Demographic Information Impact on Pluralistic Offensiveness Modeling
October 30, 2026
Analysis of LLM offensiveness modeling shows that socio-demographic data provides small, inconsistent gains for individual predictions but significantly improves accuracy at group and population levels. Explicit reasoning helps, but demographic benefits are primarily revealed through aggregation.
HOW THIS AFFECTS YOU
●
researcherUse population-level aggregation to capture demographic signals that are masked at the individual level.
●
policyThis highlights the difficulty of using demographics to ensure individual fairness in content moderation.