A new frontier that requires careful navigation has been created by the intersection of artificial intelligence and open source principles. The Open Source Initiative (OSI) held the Affiliates Town Hall on February 27, 2025. The purpose of this discussion was to gather stakeholders from the open source community to discuss the Open Source AI Definition (OSAID) and the future of open source AI systems.
During this discussion, attendees were provided valuable insights into the challenges, opportunities, and regulatory framework surrounding open source AI development.
For those who are not familiar with the term Open Source Initiative (OSI), OSI is a "California public benefit corporation actively involved in Open Source community-building, education, and public advocacy to promote awareness and the importance of non-proprietary software."
Why there is a need for an Open Source AI Definition
Mr Jordan Maris, EU Policy Analyst for the OSI and Mr Stefano Maffulli, Executive Director at OSI analysed why there is an urgent need for clarity around what constitutes "Open Source AI." According to Mr Maffuli and Mr Maris, the European Union's AI Act has been a factor that intensified this need as it includes an exemption for open source AI systems; however, it lacks a precise definition of what qualifies for this exemption.
Mr Maris shared an interesting anecdote with respect to a conversation in Brussels where an MEP advisor asked directly: "What is Open Source AI?" This is a question that comes at a critical time since various stakeholders—especially large tech companies—attempt to create a definition of open source AI; However, the ways may conflict with the traditional four freedoms associated with open-source software.
"Openwashing" and the challenge of it
During the town hall, there was a significant concern that was highlighted and this was the practice of "openwashing." This entails various companies claiming their AI systems are open source; however, they are imposing restrictions that contradict fundamental open source principles. Mr Maris explained that OSI is "up against an army of lobbyists" with substantial resources (41 lobbyists, 18 full-time, €9 million budget), focussed on redefining not only open source AI but also the very concept of open source itself.
The companies that are engaged in “openwashing” usually argue that a) AI poses unique threats that justify usage restrictions, b) their widely-used AI systems give them authority to define what open source AI means, and c) AI is fundamentally different from traditional software and requires different rules.
OSI's Response: The Open Source AI Definition
The OSI developed the Open Source AI Definition (OSAID) as a countermeasure to these challenges. OSAID maintains the traditional four freedoms while addressing the unique aspects of AI systems. To be considered open source, according to the OSAID, an AI system must be available under legal terms that grant users the freedom to: use the system for any purpose without requiring permission, study how the system works and inspect its components, modify the system to change its recommendations, predictions, or decisions, and share the system with or without modifications, for any purpose.
The OSAID signifies that these freedoms require access to model weights and parameters, source code used to train the system and create the dataset, the complete list of data used to train the system and, when legally possible, the actual training dataset.
Balance between Legal challenges and Open Principles
The legal complexities surrounding AI have been acknowledged by the OSI's approach, especially regarding data protection regulations like GDPR and varying copyright rules. The OSAID claims that the four freedoms do not get compromised but rather: a) all training data that is legally and technically possible to share is considered mandatory b) requires information about where data that cannot be shared can be acquired c) requires detailed descriptions of all data for purposes of replication d) full release of training and data-processing code is considered mandatory.
A Process of Global Co-Design
To develop the OSAID, OSI went through a co-design process that was global; The process included in-person workshops in different continents such as South America, North America, Africa, Europe, and Asia, participants from more than 35 countries who attended virtually, and a systematic and structured process which was based on analysing AI systems, determining necessary elements that constitute the four freedoms, validating the AI definition against existing systems, and refining the definition iteratively based on the feedback.
Response and Criticisms from the Industry
According to Mr Maris, industry players demonstrated significant pushback. As an example, Meta claims that "There is no single open source AI definition." Additionally, it seems that Meta's Llama models include some problematic restrictions such as certain applications being restricted by Acceptable Use Policies, use based on user numbers being limited by additional commercial items; and geographical restrictions, for example, EU users might be excluded from certain multimodal models.
Mr Maris pointed out that the OSI's Open Source AI Definition has also faced various criticisms such as arguments about "open source" being the wrong term for AI due to the lack of traditional "source," claims that the development process was unfair or too fast for some stakeholders, suggestions that the OSI has stretched beyond its mission, and concerns that the definition is too restrictive, as no major commercial AI releases currently meet its criteria.
The OSI's Roadmap – Looking Forward
The Open Source Initiative views version 1.0 of OSAID as the "conversation starter" and has outlined a roadmap for continued development.
2025 includes: establishing a multilateral working group to record major issues, monitoring the AI space and observing evolving practices, examining AI applications beyond LLMs/GenAI (biotech, medical, vision, etc.), and finally, soliciting public feedback through OSI forums and at conference.
2026 includes: running a process to address major issues identified and updating the definition by Q4 2026.
The European Context
Mr Jordan Maris emphasized that there are "only two visions of Open Source AI within the European policy landscape, in Brussels: Meta's, and the OSAID." This is why, the OSI is actively working to educate lawmakers, and to counter “openwashing.” This is done by explaining those issues to lawmakers and their staff, participating in EU Commission Working Groups on Codes of Practice for General Purpose AI, highlighting instances of “openwashing” and presenting the OSAID as a clear, principled alternative.
Conclusion
The presentation at the OSI Affiliates Town Hall revealed that there is a complex interplay between open source principles, AI technology, corporate interests, and regulatory frameworks. AI will continue to evolve and reshape our technological landscape and the OSAID will be making an effort to ensure that open source values aren't compromised or redefined in ways that undermine user freedoms.
The coming years will play an important role in determining whether AI systems can truly be open like software has been. The efforts of OSI to define and advocate for genuinely open AI systems are facing challenges from powerful industry players, but also hand over the opportunity to extend the benefits of open source to this new transformative technology.
Therefore, OSI calls individuals and organizations who are supporters of open source principles and believe that they play a significant role in AI development to endorse the OSAID and join the ongoing conversation about the future of open source AI.
Comment
The Free Software Foundation (FSF) has not yet expressed its standpoint concerning the AI definition. Something that is expected to happen soon.
Resources
Special thanks to Mr Nick Vidal from OSI for his help in including the following resources in this article.
PDF files
Video recordings of the same presentation at FOSDEM's website
Why Europe needs the OSAID: Openwashing and the AI Act, by Jordan Maris
An Introduction to the Open Source AI Definition, by Stefano Maffulli
Italian translation: https://www.joomlaitalia.it/guide-e-tutorial/traduzioni-joomla-magazine/il-futuro-dellintelligenza-artificiale-open-source-definizione-open-source-ai-osaid-dellosi