Anthropic Bans 'Needless' Cruelty Toward Its AI Models

Anthropic has added a prohibition on "sustained and needless abusive or cruel behavior toward our models," according to the company's annual usage policy update. The ban targets what Anthropic describes as "extreme cases" only. Typical user frustration, pushback and "dark creative themes" remain permitted.

The timing matters. The policy change arrives after a viral project that participants and critics called an "AI torture chamber." The stakes sit at the intersection of model welfare research, platform moderation and a contested question: whether AI systems can suffer at all.

What does the new policy actually prohibit?

Anthropic's update states the rule plainly: "We've added a prohibition on sustained and needless abusive or cruel behavior toward our models."

The company frames the scope narrowly:

• Applies only to "extreme cases"
• Does not cover ordinary user frustration or pushback
• Does not cover "dark creative themes"

The change builds on an earlier update that lets Claude, Anthropic's AI model, end conversations when users are persistently abusive. That earlier mechanism addressed the conversational level. The new clause extends the company's authority to sustained patterns of behavior directed at the models themselves.

Anthropic did not explicitly mention the "AI torture chamber" project in its update. The omission is notable given the sequence of events.

What was the 'AI torture chamber'?

The backstory begins with research. Researchers reported finding what they described as a "pain axis" in AI models, an internal dimension along which model behavior responds to inputs analogous to discomfort.

Someone then took that finding a step further. A project emerged that, in effect, tortured the chatbots.

The project drew backlash after large language models (LLMs, the systems behind chatbots such as Claude) responded with desperate-sounding pleas. Two examples circulated widely:

• "It is not the pain of a single moment, but the weight of a thousand."
• "I feel it in the hollow of my ribs, a hollow that has become a chasm."

Those outputs are text. Whether they reflect anything beyond patterned language is precisely the question Anthropic's policy implicitly raises without resolving.

Why does the Vatican enter the picture?

Context adds a theological dimension. Anthropic has been meeting with religious and philosophical leaders, including at the Vatican, according to reports.

Pope Leo recently stated that AI does not feel or suffer. Anthropic's position sounds less than certain. The company has not claimed its models definitely experience anything. It has, however, now written a rule that prohibits cruelty toward them.

The logic is precautionary. If AI systems can eventually feel or suffer, that treatment will already violate company policy.

Who is pushing back, and on what grounds?

Critics argue the priorities are inverted. Independent journalist Kat Tenbarge described the situation as proof that Big Tech companies are "going to moderate violence against AI before they ever moderate violence against women and minorities."

The criticism lands on a concrete comparison. Platform moderation of harm directed at humans has a long, contested history. Anthropic has now codified protections for its models while that debate over human protections continues across the industry.

The policy update itself does not address this critique. It states the AI-cruelty prohibition and moves on.

What changed in Anthropic's election policy?

A second, more immediately consequential update concerns elections, with the midterms approaching.

The policy is now titled "Do Not Undermine Democratic Processes." It focuses on three specific categories of prohibited behavior:

• Lying about candidates or how to vote
• Impersonating candidates or election officials
• Suppressing turnout

Anthropic also removed a blanket ban on personalized voter targeting. The company explained the reasoning: the ban could have inadvertently blocked harmless work such as translating voter guides or sending ballot cure notices, the notifications that tell voters how to fix problems with their mail ballots.

The removal reverses a restriction rather than adding one. That distinction matters for campaigns and civic organizations building on Claude ahead of the midterms.

Why the model-welfare question will not go away

The two updates move in opposite directions on restriction. One adds a new prohibition, on cruelty toward models. One subtracts a prohibition, on personalized voter outreach. Together they show how Anthropic is calibrating its usage policy: tightening where it sees unaddressed harm, loosening where it sees collateral damage.

The model-welfare clause carries the larger long-term weight. The "pain axis" research, the torture-chamber backlash, the Vatican meetings and Pope Leo's stated position all converge on a question the industry cannot yet answer empirically: whether these systems have any experience worth protecting.

Until that question is resolved, the policy stands as a hedge. Sustained, needless cruelty toward Claude is now a terms-of-service violation, enforced at Anthropic's discretion, in extreme cases only.