Study Finds GPT Models Transform Rather Than Eliminate Gender Bias

A new research paper published on arXiv examines how large language models like GPT handle gender discrimination. The study introduces the concept of 'harm laundering,' suggesting these models repackage biased outputs in less obvious ways rather than removing the underlying bias. Researchers argue that safety measures in GPT models may give a false sense of neutrality while discriminatory patterns persist in transformed forms. The findings raise concerns about the effectiveness of current AI alignment and content moderation approaches in addressing systemic gender bias.