Researchers at Washington University in St. Louis have introduced Clinical Uncertainty Risk Alignment, or CURA, a novel framework designed to calibrate confidence levels in clinical artificial intelligence models. Developed by first-year doctoral student Sizhe Wang under the supervision of Chenyang Lu, Fullgraf Professor and director of the AI for Health Institute, CURA addresses a critical limitation in healthcare AI: systemic overconfidence that undermines safe human-machine collaboration. The research will be presented at the Association for Computational Linguistics annual meeting, scheduled for July 2 to 7 in San Diego. Current clinical language models frequently generate highly confident predictions even when erroneous, creating dangerous blind spots when integrated into physician workflows. Studies indicate that uncalibrated AI outputs can degrade clinical outcomes when clinicians either incorrectly defer to flawed algorithmic suggestions or reject accurate AI recommendations. CURA resolves this misalignment by training models to quantify their own predictive uncertainty with mathematical precision. Rather than defaulting to near-zero uncertainty, the framework dynamically adjusts confidence scores to reflect the actual probability of error. To validate the approach, the research team fine-tuned three existing clinical language models using the MIMIC IV critical care dataset, which contains extensive electronic health records and corresponding clinical labels. The models underwent individual uncertainty calibration, a process that aligns prediction confidence with empirical error rates. During evaluation across five distinct clinical risk-prediction tasks, CURA demonstrated consistent calibration improvements without compromising the models ability to differentiate between high-risk and low-risk patients. The calibrated uncertainty output creates a practical triage mechanism for clinical settings. Predictions exhibiting low uncertainty are safely routed for automated processing, while high-uncertainty outputs are flagged for mandatory physician review. This structure effectively eliminates the near-zero uncertainty baseline that characterizes standard fine-tuned clinical models, particularly for complex, high-risk cases that historically required intensive manual interpretation. By providing reliability metrics that mirror actual performance, CURA enables clinicians to allocate attention efficiently and maintain oversight over algorithmic decision-making. The development underscores a broader shift in medical AI research toward trustworthiness and calibrated transparency rather than raw predictive accuracy alone. Lu emphasized that aligning algorithmic confidence with clinical judgment is essential for scalable, safe AI integration in healthcare environments. Wang confirmed that the framework successfully reduced overconfidence across all tested tasks and model architectures, establishing a reliable foundation for clinical deployment. The research team plans to expand CURA to diverse patient demographics and evaluate its operational impact in live healthcare settings. Future iterations will focus on validating whether calibrated uncertainty metrics directly improve clinical decision-making workflows and reduce adverse events in real-time care delivery. The complete paper is publicly accessible on the arXiv preprint server under the identifier arxiv.2604.14651.