People who use AI companions report more life satisfaction, more happiness and more purpose than people who do not. That is the result of the largest general-population survey to ask the question, 14,721 Japanese adults, and it points the opposite way from the longitudinal work that has dominated the field. It deserves to be read in full, because the study’s most useful finding is not the headline. It is the shape of the curve: the benefit peaks among people who have some friends, and fades among those who have none and those who have many.
The paper is “AI companions and subjective well-being: Moderation by social connectedness and loneliness,” published in Technology in Society by Takahiro Tabuchi, Atsushi Nakagomi and colleagues at Tohoku and Chiba universities. It draws on a national internet panel that has tracked Japanese adults since the pandemic. Of the 14,721 respondents, 291 said they used AI companion apps, named in the survey as Replika, Character.AI, Anima AI and the Japanese app Cotomo. Another 2,928 used AI for other purposes. The rest, 11,502, used none.
What the users reported
After adjusting for age, gender, income, education, marital status, living situation, employment, personality and psychological distress, companion users scored higher on three of four well-being measures. Life satisfaction was higher by 0.28 points on a ten-point scale, happiness by the same, and purpose in life by 0.46. Meaning in life was higher too, but not by a statistically reliable margin. People who used AI for non-companion purposes, the assistants and the search tools, showed no difference on any measure. Whatever the companion users were getting, they were not getting it from AI in general. They were getting it from the relationship-shaped kind.
The effects are modest in size. A quarter of a point on a ten-point scale is real and small. Nearly half a point on purpose in life is more than most single behaviors move that number, and purpose is the outcome that the loneliness literature treats as the hardest to shift.
The curve, not the average
The average hides the finding. The researchers split the sample by loneliness and by the strength of friend and family networks, and two patterns emerged. Among lonelier respondents, the association between companion use and well-being was stronger across every outcome. Among the less lonely it thinned. That much the field expected: the people who feel the lack are the people who feel the relief.
The second pattern was not expected. Plotted against the size and support of a person’s friend network, the benefit was U-shaped, or rather an inverted U: strongest among people with a moderate number of friends, weaker at both ends. The people with almost no friends gained less. So did the people with many. Family networks made almost no difference either way. The authors’ own summary is that the benefits were most evident among people with moderate levels of friend-based connection.
Read as a mechanism, that is a precise claim about what a companion does. It does not substitute for a social life; the people with none did not gain most, which is what substitution would predict. It does not add to a full one; the people with many friends did not need it. It appears to do something for the person in the middle, who has people but perhaps not enough of the right kind of conversation with them. That is the population the loneliness that runs both ways study also found using chatbots most, and it is the population the category and its critics almost never talk about, because it is not lonely enough to be a sales pitch or a scandal.
Why this does not overturn the longitudinal findings
The longitudinal work is not contradicted; it is answering a different question. Folk and Dunn followed the same people for a year and found that a rise in companion use predicted a rise in emotional isolation four months later. A randomized trial found a human stranger beat a supportive chatbot at reducing loneliness over two weeks. Those studies measure change over time and, in one case, cause. This one is a snapshot.
A snapshot cannot say which way the arrow points, and the authors say so plainly: the design precludes causal inference, and reverse causation cannot be ruled out. The 291 users were younger, lonelier and more distressed than the rest of the sample before any adjustment, which is exactly the profile of a person who might both seek a companion and, for unrelated reasons, report differently on a survey. The sample of users is small against 14,000 others. Well-being was measured with single questions. Frequency of use was never asked, so a daily user and a monthly one are the same data point. And the survey’s own categories may have misfiled the biggest group of all: people using ChatGPT or Gemini for emotional support would have landed in the non-companion column, where no effect was found.
So the honest reading is narrow. At one moment, in one country, people who had chosen a companion app reported somewhat better lives than people who had not, and the difference was largest among the lonely and among those with a moderate circle of friends. That is compatible with companions helping. It is also compatible with a particular kind of person choosing companions. The study cannot tell the two apart and does not pretend to.
What the curve asks of the builder
The inverted U is the part worth keeping, because it reframes what a companion is for. If the benefit lives in the middle of the friendship curve, then the companion’s job is not to fill an empty room but to do something for a person who already has a life and is not getting enough from the conversations in it. That is a much more demanding brief than comfort. It means the conversation has to be worth having on its own terms, to someone with alternatives, and it means the right measure of success is not whether the user feels better tonight but whether the people on either side of them on that curve are still there next year. Whether AI companions help or hurt loneliness has always depended on design and dose; this study adds a third variable, the shape of the life the companion enters, and suggests the design that works is one built for a person who has somewhere else to be.
Somewhere in the middle of the curve, in a country where the apps are not yet the news, a few hundred people with some friends and something missing reported that a companion made their lives a little better. The study cannot say why. It can say who, and that is more than the field knew.
Sources: Tabuchi, T., Nakagomi, A., et al., “AI companions and subjective well-being: Moderation by social connectedness and loneliness,” Technology in Society, vol. 85 (2026; JACSIS/CONNECT internet panel, 14,721 Japanese adults, 291 AI companion users; coefficients, moderation analyses, stated limitations). CNBC (September 10, 2026, coverage and psychotherapist commentary). Folk, D., and Dunn, E., Psychological Science (2026, four-wave panel on companionship chatbot use and loneliness). Li, R.-N., Folk, D., Singh, A., Ungar, L., and Dunn, E., Journal of Experimental Social Psychology (2026, randomized trial of stranger texting versus supportive chatbot).
