Researchers test LLMs to rewrite abusive text while preserving intent

Researchers have developed and tested a framework designed to help large language models detect and transform abusive language into cleaner versions while keeping the underlying message intact. The work, known as Detoxify, focuses on processing hostile content from tweets and user reviews that contain hate speech or profanity.

The approach#

The study benchmarks four prominent LLMs—Gemini, GPT-4o, DeepSeek, and Groq—on their ability to recognize abusive text and then rewrite it in non-abusive form. The key challenge is maintaining semantic similarity and emotional tone after transformation, so the rewritten text conveys essentially the same sentiment and intent as the original, just without the harmful language.

Key findings#

The results revealed significant differences in how each model handled the task. GPT-4o and DeepSeek produced similar outputs and performed comparably well. By contrast, Groq stood out as distinctly different, frequently reconstructing sentences with overly positive phrasing that sometimes lost or shifted the original context.

To measure success, the team evaluated both raw and transformed datasets using sentiment analysis and semantic similarity checks, confirming that some transformations preserved meaning better than others.

Why it matters#

As LLMs become more prevalent in content moderation and user safety tools, understanding their strengths and limitations in handling toxic language is valuable. The Detoxify framework contributes to ongoing efforts to make automated text processing safer and more nuanced, bridging the gap between removing harm and preserving authentic communication.