ChatGPT Said...

What happens when AI becomes the third voice in the therapy room?

There is a new person appearing in my therapy room.  I've never met them.  They don't have an appointment.  They don't pay me.  And yet increasingly I find myself discussing their opinions.

ChatGPT.  Sometimes Claude.

A client will say: ChatGPT said this is probably because of my attachment style.  Or:  I put the messages into ChatGPT and it said my partner was being manipulative.  And before I know it, there are three of us in the room.  The client.  Me.  And an extraordinarily confident collection of computer code.  This raises a question I've become increasingly interested in. 

Whose thoughts are we actually talking about? 

Imagine someone comes to therapy after an argument with their partner.  They tell me what happened.  We can explore what they experienced.  What did you feel when they said that?  What did it remind you of?  What did you want to say but didn't?  What frightened you?  What did you assume their silence meant?  There isn't necessarily an answer waiting at the end, but that’s the point.

Then imagine the client has already given the entire exchange to an AI chatbot.  Now they arrive with an interpretation.  ChatGPT said he was emotionally invalidating me.

Perhaps he was.  But something subtle has changed.  We're no longer simply exploring the client's experience.  We're also exploring an AI-generated interpretation of the client's experience.  And that interpretation hasn't appeared from nowhere.  AI systems generate responses from patterns learned from enormous quantities of information, combined with their design, training and the particular conversation they're having with the user.  They don't know your partner.  They don't know what happened last Tuesday.  They weren't in your childhood.  They can't see the expression on your face when you tell me something doesn't matter while your eyes suggest that it matters enormously.  And yet their answers can sound remarkably authoritative.

AI is incredibly validating, and being validated feels good.  Which is why so many people return to AI for advice and opinion even if it is damaging.  The other attractions of talking to AI are obvious.  It's available.  Immediately.  At two in the morning.  It doesn't look bored.  You don't have to worry about burdening it.  It’s free, mostly.

There is something powerful about putting an experience into words and receiving a response that says, essentially:  Yes. I can understand why you feel like that.

But therapy isn't always validation.  Sometimes I agree with my clients.  Equally I challenge them.  Often I pay attention to what’s not being said. Body language tells me a lot.  AI misses most of your authentic communication.  A therapist who agreed with everything you said would be very comforting.  They’d also be utterly useless. 

But make no mistake, the stakes are considerably high.  People are already bringing serious psychological distress to AI.  Thoughts of suicide, harming themselves or others.  Abuse.  Trauma. 

Research published in 2026 found that 86% of surveyed 9–17-year-olds were using AI. More than one in three AI users had discussed feelings or personal problems with it, and among those children, a quarter said AI understood them better than most people.

Earlier research found that nearly three-quarters of teenagers had used AI companions, and around one in three users had chosen an AI companion rather than a human being for an important or serious conversation.  Young people aren't simply asking AI to explain algebra anymore, they're telling it things they may not be telling us.  And researchers have found reasons to be cautious.

Tests of AI mental-health systems have found examples of chatbots responding inappropriately to suicidal thinking and delusional beliefs. Other research has raised concerns about systems being overly agreeable or validating when what somebody actually needs is challenge, reality-testing or urgent human intervention. 

The technology is changing quickly and safeguards are improving.  But the fundamental difficulty remains.  A chatbot can produce something that sounds deeply understanding without actually understanding you.

I use AI for admin tasks, research, scheduling.  Many of my clients use AI.  It isn't going back into the box because therapists disapprove of some of the ways its used.  There are useful ways of using it.  It can help find words for an experience.  It can suggest questions to think about.  It can help organise thoughts before a difficult conversation.  It may even help somebody recognise that perhaps they need to talk to another human being.  The main concern is when a possibility becomes a conclusion.

My partner didn't reply becomes My partner is avoidantly attached.

I feel terrible becomes I have depression.

There is an enormous amount of human life hiding between those sentences. 

I don't want to dismiss AI.  Increasingly, I become curious about it.  What exactly did you ask?  What did it tell you?  How did you feel when you read the answer?  Did you feel relieved?  Vindicated?  Frightened?  Understood?  Did part of you already believe it before you asked?  These are other areas where therapy can make use of client engagement with AI

AI can give us an astonishing amount of information about being human.  But information isn't the same thing as understanding.  And sometimes what happens between two actual people, sitting in a room together, wondering rather than knowing, is where healing happens.