Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Not ready for the bench: LLM legal interpretation is unstable and out of step with human judgments

Abhishek Purushothama, Junghyun Min, Brandon Waldon, Nathan Schneider

TL;DR

Not ready for the bench argues that LLMs produce unstable and human-inconsistent interpretations of ordinary meaning in legal text. The study conducts controlled experiments with 138 insurance-contract scenarios, nine prompt variants, and 15 LLMs (including GPT-4) to assess robustness and correlation with human judgments. Findings show pervasive prompt sensitivity, widespread variability across models, and only modest, unreliable alignment with human judgments, even for the largest models. The work cautions against deploying generative AI for legal interpretation, notes data contamination risks, and underscores the need for rigorous validation before practical adoption.

Abstract

Legal interpretation frequently involves assessing how a legal text, as understood by an 'ordinary' speaker of the language, applies to the set of facts characterizing a legal dispute in the U.S. judicial system. Recent scholarship has proposed that legal practitioners add large language models (LLMs) to their interpretive toolkit. This work offers an empirical argument against LLM interpretation as recently practiced by legal scholars and federal judges. Our investigation in English shows that models do not provide stable interpretive judgments: varying the question format can lead the model to wildly different conclusions. Moreover, the models show weak to moderate correlation with human judgment, with large variance across model and question variant, suggesting that it is dangerous to give much credence to the conclusions produced by generative AI.

Not ready for the bench: LLM legal interpretation is unstable and out of step with human judgments

TL;DR

Not ready for the bench argues that LLMs produce unstable and human-inconsistent interpretations of ordinary meaning in legal text. The study conducts controlled experiments with 138 insurance-contract scenarios, nine prompt variants, and 15 LLMs (including GPT-4) to assess robustness and correlation with human judgments. Findings show pervasive prompt sensitivity, widespread variability across models, and only modest, unreliable alignment with human judgments, even for the largest models. The work cautions against deploying generative AI for legal interpretation, notes data contamination risks, and underscores the need for rigorous validation before practical adoption.

Abstract

Legal interpretation frequently involves assessing how a legal text, as understood by an 'ordinary' speaker of the language, applies to the set of facts characterizing a legal dispute in the U.S. judicial system. Recent scholarship has proposed that legal practitioners add large language models (LLMs) to their interpretive toolkit. This work offers an empirical argument against LLM interpretation as recently practiced by legal scholars and federal judges. Our investigation in English shows that models do not provide stable interpretive judgments: varying the question format can lead the model to wildly different conclusions. Moreover, the models show weak to moderate correlation with human judgment, with large variance across model and question variant, suggesting that it is dangerous to give much credence to the conclusions produced by generative AI.

Paper Structure

This paper contains 39 sections, 5 figures, 12 tables.

Table of Contents

  1. Introduction
  2. Background
  3. Legal Interpretation with LLMs
  4. Vague Contracts.
  5. Interpretation as binary QA.
  6. Formulation of prompt variants.
  7. Models.
  8. Results and Analysis
  9. Analysis of judgments for Yes/No prompts
  10. Some models can be stopped clocks.
  11. Robustness to question variation
  12. Ubiquitous inconsistency across question variants.
  13. Some questions are more likely to elicit minority judgments.
  14. Correlation to human judgment
  15. Only some LLMs are moderately correlated to human judgment some of the time.
  16. ...and 24 more sections

Figures (5)

  • Figure 1: A legal interpretation scenario represented as QA task with binary questions. The example is based on the case https://media.ca11.uscourts.gov/opinions/pub/files/202212581.pdf and constructed in the style of our task.
  • Figure 2: Llama-70B model responses across question variants, each of which results in a large shift away from values in either directions given the Yes/No variant, indicated with the dotted lines.
  • Figure 3: Llama-70B model probabilities versus human consensus across question variants. Dotted lines and the corresponding $R^2$ are best best-fit lines between human and instruction-tuned LLM. The Yes/No question variant, highlighted in red, represents the highest $R^2$ value in our study.
  • Figure 4: An example vignette from the questionnaire provided to the participants by waldon_vague_contracts_2023. The vignette corresponds to one of the 138 items. Since our study focuses on interpretative judgment, question 1 is of interest to us, and responses make up the human judgments used in correlation analysis in \ref{['sec:human-correlation']}.
  • Figure 5: GPT-4 judgment probabilities versus human consensus across question variants. Dotted lines are best best-fit lines between human and instruction-tuned LLM.