As a psychiatrist, how do you feel about a well trained AI being used to reduce the cost of talk therapy? Perhaps the patient starts with the AI and psychiatrist steps in when it gets stuck or reviews the transcript afterwards and comes up with a treatment plan or tweaks the AI for the next interaction?
Incredibly bearish. To summarise a few decades of common factors research poorly, our understanding of therapy's mechanism of action is limited, and it's not clear that any given technique or modality has superior curative power compared to another. The most important component is pantheoretical, the strength of the relationship between client and therapist. Alexa, Siri, et al. can not reliably determine whether I want the lights switched off in my bedroom or the living room, and we are seemingly decades away from the sort of strong AI required to hold meaningful conversations about the purpose of life or the nature of our interpersonal relationships. The Heart & Sould of Change is an approachable text on the topic if you're curious.
That's not to write it off completely; there have been adaptions of CBT and DBT into text and electronically-mediated forms that show promise among subclinical populations or as an adjunct to conventional therapy, but by and large a human is going to be necessary for a long, long time.
My uneducated guess is that therapy's effective mechanism is related to the placebo effect. There is a line of thought that the placebo mechanism may be related to reduction in stress, because something else will take care of the problem.
If I go to a therapist, and they put time in to help me get better, my stress response decreases, and I feel better.
Completely made up, in case that wasn't obvious. But that's my guess.
My point is that I believe even though very smart and dedicated people work in the field, and research it, the "state of the art" seems to be little more than opinion anyhow.
That is to say, it appears to me that even "dense scientific books" are either sharing opinionated descriptive analyses or just making this shit up.
While I can appreciate you trying to defend a field of research, I am also OK with keeping my own mental model of operation.
Since here we are talking about placebo, I wonder if it has a meaning to say "placebo" when speaking about psychotherapy, since there is no "substance". Of course there is going to be a placebo effect as long as people think it's going to work. I think the point is that since in mild and moderate depression we have to choose between the placebo of a drug and the placebo of psychoterapy, let's choose the one with less side effects.
On the topic of placebo, I found Irving Kirsch's talk about anti-depressants vs placebo in depression [1] very interesting.
There's a boatload of data showing that the most critical element of effective psychotherapy is the "therapeutic alliance." While this is a somewhat squishy term, it's broadly linked to the importance of compassion and relationships. For that reason, simply trying to replace a human with a robot faces huge challenges.
But that's the least creative possible solution. I think there are all kinds of behavioral interventions that are difficult or impossible for a human to do that we haven't tried yet. Almost certainly there are lots of them that would work well and would be better delivered by a bot.
Consider that a lot of the positive effect of psychotherapy is related to the patient being listened to by another human being. Presence and attention.
Good therapists offer what seem like obvious suggestions for practical accommodations that patients could have thought of if they were not impaired and distracted. Sometimes this amounts to just contradicting a thing the patient has always believed, without reason. "When you can't get it together to make dinner, you can get the same nutrition from the separate ingredients" has been a revelation to many.
Is there any modern AI that would be remotely functional for such a task, that is better than ELIZA?
I can't help but feel something like talk therapy is only going to work with all or almost nothing. You either have a full person that can actually empathize, or an AI of equivalent ability (so, true AGI). Or you become a mirror and let the patient fill in for that capability, like with ELIZA.
