Generalizing case facts from verdicts using LLMs: New horizons for sentencing research?

• In Event: Exploring horizons of sentencing research with new data and methods

Abstract

Criminal behaviour has been long described using legal classification. This has, however, resulted in a loss of information regarding within-offense variation. Using other data was long impossible due to full-text decisions - containing case facts describing the behaviour - being unavailable and the difficulty of creating typical crimes and subsuming individual verdicts within them. Large language models (LLMs) are designed to understand, summarize and subsume text. We employ LLMs to create typified crimes and to subsume individual verdicts within them. Using 785 facts and descriptions from criminal court opinions regarding thefts, we show that human-guided LLMs can produce typical crimes similar to those produced by a thematic analysis employed by humans. Similarly, LLMs can subsume verdicts within the categories. We suggest that typical crimes might be used to capture further details of offences that went unnoticed. We discuss the conclusions for sentencing scholarship and the new avenues it opens.

Authors

• Jakub Drápal, Charles University; Institute of State and Law of Czech Academy of Sciences

• Hannes Westermann, University of Montreal - Cyberjustice Laboratory

• Jaromír Šavelka, Carnegie Mellon University