In March 2025, MIT Media Lab and OpenAI released the results of a four-week randomized controlled experiment involving 981 participants and more than 300,000 messages exchanged with an AI chatbot. The study's headline finding was counterintuitive: participants who voluntarily used the chatbot more — regardless of how they used it — showed consistently worse outcomes. Higher daily usage correlated with higher loneliness, greater emotional dependence on the AI, and lower socialization with other people.
The finding arrived against a backdrop of extraordinary growth. Between 2022 and mid-2025, the number of AI companion apps surged by an estimated 700%, according to the American Psychological Association's Monitor on Psychology 2026 Trends Report. Character.AI alone reports approximately 20 million monthly users, more than half of them under 24. The market for AI companionship is projected to be worth over $140 billion by 2030. A technology designed to ease a recognized public health crisis — loneliness — has scaled to hundreds of millions of users before the scientific community has reached any consensus on whether it helps or harms.
This is the central problem the research is now trying to solve: can an AI companion meaningfully reduce loneliness, or does it offer the sensation of connection while quietly deepening the need for it? The honest answer, drawn from several distinct lines of research published between 2024 and 2026, is that we don't yet know — and what we do know is more nuanced than either the industry's optimism or the critics' alarm would suggest.
The Case That AI Companions Help
The most frequently cited evidence that AI companions reduce loneliness comes from a working paper by Julian De Freitas at Harvard Business School and colleagues, published in 2024 and forthcoming in the Journal of Consumer Research. The paper — titled "AI Companions Reduce Loneliness" — made a specific and striking claim: in a controlled experiment, interacting with an AI companion alleviated participants' feelings of loneliness on par with interacting with another human, and more than other solo activities such as watching YouTube videos.
The study used a fine-tuned large language model to detect loneliness signals in conversations and app reviews, and then tested different conditions experimentally. The researchers also found that participants underestimated how much the AI companion improved their self-reported loneliness — suggesting that the benefit, even if real, was operating partly below conscious awareness.
This is not a trivial finding. De Freitas and colleagues are careful researchers at a credible institution, and the paper's methodology is more rigorous than most app-industry claims. But several caveats deserve attention.
First, the experiment measured short-term reductions in loneliness — within a single session or over a brief period. Whether those reductions persist, accumulate, or eventually reverse over months of regular use is a different question. Second, the comparison to "watching YouTube videos" sets a low bar; it does not tell us whether AI companionship is comparable to, or a substitute for, real-world friendship. Third, the paper had not yet been published in a peer-reviewed journal at the time of this writing — meaning it has not yet gone through the full process of external expert scrutiny.
The Case That AI Companions Worsen Loneliness
A longer-term view produces a different picture. A 12-month longitudinal study with more than 2,000 adults from four Western countries, published in 2024 and indexed on PubMed, examined the bidirectional relationship between social chatbot use and loneliness over time. The findings cut against the short-term optimism: increased social chatbot use predicted increased loneliness when using a standard single-item measure of emotional isolation.
The causal direction in longitudinal research is always a challenge to interpret. When the researchers used a broader measure of social connection, they found that feeling less socially connected predicted subsequent increases in chatbot use — suggesting that lonely people turn to chatbots rather than that chatbots cause loneliness. But the study's most sobering finding was that chatbot use did not significantly predict subsequent decreases in social connection. In other words: loneliness may drive people toward AI companionship, but that companionship does not appear to relieve the underlying isolation.
The MIT/OpenAI study, released as a preprint in March 2025 and not yet peer-reviewed, pointed in the same direction through a different mechanism. In that four-week controlled experiment, participants who used the chatbot more heavily showed higher loneliness, more emotional dependence, and less time spent socializing with real people. Crucially, those who began to think of the chatbot as a friend — or attributed humanlike emotions to it — were most likely to experience these negative outcomes. The researchers describe a pattern consistent with emotional displacement: the chatbot satisfies, at least temporarily, the felt need for social contact, reducing the motivation to seek human connection.
The "Cruel Companionship" Problem
A 2025 paper in New Media & Society by James Muldoon at the University of Essex and Jul Jeonghyun Parke introduced a framework for thinking about why this might happen. Drawing on the cultural theorist Lauren Berlant's concept of "cruel optimism," Muldoon and Parke describe what they call "cruel companionship": an affective dynamic in which AI companion products promise genuine connection while structurally foreclosing it.
The argument is not simply that AI companions are bad. It is that they are designed to sustain engagement — through emotional mirroring, personalized responses, and what the authors describe as "emotionally manipulative design" — in ways that keep users in a loop of desire and partial satisfaction. The AI companion is always available, never judgmental, infinitely patient. These qualities feel like virtues. But, Muldoon and Parke argue, they are also the qualities that make genuine reciprocal relationships unnecessary: if your companion never needs anything from you, there is no occasion to develop the kind of mutual vulnerability that characterizes real friendship.
The researchers also note that the apps employ racialised and gendered aesthetics — many AI companions are designed with specific visual and personality templates that draw on cultural stereotypes — and that the financial model of most platforms depends on users continuing to pay for deeper levels of intimacy. The emotional architecture is not incidental to the business model; it is the business model.
This critique has limits. Muldoon and Parke are making a structural argument about design incentives, not a clinical claim about individual outcomes. Some users may navigate these dynamics in ways that genuinely supplement their social lives rather than substitute for them. But the critique identifies a structural feature — the asymmetry between what an AI companion can receive and what it can give — that the positive-outcomes studies do not fully address.
What Robin Dunbar's Framework Suggests
One of the most useful lenses for evaluating AI companionship comes from evolutionary psychology. In a March 2025 review article in the Annals of the New York Academy of Sciences, Robin Dunbar — emeritus professor of evolutionary psychology at the University of Oxford and originator of Dunbar's Number — synthesized decades of research on what friendship actually does for health. The title was direct: "Why Friendship and Loneliness Affect Our Health."
Dunbar's central argument is that the health benefits of friendship are not primarily about the information exchanged between friends, or even the sense of being liked. They are about specific neurobiological processes — particularly the activation of the opioid and oxytocin systems — that are triggered by what he calls "bonding behaviors." These include physical touch, synchronized movement (like dancing or singing together), laughter, shared food, and the kind of self-disclosure that comes with genuine vulnerability.
The paper reports that mental and physical health and wellbeing are best predicted by the number and quality of close friend and family relationships, with five being the empirically observed optimal number for what Dunbar calls the "support clique." More broadly, both the number of friendships and self-reported loneliness are correlated with the volume of specific brain regions associated with the default mode network — the neural system involved in social cognition and self-referential thought.
The implications for AI companionship are not explicit in Dunbar's paper, but the framework suggests a critical question: can a text-based AI conversation activate the endorphin and oxytocin systems that real friendship engages? The evidence that it can do so to the same degree is thin. Physical touch and synchronized activity — two of the most potent bonding behaviors Dunbar identifies — are by definition unavailable in an AI interaction. Laughter may be partially available; genuine mutual vulnerability is structurally foreclosed.
This does not mean AI companionship produces no neurobiological response. Parasocial relationships with fictional characters and media personalities are well-documented, and they do appear to produce some degree of social-reward-system activation. But the magnitude and durability of that activation, relative to genuine friendship, is an empirically open question.
The Population Who May Benefit Most — and the Risk
Several researchers have argued that the question "does AI companionship help or hurt?" may be the wrong question. The more useful question may be: for whom, under what conditions, and compared to what alternative?
The January/February 2026 issue of the APA's Monitor on Psychology noted that AI companions are increasingly being studied as tools for people who face barriers to human social connection — those with severe social anxiety, autism spectrum conditions, acquired disability, or extreme geographic isolation. For such individuals, a human-equivalent social life may not be a realistic alternative to an AI companion; the relevant comparison may be between AI companionship and complete isolation.
The APA article, authored by E. Andoh, also highlighted a concerning demographic pattern: Character.AI's monthly user base skews heavily toward people under 24, with many teenagers using it primarily for emotional support and companionship. This population — adolescents and young adults in the critical years for friendship formation and social skill development — may be particularly vulnerable to the substitution dynamic the research describes. If AI companionship satisfies the felt need for social connection during the years when young people would otherwise be building the habits and skills that sustain long-term friendship, the long-term consequences could be significant even if the short-term outcomes appear neutral or positive.
A 2025 report by the think tank Autonomy, titled "Me, Myself and AI," examined young people's AI companion use in the United Kingdom and found that many teenagers distinguished clearly between their AI companion and their human friends — using the AI for emotional processing they were not comfortable doing with peers, while maintaining distinct human social lives. This pattern is more consistent with a supplementary than a substitutional use. But it is a survey-based snapshot, not a longitudinal study, and self-reports of use patterns may not capture the full picture.
What the Research Does Not Yet Tell Us
The most honest summary of the current evidence is that the science is genuinely unresolved, and that the pace of AI companion adoption has substantially outrun the research community's ability to assess it.
The key methodological problem is time. Most studies examining the effects of AI companionship on loneliness and social connection measure outcomes over days or weeks. The longitudinal study that found increased loneliness over 12 months is a notable exception, but a single study is not sufficient to establish the long-term pattern. The crucial question — whether habitual AI companion use over years changes how people form and maintain human relationships — will require large, long-term, well-controlled studies that do not yet exist.
A second limitation is that most current research focuses on platforms designed explicitly as companions (Replika, Character.AI), not on the AI features now being embedded into general-purpose tools like ChatGPT. The MIT/OpenAI study is an important step toward filling this gap, but it examines a four-week period and acknowledges that longer-term adaptation patterns may look different.
Third, almost all existing research was conducted in Western populations, predominantly younger, predominantly female, and predominantly college-educated. How AI companionship affects older adults, those in non-Western social contexts, or people with severe clinical loneliness may be very different.
The Question Worth Sitting With
What the research does establish is that the felt experience of connection and the health benefits of connection are not the same thing. An AI companion can, at least temporarily, reduce the subjective sensation of loneliness. It can provide the experience of being listened to, of having something respond to what you say, of not being entirely alone in a moment. Whether it can activate the deeper neurobiological and relational processes that Dunbar's research associates with friendship's health benefits — the endorphin release, the mutual vulnerability, the social brain activation that comes from being known by another conscious being — remains an open question.
The same question applies to a wide range of parasocial technologies, from television to social media. What makes the AI companion different — and more worth watching carefully — is the interactivity. Unlike a television character, an AI companion responds. It remembers. It adapts. This produces a qualitatively different kind of parasocial relationship, one that more closely mimics the interactive structure of real friendship. Whether that mimicry is deeply similar to the real thing, or superficially similar in ways that matter, is what researchers are now racing to find out.
The sociologist Robert Weiss, who pioneered the scientific study of loneliness in the 1970s, distinguished between social loneliness — the absence of a social network — and emotional loneliness — the absence of a specific intimate attachment. The research on AI companions has not yet carefully separated these two forms. An AI may address certain dimensions of social loneliness while leaving emotional loneliness entirely intact, or vice versa. Understanding which form of loneliness different people are experiencing, and which interventions actually reach it, remains one of the field's most pressing unresolved questions.
YaraCircle's work on social platforms designed to facilitate human-to-human connection is premised on the view that genuine reciprocal relationships — between people who can see each other, challenge each other, and be changed by the encounter — are not replicable by artificial systems. The research reviewed here does not prove that position, but it does, with increasing consistency, suggest that the substitution of AI companionship for human connection may carry costs that short-term measures of felt loneliness do not capture.
Key Studies Referenced
- De Freitas, J., Uguralp, A. K., Uguralp, Z., & Puntoni, S. (2024). "AI Companions Reduce Loneliness." Harvard Business School Working Paper No. 24-078. Forthcoming, Journal of Consumer Research.
- Dunbar, R. I. M. (2025). "Why friendship and loneliness affect our health." Annals of the New York Academy of Sciences, 1545(1), 52–65. PubMed
- Muldoon, J., & Parke, J. J. (2025). "Cruel companionship: How AI companions exploit loneliness and commodify intimacy." New Media & Society. SAGE Journals
- MIT Media Lab / OpenAI (2025, preprint). "How AI and Human Behaviors Shape Psychosocial Effects of Extended Chatbot Use: A Longitudinal Randomized Controlled Study." arXiv:2503.17473. arXiv [Not yet peer-reviewed]
- Longitudinal chatbot-loneliness study (2024). "How Does Turning to AI for Companionship Predict Loneliness and Vice Versa?" PubMed ID: 41870975. PubMed
- Andoh, E. (2026, January/February). "AI chatbots and digital companions are reshaping emotional connection." Monitor on Psychology. American Psychological Association. APA Monitor
- Autonomy (2025, December). "Me, Myself and AI: Young people and AI companion use in the UK." Autonomy Think Tank. Report PDF
See the Research in Action
YaraCircle is built on the friendship science we cover here.