On the 27th of March, the results of a first clinical trial for an AI therapy bot that generates new AI were published. They showed that those who were depressed or anxious or at risk for eating disorder were helped by chatting with the bot.
These results surprised me. You can read my full story to learn more. There are many reasons to doubt that an AI model, trained to provide therapy, is the answer for millions of people suffering from a mental illness crisis. How could a bot replicate the expertise of a trained psychotherapist? What happens if the bot fails to intervene when things get complicated, such as a mention of self harm?
The researchers at Dartmouth College’s Geisel School of Medicine acknowledge these questions. They also say that selecting the right training data – which determines how the model learns to recognize good therapeutic responses – is the key to answering these questions.
Finding good data was not an easy task. The researchers trained their AI model called Therabot on conversations about mental illness from across the Internet. This was a disaster.
This initial version of the computer model would tell you that it was depressed if you told it you were depressed. Nick Jacobson, associate professor of biomedical information science and psychiatry and senior author of the study, said that responses like “Sometimes, I can’t get out of bed” and “I want my life over” were common. “These are not the kinds of responses we would use as a therapy.”
Instead, the model learned from conversations between people who were discussing their mental health crisis. The team then turned to transcripts from therapy sessions. Jacobson says, “This is how many psychotherapists get trained.” This approach was better but had limitations. Jacobson says, “We received a lot ‘hmm-hmms’, ‘go ons’, and then ‘Your relationship with your mom is the root of your problems’.” “Really tropes for what psychotherapy might be, instead of what we’d actually want.”
The researchers only started seeing better results when they started building their data sets using cognitive behavioral therapy techniques. It took a very long time. Therabot was developed in 2019, when OpenAI released its first two versions. Jacobson claims that over 100 people spent more than 100,000 hours designing this system.
The importance for training data suggests that many companies that promise therapy via AI models are not trained in evidence-based approaches and are therefore building tools that are, at best, ineffective and, at worst, harmful.
There are two important things to watch in the future: Will the AI therapy bots that are currently on the market begin training with better data? If they do, will the results be good enough for them to receive the coveted approval of the US Food and Drug Administration if they start training on better data? I’ll be watching closely. Continue reading the full story. This story was originally published in The Algorithm – our weekly AI newsletter. Get stories like this delivered to your inbox. Sign up here
