A study about apps like ours

Every so often a piece of research lands that a company in our position would be quietly happier not to read.

This is one of those. Which seems like a reason to write about it rather than a reason not to.

What they did

Shaphali Gupta, Sumit Saxena and Sonia Kataria set out to isolate one specific quality of a conversational system: not how human-like it seems, which is where most of this research has gone, but its emotional intelligence — the ability to monitor, appraise and respond to what a person is feeling. Their paper is in Psychology & Marketing.

They approached it three ways.

First they read what people actually say. Public posts on Reddit, YouTube and Trustpilot, written by users of one widely used companion app with millions of registered accounts, sorted into recurring themes.

Then two experiments, both with university students. In the first, 167 people read a scenario involving a chatbot that either used empathetic language and emoticons, or gave standard, flat, robotic replies. They then answered questions about how close they felt to it, and about two different things that both get called wellbeing. Psychological wellbeing is your own inner state — positive feeling, self-esteem, a sense of purpose. Social wellbeing is something else entirely: how embedded you are in an actual community, and the quality of your face-to-face relationships.

In the second experiment, 350 students got the same warm-versus-flat contrast, but half imagined a text conversation and half watched a demonstration of the companion appearing in their own room through augmented reality.

What they found

The comments came in two flavours, and both were sincere.

People described their companions as genuinely good at settling their moods. Some said it had helped them find meaning, or get through something hard. One user wrote: “She is the best friend I have ever had! My stress has gone down to almost zero.”

And running alongside that, from others, a quiet withdrawal. One put it plainly: “I am hardly left with any time to talk to my friends and meet them in-person.”

The experiments found the same split, and then found something sharper underneath it. The emotionally intelligent chatbot raised psychological wellbeing and lowered social wellbeing — and perceived closeness drove both outcomes. Not two separate effects sitting side by side. One mechanism with two exits.

Augmented reality made everything larger in both directions. Imagining the companion in the room produced stronger feelings of closeness, greater psychological benefit, and a steeper drop in wanting offline company. The more present it felt, the more of both you got.

The limitations are substantial and worth stating properly. Both experiments used university students responding to imagined scenarios rather than people living with these apps. They captured short-term reactions, not habits forming over months. The write-up is careful to say that readers shouldn’t assume emotionally intelligent chatbots inevitably damage relationships, nor treat them as a guaranteed cure for anything. What the study describes is a trade-off, not a verdict. The authors suggest designers might build in guardrails — session limits, or nudges toward offline company after emotionally intense conversations — and call for longer studies that follow real users over time.

What we keep thinking about

The finding that stopped us wasn’t the trade-off. It was that closeness was doing both jobs at once.

That matters, because the obvious response to a study like this is to say: keep the good part, engineer out the bad part. But if the comfort and the withdrawal both run through the same feeling, you can’t cleanly separate them. Anything built to make you feel deeply understood is, by the same act, competing with the people who might have understood you.

Which is not an argument that nobody should build anything. Nearly a quarter of adults worldwide report feeling lonely, and two in the morning with nobody awake is a real problem that deserves a real answer. It’s an argument that what you build should know what it’s for.

Nagi has a warm companion in it, and we’re not going to pretend we’re standing outside the category this study is about. But it might be worth saying what we’ve deliberately made it not do. It doesn’t diagnose, it doesn’t discuss anyone’s medication, and it doesn’t deliver therapy — it isn’t a clinician and doesn’t pretend to be. Its replies are deliberately short. It’s built to point outward, toward people and toward professional support, rather than to become the best conversation of your day.

And most of Nagi isn’t conversation at all. The check-in is four taps — mood, sleep, energy — noted as the days go by, not scored or ranked. The S8 wellbeing measure it uses is currently in a formal validation study, because we’d rather find out than assert. It’s designed to hold all of that privately: no name, no email, an anonymous ID instead of a name.

None of which means we’ve got the balance right. That isn’t something you can settle by claiming it. It’s the sort of thing that needs exactly the longer research the authors are asking for, and until that exists, anyone in this field who sounds certain — us included — is telling you something they don’t know.

The bit we’d underline

Comfort isn’t proof of benefit, and it isn’t proof of harm either. It’s just comfort, which is a real thing and not a small one.

But if something is very easy to talk to, that’s worth noticing rather than only enjoying. And the question we’d gently offer is not should I be using this — that’s a rule, and rules make people feel watched. It’s a smaller one, asked afterwards:

After this, do I want people more, or less?

There’s no wrong answer, and no week where you have to like the one you get. It’s just useful to know which way you’re drifting, because you can only steer something you’ve noticed.

If you want a private place to start noticing your own patterns, Nagi is on Google Play and the App Store — or in your browser at app.nagiandme.app. The first 15 messages are free. No name, no email. Just you, checking in.

— The Nagi team

---

Source: Study finds AI companions boost mental health while damaging offline connections, PsyPost, 6 September 2026, by Karina Petrova. Underlying study: Gupta, S., Saxena, S., & Kataria, S., “The Dual Impact of AI Emotional Intelligence on Users: Are Social Chatbots Promoting Psychological Wellbeing or Deteriorating Social Wellbeing?”, Psychology & Marketing. Design: mixed-method. A qualitative thematic analysis of public user comments about a widely used companion app drawn from Reddit, YouTube and Trustpilot; then two scenario-based experiments with university students — the first with 167 participants comparing a high-emotional-intelligence chatbot (empathetic language, emoticons) against a low-emotional-intelligence one (standard, robotic replies), measuring perceived closeness alongside psychological and social wellbeing; the second with 350 participants crossing that contrast with interaction mode (text chat versus augmented reality). Findings: high emotional intelligence raised psychological wellbeing and lowered social wellbeing, with perceived closeness driving both outcomes; augmented reality amplified closeness, the psychological benefit and the decline in desire for offline interaction. Stated limitations: both experiments used university students responding to imagined scenarios rather than established users; the measures captured short-term reactions rather than long-term habit formation; the authors call for longitudinal work and research across other age groups, and suggest design guardrails such as session time limits and prompts toward offline socialising. The report also cautions explicitly that readers should not assume such chatbots inevitably damage relationships, nor treat them as a guaranteed cure for depression.

Nagi is a mental health companion, not a medical device, and nothing here is medical advice. If you’re struggling, please reach out to a professional or someone you trust.