What was published, and by whom
The paper is Living with AI Companions: Sustained AI Companionship Predicts Lower Well-Being Through Lower Human Interaction, posted to arXiv on 7 September 2026 as arXiv:2609.07243 under a CC BY 4.0 licence, so every figure below can be checked at the source. The authors are Yutong Zhang, Dora Zhao, Yixin Wang, Rebecca Anselmetti, Jeffrey T. Hancock, Robert Kraut and Diyi Yang, across Stanford University, the University of Michigan, the University of Oxford and Carnegie Mellon University.
Participants were recruited through Prolific: US-based, native English-speaking Character.AI users who had used the platform for more than a month and talked to at least three different chatbots. Baseline ran from September 2024 to August 2025 (N = 1,182); follow-up ran from August 2025 to June 2026 (N = 439), a mean of 362.5 days later (SD = 88.2). Well-being was measured with the Comprehensive Inventory of Thriving. The study was approved by the Stanford IRB (protocol 76072).
One design detail matters for reading the results: of the 439 who returned, 108 had stopped using chatbots altogether, and their follow-up engagement was coded as zero. The authors re-ran everything on the 331 continuing users as a robustness check and report consistent findings.
Engagement did not stay where it started
The first result is about trajectory rather than harm. How intensively someone used chatbots at baseline predicted not only continued intensity a year later but also more companionship-oriented use and deeper self-disclosure.
| Baseline measure | Predicts at follow-up | Standardised estimate |
|---|---|---|
| Interaction intensity | Interaction intensity | β = 0.43, 95% CI [0.31, 0.54] |
| Interaction intensity | Companionship use | β = 0.32, 95% CI [0.20, 0.43] |
| Interaction intensity | Self-disclosure | β = 0.28, 95% CI [0.16, 0.40] |
| Companionship use | Companionship use | β = 0.49, 95% CI [0.22, 0.75] |
| Companionship use | Interaction intensity | β = 0.40, 95% CI [0.13, 0.66] |
The authors' reading is that social engagement develops through repeated use rather than being fixed by why someone downloaded the app. People who arrived for something instrumental could still end up in a relational pattern a year later. That is a meaningful point for general-purpose assistants too, not just products marketed as companions.
How users described the relationship shifted in the same direction, if modestly. Among the 331 still using chatbots at follow-up, "supportive companion" rose from 40.2% to 42.3% and became the single most common description, while "tool, assistant, system" fell from 43.8% to 40.5%. Romantic or intimate framing stayed small and nearly flat, 3.3% to 3.6% — a useful corrective to coverage that treats AI romance as the dominant use. More than half of participants described several kinds of use at once in both waves.
The mechanism: displaced time, not the chat itself
This is the part that is genuinely new. The team tested two competing explanations for the well-being association and only one held up cleanly.
Baseline engagement did not directly predict follow-up well-being once baseline well-being was accounted for — the coefficients sat around β = −0.03 to −0.04 and none reached significance. What did carry the association was what people were still doing a year later, and specifically what that use had replaced.
| Follow-up engagement dimension | → time with people in person | Indirect link to well-being | Direct link, with in-person time in the model |
|---|---|---|---|
| Interaction intensity | β = −0.17 (p < 0.001) | −0.02 (p < 0.001) | β = −0.06, not significant |
| Companionship use | β = −0.14 (p = 0.002) | −0.02 (p = 0.004) | β = −0.05, not significant |
| Self-disclosure | β = −0.15 (p = 0.002) | −0.02 (p = 0.002) | β = −0.05, not significant |
In-person time in turn predicted higher well-being at β = 0.14 to 0.15 across all three models. So the chain the data supports is: heavier sustained use, less face-to-face contact, lower well-being. Once you hold in-person time constant, the chatbot variable stops doing statistical work of its own.
The effect sizes are small. An indirect coefficient of −0.02 is not a crisis in any individual's life, and the paper does not present it as one. What makes it interesting is that it is consistent across three independent ways of measuring engagement, and that it names a variable you can actually check on yourself.
Independent evidence pointing the same way
A single preprint is not a conclusion. The reason this one is worth taking seriously is that separate teams using different methods keep landing on the same trade-off.
| Study | Design | What it found |
|---|---|---|
| arXiv preprint, 7 Sep 2026 (Zhang et al.) | Two-wave panel, 1,182 → 439 Character.AI users, about 12 months | Sustained engagement → less in-person time → lower well-being; no significant direct effect |
| Psychology & Marketing (Gupta, Saxena & Kataria) | Two experiments, 167 and 350 students | Emotionally intelligent chatbots raised psychological well-being while lowering social well-being, via perceived closeness; an AR interface amplified both sides |
| Stanford HAI report, 4 Aug 2026 | Survey of 1,131 Character.AI users plus chat transcripts | Only 12% named companionship as their main reason for using the bot, yet more than half described it as a friend, companion or romantic partner |
The experimental work matters because it tests the direction a panel study cannot: when researchers manipulated how emotionally intelligent a chatbot appeared, the psychological benefit and the social cost moved together. That is the same shape as the longitudinal finding, arrived at from the opposite end.
The August 2026 numbers come from the earlier wave of this same cohort, which we covered in our write-up of the Stanford loneliness study. A three-wave student panel published in September reached a compatible conclusion from yet another angle — the benefit of AI support showed up mainly for people who already had human support, as we covered in that study's write-up.
A five-minute check you can run on your own use
None of these studies tested a specific product, and nothing here is a verdict on any app. But the displacement result points at one variable that is easy to observe without any instruments.
- Estimate your in-person hours this week, not your chat hours. The chatbot number is not what predicted well-being in this data. The other one was.
- Ask what a long session replaced. An evening that was already empty is a different case from an evening where you cancelled something.
- Compare against a year ago, not yesterday. The trajectory result says engagement tends to grow quietly. A monthly glance at your usage screen catches what a daily one misses.
- Notice if the framing has shifted. In the data, "tool" drifted toward "supportive companion" over a year. That drift is normal; it is worth being aware of rather than alarmed by.
- Watch disclosure depth separately from time. Self-disclosure predicted displacement on its own, independent of how many hours were logged.
- Keep one reciprocal channel live. The authors' own design suggestion is companions that support human relationships rather than compete with them; you can apply that at the user level before any product does.
If you want the feature-level view rather than the research view, our notes on companion apps and wellbeing and on memory and continuity cover how the design choices that drive sustained engagement actually work.
What this study cannot tell you
- It is a preprint. arXiv:2609.07243 has not completed peer review, and numbers can change before formal publication.
- Two waves are not many. The authors are explicit that they cannot characterise reciprocal dynamics over longer periods, and call for three or more waves or continuous behavioural logging.
- Retention was 37.1%. 439 of 1,182 returned. The main driver was structural — 59.54% of baseline participants were no longer active on Prolific — and a Heckman selection correction was applied, but the people who never returned cannot be recovered.
- The companionship measure changed between waves. Baseline used a forced single choice; follow-up let people rate several use types. The authors flag this as limiting direct comparability of that construct.
- One platform, one country. US-based English-speaking Character.AI users recruited on Prolific. Other platforms and other populations may behave differently.
- Platform changes were unobserved. Character.AI's own model, moderation and design changes during the study window could not be accounted for.
- In-person time was a single self-reported item. It is a reasonable proxy, not a behavioural log.
For the population-level picture rather than the mechanism, our summary of how many people actually use AI companions in 2026 covers the survey side.
Related on CompanionRank
AI Companions and Loneliness: What the Stanford Study Found Does AI Chat Support Actually Ease Loneliness? AI Companion Apps and Wellbeing AI Companions for Everyday LifeFrequently asked questions
What does the 12-month follow-up actually show?
That heavier use at the start predicted heavier use a year later, and that sustained heavy use was associated with lower well-being at follow-up. Baseline engagement on its own did not predict later well-being (all coefficients around β = −0.03 to −0.04, none significant). The association ran through what people were still doing a year on, not through where they started.
Why does in-person time matter in this study?
It is the mechanism the data supports. Heavier follow-up engagement was associated with less time spent with other people in person (β = −0.17 for interaction intensity, −0.14 for companionship use, −0.15 for self-disclosure), and more in-person time was associated with higher well-being (β = 0.14 to 0.15). Once in-person time was in the model, the direct link between chatbot engagement and well-being was no longer statistically significant.
Does this prove AI companions cause lower well-being?
No. Two survey waves cannot establish causation, and the authors do not claim they do. The design rules out some reverse-causation stories because engagement was measured before well-being at follow-up, but people who were already withdrawing socially could also have been the ones who escalated their chatbot use.
How reliable is the sample after a year?
This is the study's weakest point and the authors say so. Of 1,182 baseline participants, 439 returned, a retention rate of 37.1%. Most of the loss was structural rather than selective: 59.54% of baseline participants were no longer active on the recruitment platform. The models used a Heckman selection correction to adjust for observable differences, which helps but cannot recover the people who never returned.
Did people become more attached to their chatbots over the year?
Modestly, yes. Among the 331 who were still using chatbots at follow-up, describing the bot as a supportive companion rose from 40.2% to 42.3% and became the most common description, while tool or assistant framing fell from 43.8% to 40.5%. Romantic or intimate framing stayed small and almost flat, 3.3% to 3.6%.
Is this the same as the Stanford study from August 2026?
It is the same research group and the same starting cohort, but a different question. The August 2026 work described who was at risk at a single point in time. This preprint follows the same users about a year on and tests how engagement develops and what explains the link to well-being.