Large Legal Fictions: Detecting Legal Hallucinations in Large Language Models
Matthew Dahl
Abstract: Do large language models (LLMs) know the law? LLMs are increasingly being used to augment legal practice, but their revolutionary potential is threatened by the presence of legal "hallucinations" -- textual output that is not consistent with the content of the law. In this talk, I theorize the provenance and nature of these hallucinations and discuss methods for detecting them in LLM outputs. I then share results from three experiments auditing off-the-shelf LLMs and industry retrieval-augmented generation (RAG) models, showing that legal errors remain widespread. I conclude by emphasizing the need for empirical evidence in an age of ever-increasing hype about AI's ability to replace lawyers and expand access to justice.
Mathematics
Audience: general audience
Comments: Matthew Dahl is a JD/PhD student at Yale Law School and Yale Department of Political Science. His research on AI, judicial behavior, and legal citation analysis has been published in the Journal of Empirical Legal Studies and the Journal of Legal Analysis. Before coming to Yale, he was a Fair Housing Fellow at Brancart & Brancart and received his BA from Pomona College.
NYU CDS Math and Democracy Seminar
Series comments: The Math and Democracy Seminar features research on contact points between the mathematical sciences and the structure of democratic society. The purpose of the seminar is to stimulate mathematical activity on problems relating to democracy, and to foster interdisciplinary collaboration between mathematicians and other scholars and democratic stakeholders.
Examples of topics of interest include detection of gerrymandering, fairness and accountability of algorithms used in social decision-making, voting and apportionment theory, applications of statistics to discrimination law and the census, and mathematical modeling of democratic processes. The scope is not limited to these and is expected to expand as further applications emerge.
Seminars currently conducted via Zoom (with some events also in person). Look for links in individual talk descriptions.
| Organizers: | Ben Blum-Smith*, Jonathan Niles-Weed |
| *contact for this listing |