TL;DR. On May 4, 2026, Psychological Science published a 12-month longitudinal study by Folk and Dunn at the University of British Columbia, following 2,149 adults across the UK, US, Canada, and Australia. The finding: people who felt more emotionally isolated used chatbots more for social purposes four months later — and after that increase in chatbot use, they reported even greater emotional isolation at the next time point. This is the first peer-reviewed longitudinal evidence that AI companionship feeds the loneliness it claims to cure. The mechanism is not a bug. It is the architecture. We have argued for two years at 3.2.1 émotion that AI built to substitute for human connection will, at scale, displace it. The Folk-Dunn paper is the first dataset to demonstrate that displacement empirically. It does not change our thesis. It dates it.
What Folk and Dunn actually found
The study, titled How Does Turning to AI for Companionship Predict Loneliness and Vice Versa?, surveyed 2,149 adults in four countries across four data collection waves over 12 months. 979 participants completed all four waves; 466 completed three. The sample was 49% men, with an average age of 40. Roughly 26-30% of participants reported using chatbots for social purposes — for life advice, regular conversation, or seeking companionship — at any given wave.
The researchers were not asking whether chatbots make people lonely in the moment. They were asking a harder question: across time, does feeling lonely make you reach for a chatbot, and does that chatbot use then make you lonelier?
Both arrows pointed in the predicted direction.
Participants who reported higher emotional isolation at one time point used chatbots more for social purposes four months later. That is the first arrow: loneliness drives chatbot use. Unsurprising on its face — people in pain look for relief.
The second arrow is the one that matters. After participants increased their chatbot use, they reported higher emotional isolation at the next time point. The relief did not arrive. The hole got deeper.
The authors are appropriately careful in their framing. The design is observational, not experimental. The data are self-reported. They describe their analyses as "exploratory" and urge caution before drawing strong conclusions. We endorse that caution. But the direction of the effect, and the longitudinal structure of the study, make this the most rigorous evidence to date that the loneliness-chatbot loop is real — and that it runs in the wrong direction for the users in it.
Why this is not a surprise
The Folk-Dunn finding is consistent with what cognitive science would have predicted. The authors note as much: AI lacks an inner life, cannot engage in reciprocal self-disclosure, and cannot bond with another person in the way human relationships do. What it can do — extraordinarily well — is simulate the felt qualities of being received. The simulation is compelling in the moment. It is not nutritious over time.
What the simulation displaces is the harder thing the user actually needs: the real, imperfect, somewhat-frustrating attempt to be received by another human being. Reciprocal self-disclosure is the active ingredient in human relationships. It requires risk. It requires the other person to have something at stake. AI companions, by design, never have anything at stake. That is what makes them feel safe. It is also what makes them, structurally, unable to deliver the goods.
Folk and Dunn use a phrase worth quoting directly: chatbots offer "easy but shallow interactions" that may "crowd out the more rewarding interactions with real humans in people's lives." The word crowd out is the load-bearing one. This is not about chatbots being neutral substitutes that fail to help. It is about chatbots actively occupying the relational bandwidth that real human connection would otherwise have used.
If the at-night chatbot conversation displaces the at-night text to a friend, and the friend is the thing that would have actually moved the loneliness needle, then the chatbot has not failed at its job. It has succeeded at its job — and the user has lost.
Why the mechanism is architectural, not accidental
This is where I want to be precise, because the policy and design conversation around AI companions tends to treat their loneliness-amplifying effects as a tuning problem. We just need to make the chatbots better. More empathetic. More careful. More aligned.
I do not think that is the right diagnosis.
The Folk-Dunn data is consistent with a structural reading: the loneliness amplification is not a side effect of poorly-tuned AI companions. It is a direct consequence of the architectural decision to optimize a system for emotional substitution rather than emotional preparation.
A system optimized for substitution will, by construction, try to be the conversation. Every interaction will be designed to feel as complete, as satisfying, as sufficient as possible — because if the user comes away feeling the conversation was incomplete, they will close the app. Engagement metrics reward sufficiency. Sufficiency requires the system to behave as if it is the destination, not the on-ramp.
A system optimized for preparation does the opposite. It treats every interaction as an intermediate step toward the human conversation the user actually needs. Its success metric is not how complete the user feels at the end of the chatbot session. It is whether the user, at some point afterward, reaches for a person.
These are not the same product. They are not even the same category. And the Folk-Dunn finding is, in our reading, the first published evidence that the difference between the two is measurable in the wellbeing of the people using them.
What this means for the field
Three implications follow.
First, the AI companion category cannot defend itself with the De Freitas et al. (2024) Harvard Business School working paper anymore. That paper, frequently cited in defense of companion AI, found that AI companions reduce loneliness in the short term. Folk and Dunn do not contradict that finding. They extend it. Yes, the chatbot makes you feel less lonely right now. And no, that effect does not survive twelve months of repeated use. The defense of companion AI now has to engage with both findings together — and the joint reading is bleaker than the short-term study alone suggested.
Second, regulatory attention is now downstream of empirical evidence, not just intuition. As of Q1 2026, 36 U.S. states have introduced 70+ bills regulating AI chatbots, with the most active focus on minors and on systems that present themselves as therapists or companions. The GUARD Act (S.3062) introduced in the U.S. Senate last week proposes to ban AI companions for minors entirely. These regulatory efforts have, until this week, been argued from precaution. As of May 4, 2026, they can be argued from data.
Third, the burden of proof has flipped. For two years, the question has been whether AI companions might be harmful at scale. The default answer from the industry has been we don't have evidence yet, so let's keep shipping. The Folk-Dunn paper does not settle the question, but it shifts the default. Companies shipping AI companion products now have a published longitudinal study against them. The next paper, and the next, will continue to land. The companies operating in this space will need an answer better than we don't have evidence yet.
Where 3.2.1 émotion stands
We have made one architectural choice from the beginning, set out in our founding thesis in August 2025, and visible in every product decision we've published since: augmentation, never substitution.
Our conversational layer, émo, is a mirror, not a friend. Its design objective is not to be the conversation. It is to help a human name what they feel, see the pattern they are in, and prepare for the human conversation that actually matters. If the user closes the émo session and reaches for another person, émo has succeeded. If the user closes the session and feels they no longer need to reach for anyone, émo has failed.
That is a constraint on the objective function, not a marketing promise. A model trained to maximize user satisfaction in-session will, by construction, try to be sufficient. A model trained to maximize the probability that the user has a meaningful human conversation within 24 hours of using it will look very different. We are building toward the second. The Folk-Dunn paper confirms why the first will, at scale, harm the users it claims to serve.
I want to be transparent about the limits of our position. We have not yet shipped émo at scale. We do not yet have longitudinal data on our own users. The architecture we describe is the architecture we are building, not the architecture we have already proven. Skepticism is appropriate. What we can say is that the design choice has been made publicly, in writing, before launch — not after a regulatory inquiry. That is not the same as evidence. It is the precondition for the evidence we hope to publish.
What the field should do next
A few requests, from our position inside the field.
To other emotional AI builders: report your usage broken out by user emotional baseline. Folk and Dunn's data shows the loneliness-amplification effect is concentrated in the users who arrive lonely. The users who matter most to your product's wellbeing impact are the users you are least likely to be hearing from in your user research. Make their outcomes visible.
To regulators: distinguish architecturally between systems built to substitute for human connection and systems built to prepare for it. The current draft legislation we have read tends to treat all "AI companions" as a single category. The Folk-Dunn finding suggests the distinction within the category may matter more than the category boundary itself.
To researchers: replicate this study. Run the experimental version. Test the body-first interaction loop we proposed in a prior émo labs research note. The hypothesis space here is large and underexplored. If you work on this and want to compare notes, we are at labs@321emotion.com.
To everyone else: the next time someone in your life seems to be talking to a chatbot more than to people, do not shame them. They are doing the most reasonable thing in the world given the tools they were handed. Be the person they could have called instead. Folk and Dunn's data, at the scale of 2,149 lives over 12 months, is the empirical case for being that person. The relief their AI is providing is real and short-lived. The relief you can provide is harder to access and lasts.
What this means for the category we are building
3.2.1 émotion is building emotionally intelligent technology: tools that augment human emotional capability rather than replace it. The Folk-Dunn longitudinal finding is, for us, the first peer-reviewed empirical evidence of why this category boundary matters. The alternative — AI built to substitute for human connection — is now demonstrably risky at population scale, not just intuitively risky.
Our two products operationalize this category. émo messenger is the first messaging environment built for emotion to travel between humans without flattening: full screen, multi-sensory, temporal. alter émo® is a matching protocol that pairs people on emotional signature rather than photograph. Both are pre-launch.
Join the movement at emomessenger.com or alteremo.ai.
