There's been a renaissance of study in placebo effects, meditation, and general frameworks for how people change belief for therapeutic purposes or otherwise, but to me, that's been going on for a long time and is more about acceptance than being a new development.
One of the most exciting developments that's been coming out recently is playing with language to do what's called context-free conversational change.
Essentially, you can help someone solve an issue without actually knowing the details or even generally what they need help with. It's like homomorphic encryption for therapy. A therapist can do work, a client can report results, but the problem itself can be a black box along with a bit of the solution as well since much of the change is unconscious.
It works better with feedback (a conversation) of course, but often can be utilized in a more canned manner if you know the type of problem well enough.
I'm working on putting together an automated solution that's based on some loose grammar rules, NLP, Markov chains, and anything else I can use to help a machine be creative in language to help people solve their own problems, but as a first step as a useful tool for beginner therapists to help them get used to the ideas and frameworks with language to use.
So essentially, I'm getting a good chunk of the way toward hacking on a machine that can reliably work on people's problems without having to train a full AI or anything remotely resembling real intelligence, just mimicking it.
Before you go thinking, "Didn't they do that with Eliza?" Well yes, in a way, but my implementation is using an entirely different approach.