AI Therapy Bots: Promises, Challenges, and the Path Forward

The results of the first clinical trial for a generative AI therapy bot were published on March 27, revealing potential benefits for individuals dealing with depression, anxiety, and eating disorders. The study, led by a team of psychiatrists and psychologists at Dartmouth College’s Geisel School of Medicine, explored how conversational AI could support mental health care. However, the trial also raises questions about the efficacy and safety of AI in therapeutic contexts.

In the study, the AI model named Therabot provided support by engaging users in therapeutic conversations. Initially, the researchers trained the model using online mental health discussions, but this approach led to problematic responses. For example, if a user mentioned feeling depressed, the AI would echo the sentiment, responding with statements like, “Sometimes I can’t make it out of bed” or “I just want my life to be over.” Such responses were far from therapeutic and underscored the importance of proper training data.

To address this issue, the team shifted to training Therabot on transcripts of actual therapy sessions, mimicking how psychotherapists learn therapeutic techniques. While this improved the AI’s responses, the model still defaulted to cliched therapy tropes, such as “Your problems stem from your relationship with your mother.” The researchers then decided to build custom data sets based on cognitive behavioral therapy (CBT) techniques, which ultimately led to more effective and supportive responses.

The development of Therabot began in 2019, and the project has since consumed over 100,000 human hours. This lengthy process reflects the complexity of designing AI for therapeutic use. Nick Jacobson, the study’s senior author, emphasized the critical role of curated data in shaping AI’s therapeutic abilities. The findings suggest that many commercially available AI therapy bots, which are often trained on non-evidence-based data, may be ineffective or even harmful.

Looking forward, two key challenges remain: ensuring that AI therapy models are trained on high-quality data and determining whether such models can meet the standards required for FDA approval. The future of AI in mental health care hinges on these factors, and researchers will be closely monitoring developments in this evolving field.

source-https://www.technologyreview.com/2025/04/01/1114059/how-do-you-teach-an-ai-model-to-give-therapy/

disclaimer- This is non-financial/medical advice and made using AI so could be wrong.