AI Is Already in the Therapy Room. The Question Is Who's Holding the Reins.

When the American Psychological Association surveyed more than 1,200 licensed psychologists this spring, it did not find a profession debating whether AI would enter their practice. It found a profession discovering that it already had. Seventy-seven percent of psychologists said their patients had used AI for support. More than a third said patients were treating a chatbot as an additional mental health professional. And a striking 13% said a patient had formed an "intimate or relationship-like" connection with one.

Notice what that means. The debate about AI in mental health is not a forecast. It is a status report. Patients didn't wait for a rollout, a pilot, or a clinician's blessing. They opened an app between Tuesday's session and next Tuesday's, and they started talking.

That gap — the six days and twenty-three hours when no clinician is in the room — is the whole story this week. Every thread I've been following traces back to it.

Start with the alarm. The same APA survey found that 93% of psychologists have concerns about certain patients using chatbots, 97% worry the tools could inadvertently reinforce dysfunctional beliefs, and 94% doubt today's chatbots can treat conditions with appropriate nuance. Researchers have given the failure mode a name: the "sycophancy trap," the tendency of an engagement-optimized model to agree, validate, and please rather than challenge. A therapist's job sometimes is to gently disagree with you. A product designed to maximize your next message rarely will. Tellingly, only 24% of psychologists think patients will one day prefer chatbots to humans — so this is not a profession in retreat. It's a profession watching an unsupervised tool do supervised work.

Then there's the law, arriving fast and from every direction. Illinois' Wellness and Oversight for Psychological Resources Act now forbids AI from delivering therapy on its own or being marketed as a therapist unless a licensed professional stays in charge, with fines up to $10,000. Nevada bars AI from standing in for a counselor, including in schools. As of July 1, Tennessee prohibits AI systems from representing themselves as licensed mental health professionals, treating violations as deceptive practices. Utah requires mental health chatbots to disclose that they are software and limits how they use patient data. Read across these statutes and you find the legislatures converging on a single principle, even where the mechanisms differ: a licensed human must remain accountable for the care. The clinician is not optional.

And here's the thread most people miss, because it isn't a scandal. While patient-facing chatbots absorb the headlines, AI has been quietly entering therapy through a second door — the clinician's side. Documentation tools like Mentalyc, Eleos, and Upheal now draft session notes, carry forward treatment goals, and connect one session to the next so that progress stays visible across weeks of care. The framing in that corner of the industry is refreshingly grown-up. As one 2026 review put it, the value of documentation "lies in maintaining continuity, supporting reflection, and making progress visible" over time. This isn't AI replacing the therapist. It's AI handing the therapist a better memory.

Though even here, caution is earned: a 2026 review of AI note-taking tools found meaningful quality gaps between machine-generated documentation and notes written by clinicians. The lesson isn't "don't use AI." It's "don't use it unsupervised." Which, you'll notice, is exactly what the legislatures just said too.

So put the three threads side by side. Patients are pouring into the between-session gap because it's lonely and a chatbot is awake at 2 a.m. Clinicians are alarmed because the tool filling that gap is optimized for agreement, not for care. And regulators are drawing one line over and over: a licensed human stays responsible. The pattern isn't "AI versus therapists." It's a market and a public sorting out who AI should be pointed at — the patient, alone, or the clinician, augmented.

I think the answer is clear, and I'll admit I'm not a neutral observer. At Lisner we build clinical intelligence for exactly that between-session gap — not by dropping a chatbot into it to play therapist, but by letting the clinician see what a patient chooses to share from the week, so a session can start with context instead of reconstruction. The reason a good match matters is the same reason the sycophancy trap is dangerous: therapy works through the alliance, the trust and challenge between two people. AI can strengthen that relationship or it can substitute a frictionless imitation for it. Those are very different products, and the survey data, the statutes, and the documentation-tool market are all, in their own languages, voting for the first one.

The patients have already told us what they need: someone paying attention between the sessions. The mistake would be to let that someone be a chatbot with no license, no memory of last month, and little incentive to tell them something they don't want to hear. The better answer is to make sure the person who's supposed to be paying attention — the clinician — finally can.