Reliability of LLMs as medical assistants for the general public: a randomized preregistered study

Nature's publication of a randomized controlled study revealing the inferior performance of humans assisted by a large language model (LLM) compared to the LLM alone in assessing medical scenarios has significant implications for the reliability of LLMs as medical assistants. The study, which involved 1,298 participants from a general sample, assessed the performance of humans and LLMs in evaluating ten medical scenarios leading to disease identification. This means that website owners, particularly those in the healthcare industry, need to reassess their reliance on human oversight for medical content generated by LLMs. As LLMs become more prevalent in providing medical information, website owners must consider the potential consequences of inferior human-LLM collaboration on their platforms. This could lead to a reevaluation of content policies and the role of human editors in reviewing LLM-generated medical content. To address these concerns, website owners can take the following actionable steps: (1) review their llms.txt files to ensure that LLMs are properly configured to provide accurate medical information, (2) implement robust testing and evaluation protocols for LLM-generated medical content, and (3) consider investing in AI bot tracking tools to monitor LLM activity on their platforms and identify potential areas for improvement.