Here’s a question worth sitting with for a second. Are we more isolated today than we were twenty years ago? Most assume yes, and they’re right. Though the relationship between technology and loneliness is more complicated than “too much screen time.”
Take Japan, for instance. Something showed up in its data between 2020 and 2021. A smaller proportion of people were basically alone (22.7% compared to 27.4% in the previous year).
Despite this, there was an increase in loneliness, from 5.03 to 5.86 on a standard scale among over 53,000 individuals who were followed over the period. Same population, more contact, and somehow, worse off.
That contradiction, feeling connected but alone, sits at the center of the digital loneliness debate, and most coverage of it leans too hard in one direction. Some of the underlying research holds up better than the panic suggests. Other parts of it fall apart the moment you check the fine print.
Quick Answer
The evidence turns out to be a lot thinner than the headlines suggest. Researchers randomly assigned 35,442 people to quit Facebook or Instagram for six weeks and found wellbeing improved by just 0.06 standard deviations, which is genuinely small. What has clearly shrunk is time spent together, as Americans now spend 35 minutes a day socializing, down from 41 back in 2015.
The Largest Experiment Ever Run Found Almost Nothing
Does social media cause loneliness, really? The largest randomized test shows it doesn’t.
Here’s how the study actually worked. Researchers recruited 35,442 adults in the U.S. and, in the run-up to the 2020 election, randomly told some to keep using Facebook and Instagram as usual while others deactivated one platform for six weeks.
Those who quit Facebook scored 0.06 standard deviations higher than the control group on an index combining happiness, depression, and anxiety. Deactivating Instagram produced a 0.041 improvement. Both effects were real and statistically detectable, but neither came close to what researchers would call a meaningful effect size, as the next section shows.
The Facebook effect was seen mostly among people over 35. The one for Instagram was concentrated in women under 25, hinting these platforms don’t affect everyone equally, if they harm most users at all.
An earlier, smaller run of this same experiment turned up something worth flagging. Participants who deactivated Facebook that time reported greater life satisfaction and more time with family. Though most logged straight back in once the study wrapped and the payments stopped.
Some participants in the deactivation trials didn’t reduce their screen time at all. They just moved to other apps.
Table 1: The effect of quitting, in context
| Effect on wellbeing (SD) | |
| Deactivating Facebook, 6 weeks | 0.060 |
| Deactivating Instagram, 6 weeks | 0.041 |
| Conventional threshold for a “small” effect | 0.2 |
| Threshold for a “medium” effect | 0.5 |
| Threshold for a “large” effect | 0.8 |
Source: Allcott et al., NBER Working Paper 33697, 2025, n=35,442. The bottom three rows are Cohen’s benchmarks shown for scale, not measured results.
Anyone claiming social media is a primary cause of the wider loneliness epidemic technology gets blamed for has to explain why the largest randomized experiment ever run on the question found an effect this modest.
The original NBER paper includes a chart of its own plotting these effect sizes against each other directly from the source.
Infographic 1: Quitting Facebook For Six Weeks Barely Moved The Needle
What a 0.06 Effect Actually Means
On its own, a number like 0.06 doesn’t tell you much. You need a scale to measure it against, and psychologists happen to have one. Based on Cohen’s benchmarks, a 0.2 standard deviation counts as a small effect, 0.5 as medium, and 0.8 as large.
Against that scale, 0.06 lands at roughly a third of what counts as “small.” Looking at the raw social media loneliness statistics, it’s a real, measurable benefit, though nowhere near large enough to explain a loneliness epidemic affecting 1 in 6 people worldwide, a burden the WHO associates with an estimated 871,000 deaths a year, roughly 100 an hour.
That global figure is also the closest thing to an answer for anyone wondering whether we’re lonelier than previous generations: harder to prove definitively, but the data point in the direction of a yes.
What Has Actually Shrunk: 41 Minutes a Day, Down to 35
Is technology making us lonelier? Social media clearly isn’t carrying the whole blame here, so something else has to be pulling digital loneliness upward. One candidate shows up clearly in federal data, and it has nothing to do with how people say they feel. It comes down to how they actually spent their hours.
In 2025, Americans spent 35 minutes a day socializing, down from 41 minutes a decade earlier. The share of people who socialized at all on any given day dropped too, from 38% to 30% over that same stretch.
That’s meaningfully different from a survey asking how connected people feel. Time-diary records track what people actually did with their day, which is a lot harder to wave away than a self-reported mood score.
Even this can’t explain everything by itself, though. Face-to-face time was already declining well before smartphones showed up in every pocket, according to a separate analysis tracking social engagement with friends, which fell by roughly 20 hours a month between 2003 and 2020.
This has been going on for two decades, in other words, and it matters for how you read every other statistic here. Phones are part of the story, but a decline that predates the smartphone can’t be blamed on the smartphone alone, even where a real link exists between screen habits and how people spend their evenings now.
72% of Teenagers Have Already Talked to an AI Companion
While researchers argue about social media, a different technology has already become normal for an entire generation, mostly without anyone noticing.
Infographic 2: Talking to an AI is Already Normal Teenage Behaviour
Approximately 72% of teens aged 13 to 17 have used an AI companion at least once, according to a nationally representative survey of 1,060 teenagers fielded by NORC at the University of Chicago. That covers general-purpose chatbots as well as dedicated AI companion apps teens download specifically for conversation. About 18% use one at least a few times a month, and 13% use one every day.
The definition matters here. This survey counts general-purpose chatbots used for personal conversation, not just dedicated companion apps like Character.AI or Replika. That broad definition is a large part of why the number looks so high.
Table 2: How US teens use AI companions
| Frequency | Share of teens |
| Never | 28% |
| Less than monthly | 20% |
| A few times a month | 18% |
| A few times a week | 21% |
| Daily | 13% |
Source: Common Sense Media / NORC, April–May 2025, n=1,060 teens aged 13–17. Published figures: 72% ever, 52% at least monthly, 21% weekly, 13% daily, 28% never. The two middle bands are derived by subtraction.
A third of the teens using these AI companions do it for something social: role-play, romance, emotional support, or practicing conversations. That romance category covers the same territory people search for under AI girlfriend loneliness. A quarter have shared personal information with an AI companion they wouldn’t necessarily share elsewhere.
And Two-Thirds of Them Say It’s Worse Than a Real Friend
So, this raises the question: “Can a chatbot be a friend?” Teenagers, it turns out, are more clear-eyed about this than most of the coverage assumes.
Two-thirds of teens who’ve used an AI companion say the conversations feel less satisfying than talking to an actual friend. Eighty percent still spend more time with real friends than with any AI companion. Roughly half say they distrust the advice these tools give them.
Interestingly, younger teens trust that advice more than older ones do, apparently growing more skeptical with age.
Common Sense Media’s own recommendation, based on this data, is direct: nobody under 18 should be using AI companions in their current form.
For a visual breakdown of usage frequency, Common Sense Media’s report includes the underlying chart showing the full spread from daily use to never having tried one.
AI Companions Do Reduce Loneliness – That’s the Uncomfortable Part
Do AI companions reduce loneliness? What do the stats show?
Well, this is where the research turns genuinely uncomfortable, because the evidence that AI companions help isn’t thin. It’s peer-reviewed and fairly consistent across multiple studies.
In six separate studies, including a seven-day longitudinal experiment, researchers found that AI companions helped about as much as talking to another person, and more than watching YouTube. The mechanism wasn’t distraction or novelty. It was whether people felt heard.
An even more striking finding sat underneath the headline result: participants consistently underestimated how much these conversations would help before trying it, and some avoid AI companions specifically because they expect the tools not to work.
The catch is time scale. This research measured loneliness over a single day and a single week, and the authors are explicit about that limit. It says nothing about whether AI companionship stays healthy over months or years of regular use, precisely the question everyone actually wants answered.
Short-term relief that’s genuinely well-documented doesn’t automatically mean long-term benefit, and it’s worth noting this research was published in a marketing journal by marketing academics. Nobody has published a study isolating dedicated AI companion apps from general-purpose chatbots used the same way. That distinction matters enormously for the next study.
The Trial Everyone Reported Found the Opposite. Both Are Right.
A widely covered study out of MIT and OpenAI seemed to confirm the opposite: that AI companions make loneliness worse. The actual results are more complicated, and more interesting, than the headlines that ran.
Researchers conducted a randomized controlled trial for four weeks with 981 participants and over 300,000 logged messages. They randomly assigned people to nine different combinations of conversation type, personality, and voice of the chatbots. The headline-grabbing conclusion was indeed true: those who interacted with the chatbot more daily found themselves to be more lonely and emotionally dependent, and less social with real people.
Almost nobody reported the part that actually matters. The randomized portion of the trial, specifically designed to test cause and effect, turned up nothing: none of the voice, personality, or topic conditions moved the outcomes in any significant way.
The usage-loneliness link was a correlation observed inside the study, not something the experiment proved caused anything. Average loneliness among participants actually fell over the four weeks. The negative signal was concentrated entirely in the heaviest voluntary users, exactly the group most likely to already be lonely before they started.
Table 3: The Two Best Studies on AI and Loneliness Disagree, and Here’s Exactly How
| MIT / OpenAI | Harvard / Wharton | |
| Design | RCT, 9 conditions | 6 studies, incl. 7-day longitudinal |
| Sample | 981 | Several hundred per study |
| Duration | 28 days | 1 day to 1 week |
| Finding | Heavier use → more loneliness, dependence, less socialising | AI companions reduce loneliness as much as talking to a person |
| Causal? | No — randomised conditions showed no significant effect | Yes, within the time frame tested |
| Average loneliness | Fell over the study | Fell |
| Peer reviewed | Preprint | Journal of Consumer Research, 2026 |
| Whose interest | Co-authored by OpenAI | Marketing academics, consumer journal |
Sources: Fang et al., arXiv 2503.17473 (2025); De Freitas et al., Journal of Consumer Research 52(6), 2026.
As we narrow things down, does ChatGPT make you lonely?
A companion analysis from OpenAI, covering close to 40 million chats and a survey of 4,076 users, found that heavy users were more likely to describe ChatGPT as a friend, edging toward an answer to the question: can a chatbot be a friend for at least a vocal minority? The authors are careful to note this group remains a small minority of total users.
One caveat matters enough to state plainly: this MIT study remains a preprint, not yet peer-reviewed, and is co-authored by OpenAI, which has an obvious commercial stake in how the results land. A published critique argues some findings have been over-interpreted in coverage. None of that disqualifies the research, but it’s useful to know whose interests sit where.
Short-Term Relief, Long-Term Dependence
Put side by side, these two studies aren’t actually contradicting each other.
The Harvard and Wharton research measures what happens during a single conversation or a single week. It helps, roughly as much as talking to a person does. The MIT research measures what happens to people who keep going back day after day for a month, and heavier use tracks with worse outcomes among that group specifically.
Short-term relief and long-term dependence can both be true of the same tool. That’s the shape of most things that soothe us in the moment, and it’s a much less satisfying headline for AI and loneliness than either study produced on its own.
Which Way Does the Arrow Point?
Every correlational finding in the research on digital loneliness has the same weakness, and it’s worth naming directly instead of burying it.
Heavy chatbot use tracks with loneliness, while increased social media use is associated with worse mood in some studies. Neither fact settles the question of whether screen time and loneliness are cause and effect or they are just neighbors.
A person already feeling isolated may simply reach for a screen more often, and it’s not because the screen made them lonely, but loneliness sent them looking for anything that might help.
The MIT researchers say this outright about their own data. It’s a limitation almost every study here shares to some degree, and current research mostly can’t resolve it, since nobody has run the experiment that would.
Researchers sometimes call the underlying pattern the displacement hypothesis: time given to a screen is time taken from someone in the room. It’s plausible, and parts of it show up in the time-use data above, though it’s not the only explanation, since someone already isolated may simply have less contact to displace in the first place.
What we do have is evidence at very short time scales for one side of the argument, and a large randomized experiment showing a small effect for the other. Between them sits a gap: no study has tracked AI companion use, specifically, for more than a few weeks.
That gap is where most of the public argument over technology contributing to loneliness lives, and neither camp can currently close it with data. There’s also no adult version of the teen survey above, so nobody actually knows how common this behavior is among grown-ups, only among teenagers.
What Governments Are Actually Building
Set against all this uncertainty, a number of governments and cities have decided not to wait for it to resolve.
UNDP researchers mapped more than 100 programs worldwide. They sorted them into three categories: digital literacy training for older adults, hybrid programs pairing an app with in-person facilitation, and AI-based outreach that flags people at risk and connects them to real services.
Barcelona’s VinclesBCN app links older adults into a network of family, friends, and shared-interest groups, with a human facilitator guiding people toward in-person meetups. Madrid’s PALOMA system uses AI to detect isolation risk and route people to professionals rather than substitute for a relationship. Seoul runs a similar outreach model called Knock Knock 24.
None of these programs claim the technology replaces human contact. The people who built them were explicit that it’s meant as a bridge toward it instead.
The UNDP mapping also raises questions that get less attention than the interventions themselves. Detecting who’s at risk through AI means someone has decided what loneliness looks like in the data, raising real concerns about algorithmic bias in who gets flagged and privacy around emotional data collected this way.
What’s missing is uncomfortable to say plainly. The UNDP mapping itself acknowledges that some of these interventions are backed by evidence and others still need testing.
Nobody has published a rigorous evaluation showing that PALOMA, VinclesBCN, or Knock Knock 24 actually reduce loneliness at scale. They may well work. That hasn’t been measured yet.
Questions People Ask About Technology and Loneliness
Is Social Media Making Us Lonelier?
When researchers randomly asked 35,442 people to stay off Facebook or Instagram for six weeks, their wellbeing ticked up by 0.06 and 0.041 standard deviations, which shows social media doesn’t necessarily make us lonelier.
Do AI Companions Actually Reduce Loneliness?
Within a single conversation, the peer-reviewed evidence leans yes, a genuinely encouraging short-term answer to the question. People in trials testing this came away about as comforted as they would from talking to a real person, mostly because they felt genuinely listened to. What nobody has tested yet is whether that holds up over months instead of days.
Does ChatGPT Make You Lonely?
Not according to the part of the study built to actually test it. The four-week trial behind that claim did find heavier chatbot users reporting more AI chatbot loneliness, but its randomized conditions moved nothing, and loneliness across the whole group actually dropped. A simpler explanation fits just as well: people who are already lonely tend to reach for a chatbot more often.
How Many Teenagers Use AI Companions?
A 2025 survey of 1,060 teens put the number at 72% who had tried one, with 21% checking in weekly and 13% talking to one every single day. That’s a lot closer to normal teenage behavior than most parents probably assume.
Should Teenagers Use AI Companions?
Common Sense Media says no, not in the form these tools currently take. Teens themselves don’t seem especially convinced either: two out of three say the conversations feel like a poor substitute for a real friend.
Does Quitting Social Media Make You Happier?
Slightly, on average. The gains land unevenly, showing up most clearly in people over 35 who left Facebook and women under 25 who left Instagram. When an earlier version of this experiment ended, most participants drifted right back to the app anyway.
Editorial note: Nothing here makes technology the villain or the cure. There’s no clean either-or, just a handful of small pressures pulling the same way.
Worried about a teenager? Most still prefer real friends; two-thirds find AI less satisfying. The worry is fair; panic isn’t.
Struggling yourself? A conversation with a doctor, a counselor, or someone you trust tends to do more than any tool discussed here.