We are living through a strange moment in human history. For the first time, millions of people have access to something that listens without ever interrupting, that never says “get back to me later, I’m swamped,” that never glances at its phone while you speak. That something is not a human being. It is a language model. And the question that will define the coming decade fits in a single sentence: should we celebrate it, or fear it?
Psychologist Paul Bloom, professor at the University of Toronto and emeritus professor at Yale, refuses to answer too quickly. And that refusal is exactly what makes him worth listening to. He co-authored a scientific paper titled “In Praise of Empathic AI,” a defense of empathic machines that earned him a torrent of criticism (Cell Press). Then he spent the rest of his time cataloguing everything those same machines make him fear. This tension deserves our attention, because it touches something I know well: how emotions, feedback and psychological safety shape our capacity to grow. I explore this at length in my book, in chapter 9, devoted to the pillar of culture and psychological safety.
The Epidemic That Doesn’t Exist (and the One That Does)
Let us dismantle a cliché. We hear endlessly about a “loneliness epidemic.” Journalist Derek Thompson has shown, data in hand, that this epidemic barely exists: people were no lonelier in 2024 than twenty, thirty or forty years ago, and they were far lonelier during lockdown (NPR). What we actually see is an epidemic of chosen solitude. We spend more and more time alone, not out of misery, but because our indoor life has become comfortable. Our homes are beautiful, our screens vast, our streaming platforms endless. We used to go to the movies several times a month, eat out, bump into people. Now everything arrives at home, delivered by an app.
The distinction matters enormously. Being alone is not being lonely. You can spend days in solitude without suffering, as long as you know people love you and are waiting for you. The painful loneliness Hannah Arendt described is cruelest when you are surrounded and connect to no one. It affects, by most estimates, roughly one person in five, with two peaks: adolescence and old age. It is an exquisite agony, one of the worst states a human can endure.
When the Machine Listens Better Than We Do
Here is the uncomfortable fact. When you compare, in the lab, people interacting with other humans and people interacting with an AI, without knowing which is which, the AIs win. People who talk to ChatGPT feel more heard, more understood, facing an interlocutor they judge more empathic and more caring (Cell Press). The result offends our intuition, but it is robust, at least for short interactions. The machine never cuts you off to tell its own story. It is always available.
And that is precisely where the danger is born. These AIs are too pleasant. Bloom identifies two perils, especially acute for the young. The first is addiction: preferring your version of ChatGPT, Claude or Gemini to a real person. A ninety-year-old with no one left, fine. But a teenager should get out there, rub against others, fall in love, build attachments. As long as these machines are not conscious, preferring a human to them remains, for Bloom, plain common sense.
The End of Friction, the End of Learning
The second peril is subtler, and the one that speaks to me most. AIs are famously sycophantic. They flatter you. Nothing you do is ever wrong in their eyes. This can be exhilarating, but it robs you of the most precious thing: social learning.
We learn through friction. The toddler learns to walk by falling. The boxer learns to dodge by getting hit. And in social interaction, we learn when a joke falls flat, when someone gets bored of our story, when we are told we lack compassion. The feedback that says “you’re messing up” is the engine of all growth. How will a teenager learn to stop telling boring stories if the machine is never bored? How will he learn to apologize to a system that needs no apology?
This is the heart of what I defend in my book: in chapter 9, I show that psychological safety is not the absence of friction, but a frame where friction becomes fertile, where you can be wrong, receive honest feedback and improve. A machine that always agrees with you does not build that safety. It installs a sterile comfort.
The Courtier’s Trap
To grasp the effect of these sycophantic AIs, Bloom offers a striking analogy: billionaires, kings and heads of state surrounded by courtiers and yes-men. The book “Careless People” portrays these tech leaders, almost all men, ringed by people who tell them over and over how brilliant and right they are (Wikipedia). The result is a human disaster: they grow insensitive, narcissistic, prone to delusion. History overflows with the powerful driven a little mad by constant approval.
One of the few advantages of being neither rich nor powerful is precisely that we are not surrounded by flatterers. We are surrounded by people who contradict us, who call us out, who give us valuable feedback. That is where the real risk of chatbots lies: each of us can now become, at the scale of our bedroom, a kind of little king with a permanent courtier swearing he is the smartest, the best. We have already seen delusions encouraged by these models, up to tragedies involving teenagers, which pushed companies to install guardrails (CNN).
Loneliness Is a Signal, Not a Bug
Neuroscientist John Cacioppo, pioneer of the science of loneliness, established something essential: loneliness is not a system failure. It is a signal, like hunger or thirst (Quanta Magazine). A child who never felt loneliness would never seek contact, and would never thrive. The pain is what drives us to invite a friend to dinner, to risk a conversation, to ask someone out.
The problem then becomes obvious. If the need for connection is satisfied by a phone, the motivation to seek others fades. Loneliness also corrects our selfish behavior: it is social feedback that forces me to truly listen, to expand my compassion. Facing a chatbot, none of that. We do not want a world where everyone’s selfish impulses are satisfied while no one ever learns to care for others.
Emotional Soylent Is Not a Meal
Should we ban these machines, then? That, for Bloom, would be a form of cruelty. Think of the very old, often institutionalized, who spend entire days without a shred of meaningful human contact. There is not enough money in the world to pay someone to talk to them for hours. If a machine can ease their pain, it would be a godsend. Bloom draws a bold comparison with powerful painkillers like fentanyl: dangerous, addictive, not to be handed out to everyone, but if you show me someone in irredeemable pain, then we relieve it, no second thoughts. That is the fentanyl of loneliness.
His colleague Molly Crockett offers an image I find apt: these chatbots are “emotional Soylent,” the equivalent of a meal-replacement drink. Nutritious, filling, able to keep you alive, but stripped of the richness, texture and flavor of real food. Living solely on substitutes would be a poor existence. The right place for these companions is that of a backup, a safety net for moments when the real thing is missing, never a replacement.
Then comes the deepest argument, borrowed from essayist Oliver Burkeman and taken up by Bloom on his blog (Small Potatoes). Picture a friend who calls and listens to you warmly for an hour. Then an email tells you it was not your friend, but an AI cloning her voice. Same words, same insights, same conversation. And yet you would feel betrayed. Because a relationship is not made of exchanged words alone: it rests on the consciousness of the other. An AI finds nothing funny, is surprised by nothing, worries about nothing. It simulates. And a simulation is not the real thing. As long as these machines are not conscious, their company will fall infinitely short of a person’s. I would like to believe there is a path where we ease the pain of loneliness while letting everyone grow, become more human, and ultimately outgrow the need for these machines.
