Open Source AI Salons: Shaping the OSAID, Together

Two years after the release of the Open Source AI Definition (OSAID 1.0), the Open Source Initiative hosted the first in a series of Open Source AI Salons, gatherings to discuss the future of the OSAID. In collaboration with key partners in Raleigh and Barcelona, the events brought together civil society, government, industry, and academia to understand the practical application of the OSAID.

These interdisciplinary, community-driven events provide the ground to surface challenges, examine assumptions, and explore practical issues affecting the Open Source community when it comes to AI. As we head into 2027, we are developing opportunities to bring candid discussions about the OSAID to more people in more places. OSI is steadfast in its commitment to empowering the Open Source community and promoting a healthy Open Source AI ecosystem.

The OSAID and Its Application

At their core, the OSD and OSAID help people make decisions and understand what is shared, under what terms, and when a piece of software or AI model should be categorized as Open Source. The definitions are intended to help anyone using, studying, producing, procuring, sharing, modifying, or regulating software and AI understand what it means to carry the Open Source label.

Applying the OSAID has proven more difficult than applying the OSD. Key elements that apply to Open Source software don’t translate cleanly to AI models. The norms and heuristics that applied to Open Source software shifted abruptly with AI, particularly around weights and data. There was, and remains, an urgent need to build consensus on what Open Source means beyond code and OSI-approved software licenses.

OSI first set out to define Open Source AI through a participatory, co-designed process with global stakeholders in June 2023. In October of 2024, the OSAID 1.0 was announced at All Things Open in Raleigh, North Carolina. We are hosting a new series of in-person and online events to continue the conversation.

Why We’re Reconvening the Community Now

When OSI released the Open Source AI Definition in 2024, we called it version 1.0 for a reason. We knew that AI technology and the way models are developed and shared would continue to evolve, and that we would need to revisit the definition as the ecosystem matured. Two years later, the growing popularity of open-weight models, new regulations affecting Open Source, and changing approaches to sharing AI models are raising important questions about how the OSAID is understood and applied. The Open Source AI Salons, together with OSI’s research and policy work, are helping us explore those questions with the community.

Consider how much the AI ecosystem has grown. In 2024, Hugging Face, one of the largest repositories of AI models, hosted approximately one million models. Today, that number has surpassed three million. At the same time, governments around the world are introducing regulations that affect how AI systems are developed, shared, and used, including provisions that affect Open Source developers.

OSI’s research into models shared on Hugging Face reveals some encouraging trends, along with areas of concern. More models are being released under OSI-approved licenses, particularly MIT and Apache 2.0, suggesting growing adoption of established Open Source licensing practices. At the same time, AI-specific licenses are emerging. Two licenses, ModelGo and OpenMDW, are awaiting OSI approval.

But we’re also seeing models released under a mix of Open Source and restrictive terms. For example, developers may share their code under an OSI-approved license while placing restrictions on how the model’s weights can be used through an acceptable use policy or terms of use agreement. These restrictions can prevent users from exercising the freedoms that are fundamental to Open Source AI. An OSI-approved license for code is important, but it doesn’t automatically mean an AI model meets the OSAID.

Perhaps the biggest challenge that remains is transparency around training data. The OSAID requires developers to provide detailed information about the data used to train an AI system. Despite the enormous growth in publicly available models, relatively few appear to meet the OSAID’s data information requirements.

Taken together, these developments show both the progress being made toward greater openness in AI and the challenges that remain. They also reinforce why continued community engagement matters. Through the Open Source AI Salons and our ongoing research and policy work, we can examine how the AI landscape is changing, identify where greater clarity is needed, and determine whether and how the OSAID should evolve.

The OSAID Conversation in Raleigh and Barcelona

Our first two events engaged different communities and approached the OSAID from different angles. While topics ranged from the applicability of a binary definition of Open Source to metadata and provenance, participants generally agreed on two fronts: the four freedoms (use, study, share, modify) need to be preserved, and data is the bottleneck.

The Raleigh Open Source AI Salon

Held in partnership with Duke University, the inaugural Open Source AI Salon in Raleigh followed the All Things Open 2026 conference, building on the expertise and community already gathered. The discussion focused on the practical application of the OSAID, defining the problem space, and taking stock of what the OSAID settles and what it leaves unresolved. A significant portion of the conversation covered the differences between open weights and Open Source, questioning if there is a need for a taxonomy of openness. In practice, this could preserve the need to have a single, binary definition of Open Source AI while allowing model developers and deployers to understand where they stand in relation to the OSAID and “walk up” the ladder to compliance.

Other topics included the need to provide machine-readable information about how a model is shared to promote decision-making at a scale and at a speed sufficient to meet governance demands. Participants noted the need for information on use restrictions, provenance, and specifics on what is shared and under what terms including, but not limited to, licenses.

The discussions also emphasized a need to work with other communities of openness like open science, open data, and open educational resources. Participants expressed that the traditional software community’s understanding of the four freedoms may not translate into other communities that AI now includes.

The Barcelona Open Source AI Salon

In Barcelona, our attention focused specifically on the data issue. The Salon was part of the annual Open Technology Research Symposium, led by Open Forum Europe in partnership with OSI and the Open Knowledge Foundation, and preceded Mozilla Festival. The goal was to better understand how to unravel the tensions between openness and transparency, privacy, provenance, and fairness. Participants made it clear that there is little agreement around what constitutes “open data,” and stressed pressures disincentivizing sharing despite willingness to share data. On one side, there is general consensus about the benefits of access to data for transparency, modifiability, and the study of models. At the same time, privacy concerns, resource constraints, lack of data provenance expertise, and the potential for mishandling data downstream are significant barriers to sharing data more openly.

Additionally, some echoed the need to provide guidance around the OSAID that equips a diverse set of stakeholders to apply the definition. While disagreements remained on whether gradients of openness are warranted or desirable, participants expressed a clear need to make distinguishing between Open Source AI and open weights models easier.

The stakes are high. In a space shaped by competing views on risk, economic value, and intense competition, the threat of open washing and eroding openness looms large.

Help Shape the Future of the OSAID

The next AI Salon will be held in November as part of the Digital Public Goods Alliance (DPGA) Annual Members Meeting in Paris. The DPGA Salon is only open to existing member organizations of the DPGA. We’re also hosting an AI Salon in January, during FOSDEM and encourage participants to RSVP through the OSI website.

We look forward to creating space for these crucial conversations. We’re seeking global partners to host AI Salons in 2027. We welcome opportunities to include a broader set of communities and stakeholders to help shape the conversation, learn from peers, surface real-world challenges, and shape OSI’s plan of action. Visit the AI Salon website to RSVP for an upcoming event or learn more about co-hosting.

Please read the briefs from the Raleigh and Barcelona Open Source AI Salons. Feedback is welcome!