After years of collaboration with academic and industry leaders, the Open Source Initiative (OSI) has released a formal definition for “open source artificial intelligence” (OSAI). This milestone, while non-binding, aims to provide clarity on what constitutes an open-source AI model, guiding developers, companies, and researchers on standards for transparency and usability in AI. The definition addresses the complexity of AI beyond traditional software, focusing on transparency in data, training algorithms, and model architecture.
A Comprehensive Approach to Open AI
The traditional open-source definition, which primarily applies to software, is not fully applicable to artificial intelligence. According to the OSI, an AI model to be deemed open source must go beyond code availability to include transparency around training data, data sources, and processing methodologies. This ensures that anyone can not only use the model but also understand, replicate, and modify it, with clear rights for adaptation and development.
“An open-source AI model allows for a complete understanding of how it’s made,” OSI’s Stefano Mafuli told TechCrunch. “This means access to all components, the complete code used to train and filter the data.” The definition underscores the importance of making the entire training and processing code available, as well as documenting the data sources and any licensing limitations on data usage. The model should also allow users the freedom to modify and upgrade it independently.
Current AI Models Fall Short of OSI Standards
According to the new standards, most existing AI models, including some from industry giants, do not meet OSI’s criteria for open-source status. For instance, Meta’s Llama model restricts access to platforms with over 700 million users unless a specific license is granted. Similarly, Stability AI requires a special license for businesses with over $1 million in revenue. Despite claims of being open source, such licensing restrictions conflict with OSI’s definition of transparency and unrestricted use.
Even companies like Google and Microsoft have removed references to “open source” on some platforms after feedback from OSI, as the term did not fully align with the foundation’s latest criteria.
Industry Reactions: Mixed Reception and Future Dialogues
The new definition has stirred mixed reactions across the AI landscape. Meta, for example, has expressed concerns about OSI’s strict criteria, which they say could hinder responsible distribution. A Meta spokesperson explained, “While we agree with our partner OSI on many things, the industry does not have a single definition of open source AI. This is challenging as previous definitions do not capture the complexities of today’s AI models.”
Luca Antiga, co-founder of Lightning AI, added that the OSI definition must prioritize clarity around training data licensing to ensure practical application for businesses, especially regarding compliance with organizational needs and regulatory frameworks.
A Step Toward Responsible AI Innovation
Despite the varying opinions, OSI’s open-source AI definition represents an important step in establishing standards for transparency and accessibility in AI. As AI models become more integral to society, defining open source AI can create a foundation for responsible, equitable, and accessible innovation, promoting collaboration and building trust among developers and users.
As the definition evolves, OSI will likely continue collaborating with industry leaders to strike a balance between accessibility and responsible usage, shaping a transparent future for AI development.
