Community Content Moderation

Content moderation practices offer general, automated solutions at scale. However, these solutions often fail in nuanced settings, such as in regards satire, language reclamation, and otherwise subversive media. This disproportionately affects groups posed to use such tools for resistance and liberation from systems of oppression. For example, content expressing gender and sexuality for members of the LGBTQIA+ community is often disproportionately removed. Performance disparities can arise from lack of proper representation across the AI/ML pipeline. This work specifically targets the training data part of the pipeline, to understand how proper representation affects data annotation for and by content on a predominantly female- and queer-identifying and allied community called lips.social. Specifically, we examine the effects of self-disclosed identities to tagging for moderation purposes. We hypothesize that posts from community members to be better attuned to providing context, resulting in an evidence in support of decentralized moderation approach.