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

We are pleased to see engagement [1] with our recent paper in the International Journal of Epidemiology [2]. Artificial intelligence (AI) is perhaps the most exciting technological development of our lifetimes, with implications for epidemiology and the populations we study. On the need to protect our discipline against deskilling, we agree. The optimal time spent invested in lower- versus higher-level tasks will require calibration. In the case of preparing literature reviews, learning how to carefully read and critically appraise papers is vital, but are there diminishing returns to manually screening, say, 10 000 papers? Where is the line drawn? Similarly, for writing code: is every manual keypress necessary, or can higher levels of abstraction be used instead? Frontier AI systems are general-purpose technologies and can thus aid skill acquisition and maintenance. Few education interventions appear effective [3], yet tutoring has a strong evidence base [4]. AI has enabled cheap or free mentoring on a global scale and initial empirical evaluations are encouraging [5]. Epidemiologists have a choice about how they engage with such opportunities.