The impact of generative artificial intelligence on clinical decision-making and patient safety: A narrative review
Abstract
Generative artificial intelligence (AI), in the form of large language models (LLMs), has entered clinical practice faster than the evidence on how it changes clinical decisions could be produced. The relevant question is therefore no longer whether these models hold medical knowledge, but whether their use makes clinicians' decisions better and patients' care safer. This narrative review examines the evidence on the impact of generative AI on clinical decision-making and patient safety. In evaluations of encoded knowledge, a medically aligned model gave answers that agreed with scientific consensus in 92.6% of cases, against 92.9% for clinicians. In diagnostic reasoning, a model optimized for diagnosis outperformed unassisted clinicians on challenging cases and improved the differential diagnoses of the clinicians who used it, whereas a randomized trial of a general-purpose chatbot found no significant improvement in physicians' diagnostic reasoning, although the model alone outperformed both physician groups. In management reasoning, physicians assisted by GPT-4 scored 6.5 percentage points higher than those using conventional resources. When tested on real patient data in a simulated clinical workflow, however, LLMs diagnosed less accurately than physicians, did not follow guidelines and were sensitive to the order and quantity of information. The safety literature identifies four recurrent hazards: fabricated content, with hallucination rates of 50% to 82% when prompts contained a false clinical detail; misjudgement of clinical acuity, with drafted patient messages judged to pose a risk of severe harm in 7.1% of assessments; automation bias, by which clinicians adopt the model's assessment in place of their own; and the reproduction of racial and gender stereotypes. Only about 5% of published evaluations used real patient care data. Taken together, the evidence supports generative AI as a supervised aid to clinical reasoning rather than as an autonomous decision-maker, and indicates that its safety depends as much on the design of the human–AI interaction as on the accuracy of the model. Practical safeguards for clinical use are also reviewed.