The problem with AI therapy services

For a few weeks, I’ve played with an AI therapy service. I have to say that I have been pretty happy with it. Especially because that service did a few things pretty well:

• It integrated with my calendar: it could add events to my calendar to help with accountability about decisions I took.

• It made writing easier to start: sometimes it’s hard to write to clear your thoughts; having something ask you questions makes it easier.

But there are two issues with this type of AI therapy services that I couldn’t overcome:

• I don’t trust the privacy: Therapy is about saying what’s really in your mind, and putting that in a place that is connected to a server that isn’t mine just feels weird.

• It takes too many resources: AI services that run on the cloud use more resources than on-device AI. 

Early explorations of on-device AI therapy

So lately, I’ve been playing more with an on-device AI tool called jan.ai. It’s a tool that lets you install on your device local large language models that are open source (or near to open source) and use them without having to be connected to the internet and without sending any of your data to a server that stores it all. 

This is great because it solves truly the issue of privacy, and solves partially the issue of the waste of resources as it’s all on-device. It doesn’t solve that problem completely because to create the models, there was still a shit load of energy that was needed. 

The end of a therapy conversation using Jan.io

What works and doesn’t work

So after having used that tool to overcome a pretty tough mental challenge, here’s what I learned:

• It feels like a safe space: knowing all is on my computer and doesn’t leave it, and I can delete it easily, makes it more like a safe space.

• It’s more hit and miss: Even after prompting a little bit the AI, it’s less good than an actual AI therapy service that put the hours to custom train the model.

• It’s cheaper: I don’t have to pay the $20 per month, which makes this an even more affordable solution for me. 

What I’ll explore next:

It’s clear to me that having this all on your device is the way to go. Now I need to fix the quality of the conversation by prompting the AI in a stronger way.

For that, I’ll try giving the AI a book or articles about motivational interviewing and prompt it to never give advice. I’ll see if this makes it better.

But why don’t you do therapy with a human?

You might say, Daniele, good for you that you’re doing this, but wouldn’t it be better with a human?

Sure! If it was something I could afford. I won’t go into the details of the health insurance system in Switzerland, but in short, it would cost me a ton of money to do therapy with a human. And I don’t have that money.

Plus, at the moment in Switzerland, it can even be hard to get to a therapist as there is so much demand and not enough therapists out there.