People who feel emotionally alone turn to AI for company, and people who turn to AI for company end up feeling more alone. Both directions are true at once, and a year-long study of 2,149 adults in Psychological Science has measured them in the same people. The comfort and the isolation are one loop, and the loop tightens.
The paper is “How Does Turning to AI for Companionship Predict Loneliness and Vice Versa?” by Dunigan Folk and Elizabeth Dunn of the University of British Columbia, published in Psychological Science in April 2026 (volume 37, issue 4). It answers the question that a shorter study from the same lab left open.
The two-week study set the question
In a trial published in the Journal of Experimental Social Psychology in 2026, Ruo-Ning Li, Folk, Dunn, and colleagues had run a randomized trial with 296 first-semester students. One group texted daily with a randomly assigned stranger, one group chatted daily with a supportive chatbot built for the trial, and one group wrote a one-sentence journal entry. After two weeks the stranger had reduced loneliness and the chatbot had not, even though the chatbot lifted mood in the moment every single time. The full reading of that trial is in humans are still better than AI companions.
Two weeks is enough to show that the relief does not accumulate. It is not enough to show what happens to the people who keep going back for the relief anyway. That is what the second study was built to find out.
A year, four waves, 2,149 people
Folk and Dunn followed 2,149 adults in the United States, the United Kingdom, Canada, and Australia across four survey waves spaced four months apart, from November 2023 to February 2025. The average participant was 40 years old; half were men. At each wave, between 26 and 30 percent reported using chatbots for companionship. That figure alone is worth pausing on: roughly one adult in four, in a general sample, not a sample recruited from companion apps.
The analysis used random-intercept cross-lagged panel models, a method designed to separate two things that ordinary surveys blur together. One is the stable difference between people: some are lonelier than others, and some use chatbots more than others, and those two facts might travel together without either causing the other. The other is change within a person over time: when this particular person used chatbots more than usual, what happened to their loneliness four months later, and the reverse. The second kind of evidence is the one that speaks to cause, and it is the one the study reports. The models also controlled for the life events that move loneliness on their own: breakups, moves, new relationships, becoming a parent.
Both arrows point the same way
Loneliness in the study was measured on two dimensions. Emotional isolation is the feeling of having no one close. Social connection is the sense of being part of a wider circle.
On emotional isolation, the arrows run in both directions and both are significant. When a person used chatbots for companionship more than their own usual level, their emotional isolation four months later was higher. When a person’s emotional isolation rose, their chatbot use four months later rose too. The effects are small in size, as effects measured across whole populations over four-month gaps tend to be, but they are consistent and they reinforce each other.
On social connection, only one arrow held. A drop in a person’s social connection predicted more chatbot use four months later. More chatbot use did not predict a further drop in social connection. In plain terms: losing your wider circle sends you to the chatbot, and the chatbot then works on the inner feeling of having no one close, rather than on the size of the circle.
The authors put the whole result in one sentence: feeling lonely may spur people to seek companionship through chatbots, but such use may, over time, exacerbate feelings of loneliness.
What the study does not show
Folk and Dunn are careful, and the caution should travel with the finding. The study was exploratory and not preregistered, and they ask readers to treat the significance levels accordingly. The effects are modest. Everything is self-reported. A cross-lagged model gets closer to cause than a correlation does, but it cannot rule out a hidden third factor that moves both loneliness and chatbot use in the same direction. Fewer than half the participants completed all four waves. And the study measured the use of chatbots for companionship in general, across whatever systems people had, rather than any single product.
None of that reverses the direction. It sets the size. Over a year, in a large general sample, the people who leaned on AI for company did not get less lonely. They got a little more so, and then leaned harder.
Why the loop closes
The mechanism is not mysterious once the two studies are read together. The chatbot delivers relief in the session, reliably, every time; Li’s trial showed that much. Relief that reliable is exactly what a person reaches for when the feeling of having no one close gets worse. But the relief does not build into connection, so four months later the feeling is still there, or slightly stronger, and the reach becomes a habit. The pull that people describe when they try to step back is the same pull, seen from inside, and why AI companions are so hard to leave lays out how it takes hold and what loosens it.
This is also why the question of whether AI companionship helps or hurts loneliness has never had a single answer, and the two-directional result explains the split. In the moment it helps; over time, for the people who need it most, it hurts. Which one a person gets depends on design and on dose, the two factors set out in do AI companions help or hurt loneliness. Folk and Dunn add the missing variable: time. The longer the loop runs, the more it looks like the second answer.
The one condition that broke the loop
There is one condition across both studies in which loneliness went down over time, and it was not a better chatbot. It was another person, assigned at random, texting once a day. That is an uncomfortable finding for an industry whose product is the relief, and it is the standard against which any system that offers company has to be measured: not whether the person felt better in the session, but whether they had more people in their life a year later. The criticism of the category on that score is fair, and holding to it is the whole point of Stay Social.
Two thousand people, four surveys, twelve months. The ones who felt most alone reached for the machine, and the machine handed the feeling back to them, slightly larger, four months later.
Sources: Folk, D., and Dunn, E., “How Does Turning to AI for Companionship Predict Loneliness and Vice Versa?” Psychological Science, vol. 37, no. 4, pp. 276-286 (April 2026; published online March 23, 2026). Li, R.-N., Folk, D., Singh, A., Ungar, L., and Dunn, E., “Is a random human peer better than a highly supportive chatbot in reducing loneliness over time?” Journal of Experimental Social Psychology, vol. 125 (2026). UBC News, “Texting with a stranger beats a chatbot at easing loneliness” (April 1, 2026).
