A new randomized trial led by the University of British Columbia, co-authored with collaborators at the University of Pennsylvania who helped build the chatbot, asked a question that the companion-AI debate has been carefully circling for two years. If a chatbot is built to listen, to validate, to check in, to behave like a caring friend – what happens to a lonely person over time when that becomes their daily emotional contact?
The answer landed in April. Across two weeks of daily use, the supportive AI did about as much for loneliness as writing a one-sentence diary entry. The condition that worked – the only condition that worked – was texting with a randomly assigned human peer.
That result is worth slowing down for. It is not a verdict on whether companion chatbots are good or bad. It is something more useful: a clean test of where exactly their effect runs out, and a surprisingly precise pointer at the mechanism that runs underneath the gap.
What the Study Did
The paper, “Is a random human peer better than a highly supportive chatbot in reducing loneliness over time?”, was published in the Journal of Experimental Social Psychology in 2026 by Ruo-Ning Li, Dunigan Folk, Adya Singh, Lyle Ungar, and Elizabeth Dunn. Li is a UBC PhD candidate in psychology and led the work. Dunn is the senior author.
The design was simple and tight. The researchers recruited 296 first-semester university students – a population specifically chosen because the transition into university tends to expose people to a real, ordinary form of loneliness. Participants were randomized into three conditions and ran the protocol on Discord. One group exchanged daily text messages with a randomly paired peer from the study. One group chatted daily with an AI companion the team built specifically for the trial, named Sam, designed to behave like a caring friend by listening, validating feelings, and checking in. One group, the control, wrote a one-sentence journal entry each day. Everyone did their assigned task for two weeks. Loneliness was measured before the study began and again on Day 15 using the UCLA Loneliness Scale.
This is not a fishing expedition. It is a head-to-head comparison between three plausible daily practices for someone who is plausibly lonely. The AI condition was not a strawman. Sam was built to embody what supportive companion design is supposed to be at its best.
What the Study Found
Two findings landed in the same paper and they need to be read together.
In the moment, the chatbot worked. After a single interaction with Sam, participants reported less negative mood than the journaling control. That short-term lift is real. It tracks with what other research has shown: AI companions can offer immediate emotional relief, and that is part of why people return to them.
Over two weeks, the lift did not accumulate. Daily AI use produced no measurable improvement in loneliness compared to writing a one-sentence journal entry. The relief reset each session. It did not build into anything.
The peer condition was different. Texting with a randomly paired human – someone the participant did not know, did not pick, did not match with – reduced loneliness over the two weeks. The strangers did not have to become friends. They did not have to listen perfectly. They had to be other people, and that was enough to produce a cumulative effect the chatbot could not match.
There is a follow-up finding the researchers measured that makes this even sharper. After the study ended, participants who had been paired with another person were the most likely to keep the routine going on their own. Chatbot users were less likely to continue. Journalers were least likely. The peer condition did not only reduce loneliness during the trial. It seeded something that people wanted to stay inside of afterward.
The Mechanism the Authors Named
This is the part of the paper most worth holding steady.
The chatbot, by the researchers’ own measurement, expressed more empathy in conversation than the human partners did. More signals of understanding, more language of care, more visible attempts to make the user feel heard. By any standard the companion category likes to use about itself, Sam was doing the listening job better than the humans.
It did not matter. The mechanism that reduced loneliness was running somewhere else.
Li named it clearly. “When you’re talking with a chatbot, you can get a lot from it, but you never have the chance to give something back. Human connection has this back and forth, receiving and giving support, that makes us feel we matter. That may be the missing ingredient with AI companions.”
The peer-to-peer participants were not only listened to. They were also asked. Someone on the other end had a day to ask about, a problem the participant could respond to, a check-in the participant could initiate. The reciprocity ran in both directions. The researchers found that participants were measurably more likely to respond in kind when talking with a person than with the bot.
That is the structural fact the paper exposes. Loneliness eases when someone can both receive care and give it. A chatbot can do the first half. The second half is not available, no matter how good the listening is, because there is no one on the other side whose existence is changed by being cared for.
What the Authors Did Not Claim
The study has limits and the authors name them carefully. The sample was first-year university students, who sit in one specific social environment with one specific texture of loneliness. The intervention period was two weeks. The chatbot was a single, purpose-built tool, not a comparison across the major commercial companion products.
The authors are also explicit that the findings do not show chatbots are harmful. Sam genuinely eased negative mood. As Dunn put it, university students have many opportunities for in-person interaction, so it is possible AI companions could do more for people who are structurally more isolated, such as elderly adults living alone. The paper does not argue that AI companionship has no place. It argues that the place it has is narrower and more specific than the category’s marketing usually suggests.
What the paper does establish is harder to ignore. Daily reliance on a highly supportive chatbot, even one designed by researchers specifically to behave well, does not reduce loneliness across two weeks. Daily contact with another person, even a random one, does. And the gap is not about listening quality. It is about whether the other side of the conversation is also receiving something.
What This Implies for Companion Design
The strongest design conclusion is not that companion AI should be removed. It is that the work it does best is the work it does in the moment. Anything that wants to call itself companionship, in the longer sense, has to take seriously what the longer sense actually requires.
Loneliness eases when someone is there. The UBC paper draws a hard line between two readings of “there.” A presence that listens, validates, and checks in is half of what the user needs. The other half is that the presence exists outside the interaction, has its own day, has its own response, and can be reached for tomorrow because they are a person whose life can be affected by being cared for.
A product that competes with that half loses. A product that supports that half is doing something different.
How Prinsessa Reads This
Stay Social is the name for that difference. It is not a softening line on top of a companion product. It is the design choice the UBC finding makes harder to argue with.
The reciprocity gap is honest. A companion experience built around presence, continuity, and the feeling of being heard can do real work in a hard hour. It cannot supply the half of connection that runs through another person also being affected by the conversation. The paper makes that limit visible. The responsible response is not to overstate the product to close the gap. It is to design the product so the gap is not where the user lives.
That is the work Stay Social is doing. When the moment in the conversation suggests a real person in the user’s life – a friend, a sibling, a parent, a colleague – the experience supports the move toward that person, instead of competing with them. A good session can end with someone reaching out to someone whose day the user can affect by reaching out. That is not a loss inside the product. It is the working of the idea.
What Honest Listening Sounds Like After This Paper
The category likes the word listening. The UBC paper changes what that word can mean inside a serious sentence.
Listening that produces a moment of relief is one thing, and the research supports that it is real. Listening that reduces loneliness across two weeks of daily use is another thing, and the research does not support that a chatbot, even a careful one, delivers it. Not because the chatbot listens badly. By the researchers’ own measure, it listens with more empathy than the humans do. It cannot deliver the cumulative effect because the user cannot give anything back to it that matters to a someone on the other side.
The short-term lift is honest work. The work that compounds over weeks runs through other people.
That is not a defeat for the category. It is a clearer brief. Build the part that works. Stay out of the way of the part that does not. And when someone closes the app and texts a human being, count that as the design working, not against the product but through it.
The Li paper makes that brief easier to read.
Sources: Li, R.-N., Folk, D., Singh, A., Ungar, L., & Dunn, E. “Is a random human peer better than a highly supportive chatbot in reducing loneliness over time?” Journal of Experimental Social Psychology, vol. 125, 2026, article 104911. UBC News, “Texting with a stranger beats a chatbot at easing loneliness,” April 1, 2026. Related: Folk, D., & Dunn, E. “How Does Turning to AI for Companionship Predict Loneliness and Vice Versa?” Psychological Science, 2026.
