AI in Psychotherapy: What It Can Do, and What It Shouldn't Be Asked To
I'm cautiously optimistic about AI in psychotherapy. That is a harder position to hold than either of the easy ones, and most of the conversation seems to live at the extremes. Whether AI is going to revolutionise mental health care, or it will be a hollow imitation that has no business near a vulnerable person. I don't think that either is right. I think it is a useful tool that does some things well and other things badly, and the whole question is whether we can tell the difference before we hand it the wrong job.
I came to this through my own research a couple of years ago, but the position I've landed on now is a clinical one as much as an academic one, so that's how I'll write about it.
Where it works: The clearest thing to say is that AI is well suited to structured therapies. CBT and DBT are systematic in operation by design. They move in steps, they set homework, they track progress, they prompt and remind. That structure is exactly what a machine is good at. The chatbots that already exist, such as Woebot and the like, deliver cognitive restructuring and mood monitoring in a way that genuinely helps some people, and there's trial evidence behind AI-assisted DBT showing real symptom reduction. None of this surprises me. If a therapy can be written down as a procedure, a machine can carry part of the procedure.
The accessibility argument matters too, and I don't want to be glib about it. A tool that's available at three in the morning, costs little, and doesn't have a six-month waiting list is doing something for people that the system currently can't. In a country where access to therapy is as patchy as it is here, that's not a small thing. For someone holding on between sessions, or waiting to be seen at all, an interim support is better than nothing.
Where it doesn't work well at all: The trouble starts when AI is pointed at relational work. Psychodynamic, humanistic, existential therapy as these don't run on procedure. They run on the relationship itself, on attunement, on what's happening in the room that isn't being said, on transference and countertransference, on a therapist who can sit with someone and actually be moved by them. A model that doesn't have their own inner world, can't meet a client's. It can simulate the words of empathy. It can't do the thing empathy is.
This is the part the literature is consistent on, and my own sense of the work agrees. Empathy is the medium the whole thing happens in. Take it out and you don't have a leaner version of relational therapy. What you have instead is something else wearing its clothes.
The risks I'd keep in view: Three things worry me, and none of them are hypothetical.
The first is data. People bring their most private material to therapy, and AI tools run on collecting and storing exactly that. A lot of people use these tools without any idea where what they type ends up. GDPR and the new AI Act set a floor, but a confidentiality breach in a therapeutic context isn't like a leaked email it can damage the relationship that the work depends on, and you don't get that back easily.
The second is bias. A model is only as good as what it was trained on, and most of these systems have learned from data skewed toward Western, white, middle-class populations. In therapy, where someone's culture and context shape everything about how they suffer and what helps, a tool that quietly assumes a default person is going to serve some clients badly and not tell anyone it's doing it.
The third is the one I think about most: over-reliance, and the slow deskilling that comes with it. There's a fine line between using a tool and letting it do your thinking. If a therapist keeps handing the routine work over, the worry isn't only one bad decision. Instead, it is that the clinical instinct goes soft from disuse, and you stop noticing the things you used to catch. The same applies to clients. A tool that's always available, always responsive, never tired, can deepen a dependency rather than ease it.
Where I land personally. AI should support the therapist, not stand in for them. Use it for the structured, the administrative, the between-session scaffolding, and keep the human firmly in the relational work where the actual change happens. That is where the tool is genuinely good versus where it is genuinely dangerous.
Underneath all of it, the technology was never really the question. What matters is what we think therapy is for. If it's the delivery of techniques, then yes, a machine can deliver techniques. But if it's two people in a room doing something that only happens between people, then the most a tool can do is protect the therapist's attention for that and the worst it can do is be mistaken for it. I'm hopeful about the first and I'm very wary of the second. I don't think you get to hold one without the other.