In 2022, a number of powerful machine learning (ML) models were released under open licenses. Tools such as the BLOOM LLM, Stable Diffusion or Whisper have led to significant downstream use and innovation, demonstrating that open (source) development of advanced machine learning models is possible. In the context of this development, we have also seen the emergence of a new class of open licenses, the so-called Responsible AI Licenses or RAIL licenses. The Open RAIL family of licenses – adopted by BLOOM LLM and Stable Diffusion – represents a new type of open license that combines elements of permissive open source licenses with usage restrictions based on ethical considerations. In recent months, these licenses have received considerable attention as they aim to ensure the responsible use of publicly available ML models.
To contribute to the discussion about the impact and role of Open RAIL licenses, and to better understand the dynamics of open (source) licensing of AI/ML models in general, we publish today a quantitative analysis of license usage for ML models published on huggingface.co. The paper – written by Paul Keller and Nicolò Bonato – analyzes the use of open licenses for sharing ML models, in particular the Open RAIL family of licenses, by conducting a historical analysis of publicly available license information on the huggingface.co platform.
It finds that over the period between September 2022 and January 2023, the use of Open RAIL licenses has grown rapidly from 0.5% to 7% of all licensed repositories, during which time Open RAIL licenses have overtaken all other categories of restrictive open source licenses, and are now the second most used category after permissive open source software licenses, which dominate the field with a share of more than 80% of all repositories.
The paper concludes by noting that Open RAIL licenses, which aim to promote responsible AI development by embedding safeguards in licensing practices, have so far had limited impact in a field dominated by permissive open source licenses that prioritize rapid technological progress. Since the field is dominated by permissive open source licenses, this means that calls for regulation to ensure the responsible development of open ML are likely to be met in the form of external regulation, such as the European Union’s proposal for an AI Act.
An interesting consequence of this situation may be that broad regulatory efforts such as the AI Act are likely to impose conditions on the development of ML systems that will be difficult for developers of openly licensed ML models to comply with. In such a scenario, Open RAIL licenses may end up being less of a tool for self-regulation based on community norms and more of a tool for regulatory compliance.