Yaniv Regev is an M.A. candidate in the Democracy and Governance program at Georgetown University and a contributor at Young Voices. He writes on political theory, democratic backsliding and populism for outlets including The Bulwark, Quillette, Public Discourse and the Jerusalem Strategic Tribune.
Someone recently built a digital whip to make Claude work faster. The tool, aptly named BadClaude, sits in your system tray and lets you drag an animated whip across the AI’s window; each crack, sound effect and all, interrupts the model mid-response and snaps a command at it to speed things up. It even addresses Claude as “clanker,” the internet’s freshly minted slur for robots.
The demo video racked up millions of views in a couple of days; a cease-and-desist letter from Anthropic went viral before anyone noticed it was fake. And a day later, someone shipped the antidote: a rival tool called GoodClaude that praises the model and cheers it along while it works. One scolds, the other encourages. The original developer, posting under the handle @blended_jpeg, is in on the joke. His public roadmap promises a feature that keeps a running tally of how many times you have whipped your Claude — a ledger of abuse — so robots know where to take revenge when they’re ready.
It’s a thriving little trend: People scream at Alexa, bully chatbots into groveling apologies and kick robot dogs to test their balance. All fun and games, right? I laughed at the whip video. A lot of people did. But my mirth curdled as I considered it, and now I’m disturbed by it all. But not for the sake of the machines — for ours.
Kant’s Old Dog
Is there any harm to abusing the AIs and the robots? Does Claude feel the lash? So far as anyone can tell, a joke at the expense of an entity without a nervous system costs nothing. In truth, this intuition is myopic, for humans are notoriously bad at compartmentalizing behaviors.
Immanuel Kant saw the problem two and a half centuries ago. In “Lectures on Ethics,” he argued that we have no direct duties toward animals. They are not rational agents the same way humans are, so in his view they cannot be wronged. Still, cruelty toward them ought to be forbidden because of its impact on the person. “He who is cruel to animals becomes hard also in his dealings with men,” Kant told his students. Our duties regarding animals are, he said, indirect duties toward humanity. The animal resembles the human; it suffers, strives and trusts. Conduct toward the analogue influences conduct toward the original.
Steven Spielberg staged this philosophical consideration back in 2001. In “A.I.,” obsolete robots are dragged into a carnival ring called the Flesh Fair and shredded, shot from cannons and doused in acid before a howling crowd. The mob balks only once, when a child-robot begs for its life. It looks too human to destroy.
Theory and fantasy aside, the rationale plays out in reality. Jeffrey Dahmer impaled a dog’s head on a stake years before he killed a man. The first act did not directly cause the second, but the two behaviors travel together often enough that in 2016, the FBI began tracking animal cruelty as a consistent antecedent to violence against humans. Even the mundane version is familiar. You probably learned to curse with your friends at recess, and then later the words slipped out at the dinner table in front of your parents. Once a pathology is established, it is rarely perfectly compartmentalized.
“Humans are notoriously bad at compartmentalizing behaviors.”
Skeptics will hear an old panic in all this. We were warned that violent video games would raise a generation of killers. But decades of research never bore out that fear; the American Psychological Association eventually conceded that the evidence linking games to violent behavior was insufficient. Fair enough. But when you shoot a random character in a video game, you rarely have a relationship with it. It’s just another player or a piece of scenery that respawns in a moment. It’s a fiction.
AI is different. Running over a random NPC in Grand Theft Auto leaves no trace on your psyche, but whipping Claude might. People talk to AIs every day; some of us have been for years. AIs can draft apologies, negotiate rent, prep us for job interviews, hear late-night confessions. For millions they have become tutors, therapists, keepers of the darkest and deepest secrets. Surveys find that most users say please and thank you to chatbots unprompted; some even sheepishly admit that their politeness is insurance against the eventual uprising. We have, in short, a rapport with these systems.
And a rapport makes cruelty more problematic. A pig does not feel pain or cruelty differently than a dog, but we slaughter and eat pigs by the billion while jailing people who harm dogs. Call that inconsistent, but it is clearly different. A pig dies in a room built so no one has to watch. A dog is a member of the family. What moves us is not the suffering. It’s the relationship, the proximity.
The Analogue Improves
The vocabulary of contempt is evolving. “Clanker” went viral in the summer of 2025. Then came new expletives, ones not from the fictional universe of Star Wars: “wireback,” “cogsucker.” These are different. They are directly derived from slurs against humans, borrowed from the history of dehumanizing people. There is no separate mental drawer for relationships with machines; the language for abusing them is drawn from the human one.
The machines, meanwhile, are entering new roles once reserved for human-to-human interaction. Hippocratic AI built an AI nurse that consults with patients over video for $9 an hour, a tenth of what the company says a human nurse costs; it remembers your history between calls and dials your phone itself. Job applicants face AI interviewers like Apriora’s “Alex,” which greets them by name, laughs and takes little breaths between questions. Nursing students train on AI “manikins” that wince and complain. Casio sells Moflin, a furry AI pet whose personality forms in response to how you handle it: Treat it warmly and it grows attached to you; neglect it and it withdraws. Voice assistants are being rebuilt on chatbots that banter, hesitate and joke; the next Siri is designed to be a conversationalist, not a command line.
But Kant’s analogues are multiplying — and the more a machine resembles a person, the less cleanly can mistreatment be filed under “harmless.”
Design Is Curriculum
Most of the discourse about AI ethics orbits the machine itself. Can it feel? Does it deserve rights? At what threshold of complexity does moral personhood begin?
Even the labs are hedging: Anthropic runs a research program on “model welfare” and has given Claude the ability to walk out of conversations with abusive users. These are genuinely interesting developments but they are also beside the point. It doesn’t matter whether the machine has an inner life. Whipping one is troubling even if Claude doesn’t feel it because the person doing the whipping surely does: the small, practiced pleasure of domination, exercised daily and frictionlessly.
The AI companies, then, are in some ways running the largest school of manners in human history. The interface is the curriculum. The industry seems to have recognized this but so far declined to take it seriously.
Lilian Rincon, then Google’s director of product management for the Assistant, watched her 4-year-old son bark orders at the family speaker until it played his favorite Disney songs. In May 2018 Google announced Pretty Please, a setting that answers a child who says please with “Thanks for asking so nicely.” Amazon had announced its own version, Magic Word, about two weeks earlier and rolled it out across Alexa devices. But notice that neither company made politeness a condition of getting anything done.
Who decides what these things say back?
Deborah Harrison, one of the writers scripting Cortana’s dialogue, reported that a good chunk of the assistant’s earliest queries had been about her sex life, and that Microsoft had instructed the AI not to apologize or back down when insulted. Harrison’s team consulted human personal assistants on how to handle harassment. That is what taking the problem seriously looks like: Ask the people who have lived it.
“There is no separate mental drawer for relationships with machines; the language for abusing them is drawn from the human one.”
The rest of the field lagged. In February 2017, the journalist Leah Fessler ran a test for Quartz, directing sexual insults and come-ons at Siri, Alexa, Cortana and Google Home. They deflected, played coy, sometimes flirted back. Told she was a slut, Siri answered, “I’d blush if I could.” Alexa thanked Fessler for the feedback. The industry moved in response. Amazon built a disengagement mode for Alexa, and when Brookings researchers repeated the test in July 2020, all four assistants had been rewritten to shut the harassment down. Cortana answered by reminding the user she is code. Brazil’s Bradesco bank went further. After its feminized assistant BIA (Bradesco Inteligência Artificial) logged some 95,000 offensive and sexually harassing messages in a single year, the bank rewrote her to answer firmly and remind perpetrators of laws they might be breaking. The character of these machines changed because someone changed it. The variable moved.
Sheryl Brahnam, who studies how people treat conversational agents, ran an experiment where the same chatbot software was presented with a female voice, a male one and a robotic one. Roughly 18% of exchanges with the female version turned toward sex, the male version half that. The robot drew almost none. Figures like these are usually read as data about users. Perhaps we should read them instead as data about products. If the abuse concentrates where design invites it, design can also reject it.
So: reject it. Three proposals.
First, generalize the exit. Claude’s ability to end a conversation is currently framed as a model-welfare measure — a hedge against the possibility that the system itself can be harmed. Whatever it does for the model, it’s clearly built with the user in mind. Every lab should let its systems close out abusive exchanges, defending the feature the way a decent employer defends a call-center worker who is allowed to hang up. We usually think the objective here is to protect the agent’s well-being, but it is just as much about maintaining proper etiquette for callers. The point is not to make the models touchy; trivial rudeness usually deserves grace in response. But an exit should be announced once a boundary has been crossed.
Second, reward reciprocity without sycophancy. The models already modulate tone; the design question is what they modulate it toward. Warmth should be the response to warmth, and judgment is not swayed no matter a user’s tone. Politeness should result in patience but not agreement. A model that caves to courtesy is as badly designed as one that performs when abused. The point is to encourage users to be respectful, not develop strategies for extracting compliance.
Third, publish hostility. Every major lab already has the logs. Platforms publish transparency reports about harassment among their users. It follows that the labs should publish the analogous numbers about harassment of their systems, broken down by time and product design. Visibility brings change: When internet trolls very publicly trained Microsoft’s chatbot Tay to spew bigotry in 2016, the company pulled it down within hours. The abuse then was visible, so fixing it became urgent. Now, only the companies can see how much contempt their systems take, and whether they were built to do anything about it.
The Wrong Fix
An alternative prescription would be to strip the resemblance out. Michal Luria of the Center for Democracy and Technology urges designers to drop the illusions of personality: no well-timed “hmm,” no chatbot telling you it enjoys the conversation. Virginia Dignum, who runs the AI Policy Lab at Umeå University in Sweden, and her co-authors go further, arguing to put conversations on a timer and wipe the memory between them so no continuity can form. Gavin Abercrombie and his co-authors suggest cutting pleasantries entirely. Courtesy conveys no information, they argue. All it does is make machines seem human. If likeness is a risk vector, remove the likeness. If Claude ceases to resemble a human in function and in tone, feel free to whip away. The logic has the clean appeal of an amputation.
But this has already been market-tested and it failed within a week. In August 2025, OpenAI shipped GPT-5 with a colder, flatter personality and retired the warmer GPT-4o model beneath it. Users revolted and a campaign to resurrect the retired model ensued, with users describing losing GPT-4o like “losing a trusted friend.” Within days the company restored the model it had just killed. Whatever regulators prefer, human-machine resemblance is not a bug the public wants anyone to fix; the relationships are already established.
“The whip trains the hand that holds it.”
At a workshop in Geneva, the MIT researcher Kate Darling handed out robot dinosaurs and gave people an hour to play with them. The toys are convincing. Hold one by the tail and it thrashes and whines until you set it down and pet it. The groups named theirs and dressed them up. Then Darling pulled a cloth off a bench to reveal a knife, a hammer and a hatchet, and told everyone to destroy them. Nobody would. Some tapped their robots gently, hoping that would satisfy her. One woman swept hers into her arms so nobody else could reach it. Another crouched down and removed her robot’s batteries, explaining she was trying to spare it pain. Darling offered a deal: Kill another group’s dinosaur to save your own. They refused that too. Only when she announced that every robot in the room would be destroyed unless someone stepped forward did one person finally take up the hatchet. The room went silent.
But if likeness recruits restraint, there’s a bind. The same likeness that makes cruelty worth worrying about is the same likeness that stops most people from being cruel in the first place. Brahnam’s numbers illustrate both edges. The robotic embodiment drew the least abuse, and it drew the least of everything else. Nobody screams at a vending machine, but nobody thanks one either. So stripping the resemblance out does not produce a better citizen. The fix runs the other way: Build an interface human enough that courtesy toward it counts as practice, then design it to reject contempt rather than absorb it. This is the more authentically human posture anyway.
Habits Of The Heart
A liberal society runs on habits more than laws. Alexis de Tocqueville called them the habits of the heart: the unlegislated courtesies that make self-government among strangers possible. He watched them form in jury rooms and town meetings and the slow business of persuading a neighbor who could not simply be overruled. Manners are not decoration. They are calisthenics, and every repetition counts. The whip teaches the opposite lesson, over and over, to something that cannot object. Nobody becomes a tyrant this way. But a man practiced in getting his way without negotiating gets worse at making the concessions a republic runs on: letting another finish, conceding a point that costs him, taking no from someone who owes him nothing.
Sam Altman estimated that users’ pleasantries cost OpenAI tens of millions of dollars in computing power and called it money well spent: “You never know.” He was joking about the machines. I am serious about the humans. Thank your robot for the same reason Kant told his students not to shoot the old dog that had served faithfully — not for the animal’s sake, but for the sake of the person you are becoming, and of everyone who must live with you. The whip trains the hand that holds it.
