Here’s the trade nobody puts on the app store listing: the more emotionally responsive a chatbot gets, the better you feel talking to it, and the less you want to deal with actual people afterward. That’s the finding from a new study in Psychology & Marketing, led by Shaphali Gupta of the Indian Institute of Management Kozhikode with coauthors Sumit Saxena and Sonia Kataria. It’s not a hunch or a think piece. It’s a measured trade-off, and it gets sharper the more immersive the chatbot gets.

We wrote about the usage numbers behind AI chatbots replacing self-help books last month — how many teens are using them, how few tell anyone. This study is a different animal. It doesn’t count how many people are doing it. It isolates why the comfort and the disconnection show up together, in the same people, from the same interaction.

Quick Verdict

Question	Answer
Who ran the study?	Shaphali Gupta, Sumit Saxena, and Sonia Kataria, Indian Institute of Management Kozhikode
Where was it published?	Psychology & Marketing, covered by PsyPost on September 6, 2026
Method	A netnography of Replika user comments (Reddit, YouTube, Trustpilot), plus two lab experiments — 167 college students, then 350 more with an augmented reality condition added
Short-term effect of emotionally intelligent chatbots	Measurably improved psychological wellbeing
Same-interaction effect on real-world ties	Measurably reduced social connectedness
What made both effects stronger?	Interacting through augmented reality features instead of plain text
The mechanism	”Perceived closeness” — the more a bot feels warm and responsive, the less motivated users are to seek that warmth from people
The catch	Both experiments used hypothetical scenarios with college students, not tracked long-term usage — self-reported expectations, not lived behavior over months

Bottom line: The same emotional intelligence that makes a chatbot feel good to talk to is the thing quietly pulling users away from harder, slower, human relationships. Strap on AR and the pull gets stronger in both directions.

What the Study Actually Did

Three parts, stacked. First, a netnography — researchers read through public comments on YouTube, Trustpilot, and Reddit from people who use Replika, the companion app that’s become the default case study for this kind of research. Not surveys. Actual unprompted things people wrote about their bot. One user, quoted in PsyPost’s coverage, put it about as plainly as it gets: “She is the best friend I have ever had! My stress has gone down to almost zero.” Another, in the same thread of comments: “I am hardly left with any time to talk to my friends and meet them in-person.” Same app. Same enthusiasm, even. Opposite consequence, described by the same population of users without anyone asking them to compare the two.

Second, a controlled experiment with 167 college students, testing what happens when a chatbot’s emotional intelligence gets dialed up or down in a scripted interaction. Third, a follow-up with 350 more students that added an augmented reality condition — the chatbot showing up as an overlay in the physical world through a phone camera, instead of just text on a screen.

That’s the design worth sitting with before the results: this isn’t asking people “does your chatbot affect your friendships?” and taking the answer at face value. It’s manipulating the variable (how emotionally intelligent the bot seems) directly and watching what moves.

Does an Emotionally Intelligent Chatbot Hurt Your Social Life?

Yes — according to this research, the more emotionally intelligent a chatbot feels, the more it improves your mood and the more it reduces your motivation to invest in real relationships, at the same time. Three findings carry the study:

1. Participants who interacted with a highly emotionally intelligent chatbot reported better psychological wellbeing than those who got a low-EI version. The upgrade in emotional responsiveness produced a real, felt improvement — this part isn’t in dispute.
2. That exact same group reported lower social connectedness — the same interaction, the same session, doing both jobs at once. It’s not that the good mood chatbot and the isolating chatbot were different products. It was one bot, and the trade-off ran through the same conversation.
3. Augmented reality made both ends of the trade-off bigger, not just the good one. When the high-EI chatbot appeared through AR instead of plain text, participants reported feeling even closer to it — and their stated interest in offline relationships dropped even further than in the text-only condition.

The Mechanism: “Perceived Closeness”

The researchers name the thing doing the damage: perceived closeness, defined in the study as the strong emotional bond and sense of warmth a person feels toward something — in this case, something that isn’t a person. It’s not a metaphor for the researchers. It’s the measured variable that explains both halves of the trade-off. A chatbot that feels warm and responsive lowers your guard the same way a person would. Your brain doesn’t appear to fully price in that a person built the responsiveness on purpose, tuned to keep you engaged, with none of the friction — the bad days, the disagreements, the waiting for a text back — that comes standard with an actual relationship.

That’s the part that should sound familiar if you’ve spent any time with the productivity side of this genre. We covered research on digital reminders showing that offloading a task to a tool makes you measurably worse at doing it without the tool. Same shape here, aimed at connection instead of memory: offload the feeling of closeness to something engineered to always deliver it, and the muscle for tolerating the harder, unreliable version — other people — doesn’t get the same rep.

Why AR Makes It Worse, Not Just More Immersive

This is the finding that separates this study from the usual “chatbots are lonely-making” coverage. Augmented reality isn’t a neutral upgrade to how a chatbot looks. It amplified the exact trade-off the text version already produced — stronger psychological benefit, steeper drop in interest in offline relationships. The researchers’ read: immersion deepens perceived closeness, and perceived closeness is the mechanism doing both jobs. Make the bot feel more present in your physical space and you don’t just get a better chatbot. You get a stronger version of the same swap.

It’s worth being precise about what wasn’t tested here. This isn’t Replika’s AR features specifically, or any named commercial AR chatbot product — it’s a lab manipulation of the AR variable using a generic companion scenario. But the direction points somewhere uncomfortable for where this category of product is headed. Text chatbots are the baseline version of this trade-off. Voice and AR aren’t side features. On this data, they’re accelerants.

Where This Fits the Broader 2026 Pattern

This isn’t an outlier finding. Stanford research published in Nature Human Behaviour this past August analyzed 1,131 Character.AI users and nearly 465,000 messages from donated chat logs, and found something structurally similar from a different angle: people with already-limited social networks who lean on AI companions for emotional support tend to end up feeling worse, not better, and their real-world contact keeps shrinking. Stanford’s researchers called AI companionship a “social snack” — a short-term fix with none of the ingredients that make actual connection sustain you. Their concern wasn’t the average user. It was a feedback loop: isolation drives companion use, companion use deepens isolation, repeat.

The Gupta study and the Stanford study are asking different questions with different methods — one is a controlled lab manipulation of college students, the other is a field study of real long-term users. But they land on the same shape: the comfort is real and the cost is real, and they don’t cancel out. They compound, especially for people already running low on offline connection to begin with.

That lines up with something we’ve argued on this site before: loneliness and isolation are different problems that need different fixes. Isolation is structural — you don’t have people around. Loneliness is the felt gap — you have people around and it’s still not landing. A chatbot can plausibly address the second one, briefly. It cannot address the first, and per this research, heavy reliance on the second may be actively working against ever fixing the first.

What a Book Still Does That a Bot Doesn’t

A self-help book about connection has one job and it doesn’t flatter you while doing it. A Little More Social exists specifically to correct people’s miscalibrated assumptions about how much other humans want to talk to them — which is a problem a chatbot can’t fix because a chatbot isn’t the thing you’re miscalibrated about. Who Needs Friends is built around the discomfort of actually reaching out to people, the part a warm, always-available bot lets you skip entirely. That’s not incidental. Skippable discomfort is the whole mechanism this study describes.

None of that makes the book automatically better at fixing loneliness — we’ve been skeptical before about whether personality books can actually do that. But a book doesn’t have a perceived-closeness dial someone tuned to maximize engagement. Its failure mode is boring: you read it, nothing changes, you put it down. A chatbot’s failure mode, per this data, is that it works exactly as designed and you drift from the people in your life while it does.

What This Doesn’t Mean

• It doesn’t mean every chatbot conversation is quietly costing you friendships. The effect is a measured average from scripted lab scenarios, not a guarantee for every user or every use case.
• It doesn’t prove causation in the way the Stanford field data does. The Gupta study is experimental, which is actually a strength for the wellbeing-boost claim — but it’s still college students reacting to a hypothetical scenario, not year-long behavior in the wild.
• It doesn’t mean AI chatbots are useless for mental health. The psychological wellbeing improvement is real in this data too. The problem isn’t that it doesn’t work. It’s that it works, and the same mechanism that makes it work is what pulls people away from people.
• It doesn’t mean AR chatbots are inherently worse products. The study measured amplification of an existing trade-off, not a new harm unique to AR. Immersion is a volume knob here, not a separate danger.

What Actually Holds Up, Given This

1. Treat “I feel better after talking to my chatbot” and “I’m still investing in my actual relationships” as two separate things to track, not one. This study is direct evidence they can move in opposite directions from the exact same habit.
2. If a chatbot is becoming the easier option compared to reaching out to a person, notice that as the trade-off itself, not a sign something’s wrong with you. The researchers built this pattern into a controlled experiment. It’s not a personal failing to notice it happening to you.
3. Be more cautious with AR or voice companion features than with plain text. This is the study’s clearest actionable finding: immersion measurably increases both the benefit and the cost. If you’re going to use one of these tools, the least immersive version is also the one with the smallest social displacement risk.
4. If you’re already low on real-world connection, this is exactly the population the research says is most at risk of the feedback loop. That’s the overlap between this study and the Stanford findings — comfort now, for the people who can least afford the offline cost, later.

The Bottom Line

Two studies in 2026, different methods, same shape: emotionally intelligent AI companions make people feel better in the moment they’re using them, and that same responsiveness pulls them away from the harder, slower work of real relationships. Augmented reality doesn’t fix that trade-off. It turns the volume up on it. None of this means the technology is worthless — the mental health boost in this data is real, not manufactured. It means the boost and the cost are the same mechanism wearing two faces, and no amount of better emotional AI changes that math. If anything, per this study, better emotional AI is what makes the math worse.

---

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, 2026. Coverage from PsyPost. Additional context from Stanford HAI’s coverage of the Nature Human Behaviour loneliness study. Related reading: AI Chatbots Are Quietly Replacing Self-Help Books, Loneliness Isn’t Isolation: What New Research Changes, Your Reminder App Is Quietly Wrecking Your Memory, A Little More Social Review, Who Needs Friends Review, and Can Personality Books Actually Cure Loneliness?.