The Enforcement and Feasibility of Hate Speech Moderation

A recent audit of hate speech moderation on Twitter (now X) found that approximately 80% of identified hateful tweets, including violent content, remained online five months after posting. The platform's removal rates for hate speech were only marginally higher than for non-hateful content, significantly lagging behind moderation of issues like scams or adult content. Current automated detection methods are insufficient for reliable classification but can effectively prioritize content for human review, suggesting that current staffing levels are a major impediment to effective enforcement.

Why it matters

The persistent and widespread presence of unmoderated hate speech on online platforms presents significant reputational, regulatory, and societal risks. Ineffective moderation undermines platform integrity, can deter users, and may invite stricter external oversight or legal action, impacting market position and public trust.

Key insights

• 80% of hateful tweets, including violent content, remained online on Twitter (now X) five months after initial posting.
• Removal rates for hate speech were only marginally higher than for non-hateful tweets.
• Moderation of hate speech was significantly less effective compared to moderation of scams or adult content.
• The severity and reach of hateful content did not significantly impact its likelihood of removal.
• Automated detection systems are currently unable to reliably classify hate speech but are effective at ranking content for human triage.
• Simulation suggests that current staffing levels for human review are insufficient to address the volume of hateful content.

Source

arXiv — Computers and Society — https://arxiv.org/abs/2604.12289