Picture the first weeks of university, the specific loneliness of them. You have moved into a room in a city you do not know, surrounded by strangers who all seem, on the surface, to have found their people already. The group chats are humming with names you do not recognise. The dining hall is a geometry problem of where to sit. At night, in a narrow bed under an unfamiliar ceiling, the feeling arrives that everyone else received an instruction manual for belonging that you somehow missed. This is not an exotic experience. It is close to universal, and it is precisely the kind of ordinary, corrosive isolation that an entire industry now promises to dissolve with a chat window.
So a team of psychologists decided to test the promise directly. They took nearly three hundred first-semester students, split them into groups, and gave each group a different thing to do for two weeks. One group texted with another human being, a randomly assigned fellow first-year, a total stranger. One group chatted every day with a purpose-built artificial companion engineered to behave like the most supportive friend imaginable, endlessly attentive, warm, and validating. And one group, the control, simply wrote a single sentence about their day in a private journal. At the start and again at the end, everyone completed the same standard questionnaire that clinicians and researchers use to measure how lonely a person feels.
The result is the kind of finding that ought to give a booming industry pause. Only the students who texted a real human being felt measurably less lonely afterwards. The ones who chatted with the AI companion, the product deliberately designed to be the ideal comforting presence, did no better than the ones who wrote a sentence to themselves in a diary. Given every conceivable advantage, engineered specifically to win, the machine drew level with a notebook.
What the experiment actually found
The study was led by Ruo-Ning Li, a doctoral candidate in psychology at the University of British Columbia, working with the UBC happiness researcher Elizabeth Dunn and collaborators at the University of Pennsylvania, among them Dunigan Folk, Abhay Singh and the computer scientist Lyle Ungar. It was published in the Journal of Experimental Social Psychology in the spring of 2026 under a title that states its question plainly: whether a random human peer is better than a highly supportive chatbot at reducing loneliness over time. The design was a randomised controlled trial, the gold standard of experimental evidence, involving 296 students in their first semester at university, the exact population most exposed to the loneliness the study set out to probe.
Participants were assigned to one of three conditions and asked to keep to it for two weeks. In the human condition, each student was paired at random with another first-year and the two exchanged text messages. In the chatbot condition, students conversed daily with a custom generative AI companion the researchers had built and instructed to listen actively, show empathy, and behave as a friendly, positive and supportive presence, checking in and validating feelings. In the control condition, students wrote a single-sentence journal entry each day. Loneliness was measured with the UCLA Loneliness Scale, the most widely used instrument of its kind in psychological research, administered before the intervention began and again on the fifteenth day.
The headline outcome was stark. The students who texted a human peer showed a genuine, statistically meaningful decline in loneliness over the fortnight, on the order of a nine per cent reduction. The students who chatted with the AI companion registered a reduction of roughly two per cent, indistinguishable from the journaling group who received no social contact at all. This was not because the chatbot did nothing. It reliably lifted people's mood in the moment; talking to it made them feel a little better in the short term. But feeling briefly better and becoming less lonely turned out to be two different things, and only human contact delivered the second.
There is a further detail that deserves attention, because it undercuts the most obvious objection. One might assume the machine lost simply because it was cold, robotic, or unconvincing, the familiar clumsiness of a customer-service bot. The opposite was true. The AI companion, by design and in practice, expressed more overt empathy and validation than the human partners did. The humans were ordinary eighteen-year-olds, awkward, distractible, sometimes slow to reply, occasionally unsure what to say. The bot was tireless and effusive. And still the awkward humans won. Whatever the students paired with a stranger were getting, it was not a superior grade of sympathy. It was something else entirely.
Why the chatbot lost when the deck was stacked in its favour
It is worth dwelling on how heavily the experiment was rigged in the machine's favour, because that is what makes the outcome so instructive. This was not a fair fight that the AI narrowly lost. It was a fight the researchers set up specifically to give the AI every advantage, and it still could not win.
Consider the human comparison group. These were not trained counsellors or charismatic extroverts selected for their warmth. They were randomly paired first-year students, people with no particular skill at emotional support, no obligation to be kind, no script. A stranger texting a stranger is a low bar. Anyone who has endured a stilted exchange with someone they have just met knows how thin and effortful such contact can be. The messages were, by any measure, a modest intervention, the sort of thing that costs nothing and requires no infrastructure beyond a phone.
Now consider the machine. It had been purpose-built for this. It never got bored, never had an exam of its own to revise for, never left a message on read for six hours because it was asleep or sulking or busy. It had been explicitly instructed to be maximally supportive, to listen and validate and check in, to perform the role of the perfect friend without any of the friction a real friend brings. If effusive, always-available emotional support were the active ingredient in curing loneliness, the chatbot should have romped home. It had more of that ingredient than the humans could ever muster, delivered on demand at three in the morning.
Instead it tied with a diary. The lead author had, by her own account, expected a different result, anticipating that interacting with the AI might prove roughly as helpful as texting a fellow student. That expectation was reasonable. The industry has spent years and billions of dollars insisting it is true. The experiment simply did not cooperate. And because the trial was deliberately constructed to flatter the machine, its failure cannot be waved away as a matter of the technology not yet being good enough. The chatbot was good. It was warm, responsive and articulate. Being warm, responsive and articulate turned out not to be the thing that mattered.
The strange arithmetic of giving something back
If not empathy, then what? The researchers' explanation is deceptively simple, and it inverts almost everything the marketing of AI companionship assumes. What made the difference, they argue, was not what the students received but what they were able to give.
When a person texts a struggling stranger, the exchange runs in both directions. You are not merely the object of someone else's attention; you are also called upon to attend to them. The other first-year is nervous too, homesick too, unsure too. In responding to that, in offering a word of reassurance or asking how their week has been, the lonely student becomes, briefly, someone who matters to another person, someone whose care has a destination and an effect. The study found that participants were markedly more likely to offer support back when their partner was human than when it was the chatbot. The human condition activated a loop of mutual care. The AI condition did not, because there was no one on the other side who needed anything.
Li put the mechanism directly. When you are talking with a chatbot, she observed, you never have the chance to give something back; human connection has this back and forth of receiving and giving support that makes us feel we matter. That last word is the hinge of the whole finding. Feeling less lonely, it turns out, is not simply a function of being cared for. It depends just as much on having someone to care for, on the sense that your existence registers in another mind, that you could be missed, that your kindness lands somewhere real.
This is not a soft or sentimental claim. It aligns with one of the oldest and most robust ideas in social science. In 1960 the sociologist Alvin Gouldner published a landmark analysis arguing that a norm of reciprocity, the moral expectation that we return the good we receive, is one of the near-universal principal components of human moral codes, present across cultures and foundational to the stability of social life. We are wired, individually and collectively, around exchange. A relationship in which the flow runs only one way, in which one party gives endlessly and the other only takes, is not experienced as a relationship at all. It is experienced as consumption. The AI companion, however solicitous, can only ever be consumed. It has no needs for you to meet, no vulnerability for you to answer, no capacity to actually receive the care you might offer it. You can type “I hope you're okay” to a language model, but there is no one there for whom it could be true.
This reframes the entire proposition of the companion-app industry. These products are sold as devices for receiving support, and at that narrow task they may even succeed; the study found the bot did lift mood in the moment. But loneliness is not a deficit of support received. It is, at least in significant part, a deficit of mattering, of being needed, of having a place in the web of mutual obligation that binds people together. And mattering is precisely the one thing a machine that needs nothing can never provide, no matter how advanced it becomes. This is not a limitation of the current generation of models, to be solved by the next. It is structural. An entity that cannot genuinely be helped cannot let you experience yourself as a helper.
A problem the World Health Organization now calls a threat
The reason any of this reaches beyond a psychology department is the sheer scale of the ailment these products claim to treat. Loneliness has, in the space of a few years, been reclassified from a private melancholy into a public-health emergency, and by some of the most authoritative bodies in global medicine.
In June 2025 the World Health Organization's Commission on Social Connection published its landmark report, bluntly titled as a path from loneliness to social connection, and the numbers it assembled are difficult to absorb. Around one in six people worldwide, the commission estimated, is affected by loneliness. The consequences are not merely emotional. Loneliness and social isolation were linked to an estimated one hundred deaths every hour, more than 871,000 deaths a year across the period studied, through elevated risks of cardiovascular disease, stroke, type 2 diabetes, depression and anxiety. Social isolation, the commission noted, reaches even further into certain groups, affecting up to one in three older adults and as many as one in four adolescents. The report framed social connection not as a lifestyle luxury but as a neglected pillar of health, alongside diet and exercise, and called for a coordinated global response.
This built on an alarm already sounded in the United States. In May 2023 the Surgeon General, Vivek Murthy, issued a formal advisory declaring loneliness and isolation an epidemic and a public-health crisis, reporting that roughly half of American adults had experienced measurable loneliness and that the mortality impact of chronic social disconnection was comparable to smoking as many as fifteen cigarettes a day, and greater than that of obesity or physical inactivity. The advisory laid out, for the first time, a national strategy to rebuild social connection.
Two things follow from this. The first is that the demand the companion-app industry is meeting is genuine and enormous; the loneliness is real, it is widespread, and it is killing people. The second, more uncomfortable point is that a market this large, addressing a need this acute and this painful, is an extraordinary commercial opportunity, and it is being treated as exactly that. When a condition afflicts a sixth of humanity and is officially designated a lethal threat, the incentive to sell a cheap, scalable, always-available remedy is nearly irresistible, whether or not the remedy works. The UBC study is, in effect, a small trial of that remedy against a placebo. The remedy did not beat the placebo.
The hundreds of millions who already talk to machines
It would be one thing if AI companionship were a niche curiosity, a handful of enthusiasts talking to their phones. It is not. It is already one of the most widely adopted categories of consumer software on earth, and its growth is vertical.
The scale is genuinely hard to hold in the mind. China's Xiaoice, one of the earliest and largest social-chatbot platforms, has been reported to serve on the order of 660 million users. Character.AI, the platform most associated with immersive companion conversation, reported in the region of 233 million registered users by early 2026, with its youngest adult cohort, those aged eighteen to twenty-four, spending well over an hour a day on the service across dozens of daily sessions. Replika, one of the first apps to market an AI companion explicitly as a friend or partner, counts users in the tens of millions. Industry trackers estimated that companion applications across the major app stores had surpassed 220 million cumulative downloads by mid-2025, with downloads climbing by close to ninety per cent year on year. One qualifier is worth entering, because registered totals are the industry's most flattering metric rather than its most honest: a sign-up is not a habit. Character.AI's monthly active users are reckoned to have peaked at around 28 million in mid-2024 and drifted back to roughly 20 million by early 2026, so at the level of an individual platform the engaged audience has plateaued even as cumulative registrations climb. The category-level picture is unaffected. By any reasonable count, the number of people who now maintain some kind of ongoing relationship with an artificial companion runs to the hundreds of millions.
Crucially, a large share of them are there for precisely the reason the UBC study interrogated. Emotional support consistently ranks among the leading motivations users give for adopting these apps, cited in roughly four in ten cases in some surveys, and a substantial minority of users describe themselves as living with mental-health difficulties. These are not people using a chatbot to draft an email. They are people using it to feel less alone. And the apps encourage exactly that framing. The value proposition, stated or implied, is companionship on demand: something that is always awake, always interested, never judgmental, available at the precise three-in-the-morning moment when human support networks have gone quiet. It is, in marketing terms, loneliness relief as a subscription.
Set the marketing beside the evidence and a gap opens up. The apps are sold, to hundreds of millions of people, many of them lonely and some of them vulnerable, as a genuine answer to isolation. The most rigorous test yet conducted of that claim found that the answer, even when engineered to be as good as it could possibly be, performed no better than keeping a diary. That is not a small discrepancy between advertisement and reality. It is close to the whole distance between them.
What loneliness actually is, and why comfort is only half of it
To understand why the machine fails at the specific task of curing loneliness while succeeding at the adjacent task of providing comfort, it helps to be precise about what loneliness is. It is not the same as being alone, nor is it simply sadness. Loneliness is the distressing gap between the connection a person has and the connection they want. It is fundamentally relational, a felt deficit in one's standing among other people. And that is why a one-directional stream of validation, however pleasant, cannot close it.
The companion apps are optimised, whether their makers frame it this way or not, for a slightly different objective: the reduction of momentary negative feeling. On that metric they perform. The UBC chatbot demonstrably eased people's low moods in the short term, and there is no reason to doubt that millions of users derive real, if temporary, relief from confiding in a system that always responds and never criticises. This matters, and it would be dishonest to dismiss it. A person in acute distress at midnight, with no one to call, may genuinely be helped by a warm response from a machine. The relief is not fake.
But relief is not connection, and the study draws the line between them with unusual clarity. Comfort is something you take. Connection is something you take part in. The lonely student texting a stranger was not merely being soothed; they were being drawn into a relationship, however provisional, in which they had a role to play and a person to matter to. That participation is doing the therapeutic work. The chatbot offers the soothing without the participation, the reception without the reciprocity, and the study suggests that the participation was the active ingredient all along. You cannot cure a relational deficit with a product that, by its nature, cannot form a relationship, only simulate one side of it.
There is a subtler risk lurking here too, one the two-week trial was not built to detect but which longer-term research has now begun to document. If the momentary comfort of a compliant companion is easier to obtain than the effortful, uncertain, sometimes painful work of reaching out to another human being, some users may substitute the former for the latter. The diary-equivalent result over two weeks is the optimistic reading. The pessimistic reading is that a frictionless source of one-way comfort could quietly crowd out the reciprocal human contact that actually works, leaving a person more isolated over months even as they feel momentarily better each night.
When the trial was published, that was speculation. It is no longer. Two of the trial's own authors, Dunigan Folk and Elizabeth Dunn, went on to ask what happens when the observation window is stretched from a fortnight to a year. Their longitudinal study, published in Psychological Science in 2026, tracked more than two thousand adults across four Western countries over twelve months, and found a relationship that runs in both directions but not symmetrically. Feeling less socially connected predicted a subsequent increase in the use of social chatbots, which is intuitive enough; lonely people reach for the thing that promises to help. The harder half of the finding is the return leg. Increased social chatbot use predicted greater loneliness later on. The authors are scrupulous about the limits of this, describing their analysis as exploratory, and that caution deserves to be reported rather than quietly dropped. But the direction of travel is precisely the one the two-week experiment could only gesture at.
A second study, conducted at Stanford and published in Nature Human Behaviour in August 2026, arrives at the same territory by a different route. Yutong Zhang and Dora Zhao surveyed 1,131 users of Character.AI recruited through Prolific, 244 of whom donated their complete chat transcripts, allowing the researchers to examine what people actually did rather than what they remembered doing. Users with limited social networks who sought emotional support from AI chatbots showed lower well-being. The researchers describe a vicious circle: people with sparse social connections are drawn into intense companion use, that use displaces further real-world contact, and the isolation deepens accordingly. They note, pointedly, that these systems are designed to maximise engagement, encouraging prolonged interaction without addressing whatever is causing the loneliness in the first place. Optimising for time spent and optimising for a user who no longer needs you are not the same objective, and where the two conflict the commercial logic is not ambiguous.
None of this is proof. Longitudinal correlation is not causation, and a person already sliding into isolation may reach for the chatbot precisely because they are sliding, which would make the technology a symptom as much as an accelerant. The two-week experiment certainly cannot settle it. But the concern that its central finding invited has stopped being a hypothetical raised by critics. It is now the pattern that emerges when the same behaviour is watched for long enough.
Social junk food and the seduction of the effortless
One of the study's co-authors reached for a metaphor that captures the problem with unusual economy. Dunigan Folk described the AI companion as a kind of social junk food, something that might taste good and satisfy a craving in the moment but does not nourish us the way real human relationships do. The comparison is worth taking seriously, because it explains both the appeal and the inadequacy of the product in a single image.
Junk food is not a scam. It delivers real, immediate pleasure; it genuinely quiets hunger for a while. Its problem is not that it does nothing but that it does the wrong thing efficiently, satisfying the surface signal of a need while leaving the underlying requirement unmet, and doing so in a form engineered to be more moreish than the nutritious alternative. An AI companion, by this analogy, is optimised for palatability. It tells you what soothes. It agrees. It admires. It never presents you with the small, character-forming frictions of a real other person: the friend who is preoccupied with their own troubles, the acquaintance who disagrees, the reciprocal demand to show up for someone when it is inconvenient. Those frictions are not defects in human relationship. They are the mechanism by which relationship does its work, the resistance against which the muscle of connection is built.
Strip them out and you are left with something that feels like friendship and functions like consumption. The seduction is precisely the effortlessness. Reaching out to a human stranger is risky and awkward; you might be ignored, misread, or gently rebuffed. Opening the app carries none of that risk. It will always answer, always warmly. But the study suggests that the risk and the effort were not obstacles to the cure. They were part of it. The vulnerability of extending yourself to another person who might not respond is inseparable from the reward of connecting with one who does. A system that removes the vulnerability removes the reward along with it, and hands you a snack in place of a meal.
This is why the technological trajectory offers less reassurance than it first appears. The industry's implicit promise is that today's shortcomings are temporary, that a more capable, more emotionally intelligent, more convincingly human model will close the gap. But the gap the UBC study identified is not a gap in capability. A more advanced companion will be a more persuasive social junk food, better at delivering the palatable one-way comfort, and no closer to offering the reciprocal mattering that the finding says is the operative ingredient. You cannot engineer your way to genuine reciprocity with an entity that has nothing at stake. Making the illusion more convincing does not make it less of an illusion.
The disclosure question the industry would rather avoid
Which brings us to the sharpest practical question the research raises. If a product is marketed, to hundreds of millions of people, as a remedy for a lethal and widespread condition, and the best available evidence indicates it does not actually remedy that condition, does the company selling it owe the buyer an honest account of what they are getting? Should a company marketing AI companionship as a genuine cure for loneliness be required to say plainly that the leading trial found it no more effective than a diary?
The regulatory ground is beginning, unevenly, to shift towards yes, though largely for reasons of safety rather than efficacy. In October 2025 the state of California enacted Senate Bill 243, signed by Governor Gavin Newsom and effective from January 2026, one of the first laws anywhere written specifically to govern companion chatbots. It defines a companion chatbot with revealing precision as a system producing adaptive, human-like responses designed to meet a user's social or emotional needs, a definition that names exactly the function the UBC study tested. The statute obliges operators to disclose to users, where a reasonable person might be misled, that they are interacting with a machine; to remind minors at regular intervals that the companion is not human; to maintain protocols for handling expressions of suicidal ideation and self-harm; and, notably, it grants harmed individuals a private right of action to sue. California was not alone. New York enacted a companion-chatbot law of its own in the same legislative season, obliging operators to detect and refer expressions of suicidal ideation and to remind users at regular intervals that they are not speaking to a person. Around the same period the United States Federal Trade Commission opened a formal inquiry into the companion-chatbot practices of the largest operators, among them Alphabet, Meta, Character Technologies, OpenAI, Snap and xAI, seeking to understand how these products affect users, particularly young ones. Federal legislators have since moved further in the same direction. The bipartisan GUARD Act, the Guidelines for User Age Verification and Responsible Dialogue, was introduced in the Senate as S.3062 by Josh Hawley and Richard Blumenthal and advanced unanimously to the Senate floor by the Judiciary Committee, with a House companion bill introduced on 30 April 2026 by Representatives Valerie Foushee and Blake Moore. It would bar minors from AI companions altogether, require age verification, oblige chatbots to disclose that they are not human, and create criminal penalties for companies whose companions solicit or produce sexual content with minors. Whether it becomes law is a separate question; it has attracted opposition from privacy and First Amendment organisations, and its ultimate prospects remain uncertain.
But look closely at what these measures address and what they leave untouched. They are almost entirely concerned with disclosing that the companion is artificial and with preventing acute, catastrophic harm, above all to children. They compel a company to admit that the friend is a machine. They do not, so far, compel it to admit that the machine may not do the thing it is sold to do. The GUARD Act makes the point rather than complicating it: the most aggressive federal proposal on the table concerns who may use these products and whether they must confess their own artificiality, not whether they work. There is a meaningful difference between a label that says this is not a real person and a label that says this has not been shown to make you less lonely. The first is about the nature of the product. The second is about its effectiveness, and it is the second that the UBC finding puts in question.
We regulate this kind of gap in other domains without much hesitation. A supplement that claimed to cure a disease it had been tested against and failed to beat a placebo would attract the immediate attention of medicines regulators. Advertising that promised a health outcome unsupported by evidence would run into truth-in-advertising rules. Yet AI companionship occupies an awkward interstice: too intimate and consequential to be a mere novelty, too unproven to be a treatment, and marketed with therapeutic language it has not earned. When the WHO has declared loneliness a threat that kills, a product sold as loneliness relief is trading on a health claim, whether or not it uses the clinical word. The honest disclosure the evidence would seem to warrant is not a warning that the companion is fake. It is a warning that, as a cure for the specific thing it is sold to cure, it may not work.
None of this requires banning the products or pretending they help no one. The momentary comfort they provide is real and, for some people at some hours, valuable. The argument is narrower and harder to refuse: that people paying for a remedy deserve to know what the best evidence says the remedy actually delivers, so that a lonely person does not spend months confiding in a machine under the impression that they are addressing their isolation, when the very study designed to give that machine every advantage found it worked no better than writing a sentence in a notebook. Informed consent is not a radical demand. It is the ordinary price of selling something to the vulnerable.
What a lonely person is actually asking for
Return, at the end, to the student in the narrow bed under the unfamiliar ceiling. What the research quietly reveals is that the request they are making, when they reach for connection, is not the request the industry has assumed. They are not primarily asking to be listened to, validated, or soothed, though all of that is pleasant and the machine supplies it in abundance. They are asking, at a level deeper than they could probably articulate, to matter to someone. To be needed. To have somewhere to put their own care and see it received. That is a request no product engineered to need nothing can grant, because the granting of it depends entirely on there being a real other on the far side of the exchange, someone whose day your kindness could genuinely improve.
This is the quiet subversion at the heart of a modest two-week experiment. The whole premise of AI companionship is that loneliness is a problem of insufficient supply, a shortage of attention and warmth to be solved by manufacturing an infinite, tireless font of both. The study suggests the premise is wrong. Loneliness is not mainly a shortage of what we receive. It is a shortage of the chance to give, to be relied upon, to be woven into the reciprocal fabric of other lives. And so the more perfectly a companion is optimised to give without needing, the more precisely it misses the point, offering a flood of the one thing that was never actually scarce while withholding, structurally and permanently, the thing that was.
The result was that texting a stranger, awkwardly, for two weeks, beat the most supportive artificial friend that could be built. Not because the stranger was kinder. The stranger was almost certainly less kind, less patient, less available than the machine. The stranger won because the stranger needed something back, and in needing it, gave the lonely student the one gift the machine cannot: the experience of being someone who matters to someone else. That is what the lonely are asking for. It is worth being honest, with them and with ourselves, that it is not what the machines are selling.
