Summary
People often disagree on what is offensive because of their different backgrounds. The authors studied how adding information about people's socio-demographic groups helps AI models predict these differences. They found that while demographic info only slightly improves predictions for individuals, it makes a clearer difference when predicting group-level or population-level judgments. They also showed that models work better when they explicitly break down predictions into a common base plus differences tied to demographics, rather than just using demographics as extra data. This explains why previous studies found mixed results about the usefulness of social group info.
What this means in practice
- •For content moderation teams: Improve group-level content offensiveness assessments by incorporating socio-demographic insights with reasoning-aware models.
- •For social media platform designers: Design better tools for pluralistic offensiveness prediction that capture user diversity by modeling demographic-specific differences explicitly.
Authors
Yifan Wang, Leixin Zhang, Jinyi Mu, Yu Wang, Mayank Jobanputra, Isabel Valera, Vera Demberg
Abstract
Subjective judgments such as offensiveness often vary substantially across individuals, motivating pluralistic modeling approaches that incorporate annotators' socio-demographic information to capture this variation. However, prior findings on the benefits of socio-demographic information for pluralistic modeling are mixed, calling for a systematic analysis of both its effectiveness and the mechanisms through which it improves prediction. In this work, we study $\textbf{when}$ and $\textbf{where}$ socio-demographics benefit pluralistic offensiveness modeling, as well as $\textbf{the role of explicit reasoning}$ in leveraging this information. Across three multi-annotator datasets with prompted and trained LLMs, we first show that demographic gains are small and inconsistent for individual predictions but become clearer at the group and population levels, because aggregation attenuates the idiosyncratic variation that masks small but systematic demographic signals. Second, using a reasoning and prediction format that explicitly decomposes individual judgment into an overall baseline and a demographic-specific deviation, we show that the gains arise from both components: demographic information improves baseline estimation and enables meaningful demographic-specific shifts, and interventions show that better component estimates improve final predictions. Finally, comparing direct demographic conditioning with explicit reasoning about how the text interacts with annotator demographics, we find that reasoning yields more accurate baseline and shift estimates as well as final predictions, whereas conditioning alone rarely produces informative shifts. These results suggest that the mixed evidence in prior work may partly stem from individual-level evaluation, and that the value of socio-demographic information depends on both how it is evaluated and how models are guided to use it.