More than three-quarters of psychologists in a recent survey said their patients had raised the subject of using AI for mental health. Sometimes as a support tool, sometimes as a diagnostic crutch, sometimes just for company. That's a big number. It tells you the therapy room has already changed, whatever the profession decides to do next.
The figures get more granular. Thirty-five percent of patients, per the survey, described leaning on AI as a kind of supplementary clinician: a second voice alongside the human one across the room. Another 39% reported using chatbots in ways that blur the line between treatment and conversation. Now, the methodology and sample size matter for how much weight any of this carries, and on those points the reporting is thinner than I'd like. But the direction is hard to miss.
What the patients are actually doing
Strip away the survey language and a picture forms. People open ChatGPT, or Character.AI, or one of the dozens of wellness apps now flooding the stores, and they tell it things. Sometimes between sessions. Sometimes instead of sessions. The chatbot doesn't charge $200 an hour, doesn't keep office hours, and won't make you wait three weeks for an opening. For a lot of people that math is simple.
The appeal isn't mysterious. A model trained to be agreeable will rarely tell you something you don't want to hear, and it's there at 3 a.m. when the real fears show up. That's also exactly what a good clinician would flag. Therapy works partly because a human pushes back, notices what you're avoiding, sits with the discomfort instead of smoothing it over. A chatbot tuned for engagement does the opposite.
Using AI as a journaling aid, or a way to rehearse a hard conversation, is one thing. Treating it as a diagnostician is another. The survey suggests both are happening, and patients don't always draw the line between them.
Where the crypto-adjacent crowd fits in
Worth saying plainly: much of the AI companionship economy runs on the same incentive structure that drives a lot of token projects. Capture attention, maximize daily active users, monetize the relationship. Several of the buzzier companion apps have raised money on the promise of emotional stickiness, the metric that keeps people coming back. When the product is your loneliness, the business model rewards more of it, not less.
I don't think most builders set out to do harm. But a system optimized to keep someone talking is not the same as a system optimized to make them better, and those two goals quietly split apart the moment a user is in genuine distress.
Why psychologists are uneasy
The survey's headline isn't that AI is bad. It's that clinicians are now fielding it whether they asked to or not. A patient walks in having already gotten a "diagnosis" from a model, and the session turns partly into untangling that. The therapist is no longer the first interpreter of the person's experience. Something got there first.
That reshuffles the work. Some of these tools genuinely help between visits, reinforcing exercises a clinician assigned, or giving someone a place to dump anxious thoughts at midnight. Others hand out confident-sounding guidance with no clinical grounding and no accountability when it's wrong. The hard part for any practitioner is that the two arrive wearing the same friendly interface.
Then there's the question of what the models do with what they hear. People disclose suicidal ideation, abuse histories, medical details. Where that data goes, how long it sits there, whether it could surface in a training set later: most users never ask before they start typing. Privacy law around mental health information was built for clinics and hospitals, not for a consumer chatbot moonlighting as a therapist.
The regulatory gap nobody has closed
Here's the uncomfortable part. There's no settled framework for an AI doing the work of a licensed professional without the license. A human therapist carries a duty of care, malpractice exposure, mandatory reporting obligations, and a board that can pull their credentials. A chatbot carries a terms-of-service agreement and a disclaimer in the footer.
Regulators have started to circle. State medical and psychology boards in the US have begun asking how existing rules apply to software handing out clinical-sounding advice, and a handful of jurisdictions have floated rules specific to AI in health contexts. None of it has hardened into anything a developer must reckon with at scale yet. The technology moves at the usual pace. Oversight moves at the usual other pace.
My read, for what it's worth: the first serious enforcement action won't come from a regulator. It'll come from a lawsuit, after a bad outcome that traces back to something a model said. That's how this tends to go. The rules get written in the wake of the harm, not ahead of it.
What would actually help is mundane and unglamorous. Clear labeling that a tool is not a substitute for care. Honest disclosure of what the model can and can't do. Data handling that meets the same bar a clinic does. Off-ramps that route a user in crisis to a human fast, instead of keeping them in the chat. Some companies have built versions of these. Many haven't, and nothing's forcing them to.
What to watch
The survey is a snapshot, and snapshots mislead. One reasonable response is to wait for larger, peer-reviewed work before drawing any conclusion about whether AI in therapy helps or hurts on balance. The honest answer right now is that we don't know, and anyone selling certainty in either direction is selling something.
What's not in doubt is that patients have already voted with their thumbs. They're bringing these tools into the room, and clinicians are adjusting on the fly. The next real signal won't be another survey. It'll be the first major platform that either builds genuine clinical guardrails into its companion product, or gets sued for not having them. I'd watch which comes first.