Responsible Assessment of Beliefs Based on Computational Results: Expanding on Computational Reliabilism

In order for advanced computational systems, such as AI systems, to be successfully integrated in liberal democracies, the people who design, use or are affected by these systems in many cases must be adequately disposed to hold the results of these systems to be true. How is such belief in these results justified, given the opaque nature of advanced computational systems and the possibility of error? The theory of “computational reliabilism” (CR) outlines how such belief can be justified and lead to a genuine advancement in human knowledge. The basic idea of CR is that the belief in the results of computational systems, despite their opacity, can be justified by their positive rate of producing true beliefs. In this paper, we show that CR needs to be expanded by focusing more on the human agents who are interacting with these systems epistemically and the consequences of the human-computer interaction. The reliability of a belief-forming process based on a computational system can only be assessed by taking into account both these agents and the ethical stakes involved. Moreover, if CR is intended to guide action, a responsible assessment of reliability must rely on an internal type of justification that is relative to the respective epistemic agent and typically necessitates an institutionalized division of epistemic labor.