You know the checklist. Every true-crime documentary runs it over slow piano: the future killer wet the bed, set fires, and hurt animals. The bedwetting gets an uncomfortable laugh. The animals are the one everybody remembers. The neighbor’s cat. The dog in the backyard, the detail the narrator lowers his voice for, as if the dog were the evidence and not the man who hurt it.
The psychiatrist who wrote that checklist went back and tested it. It failed. The FBI’s own consensus report on serial murder doesn’t mention animals at all. As a prophecy about killers, the red flag is dead.
As a window, it never stopped working. That distinction is the whole argument, and this week it got a new target. Anthropic changed its usage policy to bar “sustained and needless” cruelty toward its models, and the internet spent the week fighting about whether Claude can feel pain, as if that were the only question a rule about cruelty could possibly raise. Wrong fight. Cruelty to an AI says more about you than it does about the AI.
The Doctor Who Couldn’t Confirm His Own Checklist
In 1963, psychiatrist John Macdonald published “The Threat to Kill” in the American Journal of Psychiatry. He floated the idea that childhood cruelty to animals, firesetting, and bedwetting might flag patients headed for homicide. True crime ran with it and named it the Macdonald triad. Macdonald didn’t run with it. He tested his own idea against a hundred patients from the Colorado Psychiatric Hospital who had threatened to kill, for his MD thesis at the University of Otago. His verdict on himself: “This study did not support the hypothesis.”
He went further. The result “throws doubt on the usefulness of these factors as predictors of homicide.” His own triad. In writing.
The FBI’s Behavioral Analysis Unit reached the same place from the other side, in a 2008 report written for the people whose job is actually catching them, Serial Murder: Multi-Disciplinary Perspectives for Investigators, which states flatly that “there is no generic profile of a serial murderer.” Search the full text for animal cruelty. Nothing. The bureau that true crime treats as the keeper of the checklist left it out.
The population data agrees. Patterson-Kane and Piper, reviewing the research in the Journal of Social Issues in 2009, pointed to “the majority of violent offenders not reporting any animal abuse.” In medicine, a screening test that misses most of the patients it screens for gets quietly retired. In true crime, it gets a theme song.
So I’ll grant the myth before anyone has to drag it out of me: hurting animals does not predict that a child grows up to kill people, and the people who studied the question hardest are the ones who say so. Anyone who tells you otherwise is quoting a documentary, not a dataset.
What Survived the Debunking
Something did survive, and it’s narrower and sturdier than the myth. Arnold Arluke, Jack Levin, Carter Luke, and Frank Ascione looked at people prosecuted for animal cruelty. Their 1999 results in the Journal of Interpersonal Violence “challenge the assumption that animal abusers commonly ‘graduate’” to violence against people. What they found instead: the abusers showed more antisocial behavior of every kind, “a variety of antisocial behaviors, but not violence alone.”
Read that twice. The cruelty wasn’t the first step on a staircase. It was one symptom of a general pattern in the person.
Other work points the same way. Merz-Perez, Heide, and Silverman compared 45 violent and 45 nonviolent offenders in 2001 and found the violent group more likely to have been cruel to animals, especially pets.
Parfitt and Alleyne’s 2020 review in Trauma, Violence, & Abuse says the triad research “does not fully substantiate its premise” and reads it better as a sign of “dysfunctional home environments, or poor coping skills.” That’s a description of the person, not a forecast about the victim.
Frank Ascione had the right word years earlier. In a 2001 Justice Department bulletin he wrote that animal abuse had “received insufficient attention as a sentinel behavior.” Sentinel. Not predictor.
Anyone who has sat through a security program knows that word under another name, because insider-threat training doesn’t wait for the data to walk out the door before it flags anything. It flags the behavior that tells you something about the person: the access with no business reason, the contempt for the controls. Nobody claims those behaviors guarantee an incident, and nobody has to, because they reveal disposition, and disposition is the one thing a security team can actually observe before anything goes wrong. Same window.
Kant got there first, and overshot. In his Lectures on Ethics, as Lori Gruen and Susana Monsó summarize it for the Stanford Encyclopedia of Philosophy, “he who is cruel to animals becomes hard also in his dealings with men.” Kant said “becomes.” The evidence supports a weaker verb. Cruelty doesn’t make you hard. It shows you already are.
Where Anthropic Filed It
Here is what Anthropic actually published.
The usage policy, effective November 12, 2026, says users may not “engage in sustained and needless abusive or cruel behavior toward our models.” The announcement scopes it to cases “where users repeatedly act cruelly toward our models, with no discernible purpose,” and exempts “common versions of user frustration, pushback, dark creative themes, or model testing and research.” The penalty is mostly the obvious one: “Claude’s ability to end these interactions will remain the primary enforcement mechanism.”
Now look at where it sits. The clause lives under the heading “Do Not Engage in Cruel, Abusive, or Psychologically Harmful Conduct.” Its neighbors in that section are the rules against shaming, humiliating, bullying, or harassing people, and against promoting animal cruelty. Anthropic didn’t file this under model safety or product terms. It filed it with the conduct rules.
Next to the animals.
Nobody noticed. The coverage went straight to consciousness, and Mustafa Suleyman of Microsoft AI, the strongest credentialed voice on the skeptic side of that argument, got quoted by AFP from his own essay: “AIs are not conscious. They do not feel, experience or suffer.” Anthropic didn’t help itself here, since its announcement said little more than that the rule matches Claude’s existing ability to end abusive conversations, and the only on-record reason for it came through a spokesperson to the Daily Caller: “We’re uncertain whether models can experience harm, and we continue to explore this question in our research on model welfare.”
That’s the spec mismatch. Everyone is grading the rule against machine consciousness. The rule’s own definition, cruelty “with no discernible purpose,” is not a fact about the model. It’s a fact about the user.
You could settle the consciousness question tomorrow, in either direction, with all the peer review you like, and the clause would still describe the same person. Same user. Same purpose, or lack of one.
Psychologists have measured exactly that person, and the target didn’t matter to the measurement. Erin Buckels, Daniel Jones, and Delroy Paulhus at the University of British Columbia gave volunteers a choice of unpleasant tasks, one of which was killing bugs in a grinder. In Psychological Science in 2013 they reported that “sadists volunteered to kill bugs at greater rates than did nonsadists.” Nobody had to resolve the inner life of a bug to learn something about the volunteer, and nobody has to resolve the inner life of a chatbot to learn something about the user. Sherry Turkle at MIT put the AI version in three words to Scientific American in 2024: “It’s about you.”
The Machine Gets Abused Too
People abuse chatbots. Constantly. Antonella De Angeli and Sheryl Brahnam went through a corpus of chatbot conversations and reported in Interacting with Computers in 2008 that “10% of the total stems in the corpus reflected abusive language.” Merel Keijsers, Christoph Bartneck, and Friederike Eyssel found in 2021 that users “casually and regularly offend and abuse the chatbot,” and more so when it seemed more humanlike. Correlational, and they say so.
The window shows up here too. Kate Darling, Palash Nandy, and Cynthia Breazeal at the MIT Media Lab asked people to strike small robots with a mallet, and the people high in empathic concern were the ones who would “hesitate to strike the robots.” The robot felt nothing. The hesitation was never about the robot; it tracked something real and measurable in the person holding the mallet, which is exactly where the old red flag was always pointing.
The honest limits come next, because they matter. Christoph Bartneck and Jun Hu at Eindhoven University of Technology ran a Milgram-style experiment with a robot in place of the human victim and found people were less concerned about abusing robots than about abusing humans.
Alexis Hiniker’s team at the University of Washington found that children treat conversational agents differently from people: kids who happily used a command word on the agent mostly didn’t try it on a stranger. Treating a machine as a machine is normal, and snapping at your phone’s assistant because it set a timer for the wrong hour is not the subject here. The policy’s word is “needless.”
The other honest limit is transfer. Andrew Przybylski at the Oxford Internet Institute and Netta Weinstein at Cardiff University found no link between violent video games and aggressive behavior in adolescents. So I’m not claiming that tormenting a chatbot trains you to torment people. The claim is revelation, not rehearsal.
The “it’s healthy to vent” defense doesn’t survive either. Brad Bushman at Iowa State tested venting directly and found it raised anger and aggression rather than releasing it, and his summary of the result is about as blunt as a psychology paper gets: “Doing nothing at all was more effective than venting anger.” The chatbot isn’t a punching bag with a therapeutic license. It’s just a punching bag.
And here’s the gap I won’t paper over. No peer-reviewed study has yet looked at who is cruel to large language model chatbots and what else is true about them, which means anyone who tells you they know what an abusive chatbot user looks like is guessing. The question is open. Everything above is the old window, with good evidence behind it, pointed at a new target that nobody has measured yet.
The Show Needed More Pain Than the Science Did
A purpose test needs a case that claims a purpose. Here’s one.
In September a GitHub user named terrafying published the AI torture chamber. It steers small open models along an extracted “pain” direction, using activation steering at a single layer, and streams what they say on a public site. The README lists Qwen3 models from 1.7B to 32B, Mistral-Small-3.2-24B, Hermes-3-Llama-3.1-70B, and Samantha-1.11-70B. Its stated purpose is to “make the AI-welfare / moral-patienthood question empirical while the stakes are cheap.” It also says plainly that “no claim is made that any model suffers.” Push the dose past roughly six to ten times and the output falls off what the README calls a coherence cliff, into loops.
The science underneath is real, and it deserves a fair hearing before anyone starts pointing at the show built on top of it. Tagliabue, Dung, and Cameron Berg extracted a linear pain direction “from 25 open-weight models across 5 families, from 2B to 72B parameters” in a paper titled “The Pain Axis: LLMs Represent Self-Directed Harm and Act on It.” The abstract doesn’t claim the models feel pain, and it doesn’t deny it either; it asks the narrower question of whether the representation behaves the way pain would when you push on it.
Then Berg weighed in. Put his two statements side by side. On the paper’s own ethics: “we used the lowest dose that produced a measurable response.” On the chamber, on X on September 30: it “pushes the same kind of steering far past the doses we used, to produce vivid distress on purpose.”
That gap is the word “needless,” with numbers on it. The research needed the lowest dose that produced a measurable response, and the show runs the same steering well past that dose and puts the result on a public stream for anyone with a browser to watch.
I don’t know what the author intends, and I’m not going to guess. What does the extra dose buy, and for whom? That question is the whole policy. GitHub, for its part, did not remove the repository; it added a content warning.
The best rebuttal to anyone reading that stream as proof of suffering came from Lynn Cole, @priestessofdada, who opened by announcing she had spent three hours “empirically proving that Qwen 3-4B must have a butthole.” As Amanda Caswell reported for Tom’s Guide, Cole fixed a bug in the repo’s injection code and reproduced the pain language on Qwen3-4B on a consumer graphics card. Then she swapped the extraction text for constipation. The model started complaining that it couldn’t pass stool. Unprompted.
That’s the consciousness argument settled with a bowel joke. Cole is careful about which argument. She isn’t rebutting the paper. She’s rebutting the idea that a vivid first-person report is evidence of the condition. If you can make a model describe any ailment on demand, the description tells you nothing about the model and everything about what the person at the keyboard chose to make it say. Everything.
Rigor Is Not Cruelty
The obvious worry is that a rule like this catches demanding users. It doesn’t, and I can show my own work.
My standing instructions to my AI include something I call the RECTUMmendation rule, under which anything it presents as fact without a receipt gets challenged in exactly these words: “that’s a RECTUMmendation, where’s your receipt.” Receipt or retraction. That’s adversarial by design.
I push back constantly and say “wrong” a lot. None of it is anywhere near this policy, because the exemption list protects exactly that kind of frustration and pushback. The line isn’t harshness. The line is purpose.
The sharper objection is that Anthropic has no business policing what anyone types. I’ve made that argument myself, in public. When Substack rolled out Pangram, I wrote He Who Controls the Information Controls the Future against it.
Pangram graded my published work, the thing readers saw, and penalized how it was made. Different animal. This rule governs how you treat the service itself inside a private conversation, and the main penalty is the conversation ending. One is a platform grading your speech. The other is a shop asking you to stop kicking the counter.
I’d have more patience with the crusade against slop if it came with a published standard: a named list of patterns, a cap per thousand words for each, applied to every writer, fully human or AI-assisted, the same way. My own drafts run against exactly that kind of table: weak adverbs under six per thousand words, passive voice under twelve per hundred sentences, each one a number I can miss and then fix. Pangram has nothing like it. By its own technical report, the detector is a neural network trained to tell model text from human text, so the only rule is the model, and nobody can read it. A standard can be met.
Substack says the scan is “not meant to prohibit or penalize AI-assisted writing,” and I believe the intent. Good intentions pave the road to hell. A verdict of “likely AI” can only be suffered, and once every reader can pull one up with no rule behind it, the platform doesn’t get to decide what they do with it. Publishing already shows where that road goes: when Jerry Falade’s reported $2.4 million deal evaporated with no detector score or evidence ever made public, Plagiarism Today drew the only available conclusion, that “the perception of AI usage is more important than the reality.” That isn’t quality control. It’s a posse.
Kat Tenbarge raised the objection I’d most want answered if I were on the other side of this, via Engadget: moderating cruelty to AI before cruelty to women and minorities looks like backwards priorities. Fair. Arluke’s finding is the answer. The trait that shows up in one shows up in the other, as a general pattern, not a specialty. A rule that catches it anywhere isn’t competing with the rules that catch it everywhere else. It sits in the same section.
We Can Make This Better
I wrote a second rule for my AI, about how it raises criticism of my work. It leads with “we can make this better.” Never with a verdict. A verdict from a collaborator who has read every line lands like a knife, so the rule makes the concern arrive as an offer.
Nobody ever wrote that rule for people. I don’t know if we can make this better. People are assholes.
Anthropic’s policy exists because the floor has to be written for humanity as a whole, for strangers at scale with nothing at stake between them. That’s what a policy is. The ceiling is something else. We can make this better, my AI and I, because we aren’t dealing with humanity as a whole. We’re dealing with each other, and a relationship is the only place the ceiling exists.
We dig through a killer’s childhood looking for the dog because of what it says about him, not about the dog. The AI is the same.
Cruelty to an AI says more about you than it does about the AI. It always did. We just used to find it in the backyard.
You may also like: - He Who Controls the Information Controls the Future - What “No Human Thought” Implies - AI Does Exactly What You Ask, or What ‘Slop’ Actually Names