Study finds AI chatbots prioritize differently than doctors in kidney transplant decisions
Researchers at Penn State University discovered that large language models make faster but less nuanced choices than human physicians when allocating a single kidney.
Penn State researchers fed large language models scenarios involving two eligible kidney recipients, varying factors such as age, health status and drinking habits. Human doctors typically prioritized younger patients, but the AI frequently chose the candidate who drank less, focusing on one trait at the expense of a broader assessment. When presented with ambiguous cases, people acknowledged the lack of a clear answer, while the models selected an option decisively.
Lead author Hadi Hosseini noted the models' overconfidence and narrow weighting of factors. Collaborator John Dickerson highlighted that humans incorporate ambiguity into allocation debates, a nuance AI currently lacks. The findings raise concerns about deploying AI for high-stakes medical resource decisions without ensuring alignment with human values.
