Detecting Bias in the Law

Biases in language models (LMs) can cause real-world financial and legal harms to targets of those biases. For instance, LMs are used to make investment decisions based on company news and earnings calls; if an investment is concerned with, say, a particular ethnic group, a biased model may misinterpret ambiguous speech in an unflattering light, leading to a decision not to invest. LMs have been used to make first pass hiring decisions and have specifically disfavored female applicants. We anticipate the use of LMs in the acceleration of legal decision-making, an especially important and high-risk domain. We believe the first step to combatting bias is to identify it, ideally before actual harm is done. In this work we will develop a technique for detecting bias in text, specifically in the legal subdomain. Our previous work showed that it is possible to alter a model's bias by simply exposing it to text from a targeted community; the resulting model is changed in bias relative to the bias of the community, as reflected by the words they use. We will develop tools that can take a step further than simply identifying bias and will leverage key explainability components of current architecture to determine specific words, phrases, and documents in the legal corpus that contribute to the bias. The identification of these documents can be the first component in a system that reveals the harms specific laws may impose on targeted communities. This work is a collaboration with USC's Gould School of Law.