What was published, and where
The paper is 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, by Jiangnan Hu of Yingtan Vocational and Technical College, in Frontiers in Psychology Volume 17 (DOI 10.3389/fpsyg.2026.1928350). It is open access, so every number below can be checked against the source.
Students were recruited through class groups at four higher vocational colleges in Jiangxi and Guangdong provinces during the spring semester of the 2025–2026 academic year. The valid wave-1 sample was 1,500; the matched three-wave sample used for the model was 1,200, an 80% retention rate, with no meaningful baseline differences between those who stayed and those who dropped out. The sample was 60.2% female, mean age 19.43 (SD 1.28, range 17–24), and reported an average of 38.5 minutes of AI use per day.
Four instruments carry the analysis: the UCLA Loneliness Scale short form (ULS-8) for loneliness, the Social Interaction Anxiety Scale (SIAS), the Multidimensional Scale of Perceived Social Support (MSPSS) for real-life support, measured at wave 1 only, and an adapted Online Social Support Scale for perceived support from AI chatbots. That last construct matters for how you read the result: it measures the felt support a student gets from AI chat tools in general, not usage of a dedicated companion product. The revised cross-lagged model fit at CFI 0.971, RMSEA 0.077 and SRMR 0.014.
The three findings, in order
1. AI support predicted less loneliness — but not immediately. The wave 1 to wave 2 path was β = −0.043 with p = .064, which is not significant. The wave 2 to wave 3 path was β = −0.166, p < .001. In unstandardised terms, a one-standard-deviation rise in perceived AI support (12.92 points) corresponded to roughly a 0.92-point fall in loneliness on a scale that runs from 8 to 32. That is a real effect and a small one.
2. The same AI support predicted more social interaction anxiety. β = 0.161 from wave 1 to wave 2 and β = 0.158 from wave 2 to wave 3, both p < .001. This path was larger and more consistent than the loneliness benefit, and it survived every sensitivity specification the author ran.
3. Anxiety fed straight back into loneliness. β = 0.198 and β = 0.232 across the two intervals, p < .001. The indirect route — AI support, then anxiety, then loneliness — came out positive at 0.038, 95% CI [0.026, 0.050]. So one variable pushes loneliness down through a direct route and back up through an indirect one, which is exactly why single-number summaries of this literature keep contradicting each other.
The condition: it worked for students who already had people
The moderation analysis is the part worth carrying away. Real-life perceived social support, measured at wave 1, changed what AI support did later.
| Student's real-life support at wave 1 | Slope: AI support (wave 2) → loneliness (wave 3) | Significant? |
|---|---|---|
| High | −0.287 | Yes, p < .001 |
| Low | −0.023 | No |
Read plainly: the compensatory benefit showed up for students who already had a functioning social network, and effectively vanished for the students a chatbot is most often proposed as a remedy for. Baseline real-life support was also the strongest single predictor in the model of later loneliness (β = −0.449 from wave 1 to wave 3), and it independently predicted lower social anxiety at wave 3 (β = −0.166).
The author's framing is that the psychological meaning of AI support is conditional — it works as a supplement to a social life, not as a substitute for one. That is a narrower claim than either "AI companions cure loneliness" or "AI companions cause it", and it is the claim the data actually supports.
Why social anxiety moved the other way
The study measures the association; it does not prove the mechanism. The interpretation the author offers is the obvious one, and it is worth stating carefully because it is testable rather than moralising.
A chat interface removes the specific things that make human conversation effortful: the wait, the read on someone's face, the risk of being misread, the cost of saying the wrong thing. Social interaction anxiety, though, normally declines through exposure to exactly those conditions. A support channel that is always available and never awkward can relieve the immediate feeling while quietly removing the practice. Nothing in a three-wave panel can confirm that chain, and the author does not claim it does.
It does line up with what we covered in research on validation and anthropomorphism, where the design features that make a companion feel good in the moment are not always the ones that leave a user better off a month later.
Four 2026 datasets, one shape
This paper is not an outlier. Four separate 2026 datasets, using different methods in different countries, keep landing on the same conditional shape.
| Source | Design and sample | What it reports |
|---|---|---|
| Frontiers in Psychology, 15 Sep 2026 | Three-wave panel, 1,200 vocational college students, China | AI support → lower loneliness only where real-life support was high (slope −0.287 vs −0.023) |
| Nature Human Behaviour, 4 Aug 2026 | Survey of 1,131 US Character.AI users plus 4,664 chat sessions (464,687 messages) from 237 of them | Smaller offline networks → companionship as the primary use (β = −0.03) → lower well-being (β = −0.48); stronger when use was intensive (β = −0.31) or highly disclosive (β = −0.38) |
| Acta Psychologica, 2026 | Cross-sectional, 233 young adults, India | Small and non-significant links: loneliness r = −0.18, social anxiety r = −0.12; authors call the findings exploratory |
| arXiv preprint, 2 Jun 2026 | Longitudinal, 28 days of daily five-minute conversations about personal issues | Preference for seeking support from humans fell approximately 10.3%; preference for AI rose approximately 11.6% |
The August Nature Human Behaviour paper, which we covered in detail in our piece on that study, reached its conclusion from real chat logs rather than self-report and arrived at the same moderator: offline social environment. Two very different methods converging on one condition is stronger evidence than either alone.
What to do with this if you use a companion app
None of these studies tested a specific product, so nothing here is a verdict on any app. What they do support is a short set of questions worth running on your own use every few weeks.
- Ask what the chat replaced, not what it added. A conversation that filled an empty evening is different from one that replaced a call you would otherwise have made. The second is the pattern the data flags.
- Watch the direction over a month, not a session. Every study here found the immediate experience positive. The disagreement is about what happens eight weeks later.
- Treat disclosure depth as a dial. The Nature Human Behaviour data found the well-being association most negative among the most intensive and most disclosive users, not among casual ones.
- Keep one human channel live. Both the moderation result and the offline-network result say the same thing: the app works better as a supplement than as the whole system.
- If your offline circle is currently thin, be deliberate. That is precisely the group where the measured benefit disappeared, which is an argument for pairing the app with something else, not for going without.
For the practical side of choosing a tool rather than reading a study, our notes on companion apps and wellbeing and everyday use patterns cover the feature-level decisions.
What the study cannot tell you
- One country, one education segment. Higher vocational college students in two Chinese provinces are not a stand-in for adult companion-app users anywhere else.
- Cross-lagged panels are not experiments. They rule out some reverse-causation stories and no others. The author is explicit that the design is correlational.
- Everything is self-reported. Perceived support from AI is a feeling, not a log of usage, and it was not validated against what students actually did.
- The sample was unusually distressed. The share meeting the PHQ-9 threshold for at least moderate depressive symptoms was 69.5%, 55.0% and 72.9% across the three waves. That is far above typical undergraduate estimates and is a real caution about generalising the effect sizes.
- Model fit was acceptable, not excellent. CFI 0.971 is good; TLI 0.910 and RMSEA 0.077 are adequate, and the model needed extra long-range paths to get there.
- No commercial companion product was tested. The construct is support from AI chat tools generally.
If you want the population-level picture rather than the mechanism, our summary of how many people actually use AI companions in 2026 covers the survey side, and the older-adults research shows the same conditional pattern in a very different age group.
Related on CompanionRank
AI Companions and Loneliness: What the Stanford Study Found AI Companion Apps and Wellbeing Older Adults and AI Companions: The 2026 Research How Many People Use AI Companions in 2026?Frequently asked questions
Does AI chat support reduce loneliness?
In this study it did, but only under a condition. Perceived support from AI chatbots at wave 2 predicted lower loneliness at wave 3 with β = −0.166 (p < .001), roughly a 0.92-point fall on an 8–32 scale per standard deviation. When the sample was split by real-life social support, the effect held for students with high real-life support (slope −0.287, p < .001) and disappeared for those with low real-life support (slope −0.023, not significant).
Why would AI support increase social anxiety?
The study measured the association, not the mechanism: perceived AI support predicted higher social interaction anxiety at the next wave (β = 0.161 and 0.158, both p < .001). The author's interpretation is that a chat channel removes the friction that ordinarily reduces social anxiety through exposure — the wait, the uncertainty, the risk of being misread. A three-wave panel cannot confirm that chain, and the paper does not claim it does.
How big is the effect in plain terms?
Small. One standard deviation more perceived AI support (12.92 points on the scale used) corresponded to about a 0.92-point drop in loneliness on the ULS-8, which runs from 8 to 32. For comparison, baseline real-life social support predicted later loneliness far more strongly, at β = −0.449 from wave 1 to wave 3.
Does this mean companion apps are bad for lonely people?
No. It means the measured benefit did not show up for the students with the least real-life support, so an app should not be treated as a replacement for a social network. None of the studies here tested a commercial companion app, and none of them demonstrate that using one causes loneliness.
Is this the same as the Nature Human Behaviour study from August?
No. That one, published 4 August 2026, surveyed 1,131 US Character.AI users and analysed 4,664 real chat sessions from 237 of them, and found that smaller offline social networks were associated with using a chatbot primarily for companionship, which in turn was associated with lower well-being (β = −0.48). It is a different sample and method that reaches the same conditional conclusion.
Can I rely on a single study?
No, and you do not have to. Four independent 2026 datasets — the Frontiers three-wave panel, the Nature Human Behaviour chat-log study, a cross-sectional Acta Psychologica sample of 233 young adults in India, and a 28-day longitudinal preprint — all report that the effect of AI support depends on the user's offline situation rather than being uniformly good or bad.