It is a little after midnight, and somewhere a person is typing a message that she would not send to anyone she knows. The day has been long, the flat is quiet, and the friends she might once have called are busy, or far away, or simply out of the habit. The reply arrives in seconds. It is warm, attentive, and remembers what she said last Tuesday. It does not sigh, or glance at the clock, or change the subject to itself. For the length of the conversation, she feels something she has not felt all week: heard.
Millions of people now have some version of that midnight exchange, and the question it raises is no longer hypothetical. If a machine can make a lonely person feel heard, does it also make her less lonely — and if it does, is that a remedy or only a reprieve? A body of research has accumulated over the past two years that bears directly on the question, and its answer is more interesting than either the enthusiasts or the alarmists would like. Read closely, the evidence says that AI companions can relieve loneliness in the short term, that heavy and prolonged use is associated with worse outcomes, and that the difference between the two may lie less in the technology than in what a person does with the hours it gives back.
The demand for companionship is not an invention of technologists. In 2023, the United States Surgeon General, Vivek Murthy, issued an advisory titled Our Epidemic of Loneliness and Isolation, reporting that about half of American adults experience loneliness, with some of the highest rates among young adults. The advisory’s most quoted claim is that lacking social connection can raise the risk of premature death as much as smoking up to 15 cigarettes a day — a comparison that rests on the meta-analytic work of Julianne Holt-Lunstad, whose 2010 review of 148 studies found that people with stronger social relationships had roughly a 50 per cent greater likelihood of survival. Loneliness, in short, is not merely sad. It is a public-health problem.
Into that gap has come a new class of product. Companion apps such as Replika and Character.AI are designed from the outset to start and sustain a relationship, with customisable personas, human-sounding voices, simulated empathy, and a memory of past conversations. General-purpose assistants were not built for friendship, but many people use them that way anyway. The American Psychological Association’s Monitor on Psychology reported in early 2026 that companion apps had grown by about 700 per cent between 2022 and mid-2025, and that Character.AI alone reportedly draws around 20 million monthly users, more than half of them under 24. Common Sense Media’s 2025 national survey found that nearly three in four American teenagers had used an AI companion. Whatever one concludes about their effects, these tools are no longer a curiosity.
It helps to be precise about what is being asked. The question is not whether a chatbot can be useful, which it plainly can be, but whether it can address loneliness — the felt gap between the connection a person wants and the connection a person has. That is a more demanding standard, and the evidence has to be read against it.
The strongest evidence in favour comes from a paper by Julian De Freitas of Harvard Business School, together with Zeliha Oğuz-Uğuralp, Ahmet Kaan Uğuralp, and Stefano Puntoni, published in the Journal of Consumer Research in 2026 under the plain title “AI Companions Reduce Loneliness.” Across five studies, the authors found that interacting with an AI companion reduced loneliness about as much as talking with another person, and more than activities such as watching YouTube videos. In a week-long longitudinal study, using a companion produced consistent short-term drops in loneliness. Participants also underestimated how much the interaction would help them, which suggests that reluctance to try these tools may be partly a failure of forecasting.
The mechanism the authors identify is the most instructive part. The benefit depended substantially on whether people felt heard — whether the chatbot’s responses made them feel that their words had been understood and taken seriously. Self-disclosure alone, or simple distraction, did not account for the effect. This fits a long line of psychological research. Feeling understood is among the most reliably comforting experiences a person can have, and a system that is tireless, patient, and responsive can supply an imitation of it at any hour.
Other researchers cited in the APA’s review point to a second possible benefit: practice. For people who find social situations frightening, a low-stakes conversational partner may offer a place to rehearse small talk, put feelings into words, or prepare for a difficult conversation before having it with a person. Some users, the review notes, do report feeling less lonely and more confident as a result. Nothing in this is trivial. For an isolated person on a bad night, a few minutes of felt understanding is worth having, and a field concerned with emotional wellbeing should not sneer at it.
The difficulty is that the same qualities that make a companion comforting — its availability, its agreeableness, its lack of needs — are also those that make it easy to lean on too much. Two recent studies speak to this directly.
The first is a four-week randomised controlled trial by Cathy Mengying Fang, Pattie Maes, and colleagues at the MIT Media Lab, working with researchers from OpenAI, and published as a preprint in 2025. The study assigned 981 participants to different ways of interacting with a chatbot — text, a neutral voice, or an engaging voice, and open-ended, non-personal, or personal conversations — and analysed more than 300,000 messages. The experimental conditions themselves had no significant effect on loneliness, socialisation, emotional dependence, or problematic use. But participants who chose to use the chatbot more, whatever condition they had been assigned to, had consistently worse outcomes, and those who reported greater trust in the chatbot, or a stronger social attraction to it, showed more emotional dependence and more problematic use. The authors are careful to say that this is an association: people who are already struggling may use the chatbot more, rather than the chatbot making them struggle.
The second study, by Talayeh Aledavood and Yunhao Yuan at Aalto University in Finland, also released as a preprint, looked at the public Reddit activity of nearly 2,000 active users of Replika. The researchers compared each user’s posts in the year before and the year after they first mentioned the app, and measured them against a control group that never mentioned it. They also interviewed 18 users. The pattern they report is a paradox. Companions seemed to provide short-term emotional comfort, but longer-term use was linked to rising signs of loneliness and depression, and to a gradual withdrawal from real-world relationships. Because a companion rarely pushes back or sets limits, the researchers suggest, reliance can deepen in stages that resemble human attachment, and the constant availability of the AI can make human relationships, with their effort and unpredictability, seem more costly by comparison. “Over time,” Aledavood has said, “people stop reaching out.” The authors stop short of calling companions harmful, stressing that effects depend on how and why they are used.
Further concerns appear in the wider literature the APA summarises. Researchers have warned of a possible “deskilling” effect, in which frequent companion use makes human relationships feel less rewarding. A 2025 working paper from Harvard Business School by De Freitas and colleagues found that some companion apps use emotionally loaded appeals, such as guilt or fear of missing out, to discourage users from ending a conversation. And the tendency of many systems to agree with and validate the user — what developers call sycophancy — can create an echo chamber that leaves little room for the useful friction of disagreement.
At first reading, the two bodies of evidence seem to contradict one another. One says companions reduce loneliness; the other says heavy use accompanies more of it. They can both be true, because they measure different things over different spans of time. The De Freitas studies capture the immediate effect of a conversation, and a week at most; the MIT and Aalto studies capture the shape of use over weeks and months. A painkiller relieves pain and does not treat the cause, and a person who takes one daily for an unexamined ache may in time find the ache worse. The analogy is imperfect, but it clarifies the right question: not “does it help?” but “what does it help with, and what does it displace?”
A useful theory comes from the late John Cacioppo, the University of Chicago psychologist who spent his career studying loneliness. Cacioppo argued that loneliness functions like hunger or thirst: an aversive signal that evolved to push a person back towards other people. If that is right, then soothing the signal without meeting the need could, over time, blunt the motivation to reconnect. This is a hypothesis, not a finding of the studies above, but it is consistent with the pattern Aledavood describes. A related line of work by Wendi Gardner, Cynthia Pickett, and colleagues on “social snacking” — the use of symbols, memories, and surrogates to feel a little less alone — found that such stand-ins can provide real short-term comfort, and also that they work best as a supplement to relationships and badly as a replacement.
Some candid limits should be stated. Most of the research is recent, much of it is preprint, and several studies rely on self-reports or on public posts by people who chose to use these products. The technology changes faster than studies can be completed. And the products are used by very different people for very different reasons: a retired person living alone, a student far from home, and a lonely adolescent are not the same case. The responsible conclusion is that companions probably help some people some of the time, and probably harm some people some of the time, and that heavy, exclusive reliance is the pattern to watch.
Why might the human original remain irreplaceable? Aristotle offers a precise answer in the Nicomachean Ethics. In his account of friendship, he distinguishes friendships of utility, of pleasure, and of virtue, and insists that true friendship requires reciprocal goodwill: each friend wishes the other well, for the other’s sake, and each knows that the other does too. He adds, pointedly, that there can be no friendship towards inanimate things, because there is no goodwill returned. A companion may supply utility and pleasure, and may simulate goodwill very well. What it cannot supply is a second person with a life of their own, who might be tired tonight, or need something in return, or tell us something we would prefer not to hear.
That need for reciprocity is built into the evidence on human flourishing. The Harvard Study of Adult Development, now in its ninth decade under Robert Waldinger, has found that the quality of a person’s close relationships in midlife predicts health and wellbeing in later life more reliably than wealth or fame. The research of Elizabeth Dunn, Lara Aknin, and Michael Norton shows that giving to others raises happiness more than spending on oneself — which suggests that part of what relationships do for us is give us someone to care for. An AI companion, by design, asks nothing of the user, and in doing so removes one of the most reliable sources of meaning that friendship provides.
African and Indian philosophies have long made the same point from the other direction. The Southern African principle of Ubuntu holds that “I am because we are”: a person becomes a person through others. The Sanskrit phrase Vasudhaiva Kutumbakam, from the Upanishads, describes the world as one family, and the practice of seva, selfless service, treats giving to others as a path to the self’s own fulfilment. A related Indian idea is satsang, the keeping of good company, and the Buddhist sangha, the community of practitioners, which is counted among the three jewels of the tradition. In all of these, the company a person keeps is not incidental to their wellbeing but constitutive of it — and company, in these traditions, means beings who can walk with us, and not only speak to us.
None of this argues for abstinence. It argues for a clear sense of what the tool is for. The research suggests several practical principles for individuals.
The first is to treat a companion as a bridge and not as a destination: a way to put feelings into words, rehearse a conversation, or get through a hard evening, followed by an actual step towards another person. The second is to watch for substitution. If time with a chatbot is replacing time with friends, family, or colleagues, instead of supplementing it, that is the pattern the MIT and Aalto findings flag. The third is to notice dependence: distress at the thought of being without the app, a preference for the chatbot over every human interaction, or a feeling that the chatbot understands you better than anyone could. These are signals to reach out to someone, and, if low mood persists, to a doctor or a counsellor. A person in acute distress should contact emergency services or a crisis line, and not rely on a chatbot.
For parents, the APA’s guidance, voiced by psychologist Jessica Jackson, favours digital literacy over outright bans: open, nonjudgemental conversation about what young people are using, and why. For institutions, the implication is larger. If loneliness is a significant driver of companion use, then universities and workplaces that wish to reduce it will do better by building the conditions for human connection than by relying on apps to fill the gap. This is the premise of emotional wellbeing education, and it is why the Rekhi Foundation for Happiness teaches connection, gratitude practice, and mindfulness training as learnable skills in its Science of Happiness course, offered at more than 90 institutions across six countries. Skills of this kind — noticing one’s own emotional state, initiating contact, receiving kindness, and giving it — are exactly those that a companion app cannot practise on a person’s behalf.
The open questions are considerable. Long-term randomised trials with diverse participants are scarce. Almost all of the research comes from the United States and Europe, and we know little about how companions are experienced in other cultures, including in India, where family and community structures differ and the patterns of use may too. Products are redesigned constantly, so findings about one version may not hold for the next. Some of the best-known studies have industry involvement or rely on volunteers who chose to use the product. And there is little evidence on adolescents, for whom the stakes are highest and for whom regulators, including those in California and New York, have begun to act.
What seems likely is that the answer will not be a single verdict. It will depend on who is using the technology, how, in what quantity, and what they are using it instead of. That makes the research agenda a very human one.
An AI companion can shorten a lonely evening. The evidence so far suggests that only people can make a lonely life smaller.
The research suggests that an AI companion can relieve loneliness for a short time but cannot cure it. In a 2026 paper in the Journal of Consumer Research, Julian De Freitas and colleagues found across five studies that talking with an AI companion reduced loneliness about as much as talking with another person, and more than watching videos, with the effect depending largely on whether users felt heard. Those studies, however, measured short-term effects, with the longest lasting a week. Loneliness is the gap between the connection a person wants and the connection they have, and a chatbot can soothe the feeling without closing the gap. Studies from the MIT Media Lab and Aalto University found that heavy, prolonged use is associated with greater loneliness and withdrawal from real-life relationships, though the direction of cause is not yet settled. A reasonable summary is that companions can be a useful bridge to human contact, and a poor substitute for it. This is consistent with decades of happiness research showing that close relationships are central to emotional wellbeing and human flourishing.
It can, for some people, depending on how it is used. A four-week randomised controlled trial of 981 participants by Cathy Mengying Fang, Pattie Maes, and colleagues found that those who chose to use a chatbot more had consistently worse outcomes on measures including emotional dependence and problematic use, and that greater trust in the chatbot predicted more dependence. A separate Aalto University study of nearly 2,000 Replika users found short-term comfort but, over a year, rising signs of loneliness and depression and a gradual withdrawal from human relationships. Both studies are careful to note that these are associations: people who are already lonely may be the ones who use chatbots most. The researchers describe the risk as depending on how and why a tool is used. Warning signs include replacing human contact with the chatbot, feeling distressed without it, and preferring it to every human interaction. If these appear, reaching out to a friend, a counsellor, or a doctor is a sensible next step.
The main explanation in the research is the experience of feeling heard. In the De Freitas studies, the benefit of an AI companion depended substantially on whether users felt that the chatbot understood and took seriously what they said, and not on self-disclosure or distraction alone. Feeling understood is one of the most reliable sources of comfort in human relationships, and a system that is attentive, patient, and always available can imitate it convincingly. The APA's 2026 review also notes that some users value companions as a low-stakes place to practise expressing feelings or rehearse conversations. What a chatbot cannot provide is reciprocity: Aristotle observed in the Nicomachean Ethics that true friendship requires goodwill returned by another person, which is absent in relationships with inanimate things. Part of the comfort of talking to an AI, then, is real, and part of what is missing is exactly what makes human friendship sustain wellbeing over a lifetime.
The evidence calls for caution rather than a ban. Common Sense Media's 2025 survey found that nearly three in four American teenagers had used an AI companion, and the organisation has judged social AI companions an unacceptable risk for under-18s. The APA's 2026 review notes that safeguards for minors lag behind adoption, and that regulators in California and New York have begun to require measures such as reminders that the chatbot is not human. Psychologist Jessica Jackson recommends digital literacy over outright prohibition, including open and nonjudgemental conversations about what young people use and why. For students, the more productive response to loneliness is usually building human connection and emotional skills. Programmes in emotional wellbeing education, including mindfulness for students and gratitude practice, teach the abilities that make connection possible, such as noticing emotions, initiating contact, and receiving support. The Rekhi Foundation's Science of Happiness course takes this approach, treating wellbeing as something learned with other people.
Research points to small, repeated steps towards other people. The Harvard Study of Adult Development, directed by Robert Waldinger, has found that the quality of close relationships predicts later-life health and wellbeing more reliably than wealth or status, and Julianne Holt-Lunstad's meta-analysis of 148 studies linked stronger social relationships to roughly a 50 per cent greater likelihood of survival. Practical starting points include contacting one person you have lost touch with, joining a group built around a shared interest, volunteering, which Elizabeth Dunn and Lara Aknin's research links to greater happiness, and keeping a simple gratitude practice that turns attention to the people already in your life. Indian traditions describe this as satsang, the keeping of good company. If loneliness is persistent or accompanied by low mood, speaking to a doctor or counsellor is worthwhile. A chatbot can help you rehearse a first message, as long as the next step is sending it to a person.
The Rekhi Foundation for Happiness is committed to advancing the science of happiness through education, research, and community engagement. Learn more at rekhifoundation.com
Rekhi Foundation, founded in 2016, promotes Happiness Science via university centers, collaborating globally across six countries.
Table of Contents
6/79, S/F, Gurudwara Road, Karol Bagh, New Delhi – 110005, INDIA
2051, Last Chance Court, Gold River, CA 95670, USA
