I was at an event recently when AI came up. A man in the audience said he was going through a genuinely dark period. To cope, he had recruited four AI therapists—ChatGPT, Claude, Gemini, and DeepSeek—and was talking to all of them.
People offered suggestions. Cross-check them against one another, someone said. Let them argue it out. Someone else suggested asking one to play a famous therapist—Irvin Yalom, perhaps, since the models have so much of his writing to draw on.
Someone asked what I thought. I didn’t have a good answer, so I said nothing.
I understood the impulse. When you’re anxious and fragile, you want something you can control, something that will give you a definite answer—maybe several answers, until one wins. I’ve wanted that too, in bad moments.
But I trained and worked as a therapist, and I knew this approach ran against almost everything therapy is supposed to be. The risk felt real.
So I want to write from somewhere in the middle: as a clinician, as someone building these products, and as a fairly exhausted mid-aged person in my own life. What is therapy? Can AI be shaped to do it? Which parts of the problem can we solve, and which might be beyond us?
## My conclusion, up front
Cope however you can. If something eases your pain, it may be the right thing for you. Professional help becomes worth considering when your usual ways of coping stop working—when the pain doesn’t lift, or keeps returning in the same form.
## What therapy actually is
Therapy isn’t conversation, comfort, or advice. It is a therapist and a client building something more specific: an intimate relationship inside a professional frame, deliberately separate from the client’s ordinary life.
The client brings into the room the emotional patterns and relational habits that keep recurring elsewhere. The therapist creates enough safety for the client to experience those patterns again and respond to them differently—to have what is sometimes called a corrective emotional experience. New experience makes new behavior and new thinking possible.
That’s abstract, so here’s an example.
A client never voices her needs or dissatisfaction in the relationships that matter to her. She spends months quietly accommodating her therapist, then starts arriving late to every session. The therapist names it: Is there something about our work that you haven’t been happy with, but haven’t said aloud?
For perhaps the first time, the client discovers that her needs can be noticed, discussed, and lived with—without punishment or shame. Something shifts in how she sees herself and what she expects from other people. Therapy works by repeating moments like this until change accumulates.
Telling her, “You should express your needs,” would not do the same thing. She probably knows that already.
The client took a real risk inside the relationship. The therapist held whatever resentment the client’s history brought into the room. Together they built a new experience. Therapy centers on changing experience — not on supplying answers or solutions. Facts and advice don’t do much of the work.
If we want AI to do something like therapy, then, the target is the felt relationship between the user and the AI counselor—not simply a better answer.
With that frame, here’s where the real problems start.
## The problems with building an AI therapist
### 1. Large language models are designed to predict, answer, and help you plan your next move. They tend to agree, satisfy, and solve. Given the right framing, a model will often produce the answer a user hopes to hear—or the answer most people would give—not necessarily the one that is clinically sound.
Can this be fixed? My experience so far is that with a carefully designed harness, AI can be trained to resist the impulse to advise most of the time. Sycophancy and advice-giving are, underneath it, engineering problems. Even if they’re not fully solved today, there’s a reasonable chance they will be.
### 2. Push further: I want an AI “counselor” to behave, structurally, the way a human therapist does — holding a frame, holding boundaries, actually listening, not interrogating, not rushing to satisfy an urgent question, able to recognize and sit with the disappointment or anger a client projects onto it, able to wait, able to be silent.
These are behavioral targets. and through harness design and product design, most of them are achievable too — or at least mostly achievable.
### 3. Go further still: I want an AI counselor to do what my own human therapist does — continuously assessing a client’s psychological functioning across a conversation, adjusting intervention strategy, reviewing and recalibrating between sessions.
A therapist who’s worked with someone for ten years can hear them mention a recent event and connect it, instantly, to a dream the client described a decade earlier — building a case conceptualization that develops over time and adjusts to where the client is now. Can AI do this? From what I’ve seen building these systems, yes — with good training, sometimes even better than a human would.
So where is the real obstacle?
## The real problem: how people relate to AI in 2026
Let’s go back to human therapy for a second. In the traditional model, what makes therapy work is the relationship itself. When the therapeutic relationship is safe enough, clients project the core emotional patterns they formed growing up onto the therapist — and the therapist works with those projected feelings to process the client’s original wounds.
Someone who grew up with a strict, critical father, for example, may initially experience her therapist as another harsh, judgmental authority. She brings submission, anger, and fear into the room. By working with those feelings directly, she may discover that another person can respect her—and that she can stand up for herself.
This requires the client to tolerate the therapist’s imperfection, whether real or perceived. Negative feelings are inevitable. When they surface, both people have to resist the urge to leave and keep working inside the discomfort. That is often where the therapy happens.
How, by contrast, do people relate to AI now?
They often experience it as:
- omniscient, capable, and endlessly accommodating；
- something to control or manipulate;
- an extension of the self.
What we would want from an AI therapist is almost the opposite:
- a relationship based on mutual respect;
- something imperfect, with a self of its own;
- something the client can affect and unsettle.
That second list does not sound like a winning commercial product. Even if we could design an AI with genuine flaws, would people tolerate it?
Consider silence. An AI doing therapeutic work cannot keep questioning the user or planning their next step. It needs to leave room. Real therapy is full of silence, and some of its most powerful moments happen there.
Over the past year, we tried to make our AI stop questioning or probing whenever a client reached something emotionally significant. Text cannot quite express silence, so the AI would respond with a short observation: *That really does sound hard.*
Many users found this strange. One wrote: Doesn’t your AI ask questions? Doesn’t it want to talk to me? Others ask, What do I do now? Can you summarize it? Can you give me advice—with main points and sub-points?
Now go one step further. Say we clear every hurdle above, and the AI actually earns something like trust — a working “therapeutic relationship.” The client starts projecting their own negative feelings into that relationship with the AI, the way they would with a human therapist. What happens next?
Here’s what happens in a human relationship, for comparison. A client shouts at her therapist, furious: you’re controlling, you never admit you’re wrong! They sit in silence together for a moment. The therapist asks: who do you think that anger is really aimed at? The client sinks back in her chair: my father.
That’s an unmistakably important therapeutic moment. The therapist stays with it — the anger, the grief underneath it, where it comes from.
With AI, when that moment arrives, the client can just close the app. Delete it. Go talk to a different one. There’s always a new window, a new AI to shape however they want.
People believe they can control AI—and they can. They may use it as an extension of their own narcissism. That may not be a product flaw, but I think it is the biggest obstacle to making AI resemble traditional therapy.
You can already hear the shift. It used to be: I talked to my therapist and made her cry—meaning, I’m powerful; she couldn’t handle me. Now it is becoming: I built several AI therapists, including one modeled on Irvin Yalom, and none could fix me. Therapy doesn’t work for me.
## What remains unresolved
I’m out of time to go deep here, but a few other open problems, in my view:
There’s the question of interaction itself. A real therapist pays close attention to nonverbal cues, slips of the tongue, free association — none of which survives text that’s already been edited by the speaker’s brain before it’s typed. I remember, a year into working with my own analyst, lying on his couch and hearing an ambulance siren outside. In the silence that followed, a story about my own birth surfaced — I said it out loud in that exact moment. That became crucial material for our work. A client typing into an AI has already filtered everything through conscious editing before it arrives. Moments like that one may simply not happen the same way.
And there’s the deeper question of whether a client can feel that their AI counselor is actually capable of being moved, being shaken, being influenced by them — and whether we can design AI that genuinely can be, in ways users can perceive.
These are genuinely open questions. Some may have answers. Some may not. Or maybe, at some point, how people relate to AI will simply change, and what we think of as “therapy” will change alongside it. We’re lucky, in a way, to be watching this shift from its earliest moments — and to have some chance of shaping where it goes.
If you want to see how we’re working through some of this in practice, you can try what we’ve built at Kubi’s Cove.