Bradley M. Kuhn, executive director and one of the founders of the Software Freedom Conservancy (SFC), criticized the recently published definition of Open Source AI by the Open Source Initiative (OSI). According to Kuhn, the OSI rushed to publish the final version of the definition and approved it without a lengthy and comprehensive discussion, at an early stage of the development of such systems. In contrast, the definition of Open Source was provided after many years of reflection and discussion. Regarding the published definition of Open Source AI, it should not be called a definition at this stage, but rather a recommendation.
Kuhn fears that the current criteria for Open Source AI will have far-reaching consequences, undermine the authority of the term Open Source, and lead to divisions within the community. The problem is that, despite objections, the OSI compromised and did not include the requirement to provide the data used to train the model. The reason for the compromise was that if such a provision were added, none of the existing large language models would qualify as open.
According to the OSI, maintaining a definition of Open Source AI that takes into account the existing realities will help prevent manufacturers from manipulating the term 'open'. In uncertain conditions, they call models open solely based on the availability of weights, even if the model’s license restricts its use (for example, many models prohibit use in commercial projects) and does not disclose implementation details.
The OSI-approved definition of Open Source AI only requires providing detailed information about the data used for training, but not the data itself. Without providing the original data on which the model was trained, it is impossible to fully reproduce the AI system, which contradicts the concept of open source. Thus, the OSI limited its consideration of AI models purely as technology and did not regard them as a comprehensive product.
The prevailing definition of an open AI system guarantees only two of the four declared freedoms of Open Source — the ability to use and distribute it, while the freedoms to modify and study it are not fully ensured. Moreover, the lack of source data complicates the detection of backdoor substitutions in machine learning models.
On the other hand, publishing source data is often impossible due to reasons beyond the AI model developer's control, such as the need to maintain confidentiality, the use of copyrighted materials, and licensing data from third-party providers. Critics of the accepted definition argue that such issues do not justify diminishing or devaluing the concept of Open Source.
Bradley Kuhn intends to participate in the upcoming OSI leadership elections and attempt to join the board of directors to work towards rescinding the accepted definition and downgrading it to a recommendation. Along with the Software Freedom Conservancy, several developers from the Debian project have expressed their disagreement with the definition of open AI, proposing a general vote to draw attention to the issue. The Software Freedom Foundation is also working on its definition of a free AI system, which intends to incorporate a requirement for the availability of all data, while acknowledging the ethical reasons that may prevent data disclosure in some cases (e.g., when training involves medical or personal data).
Source: opennet.ru
