Trust and Transparency in Artificial Intelligence

Abstract Can we trust AI? It is a question that is asked with increasing frequency and urgency as artificial intelligence becomes ever more adept at performing tasks hitherto performed almost exclusively by human beings. There are certainly reasons to be cautious, among which is the problem of understanding the inner workings of AI. Adapting themselves through machine learning, modern AIs rapidly become so complex that nobody can follow the processes by which they reach conclusions; they can be neither reliably predicted in advance nor easily examined in retrospect. It is this concern that I shall address. Transparency is often considered an important requirement for trustworthy AI and so its absence is commonly recognised as a potential objection to its adoption. Sometimes referred to as the ‘black box’ problem, it points out that we can hardly ever completely understand how an AI works and suggests that such opacity precludes trust. This is a conclusion I believe we ought to resist. The nature of AI as a black box should not entirely prevent us from trusting it. My conclusion is modest: I do not attempt to put to rest all potential concerns with trusting AI, but only deal with the specific obstacle of its opacity.