“We had to develop something that really is trained in the broad repertoire that a real therapist would be, which is a lot of different content areas. Thinking about all of the common mental health problems that folks might manifest and be ready to treat those,” Jacobson said. “That is why it took so long. There are a lot of things people experience.”
The team first trained Therabot on data derived from online peer support forums, such as cancer support pages. But Therabot initially replied by reinforcing the difficulty of daily life. They then turned to traditional psychotherapist training videos and scripts. Based on that data, Therabot’s replies leaned heavily on stereotypical therapy tropes like “go on” and “mhmm.”
The team ultimately pivoted to a more creative approach: writing their own hypothetical therapy transcripts that reflected productive therapy sessions, and training the model on that in-house data.
Jacobson estimated that more than 95% of Therabot’s replies now match that “gold standard,” but the team has spent the better part of two years finessing deviant responses.
“It could say anything. It really could, and we want it to say certain things and we’ve trained it to act in certain ways. But there’s ways that this could certainly go off the rails,” Jacobson said. “We’ve been essentially patching all of the holes that we’ve been systematically trying to probe for. Once we got to the point where we were not seeing any more major holes, that’s when we finally felt like it was ready for a release within a randomized controlled trial.”
The dangers of digital therapeutic apps have been subject to intense debate in recent years, especially because of those edge cases. AI-based apps in particular have been scrutinized.
Last year, the National Eating Disorders Association pulled Tessa, an AI-powered chatbot designed to provide support for people with eating disorders. Although the app was designed to be rules-based,…
Read the full article here
