Touching the core or scratching the surface? Background theoretical aspects of similarity and legal analogy in Artificial Intelligence

Abstract

This paper addresses theoretical discussions surrounding similarity and legal analogy, both in general and as particularly applied to artificial intelligence. It begins by outlining the main reasons why it is said that current AI systems are far from deploying the capability of forming humanlike abstractions or analogies. Problems of analogical reasoning then are divided into general problems of competing background theories specific problems of applicability of those theories in the context of AI (or the feasibility thereof). It recovers some fundamental definitions for the debate of analogy and proceeds to address the inferential model by singling out the specific problems of normative analogies. The paper intends to lay down the fundamental normative and psychological discussions as a general problem, in order to apply them to the specific context of AI. Tversky’s directionality and diagnosticity principles are framed as some good insights for the current debates about “sticky” stereotypes and discriminatory treatments attributed to AI systems. Lastly, the paper focuses on the concepts of “relevant similarity” and “sufficient similarity” in AI. It stresses that perceptual processes still prove important to current research on AI and it suggests that programming Rosch’s two psychological principles of categorization into AI Systems would helpfully provide for flexible criteria to finetune the level of abstraction of categories formed in analogy-making.