Detoxifying Toxic Communication: A Design Science Approach to Responsible AI

AI can improve rather than just punish toxic behavior: Rather than simply deleting or blocking harmful messages, this research demonstrates that AI systems can rewrite toxic language into respectful alternatives while preserving the original meaning. This constructive approach maintains communication flow in workplaces while promoting civility.

Toxic language takes many subtle forms: The study identifies that workplace toxicity extends beyond obvious slurs to include sarcasm, condescension, and subtle incivility—forms that traditional moderation tools often miss but that can significantly damage team trust and morale.

Responsible AI design requires multiple safeguards: The artifact combines two types of AI models (classifiers for detection and generative models for rewriting) to ensure both accuracy in identifying problems and quality in proposed solutions, suggesting that responsible AI systems benefit from layered technical approaches.

Meaning matters as much as safety: A key design principle emphasized is "semantic preservation"—ensuring that detoxified messages maintain their original intent and context. This prevents the system from simply sanitizing communication in ways that might distort what people actually want to say.

Fairness in moderation is a design challenge: The paper highlights that responsible AI moderation must consider fairness across different groups and contexts, indicating that one-size-fits-all content policies may not adequately serve diverse workplace environments.