Constantino Contreras

There is no doubt that epistemically opaque processes, those whose internal mechanisms cannot be surveyed by agents, are indispensable in many epistemically notable and high-stakes contexts. But when, if ever, can we secure epistemic justification in their uses? This article provides a framework for approaching this problem, particularly in the context of machine learning, by first offering a four-part taxonomy of epistemic opacity made to unify the landscape of distinctions in the current literature. Using this framework, I develop and criticize Jack Casey’s thesis that an agent's tolerance for opacity is determined by their position in the scientific realism/antirealism debate and show that this correspondence ultimately breaks down. Instead, I argue that this tolerance is determined by an agent’s affinity for an externalist or internalist theory of epistemic justification. Internalists, linked to some kind of accessibility, should be driven toward transparency programs, while externalists, prioritizing truth-conduciveness, can in principle accept opacity outright. Yet, drawing on the current literature on mechanistic interpretability, XAI and computational reliabilism, I then submit that both positions are problematic, and that the most tenable position would be given by a hybrid theory of justification outside of the classical internalist/externalist epistemological paradigm.

In this article, I propose a minimalist characterization of naturalism as the thesis that: 1) we must have a minimal, but not total, commitment to the ontology of our sciences, and 2) we must have a meta-methodological commitment to the use of methods that allow for justifications and claims that can be intersubjectively evaluated, such as formal and informal arguments, or the use of the methods of science themselves. In this way, I will first critique the classical definitions of naturalism understood as a form of physicalism and as a restriction on the postulation of supernatural entities and explanations in science. Then I will defend the priority of this meta-methodology not only in science and philosophy, but in all intellectual endeavors of research. I conclude that if philosophy intends to be an intellectual enterprise of research, then it must be oriented according to this meta-methodology; that is, it must be naturalistic.

La teoría homotópica de tipos (HoTT) es un sistema formal que ha ganado notoriedad por su propuesta como un nuevo fundamento de las matemáticas. Sin embargo, el uso de conceptos avanzados de topología y teoría de homotopía como conceptos básicos dificulta su adopción. En este artículo introduzco al lector a esta discusión, haciendo primero un repaso formal e histórico de HoTT, su teoría de tipos y sus nociones topológicas. Finalmente, discuto las dos propuestas más relevantes en la literatura como justificación informal e intuitiva de HoTT, junto a las críticas y problemas que deben enfrentar ambas propuestas para lograr su objetivo.