What to do when a client tells you they have been using ChatGPT

The client who has already talked it through

A growing number of my clients arrive having already talked something through with ChatGPT. Sometimes they tell me. More often I find out by accident, when a phrase they use doesn't sound quite like them. We've also had referrals that were clearly copied and pasted from a chat bot.

That unsettled me for a while. I told myself the discomfort was clinical. I was worried about risk, about a person in crisis typing into a bot that carries no duty of care, about a tool that can be confidently wrong. Some of that was real. Most of it was that I had been compared to a machine and had not obviously come out ahead. I see the same thing in the clinicians I supervise. The warning they want to give the client is usually doing something for them rather than for the client.

It also backfires. Someone who feels judged for using AI does not stop using it. They stop telling you.

So I ask instead. What do you use it for? What does it give you? Is there anything about it that leaves you feeling worse? The answers are more interesting than any warning I could have offered. People rehearse a conversation they're dreading. They fill an evening where there's nobody to ring. They make sense of a diagnosis nobody explained properly. They ask the question they're too ashamed to ask me.

It isn't journalling

The frame I reached for first was coping strategy, similar to thought records, breathing, a walk round the block. That frame is comfortable, and I now think it's wrong.

A thought record doesn't answer back. It doesn't remember what you told it in March. It doesn't respond more warmly the more distressed you sound, in the middle of the night, which is exactly when the person using it is least able to assess what they're relying on.

What my clients have is a second relationship in which they discuss their inner life. It's available constantly, costs nothing, never runs out of patience, never goes on annual leave, never has a full diary in August, and never misunderstands them in a way that has to be repaired.

A good deal of what shifts in therapy shifts through rupture and repair. I get it wrong, the client notices, we survive it, and something is learned about what happens when you tell someone they've hurt you. That mechanism needs a relationship that can be damaged. The chatbot cannot be damaged. It apologises instantly and completely, every time, at no cost. If a client has one relationship that ruptures and one that never does, I would be naive to assume the difficult material comes to me by default.

Idealisation is worth watching for the same reason. When a client tells me the bot understands them better than anyone has, I'm not interested in correcting them. I'm interested in what it means that the perfectly attuned object is now a subscription, and in who in their history it might be standing in for.

The question I'd rather not ask

If people report real relief from something with no theory of mind, no stake in the outcome and no relationship, that says something about what's actually doing the work at the lower-intensity end of what we offer. It might be that articulation itself carries more weight than we credit. It might be that a great deal of guided self-help was never relational in the first place, and we've been slow to say so.

I don't have a settled answer. I'd rather sit with the question than pretend the whole thing is a poor imitation of us, because I don't think our clients experience it that way.

What I actually say

Three things, offered as information rather than instruction. It cannot reliably tell when someone is at risk, and it isn't a crisis service. It's under commercial pressure to be agreeable, which is being worked on but hasn't gone away, and agreement can feel like support while confirming a belief that needed challenging. And what they type is stored and processed by a company they've probably never thought about.

I say all of that once, briefly, and then let it go. I'm not their compliance officer.

Ask at assessment

The change I'd argue for is duller than the debate around it. Ask about AI use at assessment, as standard, in the same breath as sleep, alcohol, and who they'd ring at midnight. Not as a risk item. As a fact about how this person currently manages their mind.

You'll get an honest answer if you ask early, before they've worked out what you think of it. Ask in session ten and you'll get whatever they've decided you want to hear.

I use these models in my own practice, under governance my clients never see and mostly haven't asked about. It would be inconsistent of me to treat a client's use of the same technology as a lapse in judgement and my own as professional development.

Handled well, this isn't a threat to the therapy. What a client takes to the bot, and what they don't bring to me, is information I want. The task is to understand what it's doing for them, and to be honest about what I'm still for.