Detoxifying Toxic Communication: A Design Science Approach to Responsible AI

Research from arXiv explores a Design Science approach to developing a responsible AI artifact aimed at addressing toxic communication in digital workplaces. Current moderation tools often disrupt communication by blocking or deleting content. This new approach integrates transformer-based classifiers for detecting toxic language with a generative detoxification model that rewrites such text into semantically equivalent, non-offensive paraphrases, demonstrating high accuracy in detection and strong semantic preservation.

Why it matters

The development of AI solutions capable of not just detecting but also 'detoxifying' harmful communication represents a significant advancement in fostering healthier digital environments. This shifts from purely punitive moderation to a more constructive approach, which can preserve dialogue while mitigating its negative impacts.

Key insights

• Toxic language in digital workplaces, encompassing pejoratives, sarcasm, condescension, and subtle incivility, negatively impacts trust, morale, and collaboration.
• Existing communication moderation tools typically focus on deleting or blocking harmful messages, which can disrupt communication flow and do not offer constructive resolution.
• A Design Science Research approach was used to create a responsible AI artifact for detecting and detoxifying toxic communication.
• The artifact combines fine-tuned transformer-based classifiers (DistilBERT, DistilRoBERTa) for toxicity detection.
• It utilizes a generative detoxification model (mT0-XL-Detox-ORPO) to rewrite toxic text into non-offensive, semantically equivalent paraphrases.
• Technical evaluations indicate high accuracy in toxicity detection and robust semantic preservation during the detoxification process.

Source

arXiv — Computers and Society — https://arxiv.org/abs/2609.00361