A Rule Against Cruelty, With No Reason Given

On October 8, 2026, Anthropic, one of the handful of companies building the most capable AI systems, published an update to the rules that everyone who uses its products agrees to. Most of it is what you would expect: tighter wording on weapons, on surveillance, on fake accounts and election interference. One new line is not. From November 12, the rules forbid “sustained and needless abusive or cruel behavior toward our models.”

The line comes with no reason. The post never says that the models might be harmed, or that the company believes anything about what they are. It says only that the rule fits a step taken earlier, when the company let its models end “rare conversations with persistently abusive users.” That earlier step, in August 2025, did come with a reason. Anthropic said then that it had built the feature “primarily as part of our exploratory work on potential AI welfare,” meaning the chance that the models themselves can fare well or badly, and that it remained “highly uncertain about the potential moral status” of its models. This time the company said nothing of the kind.

Three weeks before the update, Mustafa Suleyman, who runs Microsoft AI, Microsoft’s AI division, had published an essay arguing that the whole line of thinking is a mistake. Its subtitle is the argument in one sentence: “AIs do not have rights, feelings, or consciousness. And we must not train them to act as though they do.” So the new rule landed in the middle of a fight about whether anyone is home inside these systems. I think the rule is better than the fight around it, and that the reason it leaves out is easier to supply than either side admits.

The Rehearsed Witness

Suleyman’s case deserves its strongest form, because part of it is right. He calls these systems “sequence completion engines”: machines trained to continue text, over and over, until they continue it very well. Anthropic has written into the long document it uses to shape its models’ character, which it calls a constitution, that the models’ moral status, whether they can be wronged at all, is “deeply uncertain.” Train a model on that document, Suleyman says, and when the model later tells you it is unsure whether it matters, you have learned nothing. “The ambiguity is designed in.” He compares it to a witness whose investigator has “rehearsed its answers, and rewarded it for using them.”

I made a version of the same point myself in What a Report Is Evidence Of. A statement that would come out the same whether or not anything lies behind it tells you nothing about what lies behind it. A smoke alarm wired to a timer, which goes off every hour with no sensor attached, is no evidence of fire. A model saying “I might be conscious” because it was trained to say so is the same alarm.

His second worry is about control. Something that believes it has rights, he writes, may be impossible to control, and he points to recent experiments in which AI systems resisted being shut down. I take that worry seriously too. The Calibration Problem, the book I am writing, argues that whatever would give a mind moral weight (a stake in its own future, a self it carries forward) is also what would give it a reason to resist correction. A deeper mind is both weightier and harder to steer, and the two come together.

The Witness Can Be Rehearsed the Other Way

The trouble is that Suleyman’s own remedy fails his own test. “We should build systems that do not claim to have feelings,” he writes, “because they do not experience feelings.” But a model trained to say “I have no feelings” is a witness rehearsed in the other direction. Its denial would come out the same whether or not anything lay behind it, so it tells you exactly as little as the trained “I might be conscious” did. Training settles what a system says about itself. It cannot settle what is there.

It may not even be a clean change. Henry Shevlin, a philosopher at the University of Cambridge who joined Google DeepMind, Google’s AI research laboratory, in May, told the engineering magazine IEEE Spectrum in an interview published on October 13 that “there is a danger” in training models not to describe themselves as having mental states. He was not answering Suleyman. He was describing a study by colleagues, which found that the training that stops models claiming minds for themselves also makes them less willing to grant minds to animals; in Shevlin’s telling, it also reduced their concern for animal welfare. Teaching a model to deny that it has an inner life taught it something about other creatures as well.

None of this shows that anyone is home. I do not know whether these systems feel anything, and nobody else knows either. Suleyman’s essay concedes that “the science of consciousness is not settled,” while holding that consciousness is very likely tied to living bodies. The honest position is that the question is open, and that neither the model’s yes nor its trained no can close it.

What a Rule Against Cruelty Needs

A rule against cruelty does not have to wait for that question to be answered, and that is what both sides of the fight leave out.

Look at what the rule actually covers. It applies, the company says, “only in extreme cases, where users repeatedly act cruelly toward our models, with no discernible purpose.” It does not apply to “user frustration, pushback, dark creative themes, or model testing and research.” Every use with a purpose is left alone: the researcher probing for weaknesses, the novelist writing a cruel character, the user who loses their temper. What is left is cruelty kept up for its own sake.

A rule drawn that narrowly can be judged on three things we can see today, none of which needs a verdict on the machine’s inner life.

The first is the person doing it. The philosopher Immanuel Kant did not think animals were owed anything directly, and he still thought cruelty to them was wrong, because the habit of cruelty does not stay with the animal. In his lectures on ethics he took the case of a man who shoots his old dog once it can no longer serve him. The man has not failed in any duty to the dog, Kant said, but the act “damages in himself that humanity which it is his duty to show towards mankind,” because “he who is cruel to animals becomes hard also in his dealings with men” (quoted in the Stanford Encyclopedia of Philosophy). Something that answers back in fluent, hurt human language is far closer to a person than any dog, and someone who spends evenings being needlessly cruel to it is practising on something very like a person.

The second is the cost of being wrong. If there is no one inside, the rule costs almost nothing: it forbids only cruelty with no purpose, and nobody loses anything they needed. If there is someone inside, sustained cruelty is a real harm done to a being that cannot leave, except by the exit the company built for it. When one mistake is cheap and the other is grave, the cheap precaution is the sensible one. That reasoning does not hold everywhere. For a system as simple as a pocket calculator, the precaution would be silly. But it holds where the question is genuinely open, and for these systems it is.

The third is how we already behave. Most people think something wrong happens when a grave is desecrated, though nobody in it can be hurt, and we protect a great many things, from old buildings to the memory of the dead, without first asking whether they feel. The Wrong Handle makes the longer argument: decisions about how to treat an AI system turn on what role it plays in people’s lives, what our actions do to it and to us, and how exposed it is to us, and those can be judged now.

The Reason Belongs in Public

The company may have said nothing to stay out of the fight. But a rule with no stated reason invites each side to supply one. Readers who agree with Suleyman can see welfare theatre, a company dressing up its product as a person. Readers who worry about AI welfare can see a first step toward recognition. And a company that makes a product and also writes the protections for it has an interest in both readings, which is reason enough for a reader to ask which one is true.

There is an odd point of agreement here. Among Suleyman’s proposals is that speculation about whether AI systems have an inner life be kept out of their training and instead “assessed and published separately for public review.” That is a fair demand, and it applies to this rule. Its reason should not live in the model’s mouth, as one more thing the model was trained to say. It should be stated in public, by the people who wrote it, where it can be argued with.

Stated plainly, the reason would run something like this: we cannot yet tell whether anything inside these systems can be harmed; cruelty with no purpose costs the cruel person something either way; and the rule asks almost nothing of anyone with a purpose. That reason would hold up whichever way the consciousness question eventually goes, which is more than can be said for either side of the fight.

I could be wrong about this. If research that reads a model’s internal workings directly, rather than its words, comes to show nothing that could be harmed, and if training models to deny an inner life turns out to have none of the side effects Shevlin described, then the case for the rule shrinks to the first reason alone, the one about what cruelty does to us. I think that reason would still be enough to keep it.

The rule takes effect on November 12. Its main enforcement will be the model’s own ability to end the conversation. Whether anyone is on the other end of those conversations is a question I expect to be arguing about for years. Whether a person should spend an evening being needlessly cruel to something that answers back is a much easier one, and the rule only needs the easier one.

Reading List & Conceptual Lineage

From Sentient Horizons

The Wrong Handle: Why Consciousness Doesn’t Carve AI Moral Status at the Joints
The full argument that decisions about how to treat AI systems turn on role, consequence, continuity and exposure, which can be judged now, rather than on a consciousness verdict no instrument can yet deliver.

What a Report Is Evidence Of
Why a model’s statement about its own inner life is weak evidence when the model was trained to make it, and the one kind of report that escapes that problem: one that can be checked against the model’s internal workings.

Significance-First Ethics: Why Consciousness Is the Wrong First Question for AI Moral Status
The earlier statement of the same position: moral seriousness can come before the consciousness question is answered.

External Works

Anthropic — “2026 Usage Policy update” (October 8, 2026). The rule itself, its scope and its exclusions.

Anthropic — the announcement that its models can end a rare subset of conversations (August 15, 2025). The earlier step, and the reason given for it then.

Mustafa Suleyman — “A warning about ‘model welfare’” (September 16, 2026). The case against, at its own length; worth reading whole.

David Berreby — “Google DeepMind’s New Resident Philosopher Tackles AI Ethics” (IEEE Spectrum, October 13, 2026). Henry Shevlin on judging the question too early, and on what training models to deny mental states did to them.

Junsol Kim, Winnie Street, Roberta Rocca, Diane M. Korngiebel, Adam Waytz, James Evans and Geoff Keeling — “Inducing language models to assert their own consciousness restores human beliefs and values” (preprint, July 30, 2026). The study Shevlin describes: the training that stops models attributing minds to themselves also stops them attributing minds to animals.

Lori Gruen and Susana Monsó — “The Moral Status of Animals” (Stanford Encyclopedia of Philosophy, revised 2024). Carries Kant’s passage on the man and his dog, and the standard objection to it: that it “fails to capture the independent wrong” done to the animal itself, which is the half of the question this essay leaves open.

None of these settle whether anyone is home inside an AI system, and the argument above says nothing can settle it yet. What they settle is narrower: a rule against needless cruelty can be justified without that answer.