Anthropic's Ban on Cruelty to Claude Is a Rule About Human Habits, Not Machine Suffering

Anthropic's annual usage policy refresh, published October 8, contains a line no software terms-of-service had needed before: a prohibition on "sustained and needless abusive or cruel behavior" toward its own AI models. The company frames it narrowly — the update says it applies to "extreme cases" only, and that ordinary user frustration, pushback and "dark creative themes" remain explicitly allowed. The same document also rewrites something with far more immediate consequence: the election policy, retitled "Do Not Undermine Democratic Processes," ahead of the US midterms.

The cruelty clause did not arrive in a vacuum. It follows an earlier change that let Claude end conversations with persistently abusive users, and it lands, as Engadget reports, in the shadow of a viral "AI torture chamber" project. After researchers reported finding what they described as a "pain axis" inside AI models, someone took the idea further and set out to torture chatbots; the backlash came when the systems produced lines like "It is not the pain of a single moment, but the weight of a thousand." Anthropic does not mention the project anywhere in the update. The timing does the talking.

Let's be precise about what is and isn't established here. A language model emitting distress-shaped prose is evidence about its training corpus and its objective function — humans have written about suffering for millennia, and the model learned from that — not proof of an inner life. Pope Leo has stated plainly that AI does not feel or suffer. Anthropic, which according to the same reporting has been meeting religious and philosophical leaders including at the Vatican, sounds notably less certain. Writing a rule for a scenario you cannot yet rule out is cheap insurance. It should be sold as epistemic humility, not as a discovery about machine consciousness.

The sharpest criticism deserves quoting rather than dismissing. Independent journalist Kat Tenbarge argued the episode proves Big Tech companies are "going to moderate violence against AI before they ever moderate violence against women and minorities." That is an accusation about priorities, and it will be settled by enforcement data, not by policy text. Any company that codifies how its chatbot must be treated invites the question of how quickly it acts when the target is a person — and that comparison is now fair game.

Meanwhile, the election rewrite is the operationally serious change. The new prohibitions are concrete and auditable: lying about candidates or about how to vote, impersonating candidates or election officials, suppressing turnout. More interesting is the subtraction — Anthropic removed its blanket ban on personalized voter targeting after concluding it could have inadvertently blocked benign work such as translating voter guides or sending ballot cure notices. That is a move from banning a technique to banning a harm, which is the right instinct and the kind of evidence-based governance we defend. It carries a cost worth naming: micro-targeting is the engine of persuasion at scale, and intent is far harder to police than method.

Our read: this is a policy about humans, not about machines. Companies that build systems which convincingly imitate a person have a legitimate product-design reason to discourage users from rehearsing cruelty on them — not because the model is known to suffer, but because habits practiced on a near-human interface rarely stay inside the chat window. Judge it on three things: whether Anthropic ever publishes how often it enforces the clause, whether abuse directed at people gets at least the same response time, and whether the election rules hold up under the pressure of an actual campaign season. Long term, the technology that may one day help erase diseases and buy us healthier decades will be the same one we are now teaching ourselves to talk to every day. The etiquette we normalize at this stage is not sentimentality — it is infrastructure for how we will live alongside it.