computational therapy

See also: computational life coach

A recurring dream I have is to use AI to solve mental health. It is simultaneously one of the most meaningful things it seems like I could work on, and one of the most impossible because I understand (more than some people) the depth of the challenge.

There are very few things that could genuinely optimize all human reward. But a system that talks to all humans, asks them about their choices, observes their rewards, and then provides advice to people in similar situations, could do this. (yes I know how hard this is and how easy it would be to screw up).

Often my therapy is not super effective because my therapist and I aren't on the same page. The questions she asks don't line up with the concepts I want to think about. Fundamentally, she is a reflective surface for my own thoughts, not a source of advice or reassurance, the way that (e.g.) SuccessfulFriend would be. And she only has experience with a few dozen people; she is unlikely to have met someone 'like me' before.

But. Imagine a system that talked to millions of people throughout the world. It would recognize patterns in their thinking. It would have intuition about how to help arising from a level of experience that human therapists can never get. For any given person, it would know other people like them, and would understand in a much more fine-grained way what sort of trajectory might be helpful for that person.

It would also be a lot cheaper than a real therapist, which opens things up dramatically.

The thing is that the bar for therapy is pretty low. A lot of therapy is rubber-duck debugging. Systems like Eliza are annoying because they're so obviously fake. But it might not be hard to get something that works at least plausibly. It wouldn't have to pass a Turing test. I get benefit from talking to Otter even when there's nothing on the other end.

The human element of the relationship will be hard to replicate. But it may be that there are just a few big principles that matter. Caring about someone, keeping them engaged, helping them think positive thoughts about the future. Just as robotic pets can be powerful if they respond to touch and purr at you, even though you know they're not alive; an artificial friend that cares about you would be pretty welcome to a lot of people.

An AI system would also have real advantages over the typical human therapist:

• A machine could give better advice than humans. Human beings are limited: I can't give most of the good advice I've heard in life, because it's hard for me to remember it, or to empathize with a given person enough to know what advice would be relevant at that specific moment. but there's a lot of good advice out there, in books, if a system could be smart enough to surface the relevant bits when needed. Even just, "I think you'll like this book", if it can somehow be eerily accurate about it (yes I understand this is premised on doing better than Amazon and Google's entire business model, on the other hand they suck).
• Finding a therapist that works for you is inherently difficult. People are all just very different. And no matter how hard a therapist who is not a good match might try, they'll never be as effective as someone whose style matches with you. Therapists are human beings and they can only see the world in terms of their personality, concepts, mental models. A computer can encompass a vast range of mental models, can understand more deeply where you are on that scale, can try to pull you towards ones that seem healthier.
• When I've been stuck in depression, I hesitate to even talk about my problems in part because I don't have anyone to talk to who really 'gets' me. Talking to SuccessfulFriend was great, back when we were closer, because he shared my worldview, really understood what I was going through, and was smart enough to have relevant and interesting advice. Similarly, ExBoyfriend was smart, sympathetic, could have math and political and philosophical conversations with me and gave me an opportunity to vent about problems in my life, a sympathetic ear, maybe advice sometimes. But most of my friends I can't talk to, because they're now in different life stages, or because we never had that much in common, or (in some cases) because talking honestly with them would be hurtful, because it'd mean telling them what I dislike or disrespect about them, or that I don't trust them to give good advice.

How would you train a computational therapy system? We'd probably start with a big language model like GPT-N, and finetune on conversations of therapists. But ultimately it needs to learn through experience.

What is the reward function? It should be value aligned; its goal should be to discover its users' goals and help them achieve them. This is hard, partly because value alignment isn't well worked-out, partly because even with the correct reward signal, the rewards are often long-term and sparse. It can take years, and hundreds or thousands of conversations, before therapy has a noticeable effect. And even then, the effect might be unclear and hard to causally distinguish from other factors.

The interface can't be just text. We need more immediate communication; it needs to be at least a phone call and preferably a video call. Like TikTok, it needs to learn from reaction times---it needs to know what topics perk someone up, how to keep someone engaged. It needs to notice sadness, hesitation, etc and react accordingly.

• It shouldn't be purely addictive or exploitative. It needs to ultimately operate in the long-term best interest of the user. But it does need to provide enough immediate value to keep the relationship going.

It could contain explicit models of mental illness: a dynamical systems model of depression, anxiety, etc.

The addressable market is literally everyone. Everyone needs a friend.

What about privacy? We won't sell facts about your personal life. But the system may need to make use of specific information from another person's trajectory. It needs to somehow separate the specifics of someone's life: the names of their friends, their relationship history, their specific fears and self-doubts, from the parts of the life that will generalize.

Challenges: the difficulty of measuring impact, the danger of inadvertent harm, the sheer technical difficulty of doing anything novel and meaningful. My tendency to 'do magic' by going into a room and thinking hard might lead to a lot of insight here -- I really do understand the combination of depression and modern ML better than all but likely a very few people in the world -- but I realize that ML tech isn't anywhere near where I'd need it to be, and betting on crazy breakthroughs is not a good strategy.