In the goal of capturing patterns of human reasoning, the artificial intelligence (AI) community coined the idea of non-monotonic reasoning (NMR): reasoning that may retract conclusions given new information. This was proposed as more adequately modelling human-style reasoning, including legal reasoning. In this thesis, we propose non-monotonic reasoning analyses of subfields of AI not typically seen as such. The NMR view yields both new criteria of agreement to human-style reasoning for evaluation of models, and new models developed with such criteria in mind. In particular, we cover: i) classification models, focusing on case-based reasoning (CBR) systems, and; ii) interactive explanation models in explainable AI (XAI).

In CBR, we analyse abstract argumentation models for case-based reasoning (AA-CBR) as a reasoning system and prove that it does not satisfy cautious monotonicity, a property proposed in the NMR literature. We present a new alternative to AA-CBR that is provably cautiously monotonic, and satisfies as well other NMR properties, such as cut and cumulativity. This alternative also results in a principled treatment of noise in ``incoherent'' casebases. We present ways by which AA-CBR models can be mined from data, and compare with this new cautious version of AA-CBR, using the COMPAS dataset of criminal recidivism. In XAI, we analyse explanations as objects subject to reasoning and present a formal model of an interactive scenario for explanation between user and system. We analyse explanations as committing to some model behaviour, suggesting a form of entailment, which, we argue, can be non-monotonic. We illustrate the approach with arbitrated dispute trees for AA-CBR. Thus, we argue that NMR brings considerations to a wider scope of problems in AI, including AI and law, at a time in which recent systems show impressive capacities but no guarantees on behaviour.