AI and therapy
6th April 2026
The arrival of AI in mental health care is raising a fascinating question — not just about the future of therapy, but about what has makes therapy effective.
Take CBT. It has built a well-deserved reputation on measurable outcomes and a clear, teachable model: identify unhelpful thinking patterns, challenge them, change how you feel. This is a genuinely powerful modality, and the evidence base is real. But here's the interesting thing — if AI can deliver that same psychoeducational content effectively (and the evidence suggests it can, for milder presentations), what does that tell us about where the therapeutic value actually lives?
I don't think it exposes a weakness in CBT. I think it clarifies something CBT practitioners have always known but perhaps underemphasised: that the relationship does a great deal of the heavy lifting. The
techniques matter — but they may work largely because of the relational context in
which they're delivered. Attunement, trust,
feeling genuinely seen by another person: these aren't soft extras. They may be the active ingredient.
Freud's modest therapeutic ambition — to move the patient from hysterical misery to ordinary unhappiness — sounds deflationary until you sit with it. What he was pointing to is something profound: that suffering becomes more bearable when it can be brought into language, and held within a relationship rather than endured in isolation. The therapist's role is to help the patient articulate what has been unspeakable — in their personal history, and in the universal issues of mortality, love, desire. This is slow, human work. It cannot be templated.
This is where the AI moment becomes philosophically interesting. For years, CBT and psychodynamic approaches have been positioned as opposites — one rigorous and solution-focused, the other exploratory and harder to measure. But if both depend critically on the therapeutic relationship to do their work, they're perhaps less different than the debate has suggested.
Psychoanalytic therapy has long been criticised for resisting measurable outcomes — for being, as critics put it, unprovable and quasi-mystical. Some of that criticism is fair. But some of it reflects a methodological bias: the things hardest to quantify aren't necessarily the things that matter least. The relational field is difficult to reproduce with AI. That doesn't make the relational field ineffective. It makes it irreducibly human.
And that is what AI is quietly demonstrating. By taking on the teachable, structured, content-driven aspects of therapeutic work, it throws into relief what it cannot replicate: genuine presence, the therapist's own subjectivity in the room, the experience of being known by another person across time.
Whatever the model — CBT, psychodynamic, integrative — the relationship has always been doing more work than our outcome measures capture. AI won't replace therapy. But it might finally give us a clearer picture of what therapy actually is.