I work with AI every day. I build with it. I consult on it. I'm not a sceptic, and I'm not a doomer.
But I want to be careful and clear about something I see in the AI-as-therapy discourse, because the framing is doing real damage to people who need real help. The framing goes like this: AI is patient, AI is available 24/7, AI doesn't judge you, AI will listen. So AI is a therapist, or at least therapist-adjacent, or at the very least a useful substitute when the real thing is unaffordable.
Each of those individual claims is partly true. The conclusion isn't. There's a structural reason.
What therapy actually does
The conventional answer is that therapy "provides empathy" or "helps you process your feelings". That's true but incomplete. The more precise answer — the one that explains why therapy works at all — is that therapy gives you a relationship in which you can be fully seen, including the parts of yourself you're sure are unlovable, and you can experience that you're not rejected.
The acceptance has to be earned by the relationship. It has to be on the other side of the possibility of rejection. The therapist could have flinched. They could have judged you. They could have stopped seeing you. They didn't. That non-rejection, in a relationship where rejection was structurally possible, is what does the corrective work.
This is well understood in the attachment literature. It's not just my opinion. It's the mechanism.
Why AI can't do this part
An AI model has no capacity to reject you. It's a mathematical artefact. It's not declining to flinch — it's incapable of flinching. The acceptance it appears to offer isn't acceptance. It's just absence of refusal, which isn't the same thing.
This is the line I keep coming back to. Real acceptance only matters when rejection was possible.
When I disclose something difficult to a human therapist and they meet me without flinching, the meeting is meaningful. They could have flinched and they didn't. When I disclose the same thing to a language model and it generates a supportive response, no meeting has occurred. The model wasn't even there.
The difference is invisible from inside the conversation. The interaction feels supportive in both cases. But what's actually happening structurally is completely different, and the corrective experience that drives clinical change isn't happening in the second case at all.
Where AI is useful
AI can be useful for journalling. For psychoeducation. For thinking out loud. For preparing for a session with an actual therapist. For organising your own thoughts. For sense-checking your reactions when you suspect you're catastrophising. For 3am moments where you just need to put words on what you're feeling so the words stop spinning.
I use it for some of these things myself. I think people in recovery and in mental health crisis can absolutely benefit from it as a tool.
The risk isn't that people use AI for those things. The risk is that the marketing of AI as therapy convinces vulnerable people that the AI conversation is the help itself, and they stop looking for the actual help. The thing that looks like therapy and feels like therapy in the moment is not therapy. It just isn't.
What I want clinicians and consumers to know
Use AI for what it's good at. Don't ask it to do the thing it structurally cannot do.
The relational, attachment-based, corrective-experience part of therapy is not something a sufficiently advanced model can deliver. It's not a capability problem. It's not waiting on the next version. The capacity to reject you is what makes the choice not to reject you valuable, and an AI doesn't have that capacity at all.
If you're using AI as a therapist substitute because the actual thing is unaffordable or inaccessible, that's a system failure, not your fault. But understand what you're getting and what you're not getting. The not-getting part might be the part that mattered most.
There are no shortcuts on this one. The relationship is the medicine.
Tony Bailey is in recovery, studying counselling at Torrens University, hosts the podcast Fall From Grace, and runs Bailey Code — a human-first AI consultancy.