LLMs map software licenses into a partial order of permissiveness

Researchers Hamidah Oderinwale, David Atkinson, Rachel Hong, Art Abal, and Ben Laufer have published a paper on arXiv titled "Partially ordering software licenses" that tackles a long-standing pain point in open-source development: comparing licenses at scale. Currently, licenses are written and chosen in an unstructured way, making it nearly impossible to answer questions like "Is license A more permissive than license B?" without manual legal review. As platforms like Hugging Face, GitHub, and the Python Package Index host millions of projects each with its own license, developers need systematic ways to understand compatibility and restrictiveness.

The team leverages large language models (LLMs) to construct a partial ordering of licenses based on permissiveness. First, they perform pairwise comparisons across a growing list of licenses, using LLMs to judge relative permissiveness. Second, they draw on existing taxonomies to see how license selection combines shared provisions. Their analysis recovers interpretable attributes that correspond to stricter licenses—offering concrete legal implications for the open-source ecosystem. This approach not only automates what was a manual, expert-driven process but also scales to handle the ever-expanding landscape of new licenses, giving developers and platform operators a data-driven tool for navigating legal complexity.