Deep learning models (DLMs) now automate core stages of scientific inquiry, from protein-structure prediction to medical image reconstruction. Yet their opacity has led many philosophers of science to argue that they cannot genuinely contribute to scientific explanation unless they qualify as scientific models standing in fine-grained representational relations to their targets. I argue that this assumption is mistaken—and that it obscures how the deployment of machine learning techniques is actually transforming scientific practice.
The dominant view treats opacity as a threat because it allegedly blocks assessment of a model’s representational status. But this presupposes that explanatory contribution depends on representation. I challenge that presupposition. Contemporary DLMs neither function as mechanistic surrogates nor map onto the causal structure of their targets in the way traditional representational models do. Demanding such mapping misconstrues their epistemic role.
Instead, I argue that DLMs are better understood in comparison to scientific instruments in their automated epistemic role. Their contribution to explanation lies not in mirroring structure but in reliably generating results that stabilize patterns of dependence and support counterfactual reasoning. Drawing on the epistemology of experimentation, I propose robustness—established through benchmarking, error management, and cross-validation—as the appropriate normative standard.
Opacity, on this view, is not a failure of representation but a challenge of validation. Once reframed in terms of robustness rather than representational fidelity, the explanatory role of automated deep learning systems becomes intelligible without requiring explanation by representation.