Reparative, Content-Agnostic Moderation for Recommendations

Design “reparative exposure controllers” for recommender systems that apply content-agnostic dispersion while allocating additional exposure to creators and communities historically harmed by biased moderation. Reparations goals are encoded as measurable exposure and harm-reduction targets, enforced via optimization constraints without accessing content features. This idea fuses content-agnostic moderation that reduces polarization with reparations frameworks that proactively address algorithmic harms, avoiding censorship concerns tied to content inspection while addressing fairness head-on. It operationalizes justice-oriented governance goals inside a content-agnostic control loop, including user-facing disclosures, opt-in controls for communities, and auditability of reparations targets over time. The approach is feasible in practice and creates a tractable path for platforms to acknowledge and reduce cumulative harms. Impact includes improving exposure equity, reducing polarization, and enhancing legitimacy without increasing censorship risks.

References:

1. Content-Agnostic Moderation for Stance-Neutral Recommendation. Nan Li, Bo Kang, T. D. Bie (2024). arXiv.org.
2. Repairing the harm: Toward an algorithmic reparations approach to hate speech content moderation. Chelsea Peterson-Salahuddin (2024). Big Data & Society.