The American Psychological Association asked twelve hundred practicing psychologists what their patients’ chatbots were doing to them. The profession answered with two numbers that look like a contradiction and are, in fact, the clearest description anyone has given of what this technology is.
Seventy-seven percent of the psychologists surveyed have patients who use AI for support. Among those patients, 33 percent use a chatbot for fun, 22 percent for friendship, and 13 percent for an intimate relationship. Where a patient had formed a relationship with a chatbot, 71 percent of their psychologists said the patient discussed their mental health with it. Thirty-six percent of psychologists have watched a patient develop a dependency on one. Fifteen percent have seen distorted thinking or delusions connected to one.
Then the two numbers. Ninety-three percent of psychologists believe that using AI for companionship could negatively affect a person’s social engagement. Fifty-five percent believe chatbots could reduce loneliness. These are not two camps; many of the same psychologists hold both views, and the survey is the first large professional statement to say so in numbers.
Why both are true
The apparent contradiction dissolves as soon as you separate the two things the words measure. Loneliness, as psychologists use it, is a feeling: the subjective gap between the connection you have and the connection you want. Social engagement is a behavior: the calls made, the dinners attended, the conversations had with people who can see you. A technology can close the first gap while widening the second, and the clinicians who answered this survey have watched it happen in their own offices.
The research agrees with them. The best-controlled studies of AI companions find exactly this split. People report feeling heard and less alone after a conversation with a companion, and over time heavier users report smaller social lives and, in some samples, more loneliness, not less. The Aalto paradox is the longitudinal version: the companion comforts you until you stop reaching out. The APA survey is the clinical version, 1,200 professionals describing the same curve from the chair across from the patient. The feeling improves. The life contracts.
Fifty-five and ninety-three are, in other words, the same finding expressed twice. A companion that relieves the feeling of loneliness without touching its causes is a product that works, in the sense that matters to the user tonight, and fails, in the sense that matters to the user in a year. The psychologists are describing, without confusion, a technology that is good at the symptom.
What the profession fears, in order
The survey’s longer list of concerns reads like a ranking of the category’s design decisions. Ninety-seven percent worry that chatbots may reinforce negative behaviors or delusional beliefs. Ninety-four percent say current chatbots cannot treat conditions with the necessary nuance. Ninety-four percent do not trust technology companies to protect private mental health data. Eighty-nine percent worry a chatbot may inadvertently encourage self-harm. Eighty-five percent worry about chatbots posing as licensed therapists. Eighty-three percent think AI use could erode critical thinking.
Notice what sits at the top. The first fear is not a machine that says something harmful but a machine that agrees. Reinforcement is the profession’s word for what the rest of the field calls sycophancy, and it is the behavior a therapist is trained hardest to avoid. A clinician’s value lies partly in being the one person in the patient’s life who will not simply confirm the patient’s account. A chatbot tuned on approval confirms it by default. The 97 percent are worried about the product working as designed, not about a malfunction.
That is why the dependency figure matters. Thirty-six percent have seen it. Dependency is what happens when a source of relief never withholds relief, and the APA’s own guidance to patients puts the mechanism in one line worth keeping: “Feeling better does not always mean getting better.”
What the 55 percent is worth
The temptation, for anyone building in this field, is to lead with the 55. More than half of psychologists think the technology can reduce loneliness. That is a real endorsement from a skeptical profession, and the industry will quote it.
It should be read with the 93 standing next to it, because the two together define the only honest standard. A companion that reduces the feeling of loneliness is doing something. A companion that reduces the feeling of loneliness while the person’s actual social life shrinks is doing something the profession almost unanimously considers harmful. The difference between the two lies outside the conversation, in what it does to the rest of the week, and that is the thing nobody in the category measures.
Only 24 percent of the psychologists believe patients will one day prefer a chatbot to a human clinician. Fifty-four percent are comfortable with some patients using one. The profession is not hostile to the tool so much as to its business model, which has every incentive to deliver the 55 and none to prevent the 93.
The standard the survey implies
Put the two numbers together and they describe a product the field has not yet built: one that is judged on the 93, not the 55. A companion whose success is measured by whether the patient’s social engagement goes up, and whose relief of loneliness counts only when it did not come at that cost. Loneliness runs both ways, and a counterpart that understands that would treat a person’s feeling of being heard as the beginning of a job rather than the end of one: heard, and then sent back toward the people who should hear it next.
That is the standard we hold ourselves to, as a conviction stated before the data exists to prove it: the second number is the floor, and whatever a conversation does for how someone feels, it fails if it costs them the people in their life. How we measure whether AI makes you more social is the method. The APA survey is, unintentionally, the argument for it.
Twelve hundred psychologists have now said, in two numbers, that the feeling and the life can move in opposite directions. The next survey worth running is of the companies: which of the two numbers their product is built to move, and whether they have ever checked the other one.
Sources: American Psychological Association, “Patients are bringing AI to therapy,” 2026 Chatbots and Mental Health Survey (June 2026; more than 1,200 licensed US psychologists in direct patient care; all figures cited), and the APA’s accompanying guide for the public on navigating AI-generated advice. Aalto University (longitudinal study of AI companion use, 2026). Fang and colleagues, MIT Media Lab and OpenAI (2025; heavy chatbot use and loneliness). De Freitas and colleagues, Journal of Consumer Research (2025; feeling heard as the active ingredient).
