Abstract
AIs are increasingly involved in the information processing processes from which we elaborate our beliefs. Under these conditions, it is legitimate to ask under what circumstances it is legitimate to consider that such processes are reliable and therefore that we are justified in maintaining our beliefs - it is typically by clarifying this question that we can hope to describe what constitutes responsible use or provision of an AI.
This question has already been studied extensively in the context of the reliability movement in epistemology. One of the main problems encountered in this context is the so-called generality problem: to study the reliability of a particular process that produces beliefs, we need to assign it a given type. For example, a particular process may be described as involving the vision of an animal, the vision of a man, the vision of a 60-year-old man, the vision of a 60-year-old man in the evening, etc., and so the question arises as to which type should be chosen to characterize the reliability of the process. This presentation will be devoted to discussing this question in the case where AIs and, more generally, epistemic machines such as numerical simulations, are used in a process.
