From a River in Gilead to the Inference Distributions of Large Language Models: Covert Dialect Bias and Linguistic Profiling at Scale

This paper investigates whether large language models (LLMs) encode covert dialect bias that mirrors human linguistic profiling in housing markets. While alignment techniques often mask explicit racial bias in model outputs, the authors hypothesize that underlying probability distributions remain tainted by historical prejudices. Using a matched-guise paradigm, the researchers tested ten open-weight LLMs across four English varieties: Standard American English (SAE), African American Vernacular English (AAVE), Nigerian Standard English (NSE), and Nigerian Pidgin (NP). By keeping semantic content constant across 260 sentence quadruples, the study isolates the effect of dialect on the model's association with various housing-relevant adjectives across three social contexts: tenant screening, neighbor acceptance, and roommate selection.

The results reveal that LLMs systematically penalize AAVE and NP compared to SAE. Across all ten models, these varieties are associated with more negative adjectives, with NP receiving the most severe penalties. Crucially, the models do not apply a generic 'non-standard' penalty; instead, they associate each dialect with distinct stereotype clusters. AAVE is penalized through associations with unreliability and social threat, while NP is penalized through associations with corruption, danger, and foreignness.

NSE presents a unique case. As an institutionally prestigious variety, it is favored over SAE in formal tenant screening but faces increased penalties as social proximity grows (e.g., in roommate selection). This suggests that prestige operates as a mechanism independent of racial identity. The bias is consistent across models from diverse global training regimes, indicating that these prejudices are deeply embedded in web-scale pre-training data.

As LLMs are increasingly integrated into automated housing platforms and screening workflows, these covert biases pose a significant risk of automating discrimination at scale. Because these associations exist at the level of internal probability distributions rather than explicit text, they are invisible to standard auditing processes. The findings demonstrate that without dialect-aware evaluation, AI infrastructure may perpetuate historical housing discrimination, effectively acting as a modern-day 'shibboleth' that excludes individuals based on how they speak.