Home – Home – Jornals – Philosophy of Sciences 2026 number 2

The article discusses which approach can be used to solve the problem of theoretical justification of machine learning methods in the natural sciences. The authors point out that the current strategy of philosophical, epistemological and applied research related to ML (countering bias and minimizing subjective assumptions) fail to solve the “black box” problem, which makes it difficult to interpret the results and reduces their scientific value. The article suggests an alternative approach to the theoretical justification of ML based on the extension of computational powers. Using Paul Humphreys’ concept of “extending ourselves,” the authors show how computing technologies can overcome the limitations of human thinking and model complex phenomena that are inaccessible to “traditional” mathematical methods. The idea of a “bias compensation” mechanism is put forward, which can neutralize the influence of subjective factors in the framework of natural science research using ML. Special attention is paid to comparing ML with computer simulations, where the influence of assumptions/bias can be compensated by analyzing the global dynamics of the model, whereas in ML this problem remains unresolved. This entails the need to separate the “black box” problem in ML from the “epistemic opacity” that is common to both machine learning and computer simulations. It is specifically emphasized that the “black box” in machine learning arises not from the “opacity” of the model or the complexity of computational operations, but from the lack of clarity of the model’s connections with real physical processes (the “target system”). Thus, the authors demonstrate that the application of ML in natural science requires a rethinking of existing methodological prerequisites. The development of mechanisms of bias compensation in the field of ML is becoming a key task in order to overcome the “black box” problem and successfully integrate ML into scientific research.