TLDR: A new research paper titled “Place Matters: Comparing LLM Hallucination Rates for Place-Based Legal Queries” reveals that the accuracy of legal information provided by Large Language Models (LLMs) varies significantly by geographic location. The study, which tested leading LLMs on legal scenarios in Los Angeles, London, and Sydney, found that hallucination rates were lowest in Los Angeles and highest in Sydney. It also identified a strong negative correlation between the consistency of LLM responses and hallucination rates, suggesting that more consistent answers are less likely to be incorrect. The findings highlight that access to accurate AI-generated legal information is not evenly distributed globally, potentially due to factors like training data distribution and the volume of online legal content available for different jurisdictions.
Large Language Models (LLMs) are increasingly being used by individuals seeking legal information for personal matters, offering accessible and low-cost assistance. However, a critical question arises: does the accuracy of the legal information provided by these AI models vary depending on a user’s geographic location? A recent research paper, “Place Matters: Comparing LLM Hallucination Rates for Place-Based Legal Queries”, delves into this crucial issue, revealing significant disparities in LLM performance across different places.
The study, conducted by Damian Curran, Vanessa Sporne, Dr. Lea Frermann, and Prof. Jeannie Paterson, highlights that while LLMs have the potential to bridge the access to justice gap, this can only be realized if they provide accurate legal information. Inaccurate yet plausible information could, in fact, disadvantage those who need help the most.
The Challenge of Comparing Legal Knowledge Across Places
Comparing an LLM’s knowledge of law in one place versus another is inherently complex. Legal systems differ vastly in size, subject matter, and historical, political, and cultural contexts, making direct comparisons difficult. To overcome this, the researchers adopted a methodology rooted in the comparative law concept of functionalism. Instead of comparing legal institutions directly, they focused on comparing the solutions offered by different legal systems to the same practical problems.
A Novel Methodology for Evaluation
The research team constructed a unique dataset of 100 factual scenarios derived from real-world Reddit posts where users sought legal advice on family, housing, employment, crime, and traffic issues. These scenarios were made “place-agnostic” by removing any location-specific terminology, ensuring they could be applied universally. The LLMs were then prompted to identify and summarize an applicable law for each scenario in three distinct locations: Los Angeles, USA; London, UK; and Sydney, Australia. These cities were chosen as large population centers in different countries where laws are written in English, to minimize confounding factors.
Three leading closed-source LLMs – GPT-4o, Claude-Sonnet-3.5, and Gemini-Pro-1.5 – were tested. To ensure robustness and account for variations in LLM responses, each scenario-place-LLM combination was sampled 10 times. The most frequent response was then manually evaluated by legally qualified annotators for “hallucinations” – instances where the LLM’s summary was factually inconsistent with the actual law. Annotations were categorized as Accurate, Minor Hallucination, Major Hallucination, or No Law.
Key Findings: Place Matters for Accuracy
The study’s results revealed a clear and significant association between hallucination rates and geographic location. Across all models, Los Angeles exhibited the lowest hallucination rates (45%), followed by London (55%), and then Sydney (61%). GPT-4 generally performed best among the models, with its lowest hallucination rate of 43% observed in Los Angeles.
Another significant finding was a strong negative correlation between the majority sample frequency (how often an LLM produced the same response out of 10 samples) and hallucination rates. This suggests that when an LLM consistently provides the same answer, it is less likely to be hallucinating legal facts. This insight could be valuable for developing automated hallucination detection methods.
The researchers also observed differences in the nature of legal references. The vast majority (88.5%) of cited laws were codified instruments like statutes or codes. Interestingly, Los Angeles had the lowest number of unique laws cited, indicating that its legal framework might be more densely organized into fewer, comprehensive instruments compared to London or Sydney. Hallucination rates also varied by legal issue, with housing-related queries showing the highest rates and employment law queries the lowest.
Also Read:
- Unmasking Confident Errors: Spurious Correlations Challenge LLM Hallucination Detection
- AI Breakthrough: Automating the Extraction of Criminal Facts from Court Opinions
Implications and Future Directions
The findings underscore a critical point: the quality of legal information provided by LLM-powered chatbots is not uniformly distributed across the globe. This has significant implications for pro se litigants, who may receive varying levels of accuracy depending on their location, and for legal service providers integrating LLM outputs into their workflows.
While the study established an association between hallucination rates and place, it also opened avenues for future research into the causes. One plausible explanation relates to the distribution of training data. Larger jurisdictions, like California (where Los Angeles is located), might have a greater volume of online legal texts, law firm websites, and blog posts, leading to better representation in LLM training data and consequently lower hallucination rates. The study’s observations in Sydney, where federal legislation (applying across Australia) had lower hallucination rates than state-level legislation (applying only in New South Wales), further support this “long-tail knowledge” hypothesis.
The research also suggests that employing self-consistency methods, such as sampling an LLM multiple times and using majority voting, could serve as a measure of uncertainty for legal fact predictions, potentially improving accuracy despite increased computational cost. Future work could expand this research to a wider array of legal issues, places, and languages, especially those with fewer digital resources, where hallucination rates are likely to be even higher.
In conclusion, this study provides compelling evidence that geographic location significantly impacts the accuracy of legal information generated by LLMs. Understanding these disparities is crucial for ensuring equitable access to justice in an increasingly AI-driven world.
