The Free Software Foundation considers large language models (lwn.net)

At the 2025 GNU Tools Cauldron, the Free Software Foundation’s Licensing and Compliance Lab led a wide-ranging session on how free-software licensing intersects with large language models (LLMs). The FSF is surveying projects to understand current practices and plans to use that input to craft guidance; it is not preparing a GPLv4 at this time but is considering refinements to the Free Software Definition. Key legal and practical questions remain open: whether LLM-generated code is copyrightable (courts are still grappling with this), whether adding human creativity or using “creative prompts” could secure copyright, and how to handle potential infringement when models regurgitate training data. Technically and operationally, the talk highlighted major compliance risks: most models don’t disclose training corpora or preserve copyright notices (so permissive-license provenance can be lost), some model ToS claim rights over outputs, and provenance is effectively unrecoverable unless contributors publish full sessions. The FSF recommended strict metadata practices for accepting LLM-derived code—record model and version, training-data info if available, the exact prompt, any use restrictions, and clear labeling of generated code—and emphasized human-in-the-loop accountability and the tension with assistive-tech uses. The upshot: projects must adopt risk-management and documentation policies now while legal and normative answers evolve.