杜克大学 ：2026年医疗系统中的AI安全：基础设施建设与风险管理实践强化白皮书（英文版）

While AI tools can provide value to clinicians and patients, the complexity of AI technology introduces additional risks compared to traditional clinical tools, including the risk of performance drift and bias.In addition, clinical tools that rely on large language models (LLMs) also introduce the risk of hallucinations and overreliance by the user, often due to the authority with which outputs are presented regardless of the uncertainty in the calculation. The rapid acceleration of AI capabi