior” toward its models, adding a rule governing how users interact with the models themselves. Anthropic says the restriction targets extreme cases, not ordinary frustration, criticism, model testing or research.
The timing puts the policy in the middle of a wider argument over AI welfare that debates whether AI models deserve moral consideration. Anthropic has not established that Claude can suffer, leaving the new rule as a statement about acceptable user behavior rather than proof of AI consciousness.
Unpacking Anthropic’s new policy
Announced on Oct. 8, the policy will take effect from Nov. 12, and targets extreme cases in which users repeatedly act cruelly toward the models without a discernible purpose.
Anthropic maintains that the new policy does not cover ordinary frustration, criticism, pushback, dark creative themes, or legitimate model testing and research. Anthropic says Claude can already end persistently abusive conversations on Claude.ai and Claude Code, and that conversation termination will remain the primary enforcement mechanism under the updated rules.
The abuse restriction is only one part of the update. Anthropic has also clarified rules covering deceptive campaigns, election interference, weapons development, non-consensual surveillance, high-risk decisions affecting people’s lives, and AI systems connected to physical equipment that can act autonomously.
The company says many of these changes clarify existing restrictions in response to new misuse patterns and the expanding capabilities of its models.
The debate on model consciousness continues
Anthropic announced a model-welfare research program in April 2025, exploring whether advanced AI systems could have experiences that deserve moral consideration. The company acknowledges that there is no scientific consensus on whether AI systems can be conscious.
A recent developer’s experiment, drawing on an academic preprint, manipulated language models’ internal activity to produce distress-like responses. The experiment raised ethical questions, but those responses do not establish that the models consciously experienced pain
Anthropic’s decision to prohibit sustained and needless cruelty toward its models suggests the company is taking the possibility of AI suffering seriously, even as the science remains unsettled.
While the policy does not prove that Claude can feel pain, it raises the question of whether Anthropic is adopting precautionary protections before researchers can establish whether AI models are capable of subjective experiences.
Criticisms have followed the decision
Mustafa Suleyman, who has previously criticized Anthropic’s leaning on AI consciousness, reacted with a melting emoji to the new policy change announcement, the BBC reported.
Barry Scannell, a technology partner at Irish law firm William Fry, criticized the policy in a LinkedIn post, arguing that users cannot “be cruel to numbers and maths.”
According to The Guardian, OpenAI CEO Sam Altman separately warned against assigning religious authority to AI or surrendering human judgment to it, following reports of Anthropic’s discussions with religious scholars.
What the policy means for workplace Claude users
For employees using Claude, the restriction does not require accepting its answers without challenge. Anthropic says ordinary frustration, pushback, model testing, and research remain permitted. Teams can continue checking outputs, identifying errors, and evaluating performance.
The distinction also matters for developers and researchers testing difficult interactions. Anthropic describes the prohibited behavior as extreme, repeated cruelty without a discernible purpose. Its announcement does not provide a detailed set of examples defining that boundary, leaving users with a broad standard rather than a precise checklist.
Enforcement should also be understood carefully. Ending persistently abusive conversations remains the primary mechanism. Anthropic’s original explanation of that feature says an ended conversation does not prevent a user from starting another chat; ending a conversation is therefore different from banning an account.
For IT teams, a practical next step is to review the full policy before Nov. 12, particularly where Claude supports sensitive decisions or controls physical equipment. The update also addresses human oversight and operational safeguards, making its relevance broader than the debate over model consciousness. Those deployment requirements deserve attention alongside the more unusual abuse restriction.
Read more: Questions about AI safeguards extend beyond chatbot interactions, as an FTC investigation examines potential consumer harms involving OpenAI, Anthropic, and increasingly autonomous AI systems.