Abstract

One of the main challenges with the responsible use of machine learning algorithms is their opacity: we typically do not know why the algorithm produces a particular output for a given input. As a result, a wide range of research is being done into methods that explain these outputs. An upcoming trend in these methods is to try to go beyond a calculation of which inputs contributed (and to what degree) to the output, and to instead explain the output using concepts. For example, one may aim for an explanation that says the model labelled an image as a kitchen because it contains a sink, even though the input space consists of individual pixels. However, methods that extract concepts from AI systems yield compositions of pixels that are wildly different from how humans would conceptualize images. On the other hand, methods that let humans attribute concepts to AI systems introduce inaccuracies in the explanation. This chapter investigates this disconnect that occurs in concept-based XAI to determine the prospect of such explanations. I link the literature on social externalist concept attribution to AI and the current XAI tools to formulate a cautiously optimistic answer in our use of concepts to explain the functioning of AI systems over limited domains. However, the conceptual disconnect does dim the promise of explanations from current ML methods that are as good as the explanations humans can offer for their own reasoning.