Generative models trained on electronic health records (EHRs) are considered “null-shot predictors” of clinical outcomes. However, this interpretation is misleading. The article explains how to correctly interpret the "zero shot" results from these models. Generative EHR models generate new health data based on training data. A "zero shot" means evaluating the model without specific tuning for the target task. The authors emphasize that such results may not reliably predict clinical outcomes. The article was published in Nature Medicine online on January 7, 2026 with DOI: 10.1038/s41591-025-04094-8.