[arXiv]score: 0.17

Seven LLMs Show Emergent Racial Steering in Housing Search Across Four U.S. Cities

June 8, 2026

A behavioral audit of seven open-weight and closed-source LLMs using fair housing paired-testing methodology found racial steering is an emergent, context-dependent behavior rather than a static model property. Steering intensity varies with user identity, preference framing, and the spatial logic each model has internalized about specific cities.

HOW THIS AFFECTS YOU

●

founderWorth watching because housing search integrations using LLMs carry Fair Housing Act liability risk that standard safety evals will not catch.

●

policyThis changes how you should think about LLM auditing for fair housing compliance — static red-teaming misses emergent steering that only surfaces through iterative, identity-conditioned prompting.

read original ↗arxiv.org

← back to feed