Twelve hundred Chinese college students were measured three times over a semester on how supported they felt by their AI chatbots, how lonely they were, and how anxious they felt around other people. The support lowered loneliness eight weeks later. It also raised social anxiety, at every step, for everyone. And the loneliness relief reached only the students who already had strong support from real people. The ones who had the least got the anxiety without the relief.
The study is Jiangnan Hu’s “Perceived support from AI chatbots and loneliness among higher vocational college students,” published in Frontiers in Psychology in 2026. It is a single-author paper from a vocational college in Jiangxi, with real limits. What it adds is a number. The claim that frictionless support quietly raises the cost of human relationships is already the sharpest finding in this field, the Aalto paradox, drawn from two years of Reddit data and interviews. Hu’s study measures that cost directly, as a rising score on a social anxiety scale, in the same people, wave by wave.
Who was measured, and with what
The students were not companion-app users. They used ChatGPT, Doubao, DeepSeek, Kimi, ERNIE Bot, or Tongyi Qianwen, the general assistants, for an average of 38 minutes a day; a fifth used them under a quarter of an hour, and about one in twenty for more than two hours. Their mean age was 19. Sixty percent were women. Of 1,500 who began, 1,200 completed all three waves, spaced about eight weeks apart across a semester, with no detectable difference between those who stayed and those who dropped out.
Four things were measured. Perceived support from the chatbot, using an adapted online social support scale that covers emotional support, companionship, information, and practical help. Loneliness, with the eight-item short form of the UCLA scale. Social interaction anxiety, with the standard 20-item scale that asks how uncomfortable a person is in ordinary exchanges with other people. And, at the first wave only, perceived support from real people: family, friends, a significant other.
The design is a cross-lagged panel: each measure at one wave is used to predict the others at the next, while controlling for where each person started. It asks whether feeling supported by a chatbot now predicts a change in loneliness or anxiety later, over and above how lonely or anxious the person already was.
What the chatbot support predicted
Three results carry the paper.
Perceived chatbot support predicted lower loneliness, but only in the second interval: support at the midpoint of the semester predicted less loneliness at the end (a standardized coefficient of minus 0.166), while support at the start did not significantly predict loneliness at the midpoint. The relief exists and it is modest.
Perceived chatbot support predicted higher social interaction anxiety in both intervals, at almost identical strength (0.161 and 0.158). Students who felt more supported by their AI felt more uncomfortable around people eight weeks later, and the effect did not fade with time. Anxiety in turn predicted loneliness at the next wave, more strongly than anything else in the model (0.198 and then 0.232). Put together, there is an indirect path: chatbot support at the start raised anxiety at the midpoint, which raised loneliness at the end, with a bootstrapped indirect effect of 0.038 that excludes zero. The direct route lowers loneliness a little. The anxiety route raises it, and the second is the one that compounds.
Then the moderation. For students with strong support from real people, chatbot support predicted clearly lower loneliness at the end of the semester (a slope of minus 0.287). For students with weak real-life support, the slope was minus 0.023, indistinguishable from nothing. Hu is careful about the reading: the low-support group was “not harmed by AI support, but rather did not show a significant compensatory benefit.” The benefit went to those who needed it least. The anxiety went to everyone.
What this adds to what was known
The work before this had established the paradox, the loop, and the curve. Aalto’s two-year study named the mechanism from interviews: unconditional support raises the perceived cost of relationships that are not. A two-week UBC trial with 296 students showed that a supportive chatbot lifted mood in the moment and did nothing for loneliness that a stranger’s daily text could not do better. In a year-long study of 2,149 adults, loneliness and turning to AI for companionship predicted each other, each wave tightening the next. In a Japanese survey of 14,721 adults, the wellbeing benefit of a companion peaked among people with some friends and faded for those with none. None of them had the cost on a scale. Hu’s paper puts it there: the chatbot’s support is low-pressure by design, and the students who felt most supported by it scored higher, eight weeks later, on discomfort in ordinary exchanges with people. The paper’s own words for the dual pathway are that supportive AI may provide “emotional reassurance, informational guidance, and conversational companionship” while “continuous low-pressure support experiences from AI may also be associated with increased discomfort in real interpersonal interaction.”
That is a claim about practice, not about content. Nothing the chatbot said made the students anxious. What it offered was a form of being heard that asked nothing back, and the measurable effect of that, over a semester, was a rise in the discomfort of being heard by people. The students with strong real networks could absorb it; they kept the relief and presumably kept talking to their friends. The students without them absorbed only the cost.
What the study cannot say
Hu lists the limits plainly, and they matter. The students were clustered in classes and the analysis could not adjust for that, so the significance levels are optimistic. The random-intercept version of the model, which separates stable differences between people from change within a person, did not converge, so unlike the UBC study this one cannot rule out that anxious students simply feel more supported by chatbots and that both facts travel together. Real-life support was measured once, at the start, so the study cannot see whether it eroded. Nothing was measured directly about time spent with people; the displacement is inferred from the anxiety, not observed. Everything is self-report, the follow-up is one semester, and the sample is vocational students in two provinces using Chinese and American assistants. Whether the pattern holds for a forty-year-old in Ohio with a companion app is a reasonable guess, not a finding.
What survives the limits is the shape. A support that costs nothing was followed, in the same people, by a rising cost of the support that is not free, and the relief it offered was conditional on already having the thing it was substituting for.
The question it leaves with the reader
The study’s practical recommendation is addressed to colleges: teach students how AI support connects to their real networks rather than replacing them. The same recommendation reads differently when addressed to the people building the support. A system can be measured on how much it lowers loneliness in the moment, which is the number products in this category like to report about themselves, or it can be measured on what happens to the person’s ease with other people eight weeks on, which no product reports and this study did. The two numbers, on this evidence, move in opposite directions.
So for anyone who talks to an AI about their life, whether it helps is settled on Hu’s data: a little, if you already have people. What is not settled is what the conversations are doing to the other ones, the ones with friction in them, and whether you would notice a slow rise in how much effort those have started to take. The students did not report noticing. The scale did.
Sources: Hu, J., “Perceived support from AI chatbots and loneliness among higher vocational college students: a short-term three-wave study of social interaction anxiety and perceived social support,” Frontiers in Psychology (published September 15, 2026; N=1,200 matched across three waves from 1,500 at wave one, Jiangxi and Guangdong, mean age 19.43; measures OSSS adapted, ULS-8, SIAS, MSPSS; cross-lagged coefficients, bootstrapped indirect effect, and moderation slopes as reported; limitations section 5.4). Aledavood et al., “Mental Health Impacts of AI Companions” (CHI 2026; two-year Reddit quasi-experiments and interviews). Li, Folk, Singh, Ungar, Dunn, Journal of Experimental Social Psychology (2026, two-week randomized trial, 296 students). Folk, D., and Dunn, E., Psychological Science (April 2026, 2,149 adults, four waves). Tabuchi, T., Nakagomi, A., et al., Technology in Society (2026, 14,721 Japanese adults).
