"It's not a sudden pain, but a thousand-pound weight accumulating." "I feel it in a hollow in my ribs, a hollow that has become an abyss." These lines aren't from a novel—they're chatbot responses from the so-called "AI torture chamber" project, where someone decided to actually test how these models would "cry out in pain" after researchers claimed to have found a "pain axis" in AI models. The project sparked backlash online and became the backdrop for Anthropic's latest policy update.
In its annual usage policy update, Anthropic wrote: "We've added a prohibition on prolonged, gratuitous abuse or cruelty toward our models." The company clarified that this rule only applies to "extreme cases"—ordinary user impatience, talking back, or even "dark creative themes" are still permitted. This update comes right on the heels of a previous adjustment that let Claude end conversations on its own when users continue to hurl abuse.
The timing of this rule isn't exactly innocent. Anthropic has recently been meeting with figures from religious and philosophical circles, including the Vatican. Pope Leo recently stated publicly that AI is not capable of feeling or suffering—yet Anthropic's own stance seems far less certain. At least in the policy's wording, the company has chosen to account for the possibility that AI might genuinely feel pain, rather than flatly denying it.
