AI and Law Milestones: From Top-Down to Bottom-Up Computational Models of Legal Reasoning (and Back)

Authors

• Kevin Ashley, University of Pittsburgh

Document Type

Article

Abstract

This Article introduces five research and development milestones in the field of Artificial Intelligence and Law and then discusses how large language models (“LLMs”) and generative AI (“GenAI”) are affecting each of them. The milestones include legal expert systems, lessons learned in computationally modeling legal rules and statutes, knowledge-based models of legal argument, machine learning models that classify case texts and predict outcomes, and the design and evaluation of legal applications of LLMs and GenAI. These milestones illustrate how top-down knowledge-based computational models of legal expertise have rapidly given way to bottom-up models based on machine learning. Knowledge-based models explicitly represent elements of legal knowledge such as logical representations of legal rules or component features of legal analogies. Researchers built formal systems to apply these abstract representations of legal knowledge (the top) to analyze concrete facts (the bottom). By contrast, bottom-up machine learning models start from raw data of solved problems such as the texts of legal decisions and induce features that can predict their outcomes. LLMs are an extreme example of processing, classifying cases, or predicting their outcomes from statistical representations of their texts using vectors and matrices. This Article provides examples of how recent research developments are affecting each of the milestones: LLMs are beginning to extract expert systems’ rules from regulatory texts, induce legal rules automatically from cases, detect syntactic (i.e., logical) ambiguities in statutory provisions and address semantic ambiguities in their terms, apply statutes in analyzing problems, and generate case-based legal arguments. The Article explains how researchers are creating new ways to objectively evaluate the LLMs’ legal outputs. Finally, the Article illustrates new roles for knowledge-based models in hybrid systems that integrate top-down and bottom-up approaches to improve LLMs’ performance.

Recommended Citation

Kevin Ashley, AI and Law Milestones: From Top-Down to Bottom-Up Computational Models of Legal Reasoning (and Back), 128 W. Va. L. Rev. 945 ().
Available at: https://researchrepository.wvu.edu/wvlr/vol128/iss3/5