Facebook’s Hate Speech Detection Algorithms Allegedly Disproportionately Failed to Remove Racist Content towards Minority Groups

Between 2019-2021, Facebook conducted internal research called the 'Worst of the Worst' project that revealed serious flaws in its AI-powered hate speech detection algorithms. The company surveyed over 10,000 users to identify the most harmful content, finding that the worst examples were almost all directed at minority groups, particularly Black, Muslim, LGBTQ, and Jewish users. However, Facebook's algorithms were detecting and removing content critical of white people at much higher rates - approximately 90 percent of hate speech takedowns were for content directed at white people and men. The algorithms consistently failed to remove the most derogatory racist content against minorities. Internal documents showed that only 3-5 percent of actual hate speech on the platform was being removed, despite public claims of 90+ percent removal rates. When researchers proposed overhauling the system to better protect vulnerable minorities, Facebook executives including VP Joel Kaplan rejected the proposal, citing concerns about conservative backlash and appearing non-neutral. The company implemented only minor changes, continuing to leave minorities more exposed to hate speech. Internal studies showed Black users were leaving the platform, with usage declining 2.7 percent in one month to 17.3 million adults, partly due to harassment and inadequate content moderation.