Chatbot Misses Eating Disorders

The chatbot is good but not perfect according to an Mpathic study assessing artificial intelligence chatbot ability to detect subtle mental health cues and respond effectively.

The study found that the chatbots missed nuanced risk signals for eating disorders but tended to handle explicit suicide risk better. Anthropic’s Claude Sonnet 4.5 had the highest score across safety and helpfulness.  The study has me remembering the time my dad struggled to diagnose an eating disorder in an athlete with lingering knee pain. He blamed himself for weeks for his inability to improve her range of motion and lower her pain, when after 8 weeks he finally broke through, and found out the young athlete had bulimia. It was a lesson that it’s one of the tougher conditions to assess, and presents in interesting places. The widespread adoption of GLPs has ties to eating disorders and athletes – many patients develop them, some already have them going onto treatment. This discussion highlights how the healthcare community ought to approach it.

As much as 30% of GLP-1 users have disordered eating, maybe more, but many companies and frontline healthcare workers are prescribing them without screening for this and without closely monitoring the risks. Nzinga A. Harrison, M.D., Susan Wolver, MD and yours truly chat here in this 8-minute clip about the issue, offering recommendations to the healthcare community and consumers. The segment is pulled from my 50-minute conversation on the cascading and no doubt underappreciated effect of GLPs...in case you missed, reach out for the full recording.