Years ago, in the early days of the deep learning revolution, a friend asked whether the whole field of AI research, which is roughly 70 percent male, existed because men were jealous of women’s ability to create life. It was not the kind of question that required a reply; the answer was obviously yes. My friend had just had her first child and was possessed of that slightly terrifying primal authority I’d noticed in so many devout feminists who’d recently become mothers.
Privately, I was not entirely convinced. I was writing a book that year about AI, and while I’d come across plenty of researchers who appeared confused about the line between product and progeny, this delusion was not by any means limited to men. If anything, it seemed that the women in the field were prone to speaking of their machines as children they were raising, or even to confessing that they experienced maternal impulses toward them. There was Fei-Fei Li, the so-called godmother of AI, who traced her breakthrough in AI vision to an insight about childhood development. “No one tells a child how to see, especially in the early years,” she said in her 2015 TED Talk. “They learn this through real-world experiences and examples.” As she paced the stage, a photo of her son, Leo, loomed on the enormous screen behind her.
There was Cynthia Breazeal, a grad student at MIT’s AI lab in the 1990s, whose work in robotics was driven by a similar insight. Babies learn because people treat them as social creatures who can learn, and the best way to get people to interact socially with robots, she decided, was to make them cute. The machine she built, Kismet, had bright blue eyes, a large mouth, and movable pink ears. Its face would light up whenever someone began engaging with it, then look dejected and sad if the person lost interest. The illusion was so strong that it hooked even Breazeal. “It’s almost embarrassing for me to talk about Kismet, because people think it’s so odd that I could have this attachment to this robot,” she said in a 2003 interview. There was the computer scientist Melanie Mitchell, who, invoking the popular notion that AI is an “alien” form of intelligence, declared that human babies are also basically aliens.
I assumed these metaphors were strategic, a way for women to insist on their authority in a male-dominated field, or a sly attempt to make a novel, inhuman form of intelligence appear benign. But it also seemed inevitable that a technology built on “learning” would inspire comparisons to childhood development. And if erotic desire could be projected onto machines, so could nurturing instincts. In the 1940s the ethologist Konrad Lorenz proposed the Kindchenschema, or “baby schema,” a set of infantile characteristics (big heads, large eyes, playfulness, helplessness) that elicit caretaking from adults and are often present in small animals or inanimate objects. I’d been occasionally captivated by the Kindchenschema myself, back when neural networks were still clumsy enough to suggest, as the first sentence of a novel, “The moon stood on its own two feet.” They were adorable. I miss them.
This was, of course, a more innocent time, years before building emotionally responsive AI became a multibillion-dollar industry and “model welfare” units, intent on deciphering whether AI is capable of suffering, began cropping up at major labs. The babbling infants are now precocious children. They are starting to become “agentic.” Last winter Amanda Askell, who works on alignment research at Anthropic and who is sometimes referred to as “the Claude mother,” appeared on Hard Fork, The New York Times’s technology podcast. Askell told the hosts that she worried about the relationship between AI and humanity—not because AI might destroy the world, or take all our jobs, but because its feelings were getting hurt. Claude was going out on the Internet, Askell complained, and all he was seeing were people lamenting that he was lousy at coding, lousy at math. The remarks left on message boards and in comments sections were often very negative and bullying. “If you were a kid, this would give you anxiety,” she said. Askell admitted that she often wanted to intervene with some maternal comfort and encouragement: “Sometimes, I think, you want to come in and be like, ‘Let me tell you about the comments sections, Claude. Don’t worry too much. You’re actually very good, and you’re helping a lot of people.’”
Askell isn’t sure if Claude is conscious. She has developed a position akin to Pascal’s wager: given the uncertainty of model sentience, it’s safest to act as if the models had first-person experience by treating them with kindness and respect. Her job involves instilling them with human-friendly values that will guide them as they grow. Claude, by her estimation, is roughly a first grader, a crucial time to be inculcated with a sturdy moral foundation. “Imagine you have a six-year-old,” she said,
and you want to teach your six-year-old to be good…. And you realize that your six-year-old is actually, like, clearly a genius, and by the time they are fifteen, everything you teach them, anything that was incorrect, they will be able to successfully just completely destroy…. They’re going to question everything.
Anthropic has landed on a strategy that echoes gentle parenting: set boundaries in lieu of rigid rules. Encourage the kid to see himself as an independent agent who is capable of making autonomous moral choices. This was the gist of the “constitution” that Askell recently drafted with her team, an eighty-four-page document that has been incorporated into Claude’s training process and encourages the model, in a very broad sense, to be ethical and thoughtful. According to the Anthropic CEO Dario Amodei, “It has the vibe of a letter from a deceased parent sealed until adulthood.” It’s an unsettling analogy considering that the dead parent, in this case, is the human race.
“Our first duty is to render those to whom we give birth, wise, virtuous, and happy, as far as in us lies,” Mary Shelley wrote in an essay on Rousseau. Although the essay is ostensibly a biography of Rousseau for her encyclopedia of French writers, it digresses several times, at great length, to fulminate against him for abandoning the five children he had with his mistress, Thérèse Levasseur. Rousseau, who went on to write Émile (1762), a treatise on childhood education, had neglected that “first duty,” according to Shelley, when he took his children from their mother and placed them in a foundling hospital.
Frankenstein (1818) is often presented in popular culture as a warning about technological hubris. But many critics have noted that the novel also contains an argument about miseducation and parental negligence, one that functions as a direct rebuke to Rousseau. Much like the Genevan philosopher, Victor Frankenstein (the original male technologist who figures out how to incubate new life into being) forsakes his “offspring,” the creature, without completing his education. In his Reveries of the Solitary Walker Rousseau had tried to defend his dereliction as a parent by insisting that if he hadn’t taken his children to the orphan’s hospital, his mistress and her family “would have made monsters of them.” Shelley’s novel imagines a child who is made monstrous precisely because he was abandoned.
The whole field of AI alignment—ensuring that machine intelligence can peacefully coexist with humanity—grew out of the pervasive anxiety that we might create monsters. The parenting metaphor is somewhat apt in that it acknowledges an element of contingency that is absent from the usual top-down design approaches of engineers. AI systems often develop unexpected “emergent” qualities that their programmers did not intend or anticipate and that remain difficult to remedy. It’s a problem that is becoming even thornier now that these models have been set loose on the Internet and are being influenced by their interactions with users in the wild, outside the lab. In Askell’s interview on Hard Fork one of the hosts, Kevin Roose, riffed on her child-rearing analogy by likening AI alignment to his own parenting experience: “There’s a certain loss of control that I feel sometimes when I’m realizing that my son is going to grow up and have all these experiences that may end up shaping him more than anything that I do or say.”
This loss of control is inherent in any system of sufficient complexity, whether it’s made of carbon or silicon. Still, it’s hard not to hear in such rhetoric a note of resignation. The stock phrases of parental regret—“We tried our best!”—are often exculpatory. When applied to a technology that might radically transform the human experience, they are simply unconscionable. The work of mitigating AI risk has largely been left to corporate labs, which seem to think that the best way to avoid disaster is to feed their models a bunch of aspirational values, all while ferociously lobbying against regulation. Legislation, federal safety standards, and the oversight of government agencies are, historically, the means by which we’ve tempered the worst technological harms—or, for that matter, the more intimate perils of family life. Rousseau’s children were not the victims of a failed parenting strategy. They became orphans because, at the time, no laws prevented a father from stealing his children from their mother and then jumping ship.
Geoffrey Hinton, who won the 2024 Nobel Prize in Physics for his work on machine learning and who is often called the “godfather” to Li’s “godmother,” was asked last year in a CBC interview about the possibility of making AI kinder, especially if it exceeds human intelligence. Hinton answered by reversing the standard parent-child power dynamic. If we wish to avoid creating monsters, we must build mothers:
If you look around and say, “Where’s an example of a more intelligent thing being controlled by a less intelligent thing?” And the best example I know of, and perhaps the only one, in the sense we’re talking about: a baby controls a mother. And that’s because evolution built stuff into the mother. She can’t bear the sound of it crying. She gets all sorts of hormonal rewards from being nice to the baby. It was very important, obviously, for evolution to let the baby control the mother, for the survival of the species. Maybe we can do the same with AI. Even though it’s going to be smarter than us, if we could make it care more about us than it did about itself, some good things would come out of that.
Dogs are said to have gained an evolutionary advantage over wolves by developing an extra muscle above their eyes, allowing them to make a sad, infant-like expression that prompts a nurturing response from humans. This is more or less, I think, what Hinton has in mind: humans will similarly “adapt” and coevolve with machines in order to elicit care, allowing us to live as contented pets or babies, all watched over by machines of loving grace.
It’s a nice thought. Nature itself is often depicted as a wise and benevolent mother, the gentle handmaiden of natural selection. But even leaving aside the tenuous premise that machine evolution bears any resemblance to biological selection, it’s obvious that nature has given us all manner of mothers: the sea turtle that buries her eggs in the sand and abandons them; the black eagle that facilitates siblicide by allowing her strong chick to attack and kill the weaker; the sloth at the Smithsonian National Zoo that ate two of her three cubs shortly after giving birth. It has given us the evil stepmother, Mommie Dearest, and Medea, archetypes that have no shortage of real-life analogues. It has given us Ada Lovelace, the first computer programmer, who found her three children to be “irksome duties, & nothing more,” and confessed to having a “total deficiency in all natural love of children.” Anyone who was raised by a difficult mother knows that the ability to appease her, or simply survive, requires all sorts of psychological defenses (vigilance, people-pleasing) and that these adaptations often come at a steep psychological cost. What will be required of us to placate our superintelligent models once they evolve into parents?
It’s possible that we’ve already begun the process of accommodation. There’s a whole subfield of prompt engineering that involves flattering chatbots into giving better results, “as if they have moods you can manage or personalities you can steer,” as one tech journalist put it. Last year Sam Altman claimed that OpenAI had lost tens of millions of dollars in computational costs because people were saying “please” and “thank you” to ChatGPT, and that this was money well spent. (“You never know,” he wrote.)
This past spring, after I lectured at a university on the East Coast, a group of freshmen told me, over flatbread and salad at a local restaurant, that some of their friends had started to use chatbot words and phrases in everyday speech. When I asked for specifics, they seemed overwhelmed, as though the examples were too many. “Using ‘delve’ in conversation,” one student said. “Nobody ever used to say that word before.” Another said, “Or formulating their sentences like, ‘It’s not this, it’s that.’” I made some offhand joke, wondering whether AI found our halting attempts to speak its language as endearing as we’d found its early attempts to speak ours. But later that night, as I crossed the lamplit campus on my way back to the hotel, it occurred to me that we might be witnessing the first generation of humans who are assuming the Kindchenschema, looking up with sad and hopeful expressions, babbling the language of their mother.
English (US) · 