On May 29th, the G7 countries released its vision on openness in artificial intelligence (AI). The document, coordinated under the French presidency, establishes a shared understanding of AI openness. Led by the G7 Digital and Technology Ministries, the non-binding reference recognises the economic importance of open source and the transformative potential of AI to highlight the need for a shared language on openness in AI.
To harmonise discourse on the topic, contributions from the open source community were crucial; supported by the Open Source Initiative (OSI), the organisation led workshops inviting community members to exchange knowledge and identify areas to clarify and refine what open source AI means. Importantly, OSI released an open source AI definition (OSAID) in 2024 that served as a foundation for the talks and final typology of the paper. The OSAID, however, is not fixed, “[we will] continue working on the OSAID and take onboard community feedback and lessons learned from the G7.” says Jordan Maris, an EU Policy Analyst at OSI.
Maris, who represented OSI at the publication of the vision document, emphasises that “the G7 approach is not definition – it is advisory.” By establishing core principles of AI openness and, as the paper states, “a reference point” of degrees of openness, the G7 and OSI are laying the groundwork for further coordination on defining AI openness.
Openness as a spectrum
The typology of four levels of openness includes: open source AI with open data; open source AI; open weights AI; and weights available AI. What may seem like semantic differences between the terms actually underlines the degrees of openness present and the need for further exploration of the topic.
The distinction between the first two kinds of open source AI types comes from the technical and legal limitations that can arise. As Maris puts it, “providing data is easy, but providing data information is burdensome.”
AI models often are trained on copyrighted or personal data, releasing all data models could come into conflict with regulations such as the General Data Protection Regulation (GDPR). Similarly, the sheer size of data information can be difficult to store. As a result, Maris highlights that too strict terminology will “have a negative effect on the development of open source AI.”
For the second pair of open model types, the distinction is about weights; the trained parameters used by a model. Oftentimes, proprietary models promote themselves as open weight, even if they are just available for download. Open weights AI, by contrast, release the weights under an open source licence, allowing for sharing, modification, inspection, and reuse of the weights.
This process is more than splitting hairs, by clarifying and identifying degrees of openness, this shared language drives innovation in industry and research.
Government collaboration
Alongside the publication of the vision paper, the G7 Digital and Technology Ministers, reiterated their commitment to open digital innovation and the economic and scientific benefits of a shared understanding of openness in AI. With the EU as a non-enumerated member of the G7, the vision paper comes from a high degree of international collaboration and emphasises the need for further discussion.
The opportunities of open source AI remain central to geopolitical discourse. The UN Open Source Week in June dedicated a day to AI, with the goal to bridge knowledge gaps and promote responsible AI.
As proprietary models currently dominate the AI landscape, the open source, with community as a core tenet provides an alternative approach to development. By establishing degrees of openness in AI, it’s shared understanding can encourage increased adoption of open source licences.
Photo by Devam Jhabak
