HateMirage: An Explainable Multi-Dimensional Dataset for Decoding Faux Hate and Subtle Online Abuse

Subtle hate speech tied to misinformation is widespread but understudied: The research reveals that harmful online content often doesn't come from overt toxicity alone, but from misleading narratives that normalize or incite hate in indirect ways—a gap that existing safety tools largely miss.

Current AI systems struggle with nuanced harmful content: When tested on the HateMirage dataset of 4,530 comments, language models showed that explanation quality depends more on diverse training data and reasoning capabilities than simply making models larger, suggesting current AI tools may inadequately flag sophisticated forms of online abuse.

A three-part framework helps decode complex harm: The dataset introduces a multi-dimensional approach examining Target (who is harmed), Intent (what the person aims to achieve), and Implication (broader social impact)—moving beyond single-label detection to capture how misinformation, intent, and consequences interact.

Fact-checking sources can improve hate speech detection: By tracing YouTube discussions connected to debunked misinformation claims from fact-checkers, researchers created a more realistic dataset that reflects how false narratives actually drive harmful speech online, offering a practical pathway for improving content moderation.

Better explainability is essential for responsible AI moderation: Unlike previous datasets offering only token-level or single-dimensional explanations, HateMirage's multi-dimensional framework enables AI systems to provide more transparent reasoning about why content is harmful—critical for building user trust in automated moderation decisions.