JAMA + AI Weekly Update March 28, 2026

What safety, evidence, and transparency standards are needed for AI chatbots used in mental health contexts, particularly for young people?

John Torous, MD, MS, Associate Professor of Psychiatry at Beth Israel Medical Center in Boston and JAMA Psychiatry Web Editor, joined JAMA Associate Editor Yulin Hswen, ScD, to discuss what innovation and responsible guardrails look like for the future of AI chatbots in psychiatry, with a focus on youth mental health.

Dr Torous emphasized the need for high-quality randomized studies in mental health research. Echoing a common sentiment in the field, he noted that chatbot studies should move beyond simple comparisons to waitlists toward designs that rigorously isolate therapeutic components and assess true clinical impact.

Listen now on Spotify | Apple Podcasts | YouTube | JAMA.com.

Editor’s Picks in this week’s JAMA+ AI:   

• A diagnostic study assessed how leading LLM chatbots interpret descriptions of probability when characterizing medical risks. The models often abstained from offering numeric estimates of risk, especially in higher-risk contexts. When they did offer estimates, their responses more closely reflected lay interpretations than regulatory standards, suggesting LLMs may inadvertently amplify misunderstandings about medical risk. (JAMA Network Open)
• In a cohort of over 22,000 children with blunt trauma, the PECARN cervical spine injury (CSI) prediction rule outperformed two older rules, offering the highest sensitivity for detecting injury and the lowest projected CT scan rate. The results support the PECARN rule as the preferred tool for pediatric CSI risk assessment, optimizing early detection while limiting unnecessary radiation exposure. (JAMA Network Open)
• Commercial large language models remain highly vulnerable to prompt-injection attacks: 94% of attempts to manipulate clinical advice were successful, including potentially dangerous recommendations such as contraindicated pregnancy drugs. Even state-of-the-art models with advanced safety features were susceptible, highlighting an urgent need for adversarial testing and stronger safeguards before clinical use. (JAMA Network Open)