The paper “Red Teaming for Generative AI: Report on a Copyright-Focused Exercise Completed in an Academic Medical Center” discusses a structured red teaming exercise conducted by the Dana-Farber Cancer Institute regarding their internal generative AI tool, GPT4DFCI. Given the widespread use of this tool, rigorous copyright compliance testing was deemed essential. In November 2024, 42 participants from various sectors attempted to extract copyrighted content across literary works, news articles, scientific publications, and clinical notes. While successful extractions occurred for literary content, attempts to extract news articles were unsuccessful, and scientific reproductions yielded only summaries. Clinical note testing showed effective privacy safeguards. The findings highlighted vulnerabilities in copyright compliance and indicated the need for improved filtering during AI inference. As a response, a copyright-specific meta-prompt was implemented in January 2025. The study emphasizes the importance of continuous compliance testing for academic institutions utilizing generative AI.
Source link
Red Teaming Generative AI: Insights from a Copyright-Focused Exercise in an Academic Medical Center
Share
Read more