Abstract
Defeasible reasoning is a kind of reasoning which allows to retract deductively valid conclusions when encountering exceptions. Within Artificial Intelligence research addressing this phenomenon, the focus has largely been on developing methods to technically model this form of reasoning. Defeasible reasoning has always been studied also because it is a kind of reasoning that is used in everyday situations by humans. However, relatively little research relating the technical results to human reasoning has been developed. By taking into account results from fields such as knowledge representation, philosophy and cognitive sciences, we address this gap. We adopt a human-centred perspective on the topic, taking as the pivotal point of our investigation the notion of an exception. We argue that this key phenomenon underlying defeasible reasoning has often been overlooked in the research on the topic. In particular, we develop a philosophical systematisation for defeasibility allowing to compare different formalisms according to their ontological commitments. This also clarifies options to non-monotonically generalise an ontology. A core contribution is the development of a non-monotonic description logic inspired by perceptron (or ‘tooth’) logic, combining a prototype modelling of concepts with a logic of justifiable exceptions, designed to meet some of our developed desiderata for reasoning in the presence of exceptions. To analyse and validate our system’s reasoning behaviour, we designed an experiment with humans and LLMs in order to evaluate the new logic.