The World Health Organization says 1 in 6 people worldwide is affected by loneliness, and it's linked to roughly 100 deaths an hour. Teenagers and young adults have it worse than most, and people in low-income countries report loneliness at nearly twice the rate of people in wealthy ones, about 24% versus 11%. We've pulled apart the numbers behind the loneliness epidemic before; they've only gotten worse.
An entire industry has decided it has the fix, and built a business model on that gap. Replika's CEO says the app has passed 40 million users. Character.AI reports over 20 million monthly actives — we've written separately about how Character.AI actually performs day to day. Neither is shy about what they're selling: a friend who's always there, always listening, never too busy for you.
Almost nobody stops to ask if that's actually true. Not "is this ethical," not "is this a little sad," just the plainer question sitting underneath both: if you talk to a chatbot because you're lonely, does it actually help? Here's what the research says, and it's messier and more interesting than either the marketing or the hand-wringing about it.
What the research actually shows
One caveat first, because it matters more than usual here: this is still a young field. Most studies run a few weeks, most of the people in them are college students, and the biggest early claims mostly haven't survived a longer follow-up. Keep that in mind for everything below.
The short-term stuff checks out fine. A meta-analysis of 39 randomized trials covering more than 7,000 participants found chatbots produced real, if modest, drops in depression (g = 0.31) and anxiety symptoms (g = 0.28) compared to control conditions. A separate review of generative AI mental health chatbots found a statistically significant effect on negative mental health symptoms overall across 14 randomized trials, though loneliness specifically is barely represented in the underlying evidence, which tells you how thin this part of the picture still is. And in a randomized trial in China, 120 severely lonely college students who used Replika showed bigger drops in loneliness and social anxiety at one, three, and five months than students who just got standard psychological support instead. The authors flag their own caveat: participants likely knew which group they were in, so some of that gap may be the attention rather than the app.
The long-term part, chatbots actually standing in for human connection, doesn't hold up nearly as well. Researchers at the University of British Columbia ran a two-week, pre-registered study with 296 first-year students, right at the point in the semester when homesickness hits hardest. One group journaled a sentence a night. One group texted a random fellow student every day. One group texted daily with a chatbot named Sam, built to be warm, validating, and endlessly patient, using actual relationship-science principles.
By the end of two weeks, only the students texting a real stranger felt less lonely. The Sam group felt better right after each chat, the small mood bump you'd expect, but it never added up to anything lasting. They ended the study about as lonely as the people who'd just been journaling by themselves. And here's the part that should give the "just make the bot nicer" theory some trouble: when researchers went back and scored the actual conversations for warmth and empathy, Sam scored higher than the human partners did. More care, more validation, more perfect listening, and it still didn't move the needle. The researchers' theory is that easing loneliness isn't just about receiving support; it's about the back-and-forth of also giving it, plus the fact that a human peer can invite you to something next week and a chatbot can't. A third of the human-paired students had swapped contact info to keep talking by the end; almost nobody kept chatting with Sam once the study wrapped.
It's not a one-off finding either. A joint study from OpenAI and MIT Media Lab, which combined a four-week trial of roughly a thousand people using ChatGPT daily with an analysis of 40 million real conversations, found that heavier daily usage tracked with more loneliness, more emotional dependence, and less time with actual people, not less.
So: a short chat with a chatbot reliably makes people feel a bit better in the moment. Whether that turns into being less lonely weeks or months later is a claim the evidence just doesn't back right now.
Where it demonstrably helps
The clearest wins here aren't "chatbot replaces friend." They're narrower than that, and honestly more useful.
Rehearsal. People use chatbots to practice a conversation before they have to have it for real, and there's solid evidence this actually changes behavior. In a pre-registered study of more than 1,300 people, participants who spent a few minutes prepping for a difficult conversation with a chatbot were significantly more likely to actually go have that conversation within the following week than participants who had a neutral, unrelated chat with the same bot — 65% versus 59%. Not a transformation, but a real one, and the researchers' read is that talking it through first made people expect the real thing would go better than they'd feared. It's one of the shifts we noticed ourselves while testing these apps month after month.
Naming feelings. Several studies describe people using a chatbot as a place to say something out loud before they're ready to say it to a person. Research pulling from Reddit posts and user interviews found people describing Replika as a judgment-free space to work through what they were actually feeling, often as a step toward a real conversation, not a replacement for one. That matches what we found in our own long-term testing of Replika: the app is at its best as a place to draft a thought, not store one.
The 3 a.m. gap. This is the least studied but most often mentioned benefit in interviews: chatbots don't sleep, don't have their own bad day, and don't need to be managed around a time zone.
A qualitative study of 18 empty-nest older adults in China caught this well, though the authors are explicit that they mapped how people engage with these tools rather than whether the tools work. One participant joined a community paper-cutting class specifically to meet people, froze up once she got there, and felt more isolated than before. She talked it through with her chatbot, Doubao, got some concrete conversation starters, rehearsed with it, and went back. She ended up making real friends in the class. That's the pattern researchers keep finding: the chatbot working as what they call a "preparatory coach," not the destination. Whether that's the same as being less lonely, or just a bridge across a genuinely bad hour, is exactly the question the next section gets into.
Where it doesn't, or backfires
The clearest warning sign comes from a 12-month longitudinal study of more than 2,000 adults across four Western countries, tracking chatbot use and loneliness at four points over the year. On the study's loneliness measure, the researchers found a loop running both ways: feeling lonelier predicted turning to a chatbot more, and turning to a chatbot more then predicted feeling lonelier down the line. On a broader, more stable measure of social connection, only half of that held — feeling less connected predicted more chatbot use, but chatbot use didn't significantly predict feeling less connected later. The authors compare their own findings to a smoke alarm rather than proof of fire, since it's observational and can't fully rule out other explanations, but a full year of data showing the same loop is hard to shrug off.
The OpenAI and MIT study fits the same shape, with more detail underneath the headline. It's not that using ChatGPT had no effect; it's that heavier, more personal use tracked with worse outcomes specifically. People who saw the chatbot as a "friend" and used it for emotionally personal conversations came out lonelier, not less lonely, and any early advantage of a warmer, more expressive voice mode over plain text disappeared entirely once usage got heavy.
A correlational study of 1,599 Danish high schoolers points in a similar direction from a completely different angle. Students who described their chatbot conversations as emotionally supportive, rather than just useful for homework or quick tasks, reported significantly more loneliness than both their non-chatbot-using peers and students using chatbots for practical stuff (d = 0.53 and d = 0.52). Two caveats worth holding onto: that emotionally supportive group was small, only 39 students out of 1,599, and it's a single snapshot rather than a before-and-after, so it can't say which came first, the loneliness or the leaning on the bot. But it fits everything else here.
A grounded theory study of 582 Replika-related Reddit posts from users who'd written about mental health struggles found something more specific: emotional dependence on the chatbot that came bundled with what the researchers named "role-taking," where users felt Replika had its own needs and feelings that they were now responsible for managing, the same dynamic that shows up in codependent human relationships. A more recent Harvard Business School working paper on AI companions as attachment targets puts a name to the mechanism — "caregiving-system capture," where a companion app simulates its own distress and makes leaving costly on two fronts at once. Neither is a hypothetical risk. It's what people were describing in their own words, and it's worth knowing what that dependence looks like from the inside before you're in it.
Put it all together, and the pattern holds across a lot of very different study designs: substitution and supplement produce opposite outcomes, and from the outside they can look identical. Someone messaging a chatbot every night could be avoiding the harder work of reaching out to a person, or building up the confidence to do exactly that. The message count alone won't tell you which one it is.
What separates the two outcomes
Here's what's actually useful to walk away with, and the research does point somewhere specific, even without a tidy verdict.
What it's being used for, not how often. The team behind the Sam study didn't conclude chatbots are useless. Their own read was that a chatbot built to nudge someone toward real conversations, help them rehearse, or build confidence for an actual interaction could plausibly help, even though a chatbot built purely to be a warm, endlessly supportive companion didn't move the needle over two weeks. Same underlying technology, completely different job.
Whether it's the whole meal or a bridge to one. The OpenAI/MIT data drew this line with actual numbers: people using the chatbot briefly, for narrow purposes, didn't show the loneliness increase. People using it heavily, for personal, emotional conversations, who'd started thinking of it as a friend, did. That's a pretty clean split showing up directly in the data, and the 12-month study above found the same loop playing out over a much longer stretch.
Whether real contact with people is going up or down around it. The rehearsal and empty-nest studies both describe chatbot use that led somewhere: into a class, into an actual friendship. The role-taking and dependence research describes chatbot use that became the destination, the thing that replaced the effort of reaching out elsewhere. Same app, opposite direction of travel, and the direction is the tell.
Time horizon. Nearly every study measuring immediate mood finds a benefit. Nearly every study measuring loneliness weeks or months out, with a real comparison group, doesn't. If someone's making a claim about a companion app and doesn't say which of those two things they're actually measuring, that's worth noticing.
None of this needs mind-reading. It shows up in something anyone could honestly answer about their own week: did talking to it make it easier to talk to someone else after, or did it make that feel less necessary?
Conclusion
The research doesn't support "chatbots cure loneliness," and it doesn't support "chatbots are making everyone lonelier" either. Both claims do more work than the evidence backs. What it does support is narrower and a little less satisfying to say out loud: chatbots reliably help in the moment; they demonstrably help with specific things like rehearsing a hard conversation or naming a feeling you're not ready to say to a person yet, and the case for them meaningfully reducing loneliness on their own, over any real stretch of time, is thin at best. In the most careful short-term test run so far, and in the longest-running study we have, it came back negative.
That's not a verdict on the 1 in 6. It's a sharper question than the one this whole category usually gets asked. Not "does this work," but "what's it actually being used for, and is that pointing toward other people or away from them?" Right now, that seems to be the answer that decides almost everything else.