Recent advances in artificial intelligence technologies, including machine learning (ML), have brought the issue of epistemic opacity to the forefront of the philosophy of AI. In scientific contexts, the problem of opacity is significant because it ra...
Recent advances in artificial intelligence technologies, including machine learning (ML), have brought the issue of epistemic opacity to the forefront of the philosophy of AI. In scientific contexts, the problem of opacity is significant because it raises the question of how we can trust the results produced by methods that are not epistemically transparent. One response to this issue is computational reliabilism, a theory advanced by Juan M. Durán and Nico Formanek. Building on Paul Humphreys' definition of opacity and Alvin Goldman's process reliabilism, they set out the conditions under which machine learning models can be reliable. According to Durán and Formanek, epistemically opaque processes can be reliable if they exhibit sufficient predictive accuracy and fulfill what they describe as sources of reliability, including alignment with expert knowledge, empirical validation and verification, and robustness.
This paper critically examines computational reliabilism as a framework for conferring epistemic reliability on opaque machine learning models and proposes an alternative approach. I argue that the approach of Durán and Formanek faces practical challenges and underestimates the role of explanation in epistemic justification. As an alternative, I contend that machine learning models can achieve epistemic reliability if they are accompanied by surrogate models that provide how-possibly explanations for the original model’s outputs. Such surrogate models should not only consistently reproduce the predictions of the original machine learning model with a high degree of accuracy, but also be transparent and epistemically reliable in their own right. Even though the internal mechanisms of the original machine learning model are opaque, the reliability of the surrogate model can still serve as a basis for conferring epistemic reliability on the original model.
On this basis, I reexamine central case studies in the philosophy of AI to demonstrate the explanatory value of my framework. In particular, I demonstrate how my surrogate model-based conditions for epistemic reliability can clarify the epistemic differences between the melanoma detection model and the sexual orientation prediction model discussed by Emily Sullivan, as well as the case of nuclear magnetic resonance spectroscopy analyzed by Sandra Mitchell.