Open ML considerations in EU AI law

Like other members of the machine learning community, Hugging Face has been in strict compliance with EU AI laws. This is landmark legislation aimed at shaping how democratic input interacts with AI technology developments around the world. It is also the result of extensive work and negotiations between organizations representing different components of society. As a community-driven company, we are particularly sensitive to this process. In the current position paper, written in collaboration with Creative Commons, Eleuther AI, GitHub, LAION, and Open Future, we aim to contribute to this process by sharing our experiences on the necessary role of open ML development in supporting the goals of the Act and, conversely, outlining concrete ways in which regulation can better consider the needs of open, modular, and collaborative ML development.

Hugging Face is what it is today because of our community of developers. So we’ve seen firsthand what open development can do to support more robust innovation for more diverse and context-specific use cases. Here, developers can easily share innovative new techniques, mix and match ML components to suit their needs, and work confidently with full visibility of the entire stack. We are also keenly aware of the need for transparency to play a role in supporting greater accountability and inclusivity in technology. In particular, we have worked to promote transparency through improving the documentation and accessibility of ML artifacts, educational activities, and hosting large-scale interdisciplinary collaborations. Therefore, as the EU AI Act moves toward its final stages, consideration of the specific needs and strengths of open and open source development of ML systems will help support its long-term goals. Together with our co-signatory partner organizations, we propose the following five recommendations to that end.

Clearly define AI components. clarifies that the collaborative development of open source AI components and their availability in public repositories is not subject to the requirements of the AI Act, and builds on and improves on Regulations 12a-c and Article 2(5e) of the Parliamentary Instrument; Support the AI Bureau’s coordination and inclusive governance with the open source ecosystem, based on Congressional documents. Ensuring that the research and development exception is practical and effective by allowing limited real-world testing. This condition combines aspects of the Council’s approach with an amended version of Article 2(5d) of the Parliament, which sets out proportionality requirements for ‘foundational models’, recognizes and explicitly addresses different uses and modes of development, including open source approaches, and aligns with Article 28b of the Parliament.

More details and context can be found in the full paper here.