Why Automated Science Should Be Cautiously Welcomed

This paper assesses the positive and negative aspects of the use of automated methods in science. Four worries about the use of these methods are identified and discussed. A central example is the use of machine learning methods, especially with respect to image recognition. The concept of representational opacity is introduced as a supplement to the author’s previous account of epistemic opacity. Reasons are given for why some internal representations in machine learning may be permanently inaccessible to humans. Some parallels between the problem of scientific realism and the problems of internal representations in deep neural nets are given. The possibilities of using a reliabilist epistemology to address the worries are discussed and an epistemic version of the Precautionary Principle is given.