Open Source Initiative Publishes "Open Source AI Definition" v1.0, Setting New Standards for AI System Open Sourcing

2024-10-29

Recently, at the "Comprehensive Open 2024" conference held in Raleigh, North Carolina, the Open Source Initiative (OSI) officially unveiled the results of its one-year global community initiative—the Open Source AI Definition (OSAID) version 1.0. This new definition aims to establish clear standards for the open sourcing of AI systems, thereby fostering collaboration and innovation within the AI sector.

It is reported that the project seeks to establish a set of principles to rebuild a permissionless, practical, and streamlined collaboration model for AI practitioners, much like the role the open source definition plays in the software ecosystem. OSI hopes that this initiative will create a more open, transparent, and equitable development environment in the AI domain.

The development of OSAID v1.0 received broad participation and support from over 25 organizations, including industry giants such as Microsoft, Google, Amazon, Meta, Intel, and Samsung, as well as organizations like the Mozilla Foundation, Linux Foundation, Apache Software Foundation, and the International Telecommunication Union of the United Nations. Through an extensive co-design process involving in-depth discussions and negotiations, a widely recognized definition was ultimately formulated.

Percy Liang, Director of the Stanford Center for Research on Foundation Models, expressed support for OSI's new definition. He noted that while establishing an appropriate open source definition is challenging, OSAID v1.0's requirement for the complete code handling data processing (a major driver of model quality) to be open source represents a positive advancement. Liang believes that as specific examples emerge, there will be more discussions on the application of this definition, thereby promoting further progress in the AI field.

According to the Open Source AI Definition, to qualify as an open source AI system, it must meet four criteria derived from the free software definition. These criteria include the ability to use the system for any purpose without seeking permission, to study how the system works and to inspect its components, to modify the system for any purpose, and to share the system with others for use (whether or not modifications are made). These freedoms apply both to fully functional systems and their individual components, with the prerequisite for exercising these freedoms being access to the preferred form of modifying the system.

However, despite the introduction of this new definition, some issues and concerns remain. Brian Jackson, Chief Research Director Analyst at Info-Tech Research Group, pointed out that while OSI's standards are clear and consistent with previous open source software release criteria, AI possesses key differences not covered by traditional open source software licenses, such as training data, model weights, and novel architectures. He also expressed concerns about potential harms from open source AI, such as deepfakes and fraudulent activities.

Nonetheless, the co-designers remain optimistic about the release of the Open Source AI Definition. Ayah Bdeir, AI Strategy Lead at Mozilla, stated that the new definition requires open source models to provide sufficient information about their training data so that proficient personnel can reconstruct a substantially equivalent system using the same or similar data. This marks the beginning of addressing the complexities surrounding the handling of AI training data and acknowledges the challenges in sharing complete datasets, while striving to make open datasets a more prevalent part of the AI ecosystem.

OSI expressed satisfaction with the release of OSAID v1.0 and views it as a starting point for further work. Stefano Maffulli, Executive Director of OSI, noted that although the development process was filled with challenges and differing opinions, the outcome aligned with the expectations set at the beginning of this two-year process. As collaboration with the broader open source community continues to gain understanding and application of OSAID v1.0, this will be the commencement of ongoing efforts.

The release of the Open Source AI Definition marks a significant step toward a more open, transparent, and equitable development environment in the AI sector. With increasing participation and support from organizations and individuals, the AI field is poised for more prosperous and innovative growth.