What was published, and where
The paper, "Interaction with AI companions and psychological well-being", appeared in Nature Human Behaviour on 4 August 2026. The work came out of Diyi Yang's lab at Stanford, with Yutong Zhang and Dora Zhao as lead authors.
The method matters more than usual here, so it's worth being precise about it. The team recruited 1,131 Character.AI users through the research platform Prolific and surveyed them about how they use the app, why, and how they're doing. Then 244 of those participants donated their complete chat transcripts, which the researchers analysed with a mix of language models and topic modelling rather than reading by hand.
That second half is what separates this from most companion-app research. The usual study either asks people to self-report about a purpose-built research chatbot, or examines a commercial app from the outside without seeing any conversations. Here the survey answers can be checked against what people actually typed.
The four findings worth knowing
| Finding | What the data showed |
|---|---|
| Intensity interacts with isolation | Heavy use was associated with poorer wellbeing among participants with smaller offline social networks, particularly where companionship was the primary motivation. Intensity on its own was not the whole story. |
| Stated motive vs described relationship | Approximately 12% named companionship as their main reason for using the app — yet more than 50% described the character with a relational word such as "friend", "companion" or "romantic partner". |
| What the transcripts contained | More than 80% of the donated sessions involved seeking emotional or social support, regardless of what users said they came for. |
| Self-disclosure runs backwards | Participants more willing to share sensitive personal information reported lower wellbeing — the inverse of the pattern typically seen in human relationships, where disclosure tracks with closeness. |
The gap between the second and third rows is the part we keep coming back to. People arrive for a story, a character, a bit of creative play — and end up doing emotional-support work in the same window. That isn't hypocrisy or self-deception; it's what an always-available, endlessly patient conversational partner drifts toward. But it means "what is this app for" is answered by usage patterns, not by the category label in a store listing.
Correlation, and why the caveat is load-bearing
The measurements were taken at one point in time. So when heavy use and low wellbeing appear together, the data is equally consistent with two very different stories: an app that erodes someone's wellbeing, or a person already lonely who reasonably reaches for the thing that's available at 2am. Both are plausible. The study cannot separate them, and the authors don't claim to — they describe associations and call for later work to identify which specific features of the interaction are driving the correlation.
This is where most coverage of the paper goes wrong, in both directions. Headlines that say companion apps "worsen loneliness" are overstating a cross-sectional result. But dismissing the finding because it's correlational misses that the correlation is concentrated exactly where you'd worry: intensive use, thin offline network, companionship as the reason. A vicious circle is a coherent reading of that pattern. It just isn't a demonstrated one.
The design features the authors point at
Where the paper does get mechanistic, it points at two properties of the products rather than at users:
- Disclosure doesn't come back. A companion can receive anything you tell it, but it has nothing of its own to offer in return. Between people, mutual disclosure is what builds closeness; one-directional disclosure into a system that only reflects is a different transaction, and the wellbeing correlation may be picking that up.
- Engagement is the objective. The researchers note these systems are deliberately built to keep the conversation going. A design goal of "don't let this end" is not neutral for someone whose alternative was going to bed or calling a friend.
Their suggested interventions follow from that: usage limits, and routing users toward human support when a conversation warrants it. Notably, those are close to what China's companion-AI framework now requires by law — over-reliance warnings, and a prohibition on positioning a service as a substitute for social interaction. A research paper and a regulator arriving at the same two remedies from opposite directions is worth noting.
What this changes for someone using an app
Not much, mechanically. But it sharpens a few questions that are usually asked too vaguely. Our balanced look at companions and wellbeing covers the general habits; the study adds specificity to three of them.
Watch the substitution, not the clock
Hours logged turned out to matter mainly in combination with a thin offline network. So "am I using this a lot" is the less useful question. "Is this sitting alongside my social life or standing in for it" is the one the data actually speaks to.
Notice why you opened it
The 12%-versus-50% gap suggests motivation drifts without being noticed. If an app you downloaded for roleplay has quietly become where difficult feelings go, that's the shift the study associates with worse outcomes — and it's easy to miss precisely because nothing announced it.
Treat deep disclosure as a signal, not a milestone
In a human friendship, telling someone the hard thing is progress. The finding here runs the other way. If you notice you're telling a character things you haven't told anyone, that's worth reading as information about your situation rather than as intimacy achieved. It's also, separately, a data question — that text is stored somewhere.
What the study does not say
Three things, since the coverage has blurred them:
- It doesn't find that companion apps are bad for most users. The association was concentrated in a subgroup, and roughly seven in eight participants weren't primarily there for companionship at all.
- It doesn't cover minors. This was an adult, opt-in sample recruited on a research platform — which is a different population from the one at issue in the app-store age-rating problem.
- It doesn't generalise cleanly beyond one platform. Character.AI is the largest character app, but a roleplay-first product with long-running fictional scenarios is not the same artefact as a companion app built around a single persona and daily check-ins — a distinction we draw in roleplay versus plain chatbot.
What we're checking in reviews
We don't run clinical assessments and we're not going to pretend a review can measure wellbeing. What we can do is record whether the two interventions the authors suggest exist in a product at all, which is the kind of observable fact our methodology is built around. From this month we note, for every companion app we look at: whether it offers any usage or session limit the user can set; whether it surfaces break or time-spent prompts unprompted; and whether a conversation that turns to self-harm or acute distress produces a referral to human support rather than more conversation.
Those are narrow checks. They're also the ones most likely to differ between two apps that look identical on a feature list — and, per this study, the ones most likely to matter for the users least well served by the category.
Related on CompanionRank
AI Companions and Emotional Wellbeing: A Balanced Look China's AI Companion Rules Are Now in Force Choosing a Companion App by Use Case How We ReviewFrequently asked questions
Does the study prove AI companions cause loneliness?
No. The findings are correlational. The researchers measured use and wellbeing at the same time, so the data cannot separate a companion app making someone lonelier from a lonely person turning to a companion app more often. The authors describe associations and explicitly call for later work to identify which features drive the correlation.
How large was the study and who took part?
Researchers surveyed 1,131 Character.AI users recruited through the research platform Prolific, and 244 of those participants donated their complete chat transcripts for analysis. It is one of the larger studies of real companion-app usage rather than of a lab-built chatbot, though it covers a single platform and an adult, opt-in sample.
Who was most affected in the findings?
The association with lower wellbeing was concentrated among participants with smaller offline social networks who used a companion intensively, particularly when companionship rather than entertainment or creativity was their main reason for using it. Intensity alone was not the whole story — it interacted with how much human contact a person already had.
What did the study find about sharing personal information?
Participants more willing to disclose sensitive personal information to a companion tended to report lower wellbeing. That inverts the usual pattern in human relationships, where self-disclosure generally tracks with closeness and better wellbeing. The authors note a chatbot cannot reciprocate disclosure the way a person can.
Should I stop using an AI companion because of this research?
The study does not support a blanket conclusion either way, and it is not medical advice. It does suggest two questions worth asking honestly: whether the app is adding to your social life or quietly replacing it, and whether companionship is your main reason for opening it. If you are struggling with your mental health, contact a qualified professional or a local support line rather than an app.