A Fortiori Case-Based Reasoning: Formal Studies with Applications in Artificial Intelligence and Law

Abstract

Case-based reasoning (CR) involves comparing new problems with previous cases to support decision-making. In common law systems, this forms the core of the stare decisis principle: judges and lawyers rely on precedents to justify their reasoning and arguments. An influential model of CR is the result model (RM), which formulates a principle of a fortiori reasoning: if an earlier case was decided in favor of a party, then any new case that is at least equally favorable to that party must also be decided in its favor. This model and its variants are foundational to this dissertation. An analogy with artificial intelligence (AI) is central here: just as judges generalize from precedents, AI systems generalize from training data. This makes it possible to use the RM and its variants to analyze data-driven AI systems. Part I investigates how the RM can be extended into a general theory of a fortiori CR, incorporating hierarchies, dimensional values, and incomplete information. This results in variants such as the hierarchical RM and the dimensional-hierarchical RM. Part II builds on these extensions and focuses on applications of the theory in AI and law. The theory has been implemented in logical software and tested on several datasets. One case study concerns the COMPAS system, a controversial tool used to predict recidivism. The analysis demonstrates how the theory can contribute to transparency and the responsible use of AI.