Why can’t epidemiology be automated (yet)? Perhaps it shouldn’t be

In a recent issue of the International Journal of Epidemiology, Bann and colleagues describe several ways in which artificial intelligence (AI), and especially generative AI (such as large language models), may help optimize epidemiological research, ranging from hypothesis generation to the dissemination of results [1]. Although we agree that AI provides opportunities for reducing repetitive tasks and speeding up scientific work, we argue that it is undesirable to strive for the degree of epidemiological automation described by Bann et al. The authors advise keeping a human epidemiologist in the loop to avoid and correct harmful consequences in the case of mistakes and hallucinations. Notably, the epidemiologist should be sufficiently skilled to recognize and resolve issues across the research process. These skills are developed and maintained through the continued application of epidemiology in practice. By shifting the practical application away from humans to AI, junior epidemiologists may not be able to develop the necessary skills to become the epidemiologist in the loop (never-skilling) and senior epidemiologists may have their skills deteriorate too much to remain the epidemiologist in the loop (deskilling) [2, 3]. Thus, inversely, as the use of AI systems increases, the number of epidemiologists skilled enough to prevent harmful mistakes within the system decreases. Epidemiologists, trained to perform and appraise research critically, should strive to be meaningful in the loop, rather than solely serve as symbolic monitoring mechanisms [4]. If, instead, we come to rely on AI-only systems, we may also need to rethink who (or what) becomes responsible for harmful mistakes [5].