Exploring the Plausibility of Hate and Counter Speech Detectors with Explainable AI

Explainability methods vary significantly in effectiveness: The study found that perturbation-based approaches (which test how predictions change when input is modified) outperformed other methods, while prototype-based approaches failed to provide useful explanations for hate speech detection systems.

Understanding AI decisions matters for trust: User studies demonstrated that when people could see why a model flagged content as hate speech or counter speech, they better understood and trusted the model's predictions, highlighting the importance of transparency in content moderation systems.

Gradient-based and attention-based methods show promise but have limitations: These two approaches ranked second and third respectively, suggesting they can help explain model decisions but are less reliable than perturbation-based methods for this specific task.

One-size-fits-all explainability doesn't work: The research compared four fundamentally different approaches to explaining AI decisions, finding that the best method depends on the specific task—in this case, hate speech detection—rather than assuming any single explanation technique works universally.

Explainability is critical for content moderation: Since hate speech detection systems directly impact online safety and free expression, being able to explain why content is flagged as hateful is essential for responsible deployment and user accountability.