Towards Trust Requirements as Basis to Model Explainable AI Computations

We see the key challenges of artificial intelligence (AI) as the need for computational resources to implement its algorithms, the technical ability to model phenomena accurately and, as a technology that invades almost every aspect of life, trustworthiness. We see trustworthiness as communicating the correct expectations and risks. Trust, veracity and computation are its requirements, and they define the type of messages conveyed for AI solutions. Our approach then is to develop and evolve concepts until they reflect the desired target meanings before implementation. Given the myriad concerns and permutations thereto, a model of computation seems best to address them. We outline the notions of trust, modelling and computation and relate them to AI challenges. We then develop the notions of “context” as evolving environments of states, patterns and meanings for AI computations; “trialisation” to define reversible actions and invariant attributes over AI requirements and explanations; and “categorisation” as a preliminary activity in formally specifying concepts. From these, one better refines how AI problems and solutions are specified and presented in differing contexts. This paper is on how to reckon with the induced requirements for trustworthiness.