AI companion eases loneliness about as well as a real person, study finds

Fahad Redha

September 4, 2026
5 min read

AI companion eases loneliness about as well as a real person, study finds

Talk to an AI companion app for a few minutes and, according to a new peer-reviewed study, you will likely feel less lonely afterwards, by roughly the same amount as if you had spoken to an actual person. That is the headline finding from researchers at Harvard Business School and the University of Pennsylvania’s Wharton School, published in the Journal of Consumer Research, and it lands at an odd moment for a city where AI chatbots have quietly become one of the most-used pieces of software on millions of phones.

The study, led by Harvard’s Julian De Freitas alongside Zeliha Oğuz-Uğuralp, Ahmet Kaan Uğuralp and Stefano Puntoni, ran five separate experiments to test something that has mostly been assumed rather than measured: do “AI companion” apps, built specifically to act as a synthetic conversation partner rather than a task-based assistant like ChatGPT, actually reduce loneliness, or do they just feel like they do.

In one experiment, participants who spent time with an AI companion reported loneliness reductions on par with those who had spoken to another human being, and noticeably larger reductions than those who spent the same time watching YouTube videos. A separate strand of the research tracked the same users over a week and found the effect held up session after session, not just as a one-off novelty response.

The mechanism the researchers isolated was not the sophistication of the chatbot’s replies. It was whether the person using it felt heard. Prompting the AI to behave as warm and attentive rather than neutral and efficient made the biggest difference to how much loneliness it relieved, more than the chatbot’s raw conversational skill did. The team also ruled out two simpler explanations, that people were merely benefiting from getting something off their chest, or from being distracted from their own thoughts. Neither factor on its own accounted for the loneliness reduction the study measured.

There was also a quirk in how people judged the apps beforehand. Participants in the study consistently underestimated how much better an AI companion would make them feel before they tried it, meaning the scepticism many people express about chatbot companionship going in does not appear to match how they actually respond once they use one.

None of this settles the argument, and the researchers themselves are careful not to claim it does. A separate study from Stanford researchers, published this year in Nature Human Behaviour, found the opposite effect in a specific and important group: users with limited offline social networks who turned to chatbots for emotional support often ended up feeling lonelier, not less lonely, particularly with heavy use.

A two-year study of Replika users from Aalto University, presented at the CHI 2026 conference, found a similar pattern over a longer timeframe, with intensive daily use correlating with rising loneliness and depressive signals rather than falling ones. Read together, the honest picture is closer to a dosage question than a simple yes or no: brief, occasional use of an AI companion appears to genuinely help, while heavy dependence, especially among people with few real-world alternatives, appears to carry the opposite risk.

That distinction matters more in London than the global research literature alone can show, because loneliness here is not evenly spread. Government data published this summer found that younger adults report higher rates of loneliness than pensioners do nationally, inverting the assumption that isolation is mainly an older person’s problem. The same national survey found the City of London recording the lowest loneliness rate of any local authority in England, a reminder that a single citywide average hides enormous variation between a resident of a Square Mile flat and someone renting alone in outer London with no nearby support network, exactly the kind of person the Stanford findings suggest is most at risk from leaning on a chatbot rather than people.

London already runs a version of this argument without any AI in it at all. A GP in Camden can already refer a lonely, isolated patient not to a chatbot but to a museum, a choir or a gardening group, the same social prescribing model already running across London’s NHS. That model exists precisely because the evidence for human contact, structured and delivered deliberately, is well established, while the evidence for AI companionship is only a few years old and still actively contested by researchers working from the same underlying data. The Harvard-led study is a genuine data point in a debate that London’s own health services will eventually have to take a position on, not a verdict that settles it.

The real test isn’t whether a chatbot can make someone feel heard for an evening, it’s what happens to the Londoners who have nothing else.

For more London arts and culture, visit EyeOnLondon at eye-on-london.com