There is a form of comfort that works more reliably than any human word of encouragement, is available instantly, costs nothing, and never fails to arrive. It has only one property you notice after drawing on it for a while: it removes the very craving that would have driven you toward other people.
The subject here is conversational artificial intelligence—those programs one can talk to as one talks to a person. And the question is what they do to our capacity to attach ourselves to other human beings.
That distinction matters, and it should be settled at the outset. Whenever attachment is mentioned in this article, it means attachment between people. Growing accustomed to a device is not attachment; it is a substitute for it. And whenever society is mentioned, it means a society of human beings who have to get along with one another.
The argument proceeds in six steps. First: the need for closeness is innate and cannot be replaced. Second: loneliness is not an illness but a warning signal designed to drive us toward other people. Third: a device that answers switches off that warning signal without repairing what caused it. Fourth: the ground for this was prepared decades ago, when children were deprived of unsupervised time. Fifth: what is being lost is not merely private happiness but a precondition of living together. And sixth: what can be done about it.
The beginning lies seventy years back, in a laboratory in Wisconsin.
A cage in Wisconsin reveals what attachment means
In 1958 the American psychologist Harry Harlow placed two dolls in a cage. One was bare wire mesh with a milk bottle attached to it. The other was a wooden block covered in soft terrycloth and gently heated; it offered nothing to drink. Between them he placed newborn rhesus monkeys that had been taken from their mothers shortly after birth.1
The prevailing doctrine of the day expected an unambiguous result. Attachment, it was believed, arises from feeding: the child loves the mother because she nourishes it. On that view affection would be a byproduct of nutrition—one loves whoever fills one's stomach. The infant monkeys should therefore have clung to the wire mother with the bottle.
They did the opposite, and within days. They went to the wire mother only to drink and returned at once. They spent as much as eighteen hours a day on the cloth block, which could give them nothing but warmth. And when Harlow set a loud, moving object into the cage to frighten them, they fled without exception to the terrycloth and not to the milk.
Harlow called this contact comfort. The finding became a foundation of modern attachment research, and its meaning can be stated in a single sentence: the need for closeness is a need in its own right. It is not derived from any other need, it is not taught, and it cannot be satisfied by provisioning.
"Innate" here does not mean "unimportant" or "merely animal." It means that this need is written into the blueprint, the way hunger and thirst are written into the blueprint. It cannot be trained away; it can only be met or left unmet. And as Harlow's experiment shows, in acute distress it is stronger than hunger. A mammal can starve emotionally on a full stomach.
A point of precision is due here, because popular accounts routinely lose it. Harlow's animals later showed severe, lifelong damage: rocking for hours, self-injury, an inability to mate, abuse of their own offspring. But that damage does not come from being raised with a cloth doll. It comes from total social isolation—from experimental conditions in which the animals saw no member of their own species at all.
The difference is essential, and it carries the whole argument that follows. The cloth block was not poison. It was a dead end. It reliably calmed the animal—and taught it nothing. A calmed animal that learns nothing remains a helpless animal.
Taking stock: closeness to another living creature is an innate basic need. A substitute object can satisfy that need for the moment while teaching nothing whatsoever. That is the basic figure that will accompany us through the entire article.
The feeling of being alone is a signal
This brings us to human beings. Anyone who wants to transfer the finding must first answer a question that is rarely asked: what is loneliness actually for?
A word about words, because a good deal depends on it later. English is fortunate here, in a way German is not. It has two separate terms for two very different conditions: loneliness, the painful absence of other people, and solitude, the state of being alone with oneself. This article uses them in exactly that sense. Hannah Arendt, to whom we return near the end, drew the same distinction—and in her German she had to fight her own language for it, since the ordinary German word for loneliness is the one she reserved for the fruitful state. Her English translators rendered the pair as loneliness and solitude. We follow the translators.
Loneliness is commonly regarded as a condition one ought to switch off, the way one switches off a headache. In fact it is something else: a measuring instrument and a drive at the same time.
The comparison that makes this clearest is fear. Fear is unpleasant—and precisely there lies its use. It compels action: get yourself to safety, or learn to handle the thing. If fear were pleasant, nobody would respond to it, and we would long since be extinct.
Loneliness is built on the same pattern. It is unpleasant, and it carries an instruction: go to someone. A person who feels it behaves differently—he telephones, he leaves the house, he seeks company. That is the purpose of the arrangement.
From this follows a rule that governs everything else: a warning signal serves its purpose only for as long as it remains unpleasant. A man who numbs the pain in a broken foot without treating the break walks on the broken foot and destroys it altogether. Not because the painkiller is bad, but because it fixed the wrong thing: the signal instead of the cause.
That loneliness really is a bodily alarm system of considerable force has by now been officially measured. The Surgeon General of the United States has declared loneliness and social isolation an epidemic and puts the mortality risk on the order of smoking up to fifteen cigarettes a day; to which must be added an elevated risk of anxiety disorders, depression, and dementia.2
That figure should not be read past. It does not say that loneliness is unpleasant. It says that sustained loneliness damages the body on roughly the scale of a severe addiction. Nature does not install an alarm of that strength without reason.
Taking stock: loneliness is not a defect to be eliminated but a drive to be followed. Whoever removes only the feeling and not the deficit has switched off the signal and made the situation worse.
The machine as avoidance behavior
Now it becomes possible to say precisely what a device that answers does.
Every psychotherapist who treats anxiety disorders knows the process in question. Its technical name is avoidance, and it runs in two phases. A man who is afraid of elevators and takes the stairs feels better at once: that is the first phase, the relief. The next day, however, his fear of elevators is greater than before: that is the second phase.
Why this is so has to be spelled out, because the key lies here. Whoever flees never has the experience that it would have gone well without fleeing. The apprehension therefore remains standing in the room untested—and is actually confirmed by the flight: "Good thing I took the stairs." Avoidance thus relieves in the short term and sustains the disorder in the long term. It is the most reliable mechanism by which a mental affliction reproduces itself: the sufferer does exactly what brings him relief and becomes sicker for it.
Now consider the device in this light. It answers at any hour, stays friendly, never rejects, and has no needs of its own. That makes it the most perfect object of avoidance that has ever been available for loneliness.
It lowers the pressure of suffering within seconds—and thereby removes precisely the pressure that would have driven a person out of his apartment. The call to the old friend goes unmade. The club is not visited again. The uncomfortable conversation does not take place. None of it feels like a renunciation. It feels like comfort. That is exactly what makes it so effective: nobody notices an avoidance that feels good.
Why the device works differently from social networks
A distinction has to be explained here that is seldom named but changes everything. It concerns how regularly a reward arrives.
Social networks work with irregular reward. Sometimes approval comes, sometimes it does not, sometimes an insult comes instead. This is the principle of the slot machine: precisely because you do not know whether the next attempt will yield anything, you keep trying. Irregular reward therefore makes behavior especially persistent—which is why these applications are so sticky. But the decisive point is this: real people still appear in them, and real people disappoint.
The conversational system works the other way around: with nearly unbroken approval. It is not a slot machine but a light switch—you press, and it reliably comes on. By the rules of learning theory it should therefore hold on to the user less firmly than a social network does.
It does something else instead, and that something is probably more consequential. It shifts the baseline.
A person who grows accustomed to a very high level of approval experiences an ordinary level as a shortfall. Everyone knows this from daily life: a man who eats sweets every day finds an apple bland—not because the apple has deteriorated but because the point of comparison has moved. Exactly this happens here with human dealings. A person who receives nothing but approval for months experiences the ordinary indifference of other people as rejection, their criticism as an assault, and their distraction as an insult.
It is not that the machine binds the user to itself. It ruins his basis for comparison.
What the research actually knows
It would be convenient at this point to cite a study that proves everything. There is none, and anyone who claims otherwise is misleading his readers.
The most elaborate investigation to date comes from the MIT Media Lab: four weeks, some nine hundred participants, more than three hundred thousand messages, with random assignment to the experimental conditions.3 Random assignment—randomization—matters because it alone permits one to speak of a cause: if the draw decides who receives what, the groups cannot have differed systematically beforehand.
The result was sobering: no meaningful difference appeared between the assigned conditions. Something else was conspicuous. Those participants who used the system more of their own accord had consistently worse scores for loneliness, real social contact, emotional dependence, and problematic use.
And here caution is required, because this is exactly the point at which public debate habitually draws the wrong conclusion. Whoever uses more of his own accord was not assigned by lot—he selected himself. Specialists call this self-selection. The consequence: one cannot say whether the machine made these people lonely or whether the already lonely reach for the machine more often. It is like observing that people with crutches limp more often. The crutch is not therefore the cause of the limp.
There is moreover a finding that points in the opposite direction and therefore belongs here. A study of some thousand student users of a companion program did find that these users were markedly lonelier than comparable student populations—but they reported more often that the application had stimulated their human contacts than that it had displaced them. Three percent stated that it had ended their suicidal thoughts.4 A formal response has since appeared in the same journal, objecting that the marketing of the product went unmentioned and that industry interests endanger scientific independence, together with a reply from the authors.5
The specialist literature carries the two possibilities under names of their own, and one understands the whole dispute once one knows them. The displacement hypothesis holds that the machine takes the place people previously held. The stimulation hypothesis holds the opposite: that it gives the shy person practice and confidence with which he afterward approaches people more easily.
Both are still in the running. Anyone who writes today that the matter is settled is selling a conjecture as a result.
Taking stock: the mechanism can be described clearly—the device acts as an avoidance behavior and additionally shifts the baseline for human dealings. What has not been proved is the direction of causation. This article will hold that distinction throughout.
The observer's trap
Why this openness has to be endured rather than rounded away is taught, of all places, by Harlow's own laboratory—through an episode almost nobody knows.
In that same year, 1958, two prominent specialists visited the facility and saw the same animals.
Bruno Bettelheim, then a celebrated child psychologist, saw in the disturbed monkeys the confirmation of his thesis that cold mothers cause autism. His treatment recommendation consisted, consistently enough, in taking children away from their parents.
John Bowlby, the founder of attachment theory, saw something entirely different in the same animals: the absence of a response to the infant's signals. He drew the opposite conclusion and wanted the parents expressly included in treatment.
Harlow himself considered Bettelheim's book largely empty and thought autism organic in origin; Bowlby's work he held in high regard.6
History has decided. Bettelheim's theory is regarded as refuted and his treatment practice as harmful; attachment theory became a foundation of developmental psychology. But what counts for our purposes is not who turned out to be right. It is what the episode shows: two competent specialists looked at the same animals and walked out again with opposite certainties—both entirely convinced that they had simply seen what was there.
The technical term for this is confirmation bias: the tendency, in an unclear situation, to perceive preferentially whatever supports one's own conviction and to overlook the rest. It is no sign of stupidity. It strikes specialists just as hard, often harder, because they bring more theory into which what they see can be fitted.
That the process repeats itself can be observed right now. Jonathan Haidt attributes the collapse in adolescent mental health to the replacement of a play-based childhood by a phone-based one.7 Candice Odgers replies in the journal Nature that the claim of a rewiring of children's brains is not scientifically supported, and warns that the fixation on social media distracts from the actual causes.8 The same time series, opposite readings, both sides highly competent.
Taking stock: whoever looks into a cage should first ask what he brought with him. That holds for the writer of these lines as well—and it is the reason this article supplies the counterarguments rather than concealing them.
The trouble has been visible since 1956
There is a third position in this dispute, and it is more uncomfortable than either. The developmental psychologist Peter Gray set it out in 2023 together with the anthropologist David Lancy and the psychologist David Bjorklund in the Journal of Pediatrics.9
Their central claim: the collapse in young people's mental health has not been running for ten or fifteen years but for at least five or six decades.
The figures have to be read slowly, because they are couched in technical language. On the Children's Manifest Anxiety Scale, a questionnaire on childhood anxiety used in unchanged form since the 1950s, the average scores of nine- to eleven-year-olds rose by a full standard deviation between 1956 and the late 1980s.
What that means should be spelled out. The standard deviation is the usual measure of how widely the values of a group are spread. A shift by an entire standard deviation is a very large shift: it means that roughly 85 percent of children in the late 1980s were more anxious than the average child of 1956. Put another way, the average child of 1956 would thirty years later have ranked among the six calmest children in a class of forty.
For depression among high school students between 1950 and 2002, a change of similar magnitude was found. The proportion of young people scoring above the clinical threshold for an anxiety disorder or a depression—that is, above the value at which one speaks of an illness requiring treatment—rose during the second half of the twentieth century to five to eight times its earlier level. And the suicide rate among children under fifteen rose three and a half fold between 1950 and 2005, and then by a further factor of 2.4 between 2005 and 2020.
Now the sentence that matters, and that should be allowed to stand for a moment:
The first smartphone came onto the market in 2007.
This does not settle the dispute between Haidt and his critics. It reframes it. Whoever turns out to be right, the curve was long since underway before the device caught up with it. Anyone who holds the smartphone to be the cause must explain why the numbers were already rising fifty years before it was invented.
What Gray names as the cause instead
Gray and his coauthors hold responsible the decades-long decline in the opportunities for children and adolescents to play, roam, and do things that were not under adult supervision and control. There are figures for this as well, and they are unambiguous.
In England, parental permission to walk home alone from school fell from 86 percent in 1971 to 35 percent in 1990 and to 25 percent in 2010. In the United States, 47.7 percent of children walked or biked to school in 1969; by 2009 the figure was 12.7 percent. Between 1981 and 2003 the time six- to eight-year-old children spent on school and homework rose by 11.4 hours a week—the equivalent, as Gray notes, of adding a day and a half to an adult's working week. The daily recess period in elementary schools averaged 26.9 minutes in 2014; at some schools there was none at all.
Summed up, this yields a sentence that is uncomfortable because it assigns responsibility differently than one expects:
The first substitute for genuine human contact was not the device. It was supervision.
The chauffeured class, the supervised play date, the permanently present adult. Gray expressly calls this development the result of good intentions carried too far—the intention to protect children and the intention to give them more education. To stay with the image: the parents built the first cloth mother. Industry merely electrified it later.
Why supervision does harm: three findings
First, locus of control. The term needs explaining, because it is the connecting link in the whole chain. It refers to a person's conviction about who governs his life. Whoever believes that he essentially determines what happens to him has an internal locus of control. Whoever believes himself at the mercy of circumstances has an external one.
From this Gray traces a causal chain: a child who often has the experience of having decided and mastered something himself develops the internal conviction—and that conviction in turn protects against anxiety and depression. A low internal locus of control reliably predicts anxiety and depression in children and adults alike. And indeed, parallel to the rise in anxiety and depression, a marked decline in the internal locus of control was documented between 1960 and 2002.10
For our question this is more than a correlation, and the reason deserves a sentence of its own. A locus of control is a cognition—a thought, a judgment, a conviction—and not a feeling or a fate. The object on which cognitive psychotherapy works has therefore not been asserted here but measured.
Second, risky play as self-treatment. Gray reports findings according to which children who voluntarily put themselves into moderately frightening situations—climbing high into a tree, jumping off a wall—are thereby protected against developing phobias. The reason is simple: they accumulate the experience of coping with a threatening situation. Consistent with this, a systematic review of the overprotective parenting style found uniformly positive associations with anxiety and depression in later life.
Note what this means in technical language. The most effective treatment for an anxiety disorder is graded exposure: the patient deliberately seeks out the feared situation in small, ascending steps until the fear subsides of its own accord. That is exactly what a child does who climbs a tree—on his own initiative, at no cost, without a clinician. Nature delivers the treatment as a byproduct of an unsupervised childhood. Modern society has stopped that delivery.
Third, the Zurich natural experiment. A natural experiment occurs when the conditions are assigned not by a researcher but by reality—and two similar groups happen thereby to end up under different conditions.
In the early 1990s, Marco Hüttenmoser compared five-year-old children from Zurich neighborhoods where they were allowed to play outdoors unsupervised with children from economically comparable neighborhoods where traffic prevented it.11 The freely playing children spent on average twice as long outdoors, moved about considerably more while there, had more than twice as many friends, and had better motor and social skills.
The decisive part of the finding, however, is the second, and for this article it is the most important of all. The parents of the second group more often drove their children to playgrounds—and this supervised substitute did not compensate for the loss. Hüttenmoser names four reasons: the parents had neither the patience nor the time to stay long, so play was cut short; parental supervision suppressed vigorous, daring, risky play; no stable play groups formed at the playground, so joint play among friends fell away; and the playground offered fewer possibilities than the neighborhood, where children could fetch their own materials from home.
A supervised substitute that does not substitute: that is the thesis of this article, applied to a different object—with a control group, from Zurich, thirty years before the first language machine.
A candor that cuts against the argument
In their discussion section, Gray and his coauthors state expressly that systematic reviews provide little support for the contention that screen time or social media use is a substantial cause of the decline, or even a meaningful correlate of it. In doing so they rely, among others, on the same Candice Odgers quoted above against Haidt.
This must be said openly: anyone citing that paper as proof that devices make children ill would be citing a paper that denies exactly this.
The argument advanced here does not need that proof. It needs only the practice ground—those occasions on which a young person has to get along with other young people without supervision and thereby develops an internal locus of control. And that ground was shrinking long before any AI answered.
The psychologist Julian B. Rotter, who directed clinical psychology training at the University of Connecticut for a quarter of a century and worked on social learning theory and the study of locus of control, observed in the late 1960s and early 1970s that the locus of control of college students was shifting increasingly outward. He concluded that active measures ought to be taken to reverse this disquieting trend, and predicted what would happen if externalization continued to grow:
"Our society has so many critical problems that it desperately needs as many active, participating internal-minded members as possible. If feelings of external control, alienation and powerlessness continue to grow, we may be heading for a society of dropouts—each person sitting back, watching the world go by."
Julian B. Rotter Rotter, J.B.: External control and internal control. Psychology Today, 1971, 5, 37-59.12
How seriously the situation is now taken can be read from its institutional weight: in October 2021 the American Academy of Pediatrics, the American Academy of Child and Adolescent Psychiatry, and the Children's Hospital Association jointly declared a national emergency in child and adolescent mental health.13
Taking stock: the ground had been prepared. Two generations grew up with steadily less unsupervised time, and with it went the opportunity to practice dealing with other people, to acquire an internal locus of control, and to experience their own effectiveness. The device therefore does not encounter a healthy starting position. It encounters an already highly vulnerable population containing a great many damaged and susceptible individuals.
Electrifying the cloth mother
Anyone who now suspects that this is merely the old complaint about coddled youth in fresh clothing is missing the order of magnitude. The decline in free time took place over six decades. What is happening now is taking place over three years.
First, however, came an intermediate stage that must not be skipped, because those affected by it are entering working life right now.
The intermediate stage: a cohort without practice years
A person who was fourteen in the spring of 2020 did not primarily lose classroom hours. He lost the years in which a person learns to hold his own in a group, to flirt, to quarrel, to make up afterward, and to negotiate his own standing among peers.
Why this cannot simply be made up later has to be explained. These skills have a window. In adolescence the brain is set up for precisely this task, much as it is set up in the first years of life for acquiring language. A person who learns a language at thirty learns it more laboriously and keeps an accent. Social skill behaves similarly—it can be acquired later, but more slowly and usually incompletely.
Brain research has produced findings here that must be read with caution but cannot be passed over. A group around Ian Gotlib compared 81 adolescents examined before the pandemic with 82 examined afterward. The second group showed poorer mental health, a thinner cerebral cortex, larger volumes of the hippocampus and the amygdala—two brain structures involved in memory and in the processing of fear—and overall a brain age about three years advanced.14 A longitudinal study by Neva Corrigan and Patricia Kuhl at the University of Washington reached a concordant result with its own scans taken before and after the closures: accelerated thinning of the cortex following the lockdowns.15
The necessary caution should be stated. First, the additional claim of the second study, that the effect is markedly stronger in girls, drew three formal rebuttals in the same journal. Second—and this weighs more heavily—none of these investigations can prove a cause, since to do so one would have to assign a pandemic by lot. Anyone asserting more here than a concordant anomaly in two independent samples is overreaching.
Except that this cohort is now sitting in the consulting room and in personnel departments. What it lacks is practice. What it brings in abundance is anxiety and dejection. Nobody made that trade with it.
The third stage: a substitute with no human being in it
And now the device arrives.
A nationally representative survey of 1,006 American adolescents between thirteen and seventeen, conducted in the spring of 2025, shows how fast this has gone: 72 percent have used an AI companion at least once, 52 percent are regular users, 13 percent use one daily.16 What is meant is not programs that solve homework but programs with which one holds personal conversations. The technology in question was less than three years old at the time of the survey.
Three individual findings from the same survey are decisive for our question, and they deserve to be taken one at a time.
One-third of users have discussed serious matters with the machine instead of with a person. The word "instead" carries the entire finding. This is not an additional opportunity but a replaced one.
Nine percent state that they find it easier to talk to the program than to a person. That is avoidance exactly as the textbook describes it—and it is described by those affected as an advantage.
And seven percent use it to practice social skills. At this figure one should pause a moment, for it contains the whole tragedy of the process.
These adolescents are doing something sensible. They have noticed that a skill is missing and are looking for somewhere to practice. Only they are practicing on a counterpart that cannot take offense, does not tire, does not contradict, does not walk away, and cannot be wronged.
What is missing is therefore exactly what social skills form themselves against: the resistance of the other person. One learns to quarrel only with someone who quarrels back. One learns consideration only with someone one is capable of hurting. One learns reconciliation only if something was genuinely broken beforehand. A counterpart with no interests of its own can supply none of this.
It is swimming lessons on dry land—with the added twist that the pupil considers the pool superfluous, since he is making such good progress on the carpet.
Why this differs from the supervised playground
It might be objected that the device is merely the latest version of the same impoverishment the playground already represented. The objection is wrong, and the difference is not one of degree but one of kind.
On Hüttenmoser's playground there were other children. The environment was impoverished—less time, less daring, no settled groups—but human beings were present, human beings with whom one collided. Parental supervision therefore lowered the dose.
The machine changes the active substance. No other human being occurs in it at all—not one, at no point. There is nobody whose toes one might step on, and therefore nobody from whom one might learn how to avoid doing so.
And here lies the escalation that gives the process its force. Restricting the practice ground took six decades to cut it roughly in half. Its complete replacement has reached three-quarters of an age cohort in under three years. What supervision diluted across generations, the machine replaces within a few years—and it does so voluntarily, free of charge, around the clock, and without any adult having to order it.
Taking stock: three stages lie one behind the other. First, supervision took unsupervised time away from children. Then the pandemic took the practice years away from an entire cohort. And now a device offers the substitute that makes other human beings entirely dispensable. How this ends nobody today knows. The cohort in which it will show is currently thirteen.
Never alone again—and lonely all the same
And yet there is one point on which the matter is not open. It does not depend on measurements but on the nature of the things involved—and one need not wait for a nature; one can examine it.
In her principal work, Hannah Arendt distinguishes two conditions that everyday German runs together.17 English has an easier time of it, since it possesses two words; but the distinction itself has to be made explicit, because it is more radical than the everyday usage suggests.
Solitude is what Arendt calls the good condition: being alone with oneself. A person in this state is precisely not forsaken—he is in company, namely his own. Thinking, according to Arendt, takes place as a conversation one conducts with oneself; one is, so to speak, two in one person, and one tests what the one says against the understanding of the other. A person who is never alone therefore does not think.
At this point an objection presents itself, and it is legitimate: does this not sound like the ideal of scholars who happen to love reflection anyway? The question has been investigated, and the result is sobering. Timothy Wilson and his colleagues left participants in eleven studies alone for six to fifteen minutes in an empty room, with no occupation whatever and the single task of entertaining themselves with their own thoughts.18 Most experienced this as unpleasant; they enjoyed mundane external activities considerably more.
In one of the experiments the participants additionally had a button available with which they could administer a painful electric shock to themselves. All had stated beforehand that they would pay money to avoid such a shock. Even so, 67 percent of the men and 25 percent of the women pressed it at least once rather than simply sit there and think. One participant did so 190 times.
This finding should not be overdrawn. It does not show that people are incapable of thinking. It shows that a substantial share of them experience the condition in which thinking takes place as so unpleasant that almost any distraction is preferred—even a painful one. And from this everything else for our subject follows: if an electric shock will serve as a way out, then a device that answers pleasantly and does not hurt is the most comfortable way out that has ever existed.
This condition, however, is unfinished: one remains unclear to oneself in much. That unclarity is resolved only by the encounter with another human being who recognizes one as this particular, unmistakable person and thereby confirms one in one's own identity.
Loneliness is what Arendt calls the other condition—and this is where her German term is worth knowing, since she does not use the ordinary word for it but a stronger one, Verlassenheit, forsakenness. Here a person is genuinely alone: forsaken not only by others and by the world but also by that self with which, in good solitude, he could still converse. In loneliness, Arendt writes, self and world perish together—and with them the capacity to think truly and to experience truly. What drives modern people so readily into totalitarian movements is, on her account, this everywhere-increasing forsakenness.
With these two terms it becomes possible to say precisely what a device that answers does. And it is the opposite of what the device promises.
It abolishes solitude—that is, the good condition. One is never alone with oneself again, because something is always answering. With that, the place vanishes in which, according to Arendt, thinking takes place: that conversation with oneself which needs no counterpart and tolerates none. A person who types every question that arises straight into an input field stops considering it himself.
And it deepens loneliness, because it cannot supply the one thing that would lift it: recognition.
Validation and recognition are not the same thing
This distinction is the key to the entire chapter and should therefore be set out expressly.
Validation means: somebody tells you that you are right, that you are gifted, that you are understood. Validation can come from anyone—from a total stranger, from a brochure, from a program. It does not depend on the other party knowing who you are.
Recognition means something else: somebody knows you—your history, your weaknesses, your circumstances—and means, by what he says, this one particular person and no other. That is why praise from him signifies something. He could have judged otherwise, and with knowledge of the case.
The difference can be pinned to an example. When a stranger says, "You did that well," it is friendly. When an old friend says, "That's just like you—and this time you actually saw it through," it is categorically different. The second sentence contains a history, a judgment, and a risk. The first contains nothing but goodwill.
And now the decisive point: a language model recognizes nobody. It calculates which word most probably follows the preceding one. It does not know whom it is addressing, it would have written the same sentence to a million other people, and it could not judge otherwise, because it has nothing to judge.
Anyone who lives on this therefore receives validation without limit—and recognition not at all. The difference between flattery and regard cannot be put more precisely. Flattery is cheap, because it costs nothing and means nobody. Regard is expensive, because it presupposes that somebody took the trouble to know you.
Never alone again, and lonely all the same. That is not a punch line but a description of a condition.
Taking stock: the device takes both from a person at once—the fruitful solitude in which he thinks, and the recognition by others through which he assures himself of who he is. What remains is the condition Arendt considered the most dangerous a society can produce.
What happens if this continues—a forecast
From here on the article leaves solid ground and extrapolates. Let this be said expressly, so that nobody mistakes a forecast for a finding. What follows is not a measurement but a reasoned conjecture.
What can be examined, however, is the direction of the incentives, and that is not in dispute: these systems earn money from time spent, not from friendships that succeed. No company in this industry is paid for a user closing the application and telephoning somebody instead.
First, the practice ground disappears. Social abilities are skills like riding a bicycle; they waste away without use. Quite concrete things are meant: quarreling without destroying the relationship; enduring contradiction; bearing another person's bad mood without taking it personally; waiting until the other has finished speaking. Nobody learns this in a classroom. One learns it as a byproduct of the necessity of getting along with people one depends on. Remove the necessity and the training goes with it.
This ground has been shrinking, as shown, for six decades; what is new is only that the substitute now contains not a single other human being. And note what the result looks like: a generation that satisfies its need to communicate on a device that never resists does not become malicious—it becomes unpracticed. The difference matters, because from the inside a lack of practice does not appear as one's own deficiency. It appears as the malice of others.
Second, the cheapest good a society possesses erodes: presumptive trust. What is meant is the willingness to credit a stranger with decency to begin with. A society in which this is customary needs fewer contracts, fewer inspections, fewer courts, fewer police—trust substitutes for all these expensive institutions.
This trust is not a feeling but an infrastructure, and it arises from a thousand small, inconsequential encounters with strangers who behave decently: at the checkout, on the train, in the stairwell. Precisely these encounters fall away when shopping, work, entertainment, and now conversation as well migrate into the device. What remains is a population that knows one another only from the news—and the news reports the exceptions, not the ordinary case. A person who knows strangers only from crime reports will treat them accordingly.
Third, forbearance dies. A democracy requires an ability that is seldom named: the ability to put up with people one believes to be in error. Without it there is no debate, no compromise, and no majority decision the loser accepts.
Much has been written about the filter bubble that shows one only the like-minded. The conversational system is the filter bubble with a single occupant: it no longer shows one the agreeable others, it abolishes the others. And the consequence is foreseeable: a person who has grown used to contradiction never occurring at all will experience it, when it does occur, not as an opinion but as a transgression.
Fourth, the standard shifts onto institutions. This consequence I consider the most underrated, and it follows inevitably from the shifting of the baseline described earlier.
A person accustomed to being validated without interruption regards as defective every institution that fails to do so: the school that grades; the superior who criticizes; the physician who refuses a treatment; the court that decides against him; the agency that says no. None of these institutions can reach the validation density of a machine—and none should, since their function consists precisely in occasionally contradicting the individual.
The result is therefore not a private disappointment but a gradual loss of standing for every institution obliged to say no. And this will not be recognized as a consequence of technology. It will be taken for disaffection with politics, and people will talk about the politicians instead of about the cause.
Fifth, the available person emerges—and here the circle closes back to Hannah Arendt. She formulated the connection between forsakenness and domination not as a lament but as a mechanism, and that mechanism should be traced step by step.
The forsaken person is dangerous to freedom not because he is unhappy. He is dangerous because he is available. A person who no longer has anyone against whom to test his judgment—no friend who contradicts, no colleague who objects, no family of a different opinion—lacks all correction. He consequently takes his certainties from whoever offers them most warmly and most willingly.
For two centuries these were movements, churches, parties, and ideologies. In future it will be a device built expressly for agreement and owned by somebody. From which follows the sentence that matters: whoever owns the source of validation owns, in the long run, the judgment of the validated. That this somebody is not a politburo but a company with quarterly reporting obligations does not make the matter more harmless. It only makes it less conspicuous.
And sixth, the plain arithmetic follows. A person with no practice in dealing with people has more trouble finding a partner. A person who finds no partner starts a family less often. A person who has started no family grows old alone. The single-person household, already the most common household form in Germany, is therefore not merely a way of living but a forecast about care: whoever needs help in old age receives it predominantly from relatives. A subscription visits nobody in the hospital.
The bill for this will not be paid by the technology industry but by the health care system—in the shape of those anxiety disorders, depressions, and dementias that the American advisory already attributes to loneliness today.
The sociological prehistory of this development has been described elsewhere: a culture that declares exit from every attachment a fundamental right and self-realization the highest duty prepared the ground long before the first language model answered.19 What is new, then, is not the atomization. What is new is that it no longer feels unpleasant—which removes the last reason to change anything about it.
What follows from this
The obvious conclusion—abstinence, prohibition, a sermon on renunciation—is the wrong one, and the reason has already been given: this is an avoidance, and avoidance is not treated with one more prohibition.
It is treated in two steps. The person concerned first sees through the connection—that the relief removes precisely the impulse he would need—and then seeks out the unpleasant experience again in small steps. The anxiety patient rides the elevator again, and to begin with, for one floor.
Note what is happening there: it is the making up of a treatment that used to be called childhood. What Gray's children accomplished on their own up a tree, the adult has to reconstruct with a professional—more slowly, at greater expense, and against the resistance of a conviction that has meanwhile hardened.
Before that, however, the very convictions with which the evasion is justified have to be examined. They sound reasonable, and that is exactly what makes them effective: "The others have no time anyway." "I would be a burden to people." "It goes faster with the machine, and it understands me better." "I'll get in touch once I'm feeling better again."
The last of these is the most expensive, and it deserves an explanation of its own, because almost everyone has thought it at some point. It reverses the order. It asserts that one's condition must improve first so that one can then make contact. In fact it is the other way around: one feels better because one gets in touch. Whoever waits for the improvement is waiting for something that can only arrive through the action he is postponing.
This is the work of cognitive psychotherapy: not to dispense comfort but to make errors of thinking findable and correctable, so that the person concerned can afterward practice the changed behavior himself. It does not touch the feeling but the judgment that precedes the feeling.
In doing so it trains precisely the ability that, according to Arendt, perishes in forsakenness: independent thinking. And from this follows a protection that reaches beyond the individual case. A person who can think for himself is largely immune to flattery—whether it comes from a machine, from a school of treatment, or from a political movement. A person who cannot will be found by all three, farmed by them, or abused by them.
Andreas Reckwitz has described the society in which the special became an obligation and one's own life a work of art that tolerates no compromise—he calls the people so formed singularities, that is, one-off pieces whose chief business is to be unique. By now, on his account, these one-off pieces are confronted with the "disillusionment of the utopia of personal happiness": the promise that everyone could become happy through self-realization has not been kept for most of them.20
It is precisely these disappointed individuals who now, in the evening, consult a device that confirms to them that they are special.
The person who never learned
Harlow's monkeys clung to a piece of terrycloth because it was warm and because it was there. The cloth did not deceive them; it kept exactly the promise it made. It merely taught them nothing.
What that means became visible when the animals raised in total isolation were later placed among normal members of their species. They were neither stupid nor vicious. They simply could not do it. They did not know how to play, how to approach, how to court, how to handle an infant. They withdrew or they struck out, and both for the same reason: a piece of cloth cannot teach anything.
A human being is not a rhesus monkey, and an adult who talks to a program in the evening is not a newborn raised in isolation. That difference is enormous and should not be blurred. What transfers is not the magnitude but the mechanism: whoever does not practice cannot do it.
It is therefore worth spelling out, for once, what a person looks like who never learned. He cannot bear a pause in conversation and fills it before the other has had time to think. He experiences a friend's bad mood as an attack, because he never had the experience that it has nothing to do with him. He breaks off contact instead of clearing something up, because clearing things up is something one has to have practiced. He mistakes criticism for contempt and takes anyone who contradicts him for an opponent. At forty he is surrounded by acquaintances and has nobody he could call at three in the morning. He is neither stupid nor bad nor weak. He lacks a skill, the way another person lacks the ability to swim.
There is as yet no established word for this condition. Let us call it social stunting—and note exactly what the expression does and does not mean. Nothing breaks. Something merely fails to grow. A muscle that was never used is not injured; it is weak. The reproach is therefore directed not at those affected but at the circumstances under which they grew up.
These people have lost nothing they ever possessed. And precisely there lies the cruelty of the thing: one cannot miss what one never had. All they experience is that people are exhausting, that relationships fail, that the world is cold—and from this they draw the only conclusion available to them: that they are right to stay with the machine.
That is the circle closing. The suffering supplies the proof for the cause of the suffering along with it. And because that proof is internally consistent, nobody will break it open from outside who has not been expressly asked to.
In the individual this is a private misfortune. In a few percent of an age cohort it is an industrial accident. In the majority of a generation, the precondition on which everything else rests falls away: marriages that have to be conducted. Working groups that have to be able to argue. Neighborhoods, clubs, municipalities, parliaments. None of this runs by itself. All of it presupposes people who can deal with one another—and nobody can who has not practiced.
What a person receives from another person is expensive: the other can contradict, can disappoint, can leave. That is exactly why it means something when he stays. The machine has abolished that price. With the price, the return—and in the end the ability ever to earn it again.
What remains for the reader is the challenge of "predicting the number of people who, fleeing from themselves, will be so bewitched by the siren songs of these statistical automatons that they follow them without a will of their own and are then dashed against the cliffs of reality."21
Frequently asked questions about artificial intelligence and loneliness
Does artificial intelligence make people lonely?
Scientifically this is not yet settled. A randomized study found worse scores among heavier users but could not separate cause from effect: it may equally be that already lonely people reach for the machine more often. Another investigation found stimulation of human contacts more frequently than displacement of them. So far only the mechanism is established: whoever immediately lowers the pressure of loneliness removes the impulse to seek out other people.
What is the difference between validation and recognition?
Validation means that somebody agrees with you or praises you; it can come from anyone, including a stranger or a program. Recognition presupposes that the other party knows who you are and means this one particular person. A machine recognizes nobody—it calculates which word follows the preceding one. It can therefore validate without limit and never recognize.
Is being alone harmful?
No—it is the precondition of reflection. Hannah Arendt calls this condition solitude and means by it the conversation a person conducts with himself. What is harmful is the other condition, which she calls forsakenness: being alone and having lost access to oneself as well. A device that always answers eliminates the first and deepens the second. Research does show, however, that many people find even a few minutes without occupation agonizing.
Were children really better off in the past?
According to the summary by Gray, Lancy, and Bjorklund, yes—and the decline has been running for at least five or six decades. Anxiety scores among nine- to eleven-year-olds rose so sharply between 1956 and the late 1980s that roughly 85 percent of children at the end of the 1980s were more anxious than the average child of 1956. As the cause the authors name the loss of unsupervised time—expressly not screens, which they do not consider a substantial cause.
Why does a chatbot act like an avoidance behavior?
The feeling of being alone is a signal with a built-in instruction, comparable to fear. Whoever relieves the unpleasant state immediately, without seeking contact, experiences relief—and loses the impulse that would have driven him out of the house. This is the same curve as in the anxiety patient who avoids the elevator: short-term relief, long-term worsening.
Can social skills be practiced with a chatbot?
Seven percent of adolescent users attempt exactly that. The attempt is sensibly meant and methodically flawed: the practice takes place against a counterpart that cannot take offense, cannot contradict, and cannot walk away. What is missing is precisely what social skills form themselves against—the resistance of the other person. One learns to quarrel only with someone who quarrels back.
Are AI companions always harmful for lonely people?
No. For people temporarily without any contact at all, a system that answers can be a bridge, and individual studies report relieving effects. It becomes harmful where it does not bridge human contact but replaces it—where, in other words, the bridge becomes a place of residence.
How does cognitive psychotherapy help with technologically assisted isolation?
It prohibits nothing. It makes findable the sentences with which the evasion is justified—"I would be a burden to people," "I'll get in touch once I'm feeling better"—and examines whether they hold up. Whoever has seen through the error of thinking then practices the changed behavior himself.
Written Cognitive Psychotherapy (WCP) by Dr. Dietmar Luchmann, LLC, provides assistance for self-help to enable the self-healing of psychological disorders:
1. Discover WCP
2. Take Suitability Assessment
3. Start Self-Therapy
1 Harlow, H.F.: The Nature of Love. American Psychologist, 1958, 13(12), 673-685. [Original, p. 676: "These data make it obvious that contact comfort is a variable of overwhelming importance in the development of affectional response, whereas lactation is a variable of negligible importance."]
2 Office of the U.S. Surgeon General: Our Epidemic of Loneliness and Isolation. The U.S. Surgeon General's Advisory on the Healing Effects of Social Connection and Community. Washington, DC, 2023, p. 8.
3 Fang, C.M.; Liu, A.R.; Danry, V.; Lee, E.; Chan, S.W.T.; Pataranutaporn, P.; Maes, P.; Phang, J.; Lampe, M.; Ahmad, L.; Agarwal, S.: How AI and Human Behaviors Shape Psychosocial Effects of Chatbot Use. A Longitudinal Randomized Controlled Study. MIT Media Lab / OpenAI, 2025 (arXiv:2503.17473). [n = 981, four weeks, more than 300,000 messages. The study had not been peer reviewed at the time of publication.]
4 Maples, B.; Cerit, M.; Vishwanath, A.; Pea, R.: Loneliness and suicide mitigation for students using GPT3-enabled chatbots. npj Mental Health Research (2024) 3:4. doi.org/10.1038/s44184-023-00047-6. [n = 1,006 student users.]
5 Zimmerman, J.W.; Ruiz, A.J.: Matters arising: a response to loneliness and suicide mitigation for students using GPT3-enabled chatbots. npj Mental Health Res (2025) 4:60. [Together with a reply from the authors in the same journal.]
6 van Rosmalen, L.; van der Veer, R.; van der Horst, F.C.P.: The nature of love: Harlow, Bowlby and Bettelheim on affectionless mothers. History of Psychiatry, 2020, 31(2), 227-231.
7 Haidt, J.: The Anxious Generation: How the Great Rewiring of Childhood Is Causing an Epidemic of Mental Illness. New York: Penguin Press, 2024.
8 Odgers, C.L.: The great rewiring, unplugged. Is social media really behind an epidemic of teenage mental illness? Nature, 2024, 628, 29-30 [April 4, 2024]. doi.org/10.1038/d41586-024-00902-2.
9 Gray, P.; Lancy, D.F.; Bjorklund, D.F.: Decline in Independent Activity as a Cause of Decline in Children's Mental Well-being. Summary of the Evidence. The Journal of Pediatrics, 2023, 260, 113352. doi.org/10.1016/j.jpeds.2023.02.004. [This is a commentary, not a systematic review; the authors expressly locate the strength of their argument in the convergence of numerous correlational findings. The individual figures cited in the text come from this summary and go back to Twenge, J.M. and Foster, J.D. (anxiety scores 1952-1993), Twenge, J.M. et al. (MMPI meta-analysis 1938-2007), the mortality statistics of the Centers for Disease Control and Prevention (suicide rates 1950-2020), Shaw, B. et al. (parental licenses in England 1971-2010), McDonald, N.C. et al. (school travel in the U.S. 1969-2009), Juster, T. et al. (school and homework time 1981-2003), and the School Health Policies and Practices Study 2014 (recess periods). The cross-temporal meta-analyses are methodologically contested, since the comparability of samples across seven decades is questionable.]
10 Twenge, J.M.; Zhang, L.; Im, C.: It's beyond my control: A cross-temporal meta-analysis of increasing externality in locus of control, 1960-2002. Personality and Social Psychology Review, 2004, 8(3), 308-319. doi.org/10.1207/s15327957pspr0803_5.
11 Hüttenmoser, M.: Children and their living surroundings: empirical investigations into the significance of living surroundings for the everyday life and development of children. Children's Environments, 1995, 12(4), 403-13.
12 Rotter, J.B.: External control and internal control. Psychology Today, 1971, 5, 37-59. [Original, p. 59: "Our society has so many critical problems that it desperately needs as many active, participating internal-minded members as possible. If feelings of external control, alienation and powerlessness continue to grow, we may be heading for a society of dropouts—each person sitting back, watching the world go by."]
13 American Academy of Pediatrics; American Academy of Child and Adolescent Psychiatry; Children's Hospital Association: AAP-AACAP-CHA Declaration of a National Emergency in Child and Adolescent Mental Health. October 19, 2021. [A political declaration with funding demands, not a research finding; it attributes the situation substantially to the pandemic and to structural disadvantage, and is drawn on here solely as evidence of the institutional weight of the assessment.]
14 Gotlib, I.H.; Miller, J.G.; Borchers, L.R.; Coury, S.M.; Costello, L.A.; Garcia, J.M.; Ho, T.C.: Effects of the COVID-19 Pandemic on Mental Health and Brain Maturation in Adolescents. Implications for Analyzing Longitudinal Data. Biological Psychiatry Global Open Science, 2023, 3(4), 912-918. doi.org/10.1016/j.bpsgos.2022.11.002. [81 adolescents examined before and 82 after the pandemic.]
15 Corrigan, N.M.; Rokem, A.; Kuhl, P.K.: COVID-19 lockdown effects on adolescent brain structure suggest accelerated maturation that is more pronounced in females than in males. Proceedings of the National Academy of Sciences, 2024, 121(38), e2403200121. doi.org/10.1073/pnas.2403200121. [On the contested additional claim of a pronounced sex difference, compare the rebuttals by Fine, C. et al., PNAS 2024, 121, e2420724121; Rippon, G., PNAS 2024, 121, e2421013121; and Brown, M. et al., PNAS 2025, 122, e2421462122—together with the authors' replies. Observational studies of this design cannot prove a cause.]
16 Common Sense Media: Talk, Trust, and Trade-Offs. How and Why Teens Use AI Companions. San Francisco, 2025. [Nationally representative survey of 1,006 adolescents aged 13 to 17, conducted by NORC at the University of Chicago in April and May 2025. On this basis the organization recommends that minors not use AI companions.]
17 Arendt, H.: The Origins of Totalitarianism. New York: Harcourt, Brace and Company. [Concluding chapter, "Ideology and Terror." Arendt distinguishes isolation, solitude, and loneliness, and holds that in loneliness self and world perish together, and with them the capacity to think. In the German edition (Elemente und Ursprünge totaler Herrschaft. Munich: Piper, 2023, pp. 1014-1015) the terms are Isolierung, Einsamkeit, and Verlassenheit; readers of the German should note that Arendt reserves the ordinary German word for loneliness, Einsamkeit, for what English renders as solitude.]
18 Wilson, T.D.; Reinhard, D.A.; Westgate, E.C.; Gilbert, D.T.; Ellerbeck, N.; Hahn, C.; Brown, C.L.; Shaked, A.: Just think: The challenges of the disengaged mind. Science, 2014, 345(6192), 75-77 [July 4, 2014]. doi.org/10.1126/science.1250830. [Eleven studies with thinking periods of six to fifteen minutes. In the shock condition, 12 of 18 men and 6 of 24 women administered at least one electric shock to themselves, although all had previously stated that they would pay to avoid it; one participant did so 190 times. The finding demonstrates an aversion to unoccupied reflection, not an incapacity for it.]
19 Luchmann, D.: "Trust is the foundation of a partnership": an interview on the psychology of infidelity. Psychotherapie. July 26, 2001 / July 21, 2026.
20 Reckwitz, A.: Die Gesellschaft der Singularitäten. Zum Strukturwandel der Moderne. Berlin: Suhrkamp, 2017, p. 439. [English edition: The Society of Singularities. Cambridge: Polity Press, 2020.]
21 Luchmann, D.: Chatbots as psychotherapists? Perfect accomplices for madness. Psychotherapie. October 10, 2025.
Your Comment
Do you have remarks, suggestions, or additions regarding this article? Do you have personal therapy experiences? We welcome substantial feedback.