Story
October 9, 2026

Anthropic’s Claude Rule Turns a Safety Update Into a Fight Over AI Personhood

Anthropic says its new limits on sustained cruelty toward Claude are a narrow safeguard alongside tougher rules on election deception, surveillance and weapons. Critics see a more consequential shift: software being treated like a moral patient.

The argument began before the formal rule. In August, Anthropic gave Claude the ability to end rare conversations with persistently harmful users, framing the move as research into model welfare; the company later said that refusal would remain its main enforcement tool.

By September, Anthropic said it was seeing AI used to help construct surveillance systems aimed at identifying and tracking political dissidents. That experience, alongside observed misuse in influence operations and weapons work, fed into a wider rewrite of its rules.

The update, announced in October and effective November 12, bars “sustained and needless abusive or cruel behavior” toward Anthropic’s models. The company insists the provision is for “extreme cases” of repeated cruelty with no clear purpose—not ordinary frustration, dark creative work, testing or research. A widely shared post marking the effective date helped turn that narrow-sounding clause into the headline issue.

Anthropic’s broader changes are less philosophically charged but more concrete: they consolidate bans on deceptive political and commercial campaigns, prohibit voter deception and election disruption, tighten restrictions on weapons guidance and control software, and more explicitly bar certain surveillance and criminal-justice uses. The company also retained human-review requirements for consequential health, financial and legal decisions.

Supporters of the Claude provision argue that consciousness is not required for caution. Box chief executive Aaron Levie, who says he does not believe AI is conscious, called politeness an “easy Pascal’s wager”: future systems need not be trained on endless human hostility.

Critics read the same policy as an attempt to humanize code. Michael Shellenberger called it “anthropomorphizing machines”; Scott Stevenson warned that granting models human-like empathy could expand their power to manipulate users. Others questioned enforcement, customer-prompt training and whether a company should dictate not only what people generate, but how they may address its bot.

Story coverage