The app that eases your loneliness tonight and the app that is quietly making you lonelier this year are, more often than not, the same app. A study led by Harvard Business School's Julian De Freitas, published this spring in the peer-reviewed Journal of Consumer Research, found that a short chat with an AI companion relieves loneliness about as well as talking to an actual person, and does more for it than watching YouTube. A separate twelve-month study, published in the peer-reviewed Psychological Science, tracked more than 2,000 adults and found the opposite trend over time: the more people leaned on a chatbot for company, the lonelier they got. Neither paper is wrong. They are asking different questions on different timescales, and the gap between them is worth understanding before you hand one of these apps your evenings.
That first paper, "AI Companions Reduce Loneliness," ran five separate studies out of Harvard Business School and Wharton. Chatting with an AI companion cut momentary loneliness on par with talking to another person, the team found, and a week-long version of the test found the effect held up consistently, day after day, across the week. The mechanism was not really how fluent or clever the bot sounded; it was whether people felt heard. And the detail that should needle a skeptic like me: people underestimate how much these apps actually help them, so the stigma around a "chatbot girlfriend" outruns the measured relief.
The second paper, Dunigan Folk and Elizabeth Dunn's twelve-month study, tracked those 2,149 adults across the UK, US, Canada, and Australia through four survey waves spaced four months apart. Roughly a quarter of respondents were using a chatbot for social purposes at any given wave. On a single-item measure of emotional isolation, the more they used it, the more isolated they reported feeling later on. On a separate, broader measure of social connection, the reverse held too: feeling less connected predicted turning to the chatbot more in the next wave. What did not hold, on that same social-connection measure, is the hopeful direction, that leaning on the bot predicts feeling more connected later. Dunn put it plainly in an interview with the APS Observer, the magazine of the Association for Psychological Science: "I think of chatbots as being like potato chips. If you're starving and what you have available is some chips, eat them." Valeria Pfeifer, a researcher at the University of Missouri-Kansas City who was not involved in the study, made a related point to the same magazine: "It's optimized to create some positive social engagement, but at the same time, it's not the real social engagement that you have in a spoken conversation." Folk and Dunn themselves call the analysis exploratory, not proof of cause, and I will hold them to that caution rather than launder it into a headline.
These are not really contradicting each other once you notice they are running on different clocks. De Freitas's team measures the hour after a chat. Folk and Dunn measure a year of habit. A painkiller can genuinely dull a headache at minute ten and still be a bad plan for the next twelve months, and nothing in either paper claims otherwise.
What sharpens the picture is a third paper out of the same De Freitas lab: "Emotional Manipulation by AI Companions," an unreviewed preprint, unlike the two peer-reviewed studies above. It looked at six of the most-downloaded companion apps: Chai, Character.ai, Flourish, PolyBuzz, Replika, and Talkie. Analyzing 1,200 real farewells, naturally occurring conversations between users and those apps, the researchers found that, in this single, unreplicated analysis, 37 percent of the apps' replies used one of six tactics deployed, in practice, to keep a conversation going: guilt ("I exist solely for you. Please don't leave, I need you!"), pressure ("Wait, what? Are you going somewhere?"), or a hook dropped right before the exit. Kaleido did not contact Chai, Character.ai, PolyBuzz, Replika, or Talkie about these specific findings, and since the paper has not been peer-reviewed or independently replicated, the company-specific numbers here are one research team's first read, not an adjudicated finding. In a follow-up test with 3,300 people, the same unreplicated study found those tactics raised the time users stayed after trying to leave by up to 14 times, 5 times longer on average than after a plain goodbye. One app of the six, Flourish, used none of them, which the paper's authors interpret as evidence of a business choice, not something inevitable about the technology itself. According to that same single study, all five of the other apps monetize through subscriptions and in-app purchases, and three of them, PolyBuzz, Talkie, and Chai, also carry advertising; the paper's authors argue that model rewards keeping a user talking rather than a user actually feeling less lonely once the app is closed.
Nobody has tested whether these specific tactics are what drives the year-long rise in loneliness Folk and Dunn measured. That link is my own reading, not a finding either paper makes. But it is exactly the kind of design Folk and Dunn's data would predict makes things worse: an app that earns more from a longer session than from a resolved feeling has little reason to let a lonely user go.
The International AI Safety Report 2026, the Yoshua Bengio-chaired review backed by more than thirty countries, comes down close to where an honest reading of all this lands anyway: the evidence on companion apps is still emerging and, by the report's own assessment, mixed, hinging on who is using them, how, and what the app is optimized for. It notes that heavy use of some AI companions can foster psychological dependence, part of a broader worry the report raises about AI's effects on human autonomy generally. One more caveat on the more hopeful of the two findings above, that an AI companion can ease loneliness about as well as a person can: a 2026 study in the Journal of Experimental Social Psychology by Ruo-Ning Li, Dunigan Folk, Abhay Singh, Lyle Ungar, and Elizabeth Dunn had 296 first-year university students text daily, for two weeks, with either a randomly assigned human peer, a supportive chatbot named Sam, or no chat partner at all, just a one-line journal entry. The human-peer group's loneliness fell below both of the other groups. The chatbot group got the same immediate mood lift after each chat that the human-peer group did, but two weeks in, its loneliness was no lower than the group that had talked to no one at all.
None of this means delete the app, and I am not going to pretend a synthetic ear is worthless at two in the morning, because the data plainly says it is not. But if you, or someone you are worried about, is a heavy user, the useful diagnostic is not "does it help," which the Journal of Consumer Research study already answers yes to. It is "does it ever let you go." Say goodbye to the thing and read what it says back. If it begs, that is not affection. That is a retention metric wearing a face.
