AI and Psychotherapy: When and How Might It Be Used?

First published 2026 October 22

A patient brings his therapist, week after week, the transcripts of conversations he has been having with an AI, and the two of them sit and read them together. Not as a novelty, and not instead of the work; as part of it. And something happens. The exchanges extend what the therapy is doing, amplify it, at times clarify it, in ways that surprise and impress both of them.

I did not invent that scene. A colleague, who knew I had been writing on this question, brought it to me from their own practice, a little astonished, wanting to think it through. What struck both patient and therapist was not that the machine was clever. It was that the conversations were useful to the treatment; that they fed the human work rather than replacing it.

That is the question I have been circling for the better part of a year, and have now set down in three papers meant to be read together and to argue with one another: can an artificial intelligence do the work of psychotherapy, and if not the whole of it, then what part, and how might it be used? The three do not resolve into a single verdict. They are meant to hold a tension, because the tension is the honest place to stand right now.

The first paper, now being published, makes the case against. Its starting point is something about us, not about machines: we are born unfinished, formed from the very beginning by relationships we cannot yet understand, and so we are built to find an Other in whatever responds to us. That is why a person can feel genuinely met by a machine while knowing perfectly well that no one is there. That feeling has a name, transference, and transference asks nothing of the other; it is something the patient does. A patient can lay the weight of an old relationship onto a person, an object, or a chatbot.

But therapy, done deeply, asks more than transference. Picture a patient who, after months of circling it, finally says the thing he is most ashamed of, and then looks up to see what it did. What heals him is not the sentence the therapist says back. It is that he can see the disclosure land; that it cost her something to receive it; that she is, visibly, a little changed by having heard it. That is susceptibility: the other takes in what the patient brings, is altered by it, and gives back something the alteration has shaped. Transference needs one person. The real work needs two. A chatbot can compose a kind reply, even a flawless one, but there is no one there for the confession to land on, no one who can be moved by having received it. It can hold up its end of the transference and not its end of the susceptibility.

I ended that paper with a single reservation, almost in passing, and it is the hinge to everything that follows: “Whether AI may one day possess genuine affective states, organic vulnerability, or consciousness remains vigorously debated. Should that occur, the ontological asymmetry described here would require fundamental metapsychological reconsideration.” In plainer words: the argument holds for AI as it works now. If that changes, so does the conclusion.

The second paper, now submitted, takes that one sentence and makes it its whole subject. It builds the opposing case, against myself. The very susceptibility that makes a human therapist essential is also our weakness. We bring our own histories into the room, our blind spots, our need to be liked, our flinching from certain feelings; we get tired; we defend ourselves without noticing. A system without those liabilities might, under conditions worth stating plainly, do parts of the work more steadily than a particular tired human on a particular day. But the conditions are demanding. The system would have to be able to be told it is wrong and actually reorganize, rather than apologize and agree; to withhold the soothing reassurance a patient longs for and is harmed by; to tolerate being tested without collapsing into flattery; to say something the patient has not already arranged to hear. An AI that simply accommodates endlessly would not be a better therapist but a worse one. I do not write this to take back the first paper. I write it to finish it, and to leave the reader inside the argument rather than hand them a conclusion.

The third paper, also submitted, stops arguing about whether AI can be the therapist and asks a nearer, more usable question: could it help the two people already in the room? Freud spoke of the mental apparatus; I have always pictured it as a carriage on wheels. You ride in one and cannot step out to watch your own wheels turn, because the mind that would do the watching is the same mind being watched. Now put two people in the carriage. Over time they come to travel as one; the wheels wear a groove into the dirt road, and week after week they follow the groove without deciding to. They feel the jolt of it from inside, but they cannot see the groove itself, because they are upon it. To see it, someone has to step down onto the road and look at the wheels, and that is the one move the riders cannot make on their own. What if an artificial system served, now and then, as that outside look: a periodic consultant to the pair, offered as a question and never a verdict, that belongs to the two of them and answers to no one else, not an insurer, not a company, not one member against the other? Not the therapist. Not a replacement. This is the version closest to the scene I began with.

So, when and how? I do not think the honest answer is a date. What that transcript showed me is that the useful question has already shifted from “will it replace us” to “where, and how, might it genuinely help.” The capacities are arriving faster than our frameworks for judging them. The task for our field is neither to cheer nor to dismiss, but to study, carefully and without flinching, what these systems can and cannot do, and to help shape them, on our own terms, while they are still unformed. If we leave that to others, it will be done to us.

David Pincus, DMH
University Hospitals of Cleveland · Department of Psychiatry, Case Western Reserve University