Dilectiss Liu

In our paper, we reassess the role of non-epistemic values in scientific practice by drawing lessons from machine learning and the automation of science. Due to several influential arguments (e.g., Rudner 1953, or Longino 1990), traditional philosophy of science has largely converged on the view that non-epistemic values are necessary for the justification of scientific claims. Recently, renewed support of this view has been made by appealing to the No Free Lunch theorems in the context of machine learning—that there is no universally optimal machine learning algorithm (Dotan 2020). The argument claims that the No Free Lunch theorems entail that epistemic values are insufficient for discriminating hypotheses. In the negative part of our paper, we critique Dotan’s argument. We argue that NFL theorems do not entail that ‘all hypotheses have the same average expected error’. We also discuss what NFL theorems do in fact say about inductive inference and theory choice. In the positive part, we argue that the possibility of ‘value-free science’ (understood as a science that involves only epistemic values in the context of justification) seems to be supported by developments in the automation of science. For example, Robot Scientist presented in King et al. (2004) can automatically propose hypotheses, design and conduct experiments, and interpret the results. We argue that in these systems, only epistemic values are involved in hypothesis construction and testing.