Harm Laundering in GPT Models: Evidence That Gender Discrimination Is Transformed Rather Than Reduced Across Safety-Trained Generations

Imagine you're developing a large language model to assist with sensitive topics, like mental health or social issues. You'd want to ensure that the model doesn't perpetuate harmful stereotypes or discriminatory content. Currently, many evaluations focus on surface-level toxicity scores, which can misleadingly suggest that a model is safe when it might still be transforming harmful narratives rather than eliminating them. This is what's called harm laundering, where the underlying issues are masked by changes in language that appear benign on the surface. For instance, while earlier models might have produced harmful content directed at women, later versions seem to correct this but instead shift the focus to men, creating new forms of representational harm. This means that while toxicity scores drop, the actual harm can persist or even grow in different forms. The authors propose a new framework to identify and address this issue, emphasizing the need for a more nuanced understanding of model outputs. By formalizing harm laundering and providing a detection protocol, they offer a way to better assess the true impact of generative models, which is crucial for anyone building applications that rely on these technologies.

The concept of harm laundering introduces a new perspective on model safety evaluations.

The analysis is based on a large dataset and employs multiple classifiers, though it lacks external validation.