AI reasoning models boost diagnostic accuracy but raise safety concerns

Large reasoning models, a newer class of artificial intelligence systems that work through problems step by step before answering, may make more accurate primary diagnoses than standard large language models, but they come with a hidden cost: a marked increase in unsafe outputs. That is the central finding of a study published in Nature Communications, which examined how these models behave when asked to diagnose real patients from electronic health records.

The researchers built AIMS-LDR, a long-context diagnostic reasoning benchmark constructed from real-world electronic health records. Using this benchmark, they compared six matched pairs of models, in which each reasoning model was evaluated against its standard counterpart, across 12 medical specialties. The scale of the comparison allowed the team to isolate what the added reasoning capability actually changes in clinical diagnostic tasks.

The results revealed a consistent trade-off. The reasoning models generated more than 34 percent more false positive diagnoses than their standard counterparts, meaning they flagged diseases that patients did not actually have. They also showed approximately 40 percent lower reproducibility, producing less consistent outputs when repeatedly assessed. At the same time, the reasoning models made more accurate primary diagnoses and produced differentials, the ranked lists of alternative conditions clinicians consider, that were more clinically relevant.

Importantly, the team reported that this trade-off was not an artifact of one particular setup. Experiments confirmed the pattern across different prompt designs, model hyperparameters, and evaluation paradigms, suggesting the effect reflects a genuine property of how reasoning models approach diagnostic tasks rather than a quirk of a single configuration.

To address the safety problem, the researchers tested a structured collaboration model in which the reasoning model broadens diagnostic coverage while a clinician verifies the reliability of its outputs. In a validation study with 15 physicians, this collaborative approach increased diagnostic accuracy by 20 percent and reduced false positives by 44 percent compared with the models operating alone.

The authors emphasize that these results were obtained under controlled retrospective conditions, using recorded patient data rather than live clinical encounters. They note that pairing reasoning model outputs with physician oversight may support safer and more accurate diagnostic decision-making, but that prospective validation in real clinical settings is needed before any deployment.

Subject of Research: Accuracy and safety trade-offs of large reasoning models in clinical diagnostic reasoning

Article Title: The accuracy-safety trade-off of large reasoning models in clinical diagnostics

Article References: Wu, S., Chen, L., Fu, Y., Zhang, X., Luo, L., Lou, M., Huang, Y., Chen, G., Fu, Y., Wang, Y. X., Di, Q., Zhou, Y., Li, X., Wong, T. Y., Yu, Y., & Zhou, H.-Y. (2026). The accuracy-safety trade-off of large reasoning models in clinical diagnostics. Nature Communications. https://doi.org/10.1038/s41467-026-78520-x

Image Credits: AI Generated

DOI: 10.1038/s41467-026-78520-x

Keywords: large reasoning models, large language models, clinical diagnostics, artificial intelligence, electronic health records, false positives, diagnostic accuracy, physician oversight, AIMS-LDR benchmark, patient safety, accuracy-safety, trade-off