Atomic Fact-Checking Boosts Clinician Trust in AI Oncology Recommendations

A study released on arXiv reveals that atomic fact-checking markedly enhances the trust clinicians place in recommendations from large language models (LLMs) for oncology decision-making. The research involved 356 clinicians who provided 7,476 trust ratings. By breaking down AI treatment suggestions into separately verifiable claims associated with source guideline documents, atomic fact-checking led to a significant increase in trust (Cohen's d = 0.94), raising the percentage of trusting clinicians from 26.9% to 66.5%. In contrast, traditional transparency methods showed a gradual improvement over baseline (d = 0.25 to 0.50). These results indicate that disaggregating AI recommendations into verifiable claims linked to guidelines results in much greater clinician trust compared to conventional explainability methods in critical clinical scenarios.