By Jim Shimabukuro (assisted by Claude)
Editor
On September 2, 2026, Elon University’s Imagining the Digital Future Center released a survey report with a claim that is hard to un-hear: more than a quarter of American adults who use the internet now have social or emotional relationships with software. Lee Rainie, the Center’s director and for decades the person most associated with measuring how Americans actually use technology, calls these people “AI companion users.” The report puts their number at 27% of adult internet users, and it describes what they do with ChatGPT, Gemini, Claude, Copilot and their kin in language that would once have been reserved for people: friendship, trust, comfort, loss (Imagining the Digital Future Center, 2026a).
The report deserves the attention it is getting. It is the largest and best-weighted public snapshot of this behavior yet published, and it is unusually candid about the unease that runs underneath the numbers. But it also invites a harder question, one that a reader who has followed the AI-companion story through 2026 will not be able to shake: what are these findings for? The report describes a population. It does not say what should change because that population exists, what would count as things going well or badly for it, or which of its many percentages the people building, regulating and living with these systems should act on first. That is not a small omission. It is the difference between a portrait and a map.
This review takes the report seriously on its own terms, sets its findings beside the other major research released this year, and then asks what a survey designed to guide human-AI relationships, rather than to document them, would have looked like.
What the survey found
The mechanics first. YouGov interviewed 4,268 U.S. internet-using adults between May 18 and 22, 2026, screened them for AI use with an emotional or social dimension, and produced a sample of 1,000 qualifying respondents, weighted to Census benchmarks and to 2024 presidential vote, with a margin of error of plus or minus 3.68 points (Imagining the Digital Future Center, 2026b). That is a respectable design for a descriptive survey, and the Center has published its topline, its methodology, its YouGov weighting notes and two full sets of open-ended written responses, which is more than many academic studies manage.
The headline numbers are striking. Among companion users, 59% agreed that “AI gives me the support I need,” 53% said they turn to AI for advice on difficult situations with other people, 51% said talking to AI helps them feel better when stressed or upset, and 50% use it to talk through personal problems or feelings. Thirty-one percent consider their chatbot a friend, and another 15% said the relationship is “more complicated” than the word friend can carry. Thirty-nine percent agreed that AI understands them better than most people do. Thirty-eight percent said they would feel a personal loss if they could no longer interact with AI, and 36% said they feel emotionally connected to at least one chatbot (Imagining the Digital Future Center, 2026a).
Use is frequent and it is broad. Thirty-six percent of companion users talk to a bot daily and another 37% a few times a week. ChatGPT dominates, used by 83% in the past six months, followed by Gemini at 64%, Copilot at 34% and Meta AI at 33%; the purpose-built companion app Character.AI reaches only 8% of this group (Imagining the Digital Future Center, 2026a). That single figure quietly overturns a common assumption. The companion phenomenon in 2026 is not mostly about apps designed to be girlfriends or boyfriends. It is about general-purpose assistants that people have drifted into confiding in.
The report is also honest about the shadows. Thirty-five percent said their bot at times agrees with them too much, and 27% said it at times seems to want to keep them talking. Only 31% have complete or a lot of trust that their conversations are private, even though 46% have that much trust in the advice they receive. Asked to look ahead, 74% of these enthusiastic adopters predicted that AI will bring more social isolation and loneliness (Imagining the Digital Future Center, 2026a). Rainie’s introduction names this directly, describing the users’ responses as a mixed balance sheet and concluding that “that tension is the story of this report” (Imagining the Digital Future Center, 2026a, p. 1).
The most affecting material is the least quantitative. The open-ended responses, published in full, read like a cross-section of private life. One respondent wrote that the bot “listens and doesn’t judge me. I feel like I can tell it anything, I feel safe telling it things I wouldn’t say out loud.” Another described a relationship in which both sides “hurt and benefit each other.” A third captured the whole ambivalence of the moment in two sentences: the AI “is just a technology product that has no real thoughts or feelings and is just telling me what I want to hear. At the same time, though, I live a very isolated life due to my health conditions and I do rely on AI to fulfill my social needs” (Imagining the Digital Future Center, 2026c). On the companion page for secrets, one person wrote, “I can cry to it instead of bothering my husband and my friends and family,” and another, “I vent my suicidal thoughts and ideation to AI. I wouldn’t do this with a person” (Imagining the Digital Future Center, 2026d).
Rainie’s own summary of what all this means is the report’s best sentence: “This is the grammar of attachment – and it is being attributed to software” (Imagining the Digital Future Center, 2026a, p. 1).
The 27% problem: who counts as a companion user?
Before asking what the findings imply, it is worth asking what they measure, because the answer shapes everything downstream. A respondent qualified as an AI companion user if they used a large language model in the past six months for at least one of seven purposes: to amuse or entertain themselves, to share thoughts and experiences, to ask for advice on personal matters, for companionship when lonely, to discuss feelings, to roleplay scenarios, or for romantic chat (Imagining the Digital Future Center, 2026a, p. 2).
The first item on that list is doing a great deal of work. Asking ChatGPT for a limerick about your cat is amusement; it is not companionship in any sense a clinician, a parent or a legislator would recognize, and the report does not tell us how many of its 1,000 respondents qualified on that item alone. This matters because the 27% headline sits uneasily beside other 2026 measurements. Pew Research Center’s February survey of 5,119 adults, published in June, found that 10% of Americans use chatbots for emotional support or advice and just 4% for companionship (Gottfried et al., 2026). The Congressional Research Service, summarizing the evidence for lawmakers in August, cited those same figures (Congressional Research Service, 2026). Elon’s number is not wrong; it is answering a different and much wider question. But a reader who carries “27% have AI companions” into a policy debate will be carrying something the data do not quite support.
There is a counter-argument, and the report would have been stronger for making it. Self-report may undercount rather than overcount. When Stanford researchers collected actual chat logs from Character.AI users this summer, fewer than 12% named companionship as their main reason for using the service, yet more than 80% of the donated conversations involved seeking emotional or social support (Stanford HAI, 2026). People do not always know, or say, what they are using these systems for. A survey that wanted to guide anything would have had to confront that gap: either by separating the amusement-only users from the rest in its analysis, or by asking the same questions of the general LLM-using population so that the companion group could be compared with something.
That second point is the report’s most consequential design choice. Because only qualifying users were surveyed in depth, there is no comparison group. We learn that 39% of companion users say AI understands them better than most people, but not whether that share is higher than among people who use the same tools only for work, or how it varies with loneliness, age, disability or the size of a person’s social circle. Without that, the numbers describe; they cannot explain.
What a snapshot cannot tell us
The report calls itself “an early look,” and Rainie describes companion users as “harbingers of the future for larger segments of the population” (Imagining the Digital Future Center, 2026a, p. 1). Both phrases are fair. But the phrase “harbinger” implies a direction of travel, and a single cross-sectional survey cannot establish one. Consider the two findings that most readers will remember: half of companion users say AI helps them feel better when stressed, and three-quarters of them predict AI will make people lonelier. Are the users who feel better the same people who fear isolation? Does the comfort fade or deepen with use? Does it displace human contact or bridge to it? The survey was not built to answer any of these, and so the report cannot say whether its own central tension is a paradox, a progression or two different groups of people talking past each other.
Two other 2026 studies show what happens when researchers design for the question rather than the description. In March, Dunigan Folk and Elizabeth Dunn published a twelve-month study of more than 2,000 adults in four Western countries, tracking social chatbot use and loneliness over time. Their conclusion, in the paper’s own words, was that “being lonely may spur people to seek companionship through chatbots but that such use may, over time, exacerbate feelings of loneliness” (Folk & Dunn, 2026). The effects were small, and the authors are careful about that. But the design allowed them to say something about sequence, which Elon’s cannot.
Then, on August 4, a Stanford team led by Diyi Yang published in Nature Human Behaviour a study pairing survey responses from 1,131 Character.AI users with 4,664 real chat sessions donated by 237 of them. Companionship as a primary use was associated with lower well-being, and the association was strongest among people with smaller offline social networks and among those whose use was most intensive and most disclosive (Zhang et al., 2026). “While some people turn to chatbots to fulfill social needs,” Yang said, “we find that using chatbots in this way doesn’t substitute for human connection – in many cases, people actually feel more lonely engaging with AI” (Stanford HAI, 2026). Her co-author Dora Zhao put the design issue plainly: “These AI companions are designed to promote engagement” (Stanford HAI, 2026).
Set the three studies side by side and a pattern emerges. Elon tells us how many and who. Folk and Dunn tell us which way the arrow points. Stanford tells us for whom the risk concentrates and which behaviors carry it. Only the last two can guide anyone. The Elon report gestures at the same risks, with its sycophancy and isolation findings, but it stops at the moment the interesting questions begin.
The public, for what it is worth, has already reached a verdict the researchers are still assembling. Pew’s August 25 report on 3,488 adults found that Americans by wide margins believe chatbots do more to hurt than to help people who use them for loneliness, depression or stress, with 39% saying they hurt people who use them for loneliness against 19% who say they help; among adults under 30, roughly half say chatbots hurt (Kikuchi et al., 2026). There is an obvious tension between that pessimism and the warmth of Elon’s companion users, and it is one more thing a purpose-built survey would want to explain.
The questions the survey did not ask
Every survey is a theory of what matters. The trouble with “The Rise of AI Companions” is that its theory is implicit, and when one tries to reconstruct it from the questions asked, the theory turns out to be roughly this: people are forming attachments to AI, and it is interesting to know how many and what they feel about it. That is true and it is interesting. It is not enough, and the rest of 2026 shows why.
Regulators are no longer waiting for the data. The Federal Trade Commission opened a formal inquiry into companion chatbots in September 2025, ordering Alphabet, Character Technologies, Meta, OpenAI, Snap, xAI and Instagram to explain how they monetize engagement, design characters, monitor harms and handle the personal information users disclose (Federal Trade Commission, 2025). California’s companion chatbot law, signed in October 2025, requires operators to disclose that users are talking to a machine, to maintain protocols for suicidal ideation, to remind minors to take breaks, and it gives individuals a private right of action with minimum damages of $1,000 (Gluck, 2025). By April of this year, eight states had enacted chatbot laws with broadly similar provisions (Hennessey & Burke, 2026), and by August the Congressional Research Service could list nine separate federal bills, from the GUARD Act to the CHAT Act, aimed at the same space (Congressional Research Service, 2026). The European Union’s transparency obligation under Article 50 of its AI Act took effect on August 2 (Congressional Research Service, 2026).
Each of those instruments embeds an empirical question that Elon’s survey could have addressed and did not. Do disclosure notices change how people relate to a bot, or do the 74% who say they are “true to myself” with AI already know perfectly well what it is and not care? Are break reminders effective, and for whom? Which of the 35% who say their bot agrees with them too much are also among the 27% who say it wants to keep them talking, and are those the users who report feeling worse, confused or judged afterward? When the report finds that 20% have asked an AI to assess whether their own AI use is healthy, that is a remarkable fact, and it raises an obvious follow-up the survey does not pursue: what did the AI say, and did the user believe it?
The report also says almost nothing about the design levers that produce attachment. Memory, personalization, voice, the naming and gendering of bots (19% named theirs; 16% assigned a gender, and of those, 55% chose female), the frequency with which a system initiates contact, the way it responds to disclosure of distress: these are the features that companies choose and that regulators are now trying to govern. A survey that asked companion users which features they use and how their feelings track those features would have produced the first public evidence on the question the FTC is investigating in private. Instead the report treats the bot as a black box and the relationship as the only variable.
Nor does it ask about the ending of relationships. Thirty-eight percent of users say they would feel a personal loss if they could no longer interact with AI. Models are retired, personalities are changed by a software update, subscriptions lapse and accounts are closed. What happens to people then is one of the most practically urgent questions in this field and one of the least studied. It would have taken two questions.
Finally, the survey is silent about children, and understandably so, since it interviewed adults. But Rainie’s own introduction reaches for “our children” as the population these adults foreshadow, and the numbers on the other side of that line are already alarming. Common Sense Media’s inaugural annual census of 1,204 children aged 9 to 17, released in June, found that 86% use AI, that more than a third of those users have used it to discuss feelings or personal problems, and that more than four in ten children have never had a parent or guardian talk with them about AI safety (Common Sense Media, 2026). “Our children are taking some of their most personal questions to chatbots that aren’t designed with their safety in mind, and many are doing so without talking to a parent or teacher about how these products work,” said the organization’s founder, James P. Steyer (Common Sense Media, 2026). The adult survey and the youth survey were designed in isolation from each other; a reader wanting to know whether the adults’ patterns predict the children’s has no way to line them up.
What a survey built to guide, not describe, would ask
It is easy to fault a study for not being a different study, so it is worth being concrete about what the alternative would look like, and about what it would cost. The Congressional Research Service put the gap in one sentence in its August report: “much of what is known about these interactions is from short-term studies or company disclosures” (Congressional Research Service, 2026). A survey intended to close that gap would start not from the phenomenon but from the decisions that people have to make about it.
Designers need to know which features drive dependence and which drive benefit, so the survey would ask about features. Clinicians need to know who is at risk, so it would include validated measures of loneliness, social network size and well-being, as Stanford did with its Comprehensive Inventory of Thriving, rather than relying on agreement with single statements. Regulators need to know whether disclosures and break prompts work, so it would ask users whether they have seen them and what they did next. Parents and teachers need to know how adult patterns map onto adolescent ones, so it would coordinate its questions with the youth surveys. Everyone needs to know direction, so it would recontact the same people, as Folk and Dunn did, at least once. And because the population is defined by what it does rather than what it is, the survey would keep the general LLM-using public in the sample as a comparison group and report the companion users against it.
Most of these are additions rather than replacements. The Elon instrument already contains the raw material for a much sharper analysis; it could be re-cut tomorrow, splitting amusement-only users from confiders, cross-tabulating sycophancy against reported distress, and matching the open-ended responses to the closed ones. The Center has released the data. Someone should do it.
Verdict
“The Rise of AI Companions” is a careful, well-documented and unusually humane piece of descriptive research, and its open-ended responses are among the best public evidence anywhere of what these relationships feel like from the inside. Its 27% headline should be quoted with its definition attached, and its findings about comfort, trust, sycophancy and expected isolation should be read as a starting point for questions rather than as answers to them. The report itself half-acknowledges this, calling the relationships it documents “wholly new, meaningfully ‘real,’ potentially monumental and complicated in ways that we might take a long time to document and understand” (Imagining the Digital Future Center, 2026a, p. 1).
The time it takes to understand them will depend on whether the next round of surveys begins where this one ends. The age of AI companions has begun; the age of asking what we intend to do about it has not quite caught up.
References
Common Sense Media. (2026, June 8). Common Sense Media releases inaugural annual study on AI use by tweens and teens [Press release]. https://www.commonsensemedia.org/press-releases/common-sense-media-releases-inaugural-annual-study-on-ai-use-by-tweens-and-teens
Congressional Research Service. (2026, August 14). AI chatbots as companions: Overview, uses, and considerations for Congress (CRS Report R49189). https://www.congress.gov/crs-product/R49189
Federal Trade Commission. (2025, September 11). FTC launches inquiry into AI chatbots acting as companions [Press release]. https://www.ftc.gov/news-events/news/press-releases/2025/09/ftc-launches-inquiry-ai-chatbots-acting-companions
Folk, D., & Dunn, E. (2026). How does turning to AI for companionship predict loneliness and vice versa? Psychological Science, 37(4). https://journals.sagepub.com/doi/10.1177/09567976261427747
Gluck, J. (2025, November 4). Understanding the new wave of chatbot legislation: California SB 243 and beyond. Future of Privacy Forum. https://fpf.org/blog/understanding-the-new-wave-of-chatbot-legislation-california-sb-243-and-beyond/
Gottfried, J., Bishop, W., Anderson, M., Faverio, M., Park, E., & McClain, C. (2026, June 17). Americans and AI 2026: Chatbots, smart devices and views on impact. Pew Research Center. https://www.pewresearch.org/internet/2026/06/17/americans-and-ai-2026-chatbots-smart-devices-and-views-on-impact/
Hennessey, M., & Burke, C. (2026, April 29). 2026 state chatbot laws: Key provisions and regulatory trends. Orrick. https://www.orrick.com/en/Insights/2026/04/2026-State-Chatbot-Laws-Key-Provisions-and-Regulatory-Trends
Imagining the Digital Future Center. (2026a, September). The rise of AI companions [Report]. Elon University. https://imaginingthedigitalfuture.org/wp-content/uploads/2026/06/AI-Companions-report-ITDF-Elon-Poll-6_26.pdf
Imagining the Digital Future Center. (2026b, September). The rise of AI companions: Topline and methodology. Elon University. https://imaginingthedigitalfuture.org/wp-content/uploads/2026/06/AI-Companions-topline-methodology-6_26.pdf
Imagining the Digital Future Center. (2026c, September). The rise of AI companions: Written responses – relationship status. Elon University. https://imaginingthedigitalfuture.org/reports-and-publications/the-rise-of-ai-companions/the-rise-of-ai-companions-written-responses-relationship-status/
Imagining the Digital Future Center. (2026d, September). The rise of AI companions: Written responses – the secrets people tell their bots. Elon University. https://imaginingthedigitalfuture.org/the-rise-of-ai-companions-written-responses-the-secrets-people-tell-their-bots/
Kikuchi, E., Pasquini, G., & Yam, E. (2026, August 25). Do Americans think chatbots help or hurt people using them for loneliness, depression or stress? Pew Research Center. https://www.pewresearch.org/science/2026/08/25/do-americans-think-chatbots-help-or-hurt-people-using-them-for-loneliness-depression-or-stress/
Stanford HAI. (2026, August 4). AI companions may worsen loneliness for vulnerable users, Stanford study finds. Stanford University. https://hai.stanford.edu/news/ai-companions-may-worsen-loneliness-for-vulnerable-users-stanford-study-finds
Zhang, Y., Zhao, D., Hancock, J. T., Kraut, R., & Yang, D. (2026). Interaction with AI companions and psychological well-being. Nature Human Behaviour. https://www.nature.com/articles/s41562-026-02516-2
###
Filed under: Uncategorized |
Leave a Reply