Same Same, But Different: Conditional Multi-Task Learning for Demographic-Specific Toxicity Detection

Different groups face different toxicity patterns: The paper demonstrates that toxic language targeting different demographic groups varies significantly, meaning a one-size-fits-all toxicity detection model will inevitably miss harmful content directed at some groups while potentially over-flagging others.

Standard AI models favor majority groups: Traditional machine learning models trained on mixed data tend to optimize for patterns in the majority group, leading to worse performance at detecting toxicity against minority demographic groups—a concrete example of how algorithmic bias emerges even without intentional discrimination.

Specialized models within one system improve fairness: By using Conditional Multi-Task Learning, the researchers created a single system with specialized branches for different demographic groups, which improved the model's ability to catch toxic content targeting minority groups while maintaining overall accuracy.

Minority group protection improved without sacrificing accuracy: The approach achieved higher recall (catching more harmful content) for minority demographic groups specifically, while keeping overall accuracy similar to baseline models—suggesting fairness improvements don't require trading off general performance.

Labeling challenges require practical solutions: The paper addresses a real-world problem: toxicity datasets often don't have complete labels for all demographic groups in every comment, and their conditional approach handles this by only using relevant labels for each group's specialized model.