Home – Home – Jornals – Philosophy of Sciences 2025 number 1

The article aims to develop an approach to the theoretical justification of the use of machine learning (ML) methods in natural sciences. The main obstacle on this path is the problem of the “black box”, or “epistemic opacity”, which is the lack of access to all elements of the cognitive process carried out through ML. In developing the approach, the authors formulate a criterion that must be met to solve the problem. The authors point out that the reason for turning to machine learning in natural sciences is the limited applicability of traditional analytical and qualitative methods for studying nature, since human thinking has reached its limits in their use - because of the complexity and multidimensionality of the studied systems. Therefore, the solution to the “black box” problem must explain how ML can overcome these limits, i.e. how human thinking can access a domain of knowledge that is inaccessible to it due to its own internal limitations. In this regard, it is argued that the approach that spontaneously formed within computer science cannot serve as a basis for solving this task. Such an approach involves incorporating existing scientific knowledge into ML tools in order to fight against bias, which is characteristic of machine learning, i.e. the researcher’s subjective assumptions that are necessary for successful generalization beyond the training set. The authors show that incorporating scientific knowledge into ML tools has only applied value and does not solve the problem of theoretical justification, since it does not meet the criterion they propose - it does not overcome the limits of human thinking, but merely aligns ML results with existing scientific knowledge.,