Chapter 6: Non-applicability Domain. The Benefits of Defining “I Don't Know” in Artificial Intelligence

At the moment a distinction between a human expert and an in silico is that a human is allowed and able to produce the answer “I don't know”. We think that during the design of an artificial intelligence (AI) system one ought to consider adding “I don't know” as an AI answer. We argue that an introduction of such an answer may make AI systems on average more accurate and therefore more trustworthy. We submit the case with predictive models. A procedure to define “I don't know” in binary classifiers, called all leave-one-out models (ALOOM), is presented and examined. We use a simulation to show how this may benefit a drug discovery process. We refer to the research field which defines “I don't know” as the non-applicability domain.