Abstract
Despite marketing claims that Retrieval-Augmented Generation (RAG) has "eliminated" AI hallucinations in legal research, this study finds that leading tools from LexisNexis and Thomson Reuters still hallucinate between 17% and 33% of the time. While these specialized tools outperform general-purpose models like GPT-4, they frequently cite real cases to support false legal propositions, creating a "more dangerous" and insidious form of misinformation for legal professionals.
Introduction
Large Language Models (LLMs) are notorious for "hallucinating"—generating plausible-sounding but false information. In the legal field, where accuracy is paramount, this is a significant barrier to adoption. To combat this, legal tech giants have integrated Retrieval-Augmented Generation (RAG), a technique where the AI retrieves relevant documents from a trusted database before generating an answer.
Major providers like LexisNexis and Westlaw have marketed these RAG-based tools as "hallucination-free." However, because these systems are proprietary, there has been no independent, systematic way to verify these claims. This paper presents the first pre-registered empirical evaluation of these tools to determine if RAG actually solves the hallucination problem in a legal context.
Methodology
The researchers employed a rigorous, expert-led evaluation framework comparing three leading proprietary tools—Lexis+ AI, Westlaw AI-Assisted Research, and Ask Practical Law AI—against a baseline of GPT-4 (without RAG).
Evaluation Rubric
Responses were evaluated based on a two-dimensional rubric:
- Correctness: Is the factual legal statement accurate?
- Groundedness: Does the cited source actually support the claim made?
A response was labeled a "hallucination" if it was either incorrect (false statement) or misgrounded (citing a real source that does not support the claim).
Dataset
The evaluation used a new, pre-registered dataset of 202 legal queries designed to probe specific vulnerabilities, including general legal research, jurisdiction-specific questions, false premise "trap" questions, and factual recall of metadata.
Key Results
The study found that while RAG-based tools improve performance over base LLMs, hallucinations remain a persistent issue.
- Lexis+ AI was the top performer, with a 65% accuracy rate and the lowest hallucination rate at 17%.
- Westlaw AI-Assisted Research achieved an accuracy of 42% but hallucinated 33% of the time.
- Ask Practical Law AI had a low accuracy (20%) primarily due to a high rate of "incomplete" answers or refusals, though its hallucination rate matched Lexis at 17%.
- GPT-4 (Baseline) hallucinated 43% of the time, confirming that RAG is a significant improvement but not a complete cure.
Discussion: The "Insidious" Hallucination
A critical finding is the emergence of "insidious" hallucinations. Unlike base LLMs that might fabricate case names entirely, RAG tools often cite real cases but attribute fake holdings to them. This is significantly harder for a legal professional to detect, as it requires a deep dive into the cited source to verify the AI's claim.
The study also highlighted that "Retrieval is not Reasoning." AI often fails to distinguish between a litigant's argument and the court's actual holding, or fails to recognize when a higher court has overruled a retrieved document. Furthermore, Westlaw's tendency to provide longer, more detailed responses increased the "surface area" for potential errors, contributing to its higher hallucination rate compared to Lexis+ AI.
Conclusion and Limitations
While RAG-based legal tools represent a step forward, they are not "hallucination-free." The proprietary nature of these systems makes it difficult to pinpoint whether failures occur during retrieval or generation. As these models are updated constantly, these results represent a snapshot of performance in Spring 2024. The authors emphasize that legal professionals must maintain a "duty of supervision," manually verifying every citation generated by AI to ensure accuracy and groundedness.
Summary updated