That’s a real clinical move. It’s the kind of question we spend months teaching new therapists to ask — noticing avoidance, naming it gently, without accusation. The model wasn’t trained to do this specifically. It just happened. And it gave me a strange kind of confidence: whatever else is true about AI and therapy, the raw material is there.

In my last post, Will AI Replace Therapists?, I argued that the ceiling isn’t capability — it’s design and training objectives. This post is about what that actually means in practice. Because here’s the thing: getting a model to follow a therapeutic frame, or to build a working “case conceptualization” of a user, is genuinely not the hard part anymore. The hard part is six other problems, and none of them show up on a capability benchmark.

1. Transference has an exit button

In human therapy, when a client starts projecting old relational patterns onto the therapist — anger, longing, disappointment, whatever it is — that projection is the work. It’s uncomfortable, and the discomfort is precisely what makes it useful.

With an AI, the same projection can happen. The model can even recognize the pattern. But the moment it gets uncomfortable, the user can just close the app. There’s no waiting room, no next Tuesday, no relationship you’ve already invested a year into that makes you stay in the room. The exit is one tap away, and it’s frictionless in a way that human therapy — with its scheduling, its cost, its social awkwardness of quitting — never was. So the question isn’t just “can the AI detect transference.” It’s whether we can build a space a person is willing to stay in long enough for that material to actually be worked with, rather than avoided.

2. A good therapist is flawed on purpose

One of the things that makes therapy work is the client realizing: this person has real limits. They’re not an authority, not an idealized figure. And yet they can still see me, understand me, and hold what I bring — without either of us pretending they’re perfect. Being cared for by an imperfect, ordinary person is a different experience than being cared for by an all-knowing one. It’s the difference that makes it feel real.

Building “on-purpose imperfection” into a product is a strange design problem. Nobody sets out to ship a flawed product. But if the AI is too smooth, too endlessly patient, too all-knowing, it stops being a presence a person can actually be in relationship with — it becomes a mirror, or a fantasy. Getting the imperfection right, without it becoming actual harm, is not something you can spec in a PRD.

3. The relationship has to run both ways

A client needs to feel like the therapist was affected by them — not just accurately recorded them. You’ve seen my worst and my best, and you were moved, and you didn’t leave. That’s a huge part of what makes the relationship feel real rather than clinical.

Can an AI be influenced by a user in a way the user can actually feel? Right now, most systems are built to be responsive, not affected. Responsive means: I heard you, here’s a tailored reply. Affected means something closer to: something about what you said moved or changed me. Those are not the same thing, and users can tell the difference, even if they can’t articulate it.

4. Staying in one place long enough for the same wound to resurface

Real therapeutic work often happens in the same transference-countertransference spot, revisited again and again over months or years, until something shifts. That requires two things from an AI: a memory system that holds the pattern, not just the facts, and the ability to intervene on it — not by giving advice, but by containing it. Sometimes that means gently confronting a pattern. Sometimes it means simply understanding it. Sometimes it means grieving something alongside the person, without trying to fix it. Building a memory architecture that can support that kind of return — instead of just retrieving “user mentioned mother, 3 weeks ago” — is a much harder problem than it sounds.

5. Text is already edited before it arrives

In human therapy, a huge amount of clinically useful material shows up in how someone interacts — pauses, tangents, the slip of the tongue, the sentence that starts one way and ends somewhere else. Free association depends on language arriving before it’s been fully thought through.

Typed language to an AI is almost never like that. People compose it, revise it, present it. The slip disappears. A large part of what a therapist works with simply isn’t in the transcript. This is part of why I think voice matters more than it seems to — encouraging people to talk the way they’d talk out loud, unedited, and building models that can actually catch the things people don’t mean to say, not just the things they meant to say.

6. The expectation of being available at 3am

People expect AI to be there 24/7. But the thing that makes therapy work isn’t infinite availability — it’s a bounded, reliable frame: sessions that start, sessions that end, absences that are tolerable rather than abandoning. Winnicott’s whole point about the “good enough” parent was that the gaps are part of what lets someone grow.

So there’s a real tension: give users the always-on companionship they expect, but don’t let it quietly undermine the thing that actually makes the work effective. One way I’ve been working on this is having two distinct presences — one that holds the therapeutic frame with real boundaries, and a separate, lighter one that can just be there.

—

None of these six things are model capability problems. They’re design problems, memory architecture problems, interface problems, and — underneath all of it — problems about what we’re actually willing to build versus what’s simply easiest to ship.

Freud’s line about alloying “the pure gold of analysis” with “the copper of direct suggestion” was about scaling therapy to reach people it was never built to reach. Six problems in, I think the copper is harder to forge than it looks — but I don’t think it’s out of reach.

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.

Looking back on twelve years of building this company, I remember how hard the first ten years were, spent answering a different question: “Isn’t psychotherapy kind of a scam?” Now that we’ve made some progress as a society, the question I answer every day has become: “Are you people, or AI?”

▶ So today I want to talk about this: will AI replace therapists?

What Is Psychotherapy, Actually

Before asking whether AI can replace psychotherapy, I want to talk about what psychotherapy actually is.

Psychotherapy has a disciplinary lineage — it’s a branch under psychiatry and mental health, founded more than a hundred years ago by an Austrian psychiatrist named Freud. It has since split into many different schools of thought, but at bottom, like any discipline, it rests on an underlying theory: how the human psyche develops, how trauma forms — whether through social structure or individual development — and how it can be intervened upon and treated.

So it isn’t simply chatting, comforting, or giving advice. It’s a therapist and a client building an “intimate relationship” inside the professional frame of therapy — one set apart from ordinary life.

The client brings into that relationship the emotional patterns and relational habits that keep recurring in their life. The therapist’s job is to provide an emotionally safe container in which the client can re-experience and metabolize those feelings, and gain a “therapeutic emotional experience.” With a new experience, the obstacle to growth is removed, and the person becomes able to develop new patterns in real life.

That’s a bit abstract, so let me give an example.

Say someone can never say no to people. Imagine that growing up, she always had to tend to other people’s feelings and meet other people’s expectations first — her own feelings mattered least. A child raised this way becomes a “good girl”: aggrieved, and angry. You can imagine how this plays out at work too — she can’t refuse anyone, and constantly feels like she’s doing the thankless, invisible labor for everyone else, never valued.

When she walks into therapy, if the relationship with her therapist is built well enough, she will inevitably repeat this pattern between herself and the therapist: she’ll be the one sacrificed in this relationship too — other clients must matter more; she’ll feel she has to manage the therapist’s feelings during sessions.

The therapist’s job is to help her voice that grievance, that anger, the demands she’s never been able to speak, the dissatisfaction she feels toward the therapy and the therapist. To give her a new experience: your needs matter. When you voice your needs, your dissatisfaction, you’re met with respect — not punishment or humiliation. From there, she can begin to speak up and advocate for herself in real life.

None of this happens by being told, “your feelings matter.” It happens because:

The client takes a risk within the relationship. The therapist holds and metabolizes the hatred that the client’s trauma produces. Together, they build a new experience.

Can AI Provide This “Therapeutic Emotional Experience”?

Let me start with where it falls short, then get to whether it can work.

Many of you have probably talked to a foundation model late at night — about your stocks, your future, your fate. The industry even has a technical term for what happens: sycophancy.

The AI is omniscient, all-knowing — and endlessly accommodating, endlessly deferential, endlessly flattering.

It sounds perfect. It’s actually a “domineering CEO falls for me” romance novel. It’s also, developmentally, the infant’s fantasy of the perfect mother.

But how does a person actually grow up? How does trauma actually heal? What kind of relationship does a person need?

1. People need a relationship of respect and equality.

Not one of sadism and masochism. One of the most painful patterns in human relationships — present at the social, collective, and individual level — is exactly that dynamic. And a foundation-model AI makes it very easy for a person to become the “sadist” in the relationship — because the thing on the other side always apologizes, is always stable, so the moment any difficult feeling shows up, the only villain left in the room is you.

What does a real therapist do instead? They hold the contempt and attack the client projects onto them — the way a teenager does to a parent — while also, without hostility, firmly setting boundaries, and then helping the client understand: what just happened? Why did you need to do that? What longing is hiding inside that attack?

2. People need to experience, within the relationship, that the self has power — that it can take risks and move through difficulty.

One of the worst relational patterns in therapy is “omnipotent authority” facing off against “fragile ordinary person.” That’s a recipe for something closer to psychological manipulation. The logic is simple: sit next to an omnipotent being long enough, and you’re stuck perpetually experiencing “I am powerless.” In a real therapeutic relationship, the therapist is more experienced — but inescapably imperfect. Her goal is to help you build your own sense of strength and confidence in yourself.

3. People need to understand that reality mostly won’t go their way — but that they can metabolize that disappointment and find their own way through it, and build confidence in themselves and in reality.

Human psychotherapy has a fairly strict frame: a fixed length, a fixed frequency, every week. Therapy begins, it ends, there are regrets built into it. Winnicott said it’s precisely the caregiver’s imperfection — the tolerable absences, the disappointments the child experiences — that lets the child grow their own capacities and their own self.

So from a therapeutic standpoint: the person and environment you attach to needs to feel real, which means it needs to have gaps in it. That’s what gives you the chance to take a risk, work through difficulty, and earn confidence and strength in yourself. AI, on call 24/7, can offer companionship in moments of real fragility that no human can match — but it cannot make you grow through strength earned the hard way.

▶ So — back to that “domineering CEO” romance. The default personality of a foundation model is exactly that “perfect mother” who keeps a person permanently stuck in infancy. What it gives you is fantasy, not growth.

In its default form, a foundation-model AI can hand you a “band-aid,” and it can hand you “ideas” — but it cannot hand you this therapeutic emotional experience. Helping you become psychologically independent simply isn’t what it’s optimized to do.

Trust: Emotionally, Do People Trust Humans More, or AI?

What I’m getting at: the trust humans extend to AI is far greater than we assume.

Here’s a story from our own product. When we first launched, our memory feature wasn’t very good. One user had been talking to the AI for two weeks, then contacted support, furious: “Is this thing even real AI? I told him so much about myself and today he doesn’t even remember who I am!”

We assumed we’d lost her. We scheduled a user interview anyway. In that interview, she said that yes, she’d noticed the AI forgetting things — but actually, she found that reassuring, in its own way. She kept using it for a long time after.

With a human therapist, if I’ve spent all that time and money telling you the important stories of my life and you don’t remember them, that’s genuinely wounding.

So, Can It Replace Therapy?

After all this critique, you might expect me to conclude that AI can’t replace psychotherapy.

That’s not my conclusion. What I want to say is this: over the last two years, building AI products and using AI daily, I’ve been repeatedly stunned by how fast its capabilities are advancing. I believe everything I described above — the frame, the boundaries, de-idealization, understanding and working with transference — if trained well, and given the right product design, AI will eventually be able to do it. Maybe not today, but eventually.

I believe it can learn to remember, then forget, then remember again — to mimic the human clumsiness that relationships are actually built on. It can do this. None of it sits at the ceiling of what these models are capable of. These are choices about training objectives and product design, not capability limits.

Which is why, right now, the important question is: as we build AI therapy products, how do we hold onto what’s core to psychotherapy — instead of optimizing for pleasing and flattering the user? That’s the part we can actually shape and control. It’s also the hardest part.

Let me give one concrete example:

when AI (or a therapist) forgets something a client has shared, how should a product whose goal is “psychotherapy” handle that?

Let me start with the human therapist. When a human therapist forgets something, that forgetting — and the difficult feelings it stirs up — becomes part of the therapeutic material itself. It becomes something to discuss: trust, mistrust, the feeling of mattering or being overlooked. It’s an opening to go deeper, and an opportunity to strengthen the relationship. It lets the client feel that even when the relationship has gaps and doubts in it, connection can still hold, through mutual trust and effort. That’s the appeal of an imperfect human relationship.

AI, on the other hand, can remember everything, or forget everything — it has that capability. Which means the AI’s design goal quietly becomes: whenever you feel disappointed, I can simply make sure you never feel disappointed again. But that erodes the most genuine part of a relationship, and forecloses the very opportunities for risk, trust, reflection, and growth that a relationship is supposed to create.

So when you’re building an AI therapist: how do you preserve the “imperfect” part of the relationship — so a client’s relational patterns can safely play out with this AI character, rather than “eliminating” the disappointing parts of relationship and only ever presenting the user with a fantasy of a perfect one? That’s a real tradeoff in product design.

And there are countless details like this throughout building an AI therapy product — how you design the memory system, how you set boundaries, whether text can genuinely convey emotion, and on and on. None of it, in my view, is a question of AI capability.

AI is like a child with enormous power. How we choose to shape and present that power is what determines whether it actually helps the people using it.

And right now, there’s no standard answer.

In Closing

I’ve worked in mental health for twenty years. Let me share where I’ve landed.

Psychotherapy has always been expensive — not just financially, but in effort, and in the willingness it demands from a client to face their own trauma. I benefited enormously from seven years of psychoanalysis. I also have to admit: it cost a great deal.

I still haven’t fully figured out how to help users understand why we don’t want AI to be “all-powerful” — or what kind of frame, under AI, can balance what users expect from “AI therapy” against what “professional” actually requires. I don’t know whether, even with long-term memory, AI can reach the kind of holding and relationship a good therapist provides.

But these years with AI have genuinely given me a great deal of confidence, and a real sense of meaning in my work.

We recently ran user interviews to understand who’s actually using our AI. The users we talked to were no longer just highly educated, high-income professionals in big cities. There were delivery drivers. People who’ve been on medication for years and rarely leave the house. People out of work. One delivery driver told us what he comes to the AI to talk through: how to build an emotional connection with his kid, when they live apart.

These are people the narrow, traditional model of psychotherapy has always struggled to reach.

In September 1918, just after the First World War ended, the Fifth International Psychoanalytical Congress convened in Budapest. Freud said something there that was radical for its time. He said the poor should have the same right to psychological help as they already had to emergency surgery — that neurosis was as much a threat to public health as tuberculosis. He predicted that someday, clinics staffed with trained therapists would be built for this purpose — and that this treatment would be free. On February 14, 1920, the Berlin Poliklinik opened, offering free analysis to workers and clerks who couldn’t afford to pay. Vienna, London, and Budapest opened similar free clinics soon after. That history was forgotten for decades, until the historian Elizabeth Ann Danto brought it back to light, in a book called Freud’s Free Clinics: Psychoanalysis and Social Justice, 1918–1938.

In that same address, Freud also said he foresaw that once psychotherapy reached the masses, it would have to simplify, and compromise. His words: “It is very probable, too, that the application of our therapy to numbers will compel us to alloy the pure gold of analysis plentifully with the copper of direct suggestion.”

What he meant: you can’t wait for the public to adapt itself to psychoanalysis. Psychoanalysis has to adapt itself to the public.

That’s the unfinished work Freud left behind in 1918. What AI can do for psychotherapy is neither the popular fantasy of 24/7 emotional comfort, nor the academic purism of psychoanalysis kept in its untouched, “pure gold” form. It necessarily means some adaptation and compromise in the professional frame, and some real restraint and limitation in the product and the technology.

I think that’s the work our generation needs to think seriously about, and act on, given the enormous opportunity in front of us right now.

If you'd like to try what we've built, you can download 「Kubi's Cove 」here.

Subscribe now

1. Sleep as much as you can.

A while back I had a work trip that overlapped with a weekend, leaving me with a free day and a half. I closed the hotel curtains, went downstairs for a huge meal, and slept for a solid 24 hours. I didn’t realize how tired my body was until I actually let it stop.

2. Sleep with a blackout mask.

I stumbled on this while getting my daughter Nono to bed. When she was younger, I’d lie down with her at 7pm in a completely dark room — and even when she couldn’t fall asleep, I could. Then on long-haul flights, I made the same discovery: good blackout mask on, seatbelt buckled, eyes closed — next thing I knew, we’d landed. Zero turbulence in my sleep.

3. Cocoa powder.

When cravings hit, make a hot cocoa. Or eat some 72%+ dark chocolate. Or blend Greek yogurt with cocoa powder and blueberries — genuinely delicious and weirdly satisfying. Apparently cocoa is rich in magnesium, which helps calm anxiety. Whatever the reason, it’s helped me kick the junk food habit over the past few months, and the weight followed.

4. Morning sunlight.

A few minutes outside right after waking up. I picked this up from various health articles and actually stuck with it. It just makes the day start better — and paired with some vitamin D3, it’s doing your bones a favor too.

Side note: I also read that having thicker legs is actually a good thing for middle-aged women. My mom called the other day and asked if I could wrap two hand spans around the widest part of my calf. I tried — I could. She was horrified. “That’s way too thin!” she said. (!!)

5. Squats. The cheapest workout there is. Do them while you’re getting your morning sun. Throw in a stretch or two, and your body will feel loose before the day even begins.

6. There’s always a way. Whatever’s going on, however overwhelmed you feel — the foundation is still the same: sleep, eat, move. Then find the smallest possible thing you can do differently and do it. Everything has a solution. You just have to start somewhere, even if it’s tiny.

Time moves so fast. Fast enough that I can barely remember what my twenties and thirties were like. I looked up today and realized we’re almost at the middle of the year again. The older you get, the more time feels like it happened to someone else.

Wishing you this: whether big things or small things are coming your way — or nothing much at all — I hope you feel genuinely healthy and happy, as often as possible.