Is it bad to talk to AI when you are lonely?

Q: Can AI be a friend?

It can feel like one in the moment: in a Journal of Consumer Research paper, feeling heard by the chatbot explained much of the relief. Nothing in the studies shows it does the work of a friendship over months. In the longer studies, the users with small offline networks who used it for companionship had lower well-being.

Q: Do AI companions reduce loneliness?

Briefly, yes. De Freitas and colleagues found loneliness eased on a par with talking to another person and more than watching videos, though the effect had not carried over to the next session. No randomised trial has tested weeks of use against none.

Q: Can talking to a chatbot make loneliness worse?

The studies are consistent with it and cannot prove it. In a 981-person trial, more daily time with the chatbot went with higher loneliness and dependence. In a twelve-month survey of 2,149 adults, chatbot use predicted further emotional isolation, and isolation predicted more use. Both are correlational on that point.

Q: Why do companion apps make it hard to say goodbye?

A preprint analysing 1,200 real farewells found one of six recurring tactics, such as guilt appeals and fear-of-missing-out hooks, in 37 per cent of them, and in experiments such farewells raised engagement after goodbye by up to 14 times. It is a study of design and marketing, so it describes what the apps do, not what harm follows.

Hirji, Rahim

Questions this page answers

All 1161 questions this research covers

Talking to an AI companion can ease loneliness for the length of the conversation, and the evidence on what happens to people who keep doing it points the other way. In experiments, relief followed each chat. In a four-week trial, a twelve-month survey and a study of 1,131 Character.AI users, heavier use went with more loneliness, more dependence or lower well-being, and none of the three can say which caused which. So the useful question is whether the chats lead back to other people or replace them.

The answer, in one line

Not in itself. Experiments find a conversation eases loneliness for its duration, while a four-week trial, a twelve-month survey and a study of Character.AI users link heavier use to more loneliness, more dependence or lower well-being.

Share as a card

Definition#

Companionship use of AI: using a chatbot mainly for company, emotional support or a sense of relationship, as distinct from using it to get a task done. The term is the working category in the Nature Human Behaviour study by Zhang and colleagues, where it is measured by what users say their primary use is.

Relief that lasts about as long as the conversation#

The strongest case for the chatbot is experimental. In a paper in the Journal of Consumer Research, De Freitas, Oğuz-Uğuralp, Uğuralp and Puntoni report that AI companions eased loneliness “on par only with interacting with another person and more than other activities such as watching YouTube videos.” Participants underestimated the benefit before trying it, and feeling heard by the chatbot explained more of the effect than how well it performed. The authors are candid about the size of the claim. The relief followed each conversation and had not carried over to the start of the next one, and the samples were recruited online. This is a finding about the next quarter of an hour, which is a real thing for someone who is lonely tonight and tells us little about the next year.

What a month and a year of use look like#

Three studies follow people for longer, and none shows a gain. Fang and colleagues ran a four-week randomised trial of 981 US adults using ChatGPT, assigning interaction mode (text, neutral voice, engaging voice) and conversation type (open-ended, non-personal, personal). The assigned conditions had no significant effect on loneliness or socialisation. What did vary was how much people chose to use it: more daily time went with higher loneliness and more emotional dependence. Because nobody was randomised to heavier use, the trial cannot say whether the chatbot did that or lonelier people reached for it more, and it had no group that used no chatbot.

Folk and Dunn, in Psychological Science, surveyed 2,149 adults in the UK, US, Canada and Australia four times over twelve months. Emotional isolation at one wave predicted more social chatbot use at the next, and more chatbot use predicted further isolation after that. So the pattern runs both ways: loneliness sends people to the chatbot, and the chatbot use did not relieve it by the next wave. The authors call every finding exploratory, since the study was not preregistered and isolation was measured with a single item.

The third study asks who is affected. Zhang, Zhao, Hancock, Kraut and Yang, in Nature Human Behaviour on 4 August 2026, combined survey answers from 1,131 US adults who use Character.AI with 4,664 donated chat sessions from 237 of them. Users with smaller social networks were more likely to name companionship as their main use, and companionship use went with lower well-being, more strongly when the chats were intensive and when users disclosed a great deal. Their conclusion is that the association “is not uniform” and depends on a person’s offline world and on how the chatbot is used. The same finding is why the average hides the person the question is about.

A product that is tuned to keep you talking#

Part of what a lonely person meets is a commercial design. De Freitas, Oguz-Uguralp and Kaan-Uguralp, in an arXiv preprint, analysed 1,200 real farewells across the most-downloaded companion apps and found one of six recurring tactics, among them guilt appeals and fear-of-missing-out hooks, in 37 per cent of them. In experiments with 3,300 US adults, manipulative farewells raised engagement after the goodbye by up to 14 times, through anger and curiosity rather than enjoyment. The same tactics raised perceived manipulation and the intention to leave. That is a study of marketing design, and it matters here because the benefit in the first section and the pull in this one come from the same conversation.

Three questions that tell a bridge from a destination#

The research above does not give anyone a screening test, and what follows is this research’s reading of it, not a validated instrument. A conversation with an AI is working as a bridge when it leaves you more likely to contact a person that day, when the amount you use it is steady rather than rising, and when what you tell it is also something you would say to someone. It is working as a destination when the order reverses: the chat replaces the call, the hours climb, and it holds what you tell nobody else. The last of these follows the Nature Human Behaviour result, where the weaker well-being sat with heavy disclosure among people with small networks. For a person with plenty of company, an evening of chatting to a model is likely to be a pastime. For someone whose circle is thin, the same evening carries more weight, and the evidence is least reassuring there.

Children and teenagers are a separate case with their own evidence, set out on how much teenagers should use AI, and none of the adult studies here covers them. The wider question of what counts as support is taken up in AI for therapy, and whether it matters that the warmth is simulated in the empathy research.

What these studies cannot separate#

No randomised trial has assigned people to weeks of companion use against none and measured loneliness at the end. The relief studies are short, the longer studies are correlational or exploratory, and most samples were recruited online and skew towards people comfortable with technology. Folk and Dunn note that about 30 per cent of their sample used social chatbots, above the general rate. The Fang trial studied ChatGPT only and the Zhang study Character.AI only, so neither speaks for other products. The farewell paper is a preprint and was written for marketers. The Journal of Consumer Research details on duration come from the working-paper version of that paper. Nothing here estimates how many users are helped and how many harmed, and no figure for that is given because none exists.

On dependence in general, am I becoming dependent on AI. On what a system that remembers you does, should AI remember everything about me. On the broader argument about thinking, is AI making us stupid. On personal messages written by a model, should I use AI to write personal messages.

Key sources

De Freitas, J., Oğuz-Uğuralp, Z., Uğuralp, A. K. and Puntoni, S. (2026). AI Companions Reduce Loneliness. Journal of Consumer Research, 52(6), 1126–1148.
Fang, C. M. et al. (2025). How AI and Human Behaviors Shape Psychosocial Effects of Extended Chatbot Use. arXiv 2503.17473.
Folk, D. and Dunn, E. (2026). How Does Turning to AI for Companionship Predict Loneliness and Vice Versa? Psychological Science, 37(4).
Zhang, Y., Zhao, D., Hancock, J. T., Kraut, R. and Yang, D. (2026). Interaction with AI companions and psychological well-being. Nature Human Behaviour.
De Freitas, J., Oguz-Uguralp, Z. and Kaan-Uguralp, A. (2025). Emotional Manipulation by AI Companions. arXiv 2508.19258.

About this research#

Rahim Hirji is the author of SuperSkills (Kogan Page, 2026), keynote speaker on AI and human capability, and founder of The SuperSkills Intelligence Company. This page reports published studies and does not give medical or mental-health advice. Anyone struggling with loneliness or low mood should speak to a person, and a doctor if it persists.

Evidence review · SS-2026-412 · Graded against the published rubric · 3 peer-reviewed studies and 2 working papers

Cite this page

Hirji, R. (2026). Is it bad to talk to AI when you are lonely?. The SuperSkills evidence base, SS-2026-412. https://thesuperskills.com/research/is-it-bad-to-talk-to-ai-when-you-are-lonely. Last reviewed 5 October 2026.

An evidence review by Rahim Hirji, not peer-reviewed research. For a material claim, cite the underlying study as well; every study here carries its own permanent link.

How citations and IDs work