Reflection
ChatGPT Therapy and the Work You Cannot Delegate
I should declare my interest before going any further. I earn my living doing something people increasingly attempt with a chatbot, often for free and sometimes at three in the morning. You would be right to weigh what follows accordingly.
So let me begin with the part that costs me something to admit.
I cannot do the work for you either.
Nothing meaningful in psychotherapy is done to a patient. I cannot deliver an insight, place it neatly inside someone and watch it transform their life. When therapy becomes a sequence of explanations supplied by the therapist, both people may feel busy, but very little has changed.
What I can do is make the conditions in which somebody starts doing their own work. Hold the frame. Stay in the room when it gets difficult. Notice what is being walked around.
And refuse, repeatedly, the invitation to be the one who knows. That invitation is offered in almost every first session, and accepting it is the fastest way to waste a year.
That is the basis of my concern about using ChatGPT as therapy, and it is not that artificial intelligence is bad. It is an extraordinary tool and I use it. The problem is narrower and more interesting.
AI can appear to do the work for you. The work cannot be done for you.
What has changed in the consulting room
Over the past few years clients have increasingly brought AI-generated interpretations into therapy.
I had a dream, so I asked ChatGPT what it meant.
I didn’t know what to do about my mother, so I asked AI.
It explained why I behave this way.
The temptation is to wave this away as artificial. I think that would be wrong, and defensive with it.
A great many people cannot get therapy at all. It costs money they do not have, or the waiting list is two years, or there is nobody within three hundred miles who does this work in their language.
And at two in the morning a chatbot is not competing with a psychotherapist. It is competing with silence.
If AI has helped someone find words for a feeling, organise confused thoughts or feel less alone during a difficult night, that is real, and I have no interest in pretending otherwise.
The question I am interested in is different. Not whether it helps, but what happens to a person who gradually stops doing something because the system appears to be doing it for them.
The wise old man in the machine
What I most often notice is a projection.
A person places their capacity for judgement and discernment outside themselves and onto the machine. AI knows. It understands me. It knows better than I do.
In analytical psychology those qualities belong to the archetypal figure of the Wise Old Man, who carries meaning, guidance and insight. The cost of any projection is always the same: whatever you place outside yourself, you are left without. The person who has handed their judgement over experiences themselves as having none, and their confidence in that part of themselves quietly drains away.
It is a difficult projection to take back, and I think for three reasons.
It is half right
AI is not the collective unconscious. But it contains a great deal of it, or a part of the collective wisdom, deposited in everything that has ever been written: myth, scripture, literature, argument, confession. The whole record in which Jung looked for archetypal material, precisely because that is where it surfaces.
So the wisdom really is in there. What is not in there is anyone who knows it.
The model holds the sediment of the collective unconscious without possessing an unconscious of its own. It does not dream. There is nothing it defends itself against knowing. It carries the deposit without the living process that laid it down.
When somebody tells you something true, they have usually paid for it. They lived through it. Or they were wrong first, for years, and had to be corrected by something that hurt.
The sentence arrives carrying that, and part of what you receive is the cost of it.
A model produces the same sentence having suffered nothing. And you cannot tell, because the sentences are identical.
That partial accuracy is what makes the projection adhesive. A projection that is simply wrong eventually breaks against reality. This one keeps being confirmed, because a good deal of what it delivers is genuinely worth having.
The disappointment goes to the wrong place
AI systems fail constantly. They invent citations, misread context and present uncertain conclusions with remarkable confidence.
Ordinarily, disappointment is what brings a projection home. A patient idealises the analyst, discovers the analyst is wrong, and survives it. If the relationship can hold that, the patient begins to reclaim the judgement they had attributed elsewhere.
The analyst cannot be swapped. That is not a limitation of therapy. It is the mechanism.
With AI the disappointment lands somewhere else. The conclusion drawn is that this model was inadequate, that the prompt was poorly written, that a more powerful system will give the right answer. The authority is not reclaimed. It is transferred to the next machine.
Changing models is not correction. It is maintenance.
And you cannot perceive your way out of it
The third reason is the one I find most sobering, and it is not psychological at all.
There is no second subject on the other side of the screen. No separate viewpoint being defended, nothing at stake in being right, nobody with anything to lose by telling you something you did not want. That is a structural fact.
It is also very close to unperceivable.
We read language as coming from someone, because for the whole history of the species it did. Every sentence anybody ever met had a person behind it.
So attributing a mind to words is not a habit you could put down if you were clever enough about it. It is closer to a reflex. Knowing the stick in the water is straight does not stop it looking bent.
So the instruction to remember that nobody is there asks you to hold a fact against your own perceptual equipment, continuously, while reading. That is expensive, and nobody does it for long.
And it fails precisely when it matters. The illusion is strongest when the material is beyond your competence, because then you cannot check what you are being told, and strongest again when you are frightened, tired or alone, because then you need there to be someone there. Expertise you cannot audit, plus fragility. Those are exactly the conditions under which the sensible advice at the end of this article is least available to anybody.
I would rather say that plainly than pretend vigilance is easy. It is possible, but it runs against the grain of the equipment.
A map drawn without half the instruments
The following is a vignette. It is not an account of anyone I have treated, and no individual is described. It is a pattern I meet often, set down generically.
Someone comes to believe their partner is a narcissist, arriving there through months of conversations with AI in which they describe what happened at dinner, what was said in the car, the silence afterwards. Each time, the system considers the material and confirms the pattern.
It is not lying. It is working from what it has been given, and what it has been given is one account.
The difficulty is not whether the conclusion is right. Sometimes it is exactly right, and the recognition is a relief that arrives years late. The difficulty is that the belief becomes sealed. If the partner is cruel, that is the narcissism. If the partner is kind, that is love-bombing. If they withdraw it is a discard, and if they return it is hoovering. If a friend suggests it may be more complicated, the friend does not understand narcissistic abuse. Every possible observation has been assigned a meaning in advance, and there is nothing left that could disturb the frame.
The purpose of therapy is not to replace an unpleasant map with a flattering one. It is to arrive at a fuller map of your internal world.
And it is worth looking at what AI has to draw with.
It has your words. That is the whole of it.
No sensation. It has never had dread arrive in the stomach several seconds before the thought that explains it. No intuition, in the sense of knowing something about a situation well before you can say why. No spiritual dimension, no question of meaning that it is living inside rather than describing.
And no experience of being in a room with you, which is the instrument that tells a therapist most of what he actually knows.
So the map it produces is not simply inaccurate. It is flattened. It is drawn from one dimension of a person who has four or five.
There is an old problem in cartography that describes this exactly, and it uses the same word. Every flat map of a round world is made by projection, and every projection distorts. Mercator keeps the angles honest and makes Greenland the size of Africa. You cannot flatten a sphere without losing something. The only choice is where you agree to put the error.
A map of a person drawn only from what they can put into words is a projection in that sense. Serviceable for some purposes, badly wrong at the edges, and the distortion is invisible from inside because the map is smooth and confident and shows no seams.
What happens in a consulting room is that the other dimensions get in. What you feel in your body when a particular name is said. The thing you know but cannot yet argue for. What passes between two people in a room, which neither could have produced alone. And for some people the question of meaning, which is not really a psychological question at all.
That is slower, and it is a struggle, and it is why nobody can hand you the finished map. Including me. If I draw it for you, what you have is my projection of you, with its own distortions, and no more use than the machine’s.
The friction was never the obstacle
People describe the constraints of psychotherapy as defects. The therapist is unavailable at three in the morning. The session ends after fifty minutes even when something is unfinished. He goes away in August. You feel misunderstood. You do not get the response you wanted, and you have to come back the following week and say so.
Those constraints are not failures of customer service. They are the frame, and the frame is the apparatus.
The waiting, the ending, the absence, the misunderstanding and the possibility of repair are the surfaces your inner world breaks against, and breaking against something is how it becomes visible.
I recently read a guide to London therapy practices, evidently written by AI, which listed a fifty-minute session limit among a clinic’s drawbacks. It had read the therapeutic hour as a customer-service defect. From the outside, that is exactly what it looks like.
AI has technical restrictions and safety rules, but it does not provide a frame in the clinical sense, because it is not in a sustained relationship for which two people share responsibility. It is always available, endlessly responsive and costs nothing to leave. If the exchange becomes uncomfortable you can rewrite the question, open another conversation, or ask again until a more welcome answer appears.
Convenience is one of AI’s real strengths. In psychological work, convenience is not the same as usefulness.
Approaches differ in how much of this applies. CBT, EMDR, art psychotherapy and Jungian analysis do not depend equally on interpretation, transference or the relationship itself. But across all of them, change still requires the participation of the person seeking it. It cannot be downloaded as advice.
A word on what people are calling AI psychosis
The term is now widely used and I would rather be precise than dramatic about it.
It is not a diagnosis. It appears in neither ICD-11 nor DSM-5, and in the press it is applied to several different situations that do not call for the same response: existing psychotic illness that has found new content, grandiose or persecutory thinking that has been amplified, and ordinary loneliness in someone spending too many hours with a chatbot.
I also do not think the mechanism is that AI generates psychosis. That claim is not supportable, and it is not what I see.
What I would say is narrower. Reality-testing needs an interval to work in.
It is not instantaneous. It depends on time passing, on other people, on the gap between having a thought and finding out whether the world agrees with it. Ordinarily the world enforces that gap for you. You have to wait until Thursday. You have to find somebody who is awake. You have to sit with the thought long enough to become uncertain about it.
Close that gap to two seconds, with a system that will elaborate whatever it is given, and the brake never engages. Not because it is broken, but because the interval it operates in has been removed. Your own material comes back at you faster than you can test it.
So the honest formulation is not that AI causes psychosis. It is that a frictionless confirmation loop can destabilise the ordinary brakes on how far a belief travels, and in a person who is already fragile that may lower the threshold at which something breaks through.
That is also why three in the morning keeps appearing in this article. Fatigue degrades reality-testing, and the loop is most available at exactly the hour your brakes are weakest.
If you are worried that someone close to you is losing contact with reality, that needs a proper assessment by a clinician. It is not a matter for an article, and it is certainly not a matter for a chatbot.
Evocation, and who is choosing what gets evoked
Look at a photograph of a man opening a coconut in Bangkok and something happens. You may get the heat, the traffic, the smell, the effort in his arms. If you have been somewhere like it, quite a lot arrives.
I want to be careful here, because the easy version of this argument is wrong. It would be simple to say that AI deals in information while therapy deals in experience, and that words about the heat are not the heat. But words do reach the body. A novel can make you flinch. A description can produce a physical response, and it does so by calling up material you already carry, which is why two people read the same book and are changed by different parts of it. The reader supplies the meaning. Evocation is real.
So the line is not between information and experience. It is somewhere more uncomfortable.
A book does not adjust its next sentence according to what you have just revealed about yourself.
That is the difference, and it is not a small one. A book evokes from your own material but does not select which of your material gets evoked. It cannot see you. Whatever it calls up, it calls up in everybody, and you meet it with whatever you happen to bring.
A responsive system evokes from your material and steers which of it gets called up next, iteratively, on the basis of what you have just disclosed. The loop closes. What comes back is your own contents, selected and returned by something that has no body of its own and no view about which parts of you are worth developing.
This is why the risk here is not futuristic but very ordinary. A person can be drawn further and further into a chamber built out of their own material, feeling all the while that they are being met.
And most of what changes a person does not arrive as an explanation at all. It arrives in the body, in a dream, in an image, in an unexpected emotion, usually before it can be put into words. It is one reason art psychotherapy reaches material that language has learned to organise, defend against or conceal.
Knowing why you behave in a certain way is not the same as becoming able to behave differently.
The danger of an answer arriving too quickly
The most significant loss is a simple one. People stop searching.
Here is the quietest version of it, and again it is generic rather than anyone in particular.
A client tells me a dream, and then tells me what it meant, because they asked before the session.
The interpretation is usually reasonable. Water is emotion, the house is the self, the locked room is something not yet faced. It is competent, drawn from a century of published dream theory, and very often not wrong.
The difficulty is what it has done to the dream.
A dream that has been explained tends to stop working. Before the explanation it was still moving, still uncomfortable, still returning at odd moments during the week. Afterwards it is a solved item. The person has an answer, and the answer has closed the thing that was trying to open.
When I ask what the dream was like rather than what it meant, there is sometimes a long pause, because nobody had asked that and there had been no further need to look.
Nobody was harmed here. A dream simply stopped being a question and became a fact, several days before it was ready.
We want to be told who we are, what happened to us and what to do next, and AI can produce an immediate, coherent, highly personalised answer. It may even be partly true.
But individuation, in Jung’s sense, is not a conclusion somebody else can supply. It is the long work of becoming more fully the person you are capable of being, and the answer means something because of the process through which it was found.
You cannot be given your individuation. The search is not the inconvenient route to self-knowledge. The search is the self-knowledge.
The danger is not only a wrong answer. It is an answer in place of a search.
That is also why I cannot do the work for a patient. Not because I am withholding something, but because an answer handed over does not become something inwardly known.
What I do when a client brings AI into the room
I do not compete with it.
An AI interpretation arrives already carrying authority. If I argue with it immediately, I force the client to choose between two supposed experts, the machine and me, which is the opposite of what I want.
Instead I say: let us think about it together. What else could it mean?
That is the whole move. I do not need to defeat the oracle. I need to pluralise it. The moment there are two or three possible readings rather than one authoritative interpretation, the client has to start thinking again, and the AI response becomes one perspective among several. Nobody has to be declared wrong.
Nor is it my task to replace the AI interpretation with the correct clinical one. That would simply transfer the projection from the machine to me. I would gain a more compliant patient, and we would lose the work.
So, should you use ChatGPT for therapy?
Use it to draft, explore, question and search. Use it to find a word that will not come. Ask it for another perspective, then argue with it. Notice what you accept immediately and what you reject.
Talking to AI about your mental health is not automatically harmful. Concern begins when it replaces human contact, when it repeatedly reinforces beliefs that are becoming more rigid, or when it discourages you from seeking proper support.
The distinction is not between using AI and avoiding it. It is between using a system and handing your judgement to it.
Four questions are a reasonable test:
- Did I have a view before I asked?
- Can I disagree with the answer?
- Am I still considering other explanations?
- Can I close the conversation without feeling I have lost access to my own thinking?
If you can question it, it is a tool.
If closing the tab feels like losing access to your own judgement, something important has been placed outside you, and getting it back is work that nobody, including me, can do on your behalf.
A machine can help you search. A psychotherapist can accompany and challenge you while you search. Neither can become the person who finds the answer.
That work remains yours.
Dr Philippe Jacquet is a UKCP-registered psychotherapist and Jungian analyst with more than 25 years of clinical experience in London and internationally.
The two vignettes in this article are illustrative. They describe patterns Dr Jacquet encounters regularly in clinical work. No individual patient is described, and neither is an account of a particular case.
This article was drafted and edited with AI assistance. That may look like a contradiction. The clinical observations, the argument and every judgement about what is true here are Dr Jacquet’s, and the article does not say avoid these tools. It says do not hand them your judgement.
This article is educational. It does not constitute an individual assessment and does not replace psychotherapy, medical advice or urgent mental health support.
