A black box approach to auditing algorithms

Abstract

AI is believed to be the most disruptive force in technology in the coming decade. The best intentions, however, can yield negative consequences, that is, the serious problems of introducing AI, especially algorithmic decision-making and its overtrust, into business and society-the resulting discriminatory and biased decisions. Quite a few studies on algorithmic decision-making have made strong claims about the causality between algorithms and biased decisions. Multiple protective measures, including the Explainable Artificial Intelligence, have also been enacted against the discriminatory and biased algorithmic decision-making practices. Nevertheless, they are persistent because of algorithmic obscurity, biased training data, the false belief that algorithms are neutral, and the public's perception that explainable and data-driven decisions are often not objective. This paper proposes a black-box approach to auditing algorithms. The approach draws on the counterfactual theories of causation. It aims at identifying obvious and obscure decision factors engendering decisions from multiple counterfactuals for a given factual.