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

Abstract

Large language models (LLMs) are increasingly being deployed in services that shape consequential life outcomes, from housing platforms to automated screening workflows. While alignment techniques have reduced the explicit expression of racial bias in model outputs, they leave untouched the attitudinal associations reflected in a model's internal probability distributions. These covert biases are invisible to users and resistant to conventional auditing that inspects generated text alone. In this paper, we build upon the matched-guise sociolinguistic paradigm to examine the covert dialect bias of LLMs in housing-related social judgments across four English varieties: Standard American English (SAE), African American Vernacular English (AAVE), Nigerian Standard English (NSE), and Nigerian Pidgin (NP). AAVE reflects the racialized dialect examined in prior covert-bias evaluations, while NSE and NP represent Black African, postcolonial varieties absent from this line of research. Using 260 meaning-matched sentence quadruples and log-probability scoring over housing-relevant adjectives, we probe ten open-weight LLMs across three evaluation contexts that gradually vary in social proximity: tenant screening, neighbor acceptance, and roommate selection. Our results show that AAVE and NP are consistently associated with more negative housing-relevant adjectives than SAE across all ten models, with NP receiving the most severe penalization overall, and that each dialect is penalized through a distinct stereotype cluster rather than as a generic non-standard category. NSE, which shares the Nigerian identity of NP but carries institutional prestige, exhibits a context-dependent pattern that diverges sharply from both: favored over SAE in formal tenant evaluation but penalized as social proximity increases. Together, these results show that LLMs inherit covert dialect bias along two distinct dimensions, racial identity and prestige, echoing housing discrimination long documented among humans and extending its reach beyond the African American community to racialized postcolonial English varieties.