Uncertainties about link uncertainty: ML models as phenomenological models

There is much debate regarding the epistemic potentials and limitations of machine learning (ML) models in science, and how best to use them to gain new scientific explanations and understanding. Emily Sullivan has drawn an analogy between ML models and scientific toy models, arguing that until the ‘link uncertainty’ between the model and target system has been reduced, they provide how-possibly explanations of their target phenomena. She takes this link uncertainty to be a significant hindrance to obtaining scientific understanding from ML models, a view which is commonly echoed in the literature. Yet, the exact nature of this uncertainty remains largely unexplored. In an attempt to clarify the uncertainties accompanying ML models, I reconsider the extent to which these models provide how-possibly explanations, and Sullivan’s analogy between toy models and ML models. My conclusion is that Sullivan generally overstates ML models’ role in providing explanations, thereby raising our epistemic expectations of them beyond what is warranted. Further analysis of the representational and explanatory power of ML models also shows that what really hinders our understanding of the target systems is an uncertainty regarding the causal mechanisms mediating the informational dependencies discovered by the ML model, which I call ‘mechanism uncertainty’. From this, I argue that a better framework for understanding the epistemic role of ML models in science is to see them as phenomenological models. These are empirically grounded models accompanied by a mechanism uncertainty, rather than link uncertainty, which hinders a deeper understanding of the target phenomena.