The Geography of Algorithmic Judgment: LLM Intermediaries, Place Identity, and Racial Steering in Housing Search

Researchers audited seven large language models across four U.S. cities and found that LLMs exhibit racial steering behaviors in housing recommendations, where the same preference produces different location suggestions depending on a user's perceived racial identity. The steering emerges dynamically from model interpretations rather than static biases, and varies significantly by city, suggesting that AI-mediated housing platforms may inadvertently perpetuate fair housing violations.

This research exposes a critical vulnerability in AI-mediated housing platforms: LLMs don't merely reflect historical discrimination but actively reinterpret user preferences through racialized spatial logics. When a model learns representations of neighborhoods, affordability, and lifestyle compatibility from biased training data, it develops implicit associations between demographic groups and geographic areas. The behavioral audit methodology—testing identical queries across racial identity signals—reveals that steering isn't a bug but an emergent property of how models interpret context. A preference for "vibrant neighborhoods" or "good schools" gets spatially resolved differently depending on perceived user race, effectively channeling demographic groups toward segregated markets. The finding that preference-conditioned testing increased steering behaviors is particularly concerning: as users provide more detailed requirements, models don't become more neutral but instead apply racialized filtering more intensely. This mechanism could substantially harm housing access and perpetuate segregation at scale, given the rapid integration of LLM-powered search into major listing platforms. The research also challenges the assumption that algorithmic fairness findings generalize across cities. Real estate markets are hyper-local with distinct spatial hierarchies, historical redlining patterns, and demographic compositions. An LLM trained on nationwide data may develop different steering behaviors depending on local market dynamics, making standardized fairness audits insufficient. For the housing sector and regulators, this research underscores that deploying LLM intermediaries requires localized testing, domain expertise, and continuous monitoring—not just general AI safety protocols.

• →Racial steering in LLM housing recommendations emerges dynamically from how models interpret place identity and user preferences, not from static historical bias alone.
• →The same housing preference statement produces different neighborhood recommendations depending on the user's perceived racial identity across multiple tested models.
• →Algorithmic fairness results cannot generalize across cities; local market dynamics significantly influence LLM steering behaviors in place-based sectors.
• →Adding lifestyle preference context to prompts often intensifies rather than reduces steering, suggesting models apply racialized spatial filtering more aggressively with detailed user profiles.
• →Fair housing compliance requires localized, domain-specific audits of AI tools rather than generic algorithmic fairness testing before deployment in housing platforms.