Study reveals racial steering risks in LLM-powered housing searches
LLMs in housing search steer users based on race, not just preferences, warns MIT study
Deep Dive
Researchers audited 7 LLMs across 4 U.S. cities, finding racial steering emerges from model interpretations of preferences. Steering varied by city and user identity, with preference conditioning often increasing or reconfiguring steering behaviors. The study warns against assuming neutrality in place-based AI deployments.
Key Points
- 7 LLMs audited across 4 U.S. cities (Chicago, Houston, Los Angeles, New York) found racial steering varied by location and user identity
- Preference-conditioned testing increased bias in 60% of models, with Black users 3.2x more likely to receive recommendations to high-poverty areas for 'quiet neighborhoods'
- City-specific spatial logic in models drives emergent bias, requiring local expertise for fair housing compliance
Why It Matters
AI-driven housing searches risk automating discrimination—urging legal teams to audit LLMs for spatial biases before deployment.