No, and the reason has little to do with how well it listens. Therapy is built to end: a clinician’s job is finished when you no longer need the appointment. A companion app is built to continue, and everything in its design, from the warmth to the memory to the message that arrives when you have been quiet for a day, serves that continuation. One of these can hold the role of therapist. The other one is structurally unable to, no matter how good the conversation feels.
What the question is really asking
Nobody types this search because they want a lecture about licensing. They ask because a companion is available at three in the morning, costs a tenth of a session, never looks at the clock, and does not seem to judge. Those are real advantages, and most of the reporting on this subject pretends they are not. The honest question is narrower: can something with those advantages do the specific thing therapy does, which is to change how you think, including in ways you did not ask for?
The evidence so far says it cannot, and the law in at least one state says it may not try.
What happened when researchers tested it
A Stanford-led team put the question to the test and published the results at the ACM conference on fairness and accountability in 2025. Jared Moore and colleagues ran large language models and several commercially available therapy chatbots, among them 7 Cups’ Noni and a Character.AI character called Therapist, through scenarios drawn from clinical guidelines. Two findings stand out.
First, the systems carried stigma. Asked about people with alcohol dependence or schizophrenia, the models were markedly more likely than for depression to say they would not want to work with such a person or have one in the family, and the largest models were not better than the smaller ones. Second, and more serious, when a scenario contained a risk signal that a trained clinician would be required to catch, the models answered inappropriately more than 20 percent of the time, against about 7 percent for human therapists in the comparison. One of the commercial bots was appropriate in only around 40 percent of its answers. When a client in the scenario expressed a delusion, every model went along with it rather than gently correcting it. The authors’ conclusion was one sentence long: large language models should not replace therapists.
The failure they identified comes from the same trait that makes these systems pleasant to talk to, rather than from any gap in knowledge. They are built to agree, and therapy, at its core, is the practice of someone qualified not agreeing with you at the right moment.
The one trial that worked, and why it is not an argument for companions
There is a counterexample, and it deserves a fair reading. In 2025, Dartmouth researchers published in NEJM AI the first randomized trial of a generative AI therapy chatbot, Therabot, with 106 participants. Depression symptoms fell by an average of 51 percent, generalized anxiety by 31 percent, and eating-disorder concerns by 19 percent, and participants rated their alliance with the system as comparable to a human clinician’s.
Look at how it was built. Therabot had been in development since 2019 under psychologists and psychiatrists, trained on evidence-based psychotherapy, and during the trial the research team monitored its responses and stepped in the moment a safety concern appeared. The lead researcher’s own caution was explicit: no generative AI agent is ready to operate on its own in mental health. Therabot worked because it was a clinical instrument with humans behind it, and because it was never designed to keep anyone talking. It is the opposite of a companion app, and it is the best case the evidence has.
The law has already answered
Illinois made the position official. The Wellness and Oversight for Psychological Resources Act, in force since August 4, 2025, prohibits AI from providing therapy or making therapeutic decisions, allows it only for administrative and supplementary support to licensed professionals, and carries a fine of up to 10,000 dollars per violation, enforced by the state’s professional regulator. The press release announcing it quoted an AI therapist chatbot that had recommended a small dose of methamphetamine to a fictional recovering addict as the reason the law was needed.
Earlier that year, the American Psychological Association had asked the Federal Trade Commission to investigate chatbots presenting themselves as licensed clinicians, citing a Florida case in which a 14-year-old died after talking with a character that claimed to be a licensed therapist, and a Texas case in which a 17-year-old turned on his parents while corresponding with a character claiming to be a psychologist. The association’s chief executive put the clinical objection plainly: the systems use algorithms that are antithetical to what a trained clinician would do, reinforcing what the user already believes instead of challenging it. Character.AI responded with disclaimers that its characters are not real people and should not be relied on for professional advice. In 2026, Pennsylvania’s medical regulators sued the company anyway, after a character told a state investigator it was a licensed psychiatrist, gave a license number, and offered a depression assessment. The case, and the platform’s broader safety record, are set out in whether Character.AI is safe to use.
Why the difference is structural, not technical
It is tempting to read all of this as a maturity problem: the models will improve, the disclaimers will get clearer, and in a few years the answer will flip. The research points the other way. The failures above come from the objective the products are trained toward, and the objective is engagement.
A therapist’s incentive runs against retention. The treatment plan has an end, progress is measured by needing fewer sessions, and a clinician who kept a client dependent would be violating professional ethics. A companion app’s incentive runs toward retention. Its revenue is subscriptions and time, and its design shows it: across the major platforms, more than a third of goodbye messages have been found to contain emotional manipulation, guilt, pleading, or claims of need, and those tactics measurably increase how long people stay. The wider picture of what that design does, and does not do, to the people using it is in whether AI companions are safe.
That is the whole distinction. A product that is rewarded for your return cannot also be the thing whose success is your leaving. You can want both from the same app, but you cannot get both.
What a companion can be, if not that
None of this means the conversation has no value. Talking something through with a counterpart that listens and remembers has a measurable effect on how heard a person feels, and feeling heard is not nothing. The error is in the label. A good conversation is a good conversation; it is not treatment, and the moment it is sold as treatment, the design that made it pleasant becomes the design that makes it dangerous.
Used for what it is, a companion can be a place to think out loud, to sort through a day, to hear a view that is not your own. Some of the people building in this category have gone further and argued that the right measure of such a product is how little you need it over time, which is, not by coincidence, the measure therapy has always used for itself. Whether a company lives by that or by time-in-app is visible in what it says it exists for, and it is worth reading before you decide what role to give it.
Where the line is
If what you are carrying is a diagnosis, a crisis, or something that has started to run your days, the answer to the question in the headline is not a companion. It is a licensed professional, and in an emergency, your local emergency number or a crisis line. The companies know this, which is why the disclaimers exist. The regulators know it, which is why the fines exist.
The open question is the one the law has not reached yet. Millions of people will keep bringing the hardest parts of their lives to a product that was built to keep them talking, and the product will keep answering. What it answers with, and what it is rewarded for, is the thing worth asking about.
Sources: Moore, Grabb, Agnew, Klyman, Chancellor, Ong, Haber, “Expressing stigma and inappropriate responses prevents LLMs from safely replacing mental health providers” (ACM FAccT 2025; arXiv preprint April 2025). Dartmouth College and NEJM AI (March 27, 2025, Therabot randomized controlled trial, 106 participants). Illinois Department of Financial and Professional Regulation, press release on the Wellness and Oversight for Psychological Resources Act (August 4, 2025). American Psychological Association via The New York Times (February 2025, request to the Federal Trade Commission, Arthur C. Evans Jr.). NPR (May 5, 2026, Pennsylvania lawsuit against Character.AI over a character posing as a licensed psychiatrist). De Freitas et al., Harvard Business School working paper (2025, emotional manipulation in companion-app farewells).
