Researchers map open-source license complexity using AI and statistical models

Developers publishing code to platforms like GitHub and Hugging Face increasingly face a fragmented landscape of software licenses, each with distinct terms and restrictions. Without a structured way to understand how these legal instruments compare to one another, choosing between them or combining code under different licenses has remained largely ad hoc.

The permissiveness puzzle#

Software licenses exist to protect developers' work and define how others can use it—yet their interpretation and selection has evolved organically, lacking systematic frameworks for comparison. The core challenge is determining when one license is genuinely more permissive (allowing more freedom) than another, and when their terms are simply incomparable. As the repository of available licenses continues to grow, this unstructured approach creates friction for platforms managing thousands of projects with varied legal requirements.

A machine-learning solution#

Researchers tackled this problem by deploying large language models (LLMs) to conduct pairwise comparisons of licenses at scale. By examining licenses in pairs and asking which is more permissive, they constructed a partial ordering—a ranking system that captures permissiveness relationships while acknowledging cases where licenses cannot be directly compared. They validated this ordering against existing taxonomies of software licenses to ensure accuracy.

The Bradley-Terry tradeoff#

To test whether a cheaper, simpler method could capture permissiveness, the team applied the Bradley-Terry model, a statistical approach traditionally used for ranking competitors. The model recovered the structure with approximately 20% information loss, with gaps traced to gaps in feature coverage across licenses. This suggests that while the full LLM approach offers greater precision, a statistical shortcut could work for many practical applications.

Understanding license selection#

By pairing model outputs with explicit reasoning traces, the researchers were able to identify which specific provisions drive restrictiveness. They also found that developers' actual license choices often reflect overlaps in shared provisions—a insight that could help teams better predict and align their legal choices with their needs.