Adverse drug reactions (ADRs) from psychiatric medications are a significant concern, often leading to hospitalizations. Many individuals turn to online communities for support, but reliable information can be scarce. Large Language Models (LLMs) offer a potential solution, but how well do they truly understand the nuances of ADRs? A new study using the Psych-ADR benchmark and the Adverse Drug Reaction Response Assessment (ADRA) framework reveals that LLMs still struggle. They tend to overestimate the likelihood of ADRs, exhibiting a “risk-averse” behavior. While they can mimic the emotional tone of expert responses, they fall short when it comes to providing specific, actionable, and contextually relevant advice. Interestingly, larger LLMs don't necessarily perform better on these specialized tasks. The research highlights the importance of incorporating “lived experience” into LLM training to bridge the gap between AI-generated advice and real-world patient needs. This points towards a future where AI assists, rather than replaces, healthcare professionals, providing support and expanding access to mental healthcare, particularly in underserved areas.