Explainable AI for higher cognitive functions: How to provide explanations in the face of increasing complexity

Since the introduction of the term explainable artificial intelligence, many contrasting definitions and methods have been proposed. A key issue is that most of the extant explanations are not sufficiently clear either to practitioners or to users. While some researchers use interpretation algorithms as post-hoc explanations (Samek et al., 2021; Ribeiro, 2016), others argue that we should use models which are interpretable in the first place (Rudin, 2019). Although the latter is important, developers are not always willing to sacrifice accuracy by choosing a less complex interpretable model. Here, we propose a working definition of what explaining an AI model means, focusing on robustness, representativeness, and comprehensibility as central properties, and on the importance of causal links (Miller, 2019). In addition, we suggest starting with simple models and gradually increasing complexity if necessary, whilst setting a case-specific threshold for its trade-off with accuracy and ensuring that we obtain good explanations of models of human cognition.