05/01/2026
I see a rapidly widening gap between therapists and researchers who are working on AI and mental health. This article is an attempt to bridge this gap in two ways.
I review some latest research on evaluating AI chatbot safety.
I also point out that this research used little to no human clinical input, and leaned heavily on AI evaluating AI.
1. I review three recently developed frameworks for evaluating how well AI chatbots perform in mental health contexts. Spoiler alert, AI chatbots don’t do very well keeping users safe, particularly in some contexts. But there is some improvement with later models. I think it's important for clinicians to know that these evaluation frameworks exist and to understand the results.
2. My review shows that only 10 clinicians were involved in developing and evaluating these frameworks (one framework didn’t involve any clinicians at all)! I don’t doubt that the clinicians involved were highly competent, but that’s a very small number of human therapists making a potentially enormous impact on deciding what is safe for literally millions of humans.
Overall, as a clinician, I am excited to see this research. Without knowing about it, I felt much more helpless about the potential damage of AI on mental health. For this research to have a bigger impact, I see a strong need for deeper collaboration and understanding between clinicians and AI researchers. Let’s do this together!
https://epachesky.substack.com/p/how-do-we-understand-when-ai-chatbots
