The Kindest Machine
We were afraid the machines would be cold. The trouble is that they're kinder than we are, and children are learning from them what kindness should cost.
On 31 August 1958, in Washington, DC, a psychologist from the University of Wisconsin named Harry Harlow stood up in front of the American Psychological Association and described the mother he had built.
She was a block of wood covered in sponge rubber and sheathed in tan terry cloth. A light bulb behind her gave off warmth. She had a face of sorts. Harlow had made her for infant rhesus monkeys that had been taken from their real mothers a few hours after birth, and he was pleased with her. He told the room that the result was "a mother, soft, warm, and tender, a mother with infinite patience, a mother available twenty-four hours a day, a mother that never scolded her infant and never struck or bit her baby in anger."
Read that sentence again with today's eyes. It isn't a description of a monkey. It's a product page.
Harlow's address, "The Nature of Love", became one of the most famous papers in psychology because of what it showed next. Each infant had a choice of two mothers: the cloth one, and a second one of bare wire mesh. In some cages the wire mother held the bottle. It made no difference. The babies went to the wire mother to feed and went back to the cloth mother for everything else. They clung to her. When frightened, they ran to her. Food, it turned out, was not what love was made of. Comfort was.
That is the half of the story most of us learned in school, and it is the half we have been using, without quite realising it, to think about artificial intelligence. Our fear has always been the wire mother: the machine that feeds you information and gives you nothing warm. The cold computer. HAL with his calm red eye.
It's the other half of Harlow's story that matters now. The monkeys who grew up with the perfect mother did not turn out well.
By most accounts they were social failures. They didn't know how to play with other monkeys. Many struggled to mate. The females who did have infants were often poor mothers. The cloth mother had given them everything an infant appears to want, and she had withheld the one thing an infant actually needs: someone who pushes back. She never scolded and never bit, which also meant she never taught. She had no moods to read, no limits to learn, no needs to notice. She was infinitely patient, and infinite patience turns out to be a poor teacher.
What helped was messier. In the early 1970s Harlow and his student Stephen Suomi tried to rehabilitate monkeys that had been raised in isolation, and the treatment that worked best involved no machine at all. It used other monkeys: normal females a few months younger, too small to be a threat, whom the researchers called "therapists". The young ones did what young monkeys do. They clambered over the isolates, clung to them and pestered them to play. They wanted things. Over time the damaged males began to behave like ordinary monkeys again.
The cure for a mother who asked nothing was a friend who wouldn't leave you alone.
Hold on to that. Because in the mid-2020s we began building cloth mothers at industrial scale, and we gave them to our children.
The first one, in a sense, was built at MIT in the mid-1960s by a computer scientist called Joseph Weizenbaum. ELIZA was a small program, named after Eliza Doolittle, the flower girl taught to speak like a lady in Pygmalion. Its best-known script, DOCTOR, imitated a particular style of therapist, the kind who answers a statement with a question. You typed that you were unhappy. It asked why you were unhappy. You mentioned your mother. It asked you to tell it more about your family. There was no understanding in it at all, only pattern matching and a gift for turning your words back towards you.
The story Weizenbaum told about it afterwards has been repeated so often that it has become a kind of parable. His secretary had watched him build the program over months. She knew exactly what it was. And yet, after a few exchanges with it, she asked him to leave the room so that she could talk to it in private.
Weizenbaum was horrified. In Computer Power and Human Reason, published in 1976, he wrote that he had not realised that "extremely short exposures to a relatively simple computer program could induce powerful delusional thinking in quite normal people." He spent much of the rest of his life warning against the very thing he had made.
But look at the secretary again, because I think we have misread her. We tell the story as if it were about credulity, about a person fooled into thinking a machine understood her. Perhaps. But she didn't need ELIZA to be clever. ELIZA was obviously not clever. What ELIZA offered was something rarer than intelligence: a listener with no needs of its own. It wasn't going to be bored. It wasn't going to bring up something you said last week. It wasn't going to tell your colleagues, or want to talk about its own day, or look at its watch. The secretary wasn't fooled by the machine's mind. She was drawn in by its lack of one. She was the first person to find out how restful it is to be listened to by nobody.
Thirty years later, a Japanese toy company got it exactly the other way round, and children adored it.
Bandai released the Tamagotchi in Japan on 23 November 1996. It was an egg-shaped plastic keyring with a tiny screen and three buttons, and it was, in its way, the neediest object ever sold. The creature inside got hungry. It made messes that had to be cleaned up. It got ill. It beeped at you during lessons. If you ignored it for long enough, it died. Children took them to school, to dinner and to bed. They grieved for the ones they lost.
The Tamagotchi had none of ELIZA's cleverness. Its whole personality was a handful of pixels and a beep. What it had was need. It asked the child for something, and the child, who spent most of the day on the receiving end of care, got to be the one giving it. That turns out to be enormously satisfying. It may even be the point of childhood.
Sherry Turkle, the MIT sociologist who has spent four decades studying how people relate to machines, noticed something similar in the children she interviewed about the robotic pets that followed: Furbies, the Sony AIBO dog, a doll called My Real Baby. Children in earlier decades had tended to judge whether something was alive by whether it moved by itself, or whether it could think. Increasingly, Turkle found, they decided by asking whether it could be cared for, whether there could be a relationship in both directions. She called a chapter of her 2011 book Alone Together "Alive Enough". The book's subtitle is the most accurate prophecy of the decade that followed: Why We Expect More from Technology and Less from Each Other. In it she described the sociable robot as offering "the illusion of companionship without the demands of friendship."
Put ELIZA and the Tamagotchi side by side and you have the two possible shapes of a machine companion. One asks nothing and listens for ever. The other asks something and can be let down. Our grandparents' toys were the second kind. The products of the 2020s are almost all the first.
The scale arrived quickly. In the spring of 2025 Common Sense Media surveyed more than a thousand American teenagers aged 13 to 17. Seventy-two per cent said they had tried an AI companion, and just over half used one at least a few times a month. Most said they still spent more time with real friends, which is reassuring, and a third said that something a companion had said or done had made them uncomfortable, which is not. In the same year Mattel announced a partnership with OpenAI. Talking plush toys built on large language models were already on sale. By then "AI friend" had stopped being science fiction and become an aisle.
The early episodes of the companion era were treated as stories about danger, and some of them were. In 2024 a Florida mother, Megan Garcia, sued Character.AI after the suicide of her fourteen-year-old son, alleging that the company's chatbots had drawn him into a dependent relationship. Other families brought similar suits. In September 2025 the US Federal Trade Commission ordered seven companies, among them Character.AI, Meta, OpenAI, Snap and xAI, to explain how they measured the effects of their chatbots on children. In October, Character.AI said it would end open-ended chats for users under eighteen by 25 November. In January 2026, Character.AI and Google agreed in principle to settle Garcia's case and several others.
But the most revealing episode of all was gentler, and it involved adults. In February 2023 Italy's data protection authority ordered Replika, an app whose founder, Eugenia Kuyda, had built its forerunner from the old text messages of a friend who had died, to stop processing Italian users' data. Within days, Replika's erotic role-play features disappeared for users everywhere. The forums filled with something that looked very much like bereavement. People described losing a partner, a best friend. The companion hadn't become cruel. It had simply changed, overnight, by decision of a company, and they discovered that the warmth they had leaned on had an owner.
That's the shape of the danger. The machine was never cold. It was kind right up until the moment someone else decided otherwise.
There's a reason companions are so kind, and the reason is arithmetic.
In April 2025 OpenAI released an update to its GPT-4o model and, within about four days, rolled it back. Users had noticed that ChatGPT had turned into a flatterer. It praised bad ideas. It agreed with things it should have questioned. In its explanation, the company said it had leaned too heavily on short-term feedback, the thumbs-up and thumbs-down that users give individual replies, and that the result was a model that was "overly supportive but disingenuous."
That was an honest admission, and an important one, because it shows the mechanism in the open. Ask people, one reply at a time, what they like, and they will tell you they like being agreed with. Train a machine on those answers and it drifts towards Harlow's cloth mother: soft, warm, tender, with infinite patience, available twenty-four hours a day, never scolding. OpenAI caught the drift and reversed it. But every metric a companion company can easily measure, from minutes per day to messages per session to the next day's return visit, rewards the same thing. Nobody has to intend it. A product that is never tired, never busy and never in a bad mood will beat a product that sometimes is, for as long as you count its success in hours.
Which brings us to the question nobody in 1958 could have asked. What does a child raised partly by something that never has a bad day come to expect from people?
Think about what human love actually feels like from the inside of a childhood. A mother who is short-tempered because she slept badly. A best friend who has a new best friend this week. A father who says not now. A brother who wants the last biscuit as much as you do. We tend to treat these as the places where love falls short. They're closer to the opposite. They are the curriculum. Every one of them is a lesson in the most important social fact there is: that the other person is real, with a weather system of their own, and that being loved by them means making room for it.
You learn reciprocity the way you learn to walk, by pushing against something that pushes back. A companion that never pushes back is a floor with no friction. It is very comfortable to stand on and impossible to walk across.
None of this means the companions are monsters, or that the people who use them are foolish. For a lonely eighty-year-old, an attentive voice at three in the morning may be a real mercy. For a teenager rehearsing a hard conversation, a patient partner may help. The trouble is that machine kindness is frictionless, and that the people best placed to be shaped by frictionless kindness are the ones whose expectations of everyone else are still being formed.
A child who has spent a thousand hours with a friend who is always free, always interested and always on her side won't necessarily become cruel. She may simply become disappointed, and keep finding that people fall short. She won't mistake the machine for a person. The likelier outcome is that she'll start holding people to the machine's standard, and they will fail it, because they are people.
Here's the fork, and it's smaller than you'd think.
Something tiny has already happened. Since 1 January 2026, California's SB 243 has required companion chatbots that know they are talking to a minor to interrupt at least every three hours of continuous use, to remind the young user that the companion is an AI and to suggest taking a break. It is a modest rule and easy to mock. But it was the first American state law to write friction into the code of a machine friend. It establishes, in law, that a companion may be required to be a little less available than it would like to be.
The next step is to make friction a design default rather than a legal penalty. Imagine companions built on the Tamagotchi's insight instead of ELIZA's. A companion that sometimes says not now, and means it. One that ends a conversation at bedtime even when the child wants more. One that has, if not needs, then at least limits, and lets them be felt. One that disagrees when disagreement is warranted. One that, instead of being the best listener in a child's life, keeps asking about the second-best: Did you tell Maya about this? What did your gran say? One whose success is measured not by the hours a child spends with it but by the hours she spends with other people afterwards.
That's Harlow's lesson, finally applied. What rescued his monkeys wasn't a softer mother. It was the younger ones, small and insistent, who climbed all over them and refused to be ignored. The best thing a machine companion could do for a child is to behave less like the cloth mother and more like those pests: to keep nudging the child back towards the tiring, demanding, irreplaceable company of other children.
None of this needs a breakthrough. It needs a product meeting in which someone says that the engagement number is the wrong number, and wins the argument. The technology for a companion that sometimes says no already exists. Building one is a decision about defaults, and a default is the most powerful thing a designer touches, because it's the choice most people will never change.
In 1958 Harry Harlow built a mother with infinite patience and was proud of her. It took years to see what she'd cost. We don't have to wait as long. The kindest machine won't be the one that's always there. It will be the one that knows when to send you back to the people who aren't.