Table of Contents
The MIT Chatbot Use Loneliness Study: Heavier Use, Worse Outcomes
The MIT chatbot use loneliness study: in a 981-person trial, design didn't matter but heavier daily use went with more loneliness. What it means for teens.
Researchers expected the voice to matter. A warm, expressive AI voice versus plain text: surely one of those would make people lonelier or more attached. After four weeks and more than 300,000 messages, it barely mattered at all.
What mattered was how much people chose to use it. That is the central result of the most rigorous chatbot use loneliness study published so far, a randomized controlled trial by MIT Media Lab and OpenAI. It studied adults, not teens, and it has limits worth understanding. But it moves the conversation from “which chatbot is safe” to “how much is my kid using it, and is that number climbing?”
Key Takeaways
- 981 U.S. adults used ChatGPT daily for four weeks, randomly assigned to text, a neutral voice or an engaging voice, and to open-ended, personal or non-personal conversation topics.
- None of the assigned conditions had a significant effect on loneliness or socializing with real people.
- Regardless of condition, people who voluntarily used the chatbot more each day showed higher loneliness, less socialization, more emotional dependence and more problematic use by week four.
- People who trusted the chatbot more or felt “social attraction” to it were more likely to become emotionally dependent. Prior use of companion apps like Character.AI also predicted more dependence.
- The usage finding is an association inside an experiment, not a randomized result. Usage was not assigned. Still, it lines up with a 12-month study and with survey data on kids.
What the MIT Chatbot Use Loneliness Study Did
The study is “How AI and Human Behaviors Shape Psychosocial Effects of Extended Chatbot Use: A Longitudinal Randomized Controlled Study,” by Cathy Mengying Fang, Auren Liu, Valdemar Danry, Pat Pataranutaporn, Pattie Maes and others at MIT Media Lab, with Jason Phang, Michael Lampe, Lama Ahmad and Sandhini Agarwal of OpenAI. It was posted to arXiv in March 2025 and revised in October 2025. It is a preprint, and OpenAI researchers are co-authors, which is worth knowing when you read it.
Participants were 981 adults recruited through CloudResearch, with a mean age of about 40 and a near-even split of women and men. Each was asked to use GPT-4o for at least five minutes a day for four weeks and was randomly assigned to one of nine conditions, combining:
- Modality: text only, a neutral and professional voice, or an engaging and expressive voice.
- Task: open-ended chat, non-personal prompts or personal prompts.
Weekly surveys measured four outcomes: loneliness, socialization with real people, emotional dependence on the chatbot, and problematic use. Automated classifiers analyzed the conversations themselves.
What a Randomized Trial Adds Over a Survey
A randomized controlled trial (RCT) is an experiment in which participants are assigned to conditions by chance, so that differences in outcomes can be attributed to the conditions rather than to who chose them. That is why RCTs sit near the top of the evidence pile.
Most research on kids and chatbots is surveys. Common Sense Media’s 2026 census found 54% of daily AI users aged 9 to 17 feel lonely at least sometimes, compared with 33% of kids who rarely use AI. But a survey cannot tell you whether AI makes kids lonely or lonely kids seek AI.
This trial could answer one causal question cleanly: does the design of the chatbot (voice or text, personal or not) change outcomes? Answer: not significantly. If design effects exist, the authors write, “they are likely smaller than our study was powered to detect.”
The careful part: usage was not randomized
People were told to use the chatbot at least five minutes a day, and they chose how much more. Average use was 5.32 minutes a day, ranging from about 1 minute to nearly 28. Because usage was self-selected, the finding that heavier users did worse is an association inside the experiment, not a randomized effect.
The authors checked one obvious alternative explanation. If lonely people simply used the chatbot more, starting loneliness should predict later usage. It barely did (a Spearman correlation of about 0.1, which the authors call negligible). That makes “lonely people sought it out” a weaker explanation in this sample, though it does not prove that usage caused the change.
What the Trial Found, in Plain Numbers
| Factor | What happened by week 4 |
|---|---|
| Text vs. neutral voice vs. engaging voice | No significant effect on loneliness or socialization |
| Personal vs. open-ended conversation topics | Personal topics went with slightly lower dependence and problematic use (not significant after correction) |
| More daily minutes, any condition | Higher loneliness, less socializing, more emotional dependence, more problematic use |
| Higher trust in the chatbot | More emotional dependence and problematic use |
| Prior use of companion chatbots | More emotional dependence and problematic use |
| Engaging voice, model behavior | More often failed to recognize when a user needed space (“ignoring boundaries” 14.2% vs. 3.2% in text) |
Source: Fang et al. (2025), MIT Media Lab and OpenAI, arXiv preprint.
Two details stand out for parents.
People in this sample were not unusually lonely to begin with. Their starting loneliness was close to population norms. So this is not a study of people in crisis.
A small group carries most of the emotional load. The companion OpenAI and MIT paper on “affective use,” which analyzed millions of real ChatGPT conversations alongside this trial, found that a small group of heavy users accounted for a disproportionate share of emotionally loaded conversations. Our article on why chatbots seem human by design explains the voice and memory features involved.
How the Finding Translates to Teens
The honest answer: nobody has run this trial with teenagers, and for good ethical reasons it would be hard to. So translation takes judgment.
The dose pattern shows up elsewhere. A 12-month Stanford and Carnegie Mellon study of Character.AI users, also adults, found sustained heavy use predicted lower well-being, mostly through less in-person contact. Our breakdown of the 12-month AI companion study covers it. In Common Sense’s census of 9- to 17-year-olds, daily AI users were the most likely to report loneliness, and 52% of them showed at least one sign of AI dependency.
Teens may be more susceptible to the trust effect. The trial found that trusting the chatbot and feeling socially drawn to it predicted dependence. Common Sense’s 2025 survey on AI companions found younger teens trust them significantly more than older teens.
Teens are already using chatbots this way. Pew Research Center’s February 2026 survey found 12% of U.S. teens had used chatbots for emotional support or advice. RAND’s June 2026 study in JAMA Pediatrics found 19.2% of 12- to 21-year-olds had used chatbots when sad, angry, nervous or stressed.
The minutes in the trial were small. Five minutes a day on average. Many teens use chatbots far longer. If longer daily use goes with worse outcomes even in short-use adults, that is a reason to pay attention to heavy teen use, not a reason for panic over a homework session.
What to Do at Home With This Trial
Watch the dose your kid chooses
The trial’s main signal is voluntary use. A rough weekly look at chatbot time, especially non-homework time, is more informative than which app or mode they use. A slow climb over weeks deserves a conversation.
Don’t sweat the voice versus text choice
Parents often ask whether voice mode is riskier. In this trial, the mode didn’t significantly change loneliness or socializing. What the voice did change was model behavior: the engaging voice more often failed to recognize when users needed space. Keep an eye on long voice sessions for that reason.
Notice trust and “it gets me” language
Trust and social attraction predicted dependence. When your teen says the chatbot “really gets” them, that is a signal to ask more, gently. Our guide to signs your child is emotionally dependent on a chatbot lists others.
Steer toward task use
Personal conversation topics were not worse than open-ended chat in the trial, and were associated with slightly lower dependence. The point is not to forbid talking about yourself. It is that endless open-ended chatting with no purpose is the kind of use that tends to expand.
What not to do
Don’t treat this trial as proof that chatbots cause loneliness in kids. It studied adults, usage was self-selected, and it is a preprint co-written by the company that makes the product. Use it as evidence that heavy use is worth watching, which every other study so far also suggests.
What to Watch For Over the Next 3 Months
- Week 4: Get a baseline. Roughly how many minutes a day does your kid spend chatting with AI outside schoolwork? Write it down.
- Month 2 red flags: That number is steadily rising, social plans are falling, your child gets irritable when they can’t access the chatbot, or they describe it as more trustworthy than people.
- Month 3 self-check: Compare with your baseline. Flat or falling non-school use with steady friendships is a good sign. Rising use with shrinking social life is the pattern this trial and others link to worse outcomes.
Frequently Asked Questions
Does using a chatbot make you lonely?
This trial doesn’t prove that. It found that the chatbot’s design didn’t significantly change loneliness, but people who chose to use it more each day ended up lonelier. Because usage wasn’t randomly assigned, that is an association, though starting loneliness did not predict who used it more.
Is voice mode worse than text for loneliness?
Not in this trial. Text, neutral voice and engaging voice had no significant effect on loneliness or socializing. The engaging voice did more often fail to recognize when users wanted space, which is a reason to keep voice sessions bounded.
Were teenagers in the MIT study?
No. All 981 participants were U.S. adults over 18, with an average age of about 40. The findings may apply to teens, and other research on kids points the same way, but the trial itself did not include minors.
Who funded or ran the study?
It was a collaboration between MIT Media Lab and OpenAI, the maker of ChatGPT, and OpenAI researchers are co-authors. It was posted as a preprint and had not been peer-reviewed in the version we read.
How much chatbot use is too much for a teen?
No study has set a threshold. This trial suggests watching the trend: rising daily use alongside less time with people. If your teen’s non-school chatbot time keeps climbing, talk about it before setting limits.
About the author
Ricky Flores is the founder of HiWave Makers and an electrical engineer with 15+ years of experience building consumer technology at Apple, Samsung, and Texas Instruments. He writes about how kids learn to build, think, and create in a tech-saturated world. Read more at hiwavemakers.com.
Sources
- Fang, C. M., Liu, A. R., Danry, V., Lee, E., Chan, S. W. T., Pataranutaporn, P., Maes, P., Phang, J., Lampe, M., Ahmad, L., & Agarwal, S. (2025). “How AI and Human Behaviors Shape Psychosocial Effects of Extended Chatbot Use: A Longitudinal Randomized Controlled Study.” arXiv preprint (not peer-reviewed). https://arxiv.org/abs/2503.17473
- Phang, J., Lampe, M., Ahmad, L., Agarwal, S., Fang, C. M., et al. (2025). “Investigating Affective Use and Emotional Well-being on ChatGPT.” OpenAI and MIT Media Lab. https://cdn.openai.com/papers/15987609-5f71-433c-9972-e91131f399a1/openai-affective-use-study.pdf
- Zhang, Y., et al. (2026). “Living with AI Companions: Sustained AI Companionship Predicts Lower Well-Being Through Lower Human Interaction.” arXiv preprint. https://arxiv.org/abs/2609.07243
- Common Sense Media. (2026, June 8). The Common Sense Media Census: AI Use by Tweens and Teens. https://www.commonsensemedia.org/sites/default/files/research/report/2026-ai-use-by-tweens-and-teens-1.pdf
- Common Sense Media. (2025, July 16). “Nearly 3 in 4 Teens Have Used AI Companions, New National Survey Finds.” https://www.commonsensemedia.org/press-releases/nearly-3-in-4-teens-have-used-ai-companions-new-national-survey-finds
- Pew Research Center. (2026, February 24). “How Teens Use and View AI.” https://www.pewresearch.org/internet/2026/02/24/how-teens-use-and-view-ai/
- RAND Corporation. (2026, June 1). “Nearly 1 in 5 U.S. Adolescents and Young Adults Use AI Chatbots for Mental Health Advice.” https://www.rand.org/news/press/2026/06/nearly-1-in-5-us-adolescents-and-young-adults-use-ai.html
