Vyoma Raman

Vyoma Raman, Catherine Vera, CJ Manna

In this study, we introduce a novel two-pronged scoring system to measure individual and systemic bias in immigration courts under the U.S. Executive Office of Immigration Review (EOIR). We analyze nearly 6 million immigration court proceedings and 228 case features to build on prior research showing that U.S. asylum decisions vary dramatically based on factors that are extraneous to the merits of a case. We close a critical gap in the literature of variability metrics that can span space and time. Using predictive modeling, we explain 58.54% of the total decision variability using two metrics: partisanship and inter-judge cohort consistency. Thus, whether the EOIR grants asylum to an applicant or not depends in majority on the combined effects of the political climate and the individual variability of the presiding judge - not the individual merits of the case. Using time series analysis, we also demonstrate that partisanship increased in the early 1990s but plateaued following the turn of the century. These conclusions are striking to the extent that they diverge from the U.S. immigration system's commitments to independence and due process. Our contributions expose systemic inequities in the U.S. asylum decision-making process, and we recommend improved and standardized variability metrics to better diagnose and monitor these issues.

Vyoma Raman, Eve Fleisig, Dan Klein

The impact of AI models on marginalized communities has traditionally been measured by identifying performance differences between specified demographic subgroups. Though this approach aims to center vulnerable groups, it risks obscuring patterns of harm faced by intersectional subgroups or shared across multiple groups. To address this, we draw on theories of marginalization from disability studies and related disciplines, which state that people farther from the norm face greater adversity, to consider the "margins" in the domain of toxicity detection. We operationalize the "margins" of a dataset by employing outlier detection to identify text about people with demographic attributes distant from the "norm". We find that model performance is consistently worse for demographic outliers, with mean squared error (MSE) between outliers and non-outliers up to 70.4% worse across toxicity types. It is also worse for text outliers, with a MSE up to 68.4% higher for outliers than non-outliers. We also find text and demographic outliers to be particularly susceptible to errors in the classification of severe toxicity and identity attacks. Compared to analysis of disparities using traditional demographic breakdowns, we find that our outlier analysis frequently surfaces greater harms faced by a larger, more intersectional group, which suggests that outlier analysis is particularly beneficial for identifying harms against those groups.