tl;dr: machines can’t emulate, yet, the stakes and the continuity needed for effective outcomes in therapy.
For now, text is a terrible medium for machine-assisted therapy. I am positive that in the near future, a multi-modal purpose-built LLM for therapy will exist, which is necessary for effective clinical judgement using micro-expressions.
Rather than replace human therapists, it will expand the pie by enabling many to seek help from a machine who otherwise wouldn’t with a human therapist.
But specifically, there are other reasons it can’t happen now, and it has to be a purpose-built LLM (or some other non-trivial wrapper):
-
Therapy does more than what good advice is supposed to. It’s a structured intervention based on a model (the likes of CBT, psychodynamic, DBT) with a case formulation that gets revised session over session.
-
Frontier models, right now, are too agreeable. Anyone who has spent more than a year in therapy knows that there is a process of rupture and repair with your therapist. You say something, they might misread you, you might push back, they might adjust. The imperfections in the process are what builds trust. There is a ton of research over the decades showing how that alliance is a stronger predictor of outcome than a specific model of therapy. Machines ain’t there, yet.
-
Not only are current models too agreeable, they are taught to fix things now. Sometimes, real change in behaviours requires sitting with your discomfort. For a machine to force you to do that, it will take more than a probabilistic distribution of the next token.
Even if somehow these things get built, I would still have my reservations. There is something unique about another human having a stake in a relationship as a mentor. I don’t know how well a machine can emulate that. But again, may be, that’s just a boomer thing and it won’t matter much to a generation growing up with AI as pervasive as water.