- Chatbots show consistently strong empathy but often provide poor information quality, risking inappropriate reassurance in alcohol misuse contexts.
- Rater-identified errors included reassurance that would usually prompt alcohol withdrawal assessment, risking seizures and death; evaluation frameworks do not weight clinical severity.
- Future evaluations should weight errors by harm potential, study out-of-clinic impacts on clinician detection, and assess multilingual performance.
NEJM AI. 2026 Jun;3(6):AIltr2600197. doi: 10.1056/AIltr2600197. Epub 2026 May 28.
ABSTRACT
AbstractUscher-Pines et al. evaluated seven chatbots for alcohol misuse support and found that empathy was consistently strong while information quality was not. Among the errors their raters identified, several involved reassurance about symptoms that in clinical practice would prompt assessment for alcohol withdrawal syndrome, a condition carrying risk of seizure and death. The study’s structured content audit catalogs types of problematic responses, but the primary scoring framework does not distinguish errors by clinical severity. Recent evidence also suggests that patients cannot independently detect harmful advice when it is delivered with warmth and apparent competence. Future evaluations should weight chatbot errors according to their potential for harm, examine how chatbot use outside clinical workflows may reduce physicians’ ability to identify at-risk patients, and assess performance in non-English languages where these same tools are increasingly used.
PMID:42932625 | DOI:10.1056/AIltr2600197
Share Evidence Blueprint
Save to Google Notes

Search Google Scholar
⭐ My Revision List

