A new protocol, "Swap, Replicate, Read the Mass," has been developed to audit demographic bias in System One decision models. The protocol involves running models on benchmarks like BBQ, CrowS-Pairs, and SilverSpoon, with options and questions systematically permuted. The audit revealed widespread stereotype alignment, order-dependent answers, and confidence scores that correlate with instability rather than actual bias. Notably, option swaps significantly altered model choices, and some models exhibited bias towards lower-SES labels. AI
IMPACT This protocol could lead to more robust auditing of AI decision models, pushing for greater fairness and transparency in production systems.
RANK_REASON The cluster contains an academic paper detailing a new protocol for auditing AI model bias. [lever_c_demoted from research: ic=1 ai=1.0]
AI-generated summary · Google Gemini · from 1 sources. How we write summaries →