Introduction
Natural Language Processing (NLP) models often "inherit" the prejudices of their training data. While bias detection is a significant area of general AI research, the legal domain presents unique challenges. Standard tools like the Word Embedding Association Test (WEAT) typically use common first names to detect bias. However, legal opinions are highly formal; individuals are usually referred to by surnames or titles, and "positive sentiment" in a legal context (such as a motion being "granted") is fundamentally different from "positive sentiment" in everyday English (like "flowers" or "laughter").
If legal AI systems—used for research, information extraction, or predicting case outcomes—rely on biased word representations, they risk amplifying historical injustices. Researchers from Thomson Reuters Labs sought to bridge this gap by creating bias detection methods specifically tuned for the linguistic nuances of the law.
Methodology
The authors trained Word2Vec (skip-gram) embeddings on a massive corpus of 12 million U.S. legal opinions spanning from 1650 to 2020. To handle the specificities of legal language, they introduced several novel steps:
- Idiomatic Phrase Extraction: Using Normalized Point-wise Mutual Information (NPMI), they identified and grouped legal jargon (e.g., pro hac vice) as single tokens, ensuring the model captures the meaning of the phrase rather than its individual components.
- Domain-Adapted WEAT: They modified the standard WEAT approach by creating new, legal-specific word lists. Instead of just first names, they used surnames categorized by race (via 2010 U.S. Census data) and a list of "Judge Given Names."
- Attribute Lists: They moved beyond "Pleasant vs. Unpleasant" to include "Grant vs. Deny" (legal outcomes) and "Positive vs. Negative Legal" terminology.
To determine if bias was merely a relic of the past, the researchers also performed temporal and topical slicing, training models on different time "cutoffs" and categorizing data by legal topic (e.g., Torts, Crimes, Property).
Key Insights and Results
The study revealed that representational biases are deeply embedded in legal text and are not easily removed by simply using modern data.
Sentiment vs. Legal Outcome
One of the most striking findings is that "positive sentiment" in general language does not always correlate with "positive outcomes" in law. For example, Asian Pacific Islander surnames showed a positive association with general "pleasant" words but a strong negative association with "positive legal" terms when compared to European surnames.
Racial and Gender Bias
The study found moderate to large effect sizes (Cohen’s $d$) indicating negative racial bias for African American and Hispanic surnames across both "positive legal" and "grant/deny" attributes. Furthermore, the "Career vs. Family" bias (associating men with careers and women with family) was extremely strong and present across all time periods and all legal topics. In gender tests, the "Grant vs. Deny" attribute consistently showed a medium negative bias against female-associated terms, even when general sentiment tests showed a positive bias.
The Persistence of History
Training only on modern cases (post-2000) did not eliminate bias. While some biases decreased slightly, racial and gender stereotypes remained remarkably stable across all time periods. Additionally, bias was found to be topic-specific; the "Grant vs. Deny" bias was significantly stronger in categories like Crimes and Property than in other areas of law.
Conclusion
The authors emphasize that this study focuses on representational bias—how groups are depicted in the underlying math of the embeddings—rather than allocative harm, which is the real-world impact of a specific tool making a biased decision. Future work is needed to see how these biased embeddings actually affect downstream tasks like legal search or case prediction. Simply filtering by date is insufficient to mitigate these biases, suggesting that more sophisticated mitigation strategies are required to ensure fairness in legal AI.
Summary updated