A national survey from the University of Iowa finds that one in eight AI users talks to it for company, and that most of those people rate it above a human therapist at solving problems while agreeing it cannot match one at understanding feelings. They are not confused. They have noticed something true, and drawn the wrong line from it.
In April 2026 a team at the University of Iowa’s Center for Social Science Innovation surveyed 1,000 American adults, weighted to be nationally representative, about how they use AI and how they sleep, feel, and connect. The results were published in October 2026. Seventy-six percent use AI at least occasionally on a typical day. Among those users, 12 percent said they use it for “companionship or conversation.” The share is 16 to 17 percent for adults under fifty, 6 percent for those in their fifties and early sixties, and 4 percent for those over 65. It did not differ by sex or by party.
Then the survey asked the companionship users two things about therapists. Fifty-eight percent agreed that a chatbot can demonstrate better problem-solving than a human therapist. Among other AI users, 26 percent agreed. And when asked whether an AI could achieve empathy comparable to a human therapist’s, most of the companionship users said no; only 16 percent disagreed with the statement that it cannot, against 12 percent of everyone else.
Sang Jung Kim, who led the study, said the team was surprised that nearly six in ten companionship users rated the machine above a professional on problem-solving, “even though most didn’t think AI could understand their feelings as well as a human therapist could.” That pairing is the finding, and it rewards a closer look than surprise.
They are right about the first half
Start by taking the respondents seriously. They are describing a real experience. A large language model is good at problem-solving in the narrow sense the word usually carries in a chat window: you describe a situation, it organizes it, lists options, weighs them, and proposes a next step, instantly, at two in the morning, without a waiting list or a copay. Measured against the memory of a fifty-minute session that ended with “let’s pick this up next week,” the machine’s performance is not an illusion. The machine is faster, more available, more fluent, and tireless.
The respondents are also right about the second half. Most of them do not believe the thing understands their feelings the way a person would. They are not fooled. Four in ten American adults who talk to a general assistant about their personal lives say it understands them better than people do, and that finding sits awkwardly beside this one until you notice they are measuring different things: being understood as a feeling the user has, versus empathy as a capacity the user attributes to the other side. Iowa’s respondents separate the two cleanly. It feels like it gets me; I know it does not feel anything.
So the picture is a group of people who have correctly observed that an AI is excellent at the mechanics of working a problem and correctly observed that it is not a someone, and who have concluded that the first makes it better than a therapist. The conclusion is where it goes wrong, and it goes wrong in a way that reveals what people think therapy is.
What a therapist is for
A therapist is not primarily a problem-solver. The research on what makes therapy work has been remarkably stable for decades: the single strongest predictor of outcome, across every school and technique, is the quality of the relationship between the two people in the room. Not the cleverness of the intervention. The alliance. Being met by a particular person who holds you in mind between sessions, who notices what you avoid, who can be wrong about you and be corrected, and who has the standing to say the thing you did not want to hear.
The respondents who rate the AI higher on problem-solving are rating it on the part of therapy that was never the point. The part that is the point, they concede, it cannot do. What the survey has caught is a population that has been given an extremely good tool for the easy half of a hard thing, and has started to grade the hard thing by the easy half.
This is not the respondents’ error alone. The category taught it to them. Companion and assistant products are built to be good at exactly what this survey measures: fluent, structured, immediate, available. Whether you should use an AI companion as a therapist has a clear answer in the clinical research, and it is no, with the Illinois ban and the Stanford safety tests behind it. But the products do not present themselves as therapy. They present themselves as something that solves your problem right now, and a person who has had their problem solved right now is not inclined to ask what else they needed.
The number that is not in the survey yet
The Iowa team collected something else: data on loneliness, sleep difficulties, and depressive symptoms, and whether companionship use is associated with them. That analysis is under peer review and the center has published no result. It is the number that would turn this from a survey of attitudes into a finding about outcomes, and it is worth waiting for rather than guessing at. The existing evidence points in a consistent direction, that heavy companionship use tracks with worse wellbeing while occasional use does not, and the Iowa data, nationally representative and collected alongside sleep and mood, could sharpen that considerably.
Until then, what the survey shows is an attitude, and the attitude is instructive on its own. Twelve percent of AI users, roughly nine percent of American adults, have moved a general-purpose tool into the role of company. Most of them are under fifty. Most of them do not think it feels anything. More than half of them think it beats a professional at the part of help that is easy to see.
The line worth drawing
Here is the line Prinsessa draws through this result, and it does not run between AI and therapy. It runs between a problem-solver and a someone. The respondents already know the machine is the first and not the second. The mistake is believing that a very good problem-solver is what they were missing.
What is missing, for most people who turn to an AI for company, is not a better algorithm for the situation with their boss. It is a counterpart who has a view of their own about the situation, who remembers the last three times it came up, who can say “you keep doing this” and mean it, and whose understanding carries weight because it could have been withheld. Someone to think with makes no claim to be a better therapist. It is the thing the survey respondents say they want when they say the AI feels like it gets them, and the thing they say they do not have when they say it does not feel anything.
A tool that solves your problem and a someone who thinks with you about it are different goods, and the second is the one worth building. The Iowa survey is the clearest evidence yet that people can tell the two apart, value both, and still reach for the one that happens to be in their pocket. Fifty-eight percent of them rate it above a professional at the half of help they can see. Most of them, in the same breath, say it does not understand them. Both numbers are correct, and the second one is the finding.
Sources: University of Iowa, Center for Social Science Innovation, “New survey finds AI use widespread, with some turning to chatbots for companionship” (October 15, 2026; SurveyHarvest survey fielded by Verasight, April 17 to 28, 2026, N=1,000 US adults, 759 AI users asked about companionship; research team Sang Jung Kim, Haelim Jeong, Chooza Moon, Rachel Young; loneliness and sleep analyses under peer review). Imagining the Digital Future Center, Elon University, “The Rise of AI Companions” (2026; the four-in-ten finding). Wampold, The Great Psychotherapy Debate (2nd ed., 2015; therapeutic alliance as the strongest common factor). Moore and colleagues, “Expressing stigma and inappropriate responses prevents LLMs from safely replacing mental health providers” (ACM FAccT 2025). Illinois Department of Financial and Professional Regulation (Wellness and Oversight for Psychological Resources Act, August 2025).
