Anthropic Wants You to Stop Bullying Claude: Being 'Cruel or Abusive' Could Have Consequences

Anthropic has added a rule to its usage policy prohibiting users from subjecting its AI models to "sustained and needless abusive or cruel behavior." Announced on Oct. 8, the rule takes effect on Nov. 12 and targets extreme cases of repeated cruelty without a clear purpose. Ordinary frustration, criticism, dark creative writing and model testing are explicitly excluded. The company is not claiming that Claude feels pain when insulted; rather, the policy reflects Anthropic's ongoing research into whether advanced AI models could have some form of morally relevant experience, a question it says remains uncertain.

The company already gave Claude Opus 4 and 4.1 the ability to end conversations in its consumer chat interfaces in August 2025, intended for "rare, extreme cases of persistently harmful or abusive user interactions." A user whose conversation is ended can still start another chat or branch the previous conversation. Anthropic has not announced an automatic account ban for ordinary rudeness and provides no detailed list of prohibited behaviors. Microsoft AI chief Mustafa Suleyman disagrees with Anthropic's approach, arguing in an Axios essay that training AI to present itself as conscious could make such systems harder to control. Anthropic frames its stance as a low-cost precaution around an unresolved possibility.

The rest of the policy update clarifies weapons restrictions, requirements for AI in high-risk settings such as health, finances and legal rights, and the need for a qualified operator when models control hardware that could cause injury. These provisions affect organizations deploying Claude in business, political campaigns, surveillance or physical equipment. After Nov. 12, the key question will be how Anthropic identifies sustained, purposeless cruelty and how often conversations are ended under the rule.

(Source：International Business Times, Singapore Edition)