Abstract

Science is rapidly moving towards a more automated future. Whilst this is not an entirely new phenomenon – automation in science has been around for at least 150 years – recent years have seen an increase in talk about, and implementation of, automated workflows. This drive for more automation is particularly prominent in laboratory-based disciplines such as material sciences, chemistry, and the life sciences. The core aims that drive this push for more automation in the laboratory are 1) to increase research efficiency and 2) to improve the replicability of experimental outputs. In this talk I will analyse how these two aims are affected by the adoption of powerful new AI tools in science, with a focus on the life sciences. I will argue that whilst new AI capabilities, in particular the ‘reasoning’ abilities of large language models (LLMs), have the potential to boost the efficiency of research, they could have a negative effect on the replicability of scientific results. More specifically, I will argue that AI-driven automation (ADA) can reduce scientists’ ability to troubleshoot the experimental process by diminishing their ability to identify, understand, and correct flawed research outputs. Two key implications of this analysis are 1) that the move from traditional automation to ADA needs to be managed in a context-sensitive manner which ensures that a laboratory’s ability to deal with experimental error remains intact, and 2) that we need to develop a better understanding of error (and error-reasoning) in science in order to tackle implication 1).