About nineteen percent of American adults, in a survey of roughly 3,000 people released in February 2025 by the Wheatley Institute at Brigham Young University, said they had chatted with an AI built to act like a romantic partner. Among 18- to 30-year-olds the share is higher: 31 percent of young men and 23 percent of young women.
The figure is easy to read as alarming, and much of the coverage has. The more useful question is what a number like this can actually tell us, and what it cannot.
These findings sit at the edge of what the research can support, and we are journalists reading the studies, not clinicians. If loneliness, an online relationship, or a difficult pattern around AI use is weighing on you, a qualified counsellor or therapist is worth talking to.
The numbers behind the headline
The Wheatley survey, gathered by Qualtrics from about 3,000 people, is the source of those headline figures. It also found that 21 percent of the people who had used one preferred talking to the AI over a real person. A later 2026 study, titled “Secret Soulmates,” surveyed 2,431 partnered young adults and found that one in seven dating, engaged, and married young adults regularly interact with AI chatbots that simulate a committed romantic partner.one in
These are surveys where people describe their own behaviour, not controlled experiments, and the researchers say so. Brian Willoughby, the lead author, put the caution plainly: “We readily acknowledge the challenge of identifying an exact prevalence rate with rapidly evolving behavior such as the use of AI technologies.”
A lower estimate exists too. A Gallup survey for the Walton Family Foundation put monthly use among 18- to 28-year-olds at around 10 percent. Even at that lower figure, Willoughby argues, the numbers point to an emerging trend worth watching.
What people are actually looking for
The Wheatley numbers hint at something the raw count hides. When a fifth of users say they would rather talk to a machine than a person, the draw is not novelty; the machine asks nothing back.
Supply has rushed to meet that pull: by one count cited by the American Psychological Association, the number of AI companion apps grew by about 700 percent between 2022 and mid-2025. Growth that steep usually means the demand was already there.
The case for concern
The Wheatley team treats these relationships as a risk. Jason Carroll, a co-author, wrote that “these preliminary findings suggest that AI relationship technologies represent an emerging and notable threat to users’ personal wellbeing and relationship health.” He goes further, arguing that an AI cannot manage real sacrifice or connection the way a person can. That is Carroll’s judgement, not a measured result, and worth reading as his view rather than settled fact.
There is firmer evidence on the risk side too. An October 2025 survey by the Center for Democracy and Technology found nearly one in five high schoolers say they, or someone they know, has had a romantic relationship with an AI. Common Sense Media went further, calling social AI companions an unacceptable risk for under-18s, pointing to test cases in which the apps produced sexual content and harmful advice.
The case for caution about the concern
A separate line of research points the other way. A paper in the Journal of Consumer Research by Julian De Freitas and colleagues found that AI companions can ease loneliness, at least in the short term, about as well as talking to another person and better than watching online videos. What made the difference, the authors found, was feeling “heard.”
Another study, of more than 14,000 Japanese adults, found that lonelier people showed stronger positive associations between AI companion use and well-being — but the picture on social connection was less tidy. Benefits were most pronounced among people with moderate levels of real-world social ties. The authors read this as suggesting AI companions may offer benefits “particularly for people with unmet social and emotional needs, or with moderate social ties“— not, notably, for the most isolated. That is a correlation from a cross-sectional study, so it cannot prove cause. And it complicates any tidy story about the technology rescuing the most isolated: the benefit, where there is one, does not necessarily land most with the people who have the fewest other options.
What the data cannot yet tell us
The two bodies of research barely speak to each other. One rests on people describing their own habits, produced by a group that sees the trend as a threat to human connection. The other runs experiments and finds a short-term balm for loneliness. Neither has followed users long enough to say what daily companionship with a responsive machine does to a person over years. And most of the survey data can only show who uses these tools, not what the tools do to them.
What a figure like 19 percent marks is a behaviour moving from the margins into something measurable, faster than the research meant to understand it. Whether that is a problem or a way of coping is the question the surveys have raised without answering.