Neural networks present an inherent component of unknowability that makes them critical from an epistemological perspective especially when they are used in connection with scientific research. Indeed, the language of a scientific field is often, if not almost always, different from the language of neural networks, which are currently viewed as “black boxes”, providing useful output but without an explanation that traces such output back to the input. The endeavour of providing an explanation of why the network has provided a certain result is called “Explainable AI (XAI)”. The general question that this work aims to tackle is whether the explanations Explainable AI is aiming for are scientific. The investigation is based on the construction of a conceptual framework that is as general as possible regarding how explanations are created in AI. It aims to verify whether scientific explanations are created at all with Explainable AI and, if so, whether such creation can be described in traditional terms from Philosophy of Science or whether new concepts are called for.

AI, particularly its subfield Machine Learning, learn and apprehend our world(s) through numerical representations of the latter. When applied in our societies, these algorithms mirror their biases and discriminations. Various strategies such as machine ethics or fairness of algorithms are employed to correct, or in computer science jargon, mitigate these "errors". In our contemporary societies where democracy is supposed to set all of us equal and free, and where justice is supposedly blind to our differences, these algorithmic biases embarrass. Yet, they seem to finally shed light on prejudices and unfairness which have long been denounced by oppressed groups. Building on feminist, post-colonialist, queer and disability studies, and the new challenges brought by intelligent machines, how can we deconstruct human morality, if the concept even exists?