OSI Opens Global Debate on the Future of Open Source AI

Why OSI is revisiting the OSAID

The salons come two years after the release of the Open Source AI Definition 1.0, which was announced at All Things Open in Raleigh, North Carolina, in October 2024. OSI said it created version 1.0 knowing that AI technology and the ways models are developed and shared would continue to change.

Applying the OSAID has proved more difficult than applying the Open Source Definition for software. Concepts that work for software do not always translate cleanly to AI models, particularly around weights and training data. The OSAID is intended to help people who use, study, produce, procure, share, modify or regulate AI understand what it means for a model to carry the Open Source label.

OSI said the growing popularity of open-weight models, new regulations affecting Open Source and changing approaches to sharing models have created questions about how the definition is understood and applied.

More models, but continuing concerns over openness

The Open Source Initiative pointed to the growth of Hugging Face as one reason for renewed community discussion. The repository hosted approximately one million AI models in 2024, according to OSI, while that number has now surpassed three million.

OSI’s research into models shared on Hugging Face found greater use of OSI-approved licenses, particularly MIT and Apache 2.0. At the same time, AI-specific licenses are emerging. ModelGo and OpenMDW are awaiting OSI approval.

The organization also identified models that combine Open Source terms for code with restrictions on model weights through acceptable use policies or terms of use agreements. An OSI-approved license for code does not automatically mean that an AI model meets the OSAID, according to OSI.

Training data remains a major issue. The OSAID requires developers to provide detailed information about the data used to train an AI system, but OSI said relatively few publicly available models appear to meet those data information requirements.

Lessons from Raleigh and Barcelona

The Raleigh salon was held in partnership with Duke University after the All Things Open 2026 conference. Participants discussed the practical application of the OSAID, the distinction between open weights and Open Source, and whether a taxonomy of openness could help developers and deployers understand their position relative to the definition.

Participants also discussed machine-readable information about how models are shared, including use restrictions, provenance and what is shared under which terms. The discussion highlighted the need to work with communities involved in open science, open data and open educational resources.

In Barcelona, the discussion focused on data. The event formed 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.

Participants examined tensions between openness and transparency, privacy, provenance and fairness. They identified privacy concerns, resource constraints, limited provenance expertise and the risk of downstream mishandling as barriers to more open data sharing. Across both events, participants generally agreed that the four freedoms, use, study, share and modify, need to be preserved, while data remains the bottleneck.

More salons planned for 2027

OSI said the next AI Salon will be held in November as part of the Digital Public Goods Alliance Annual Members Meeting in Paris. That event is open only to existing member organizations of the Digital Public Goods Alliance.

The organization is also hosting an AI Salon in January during FOSDEM and is seeking global partners to host additional salons in 2027. These events are intended to help communities surface real-world challenges, clarify how Open Source AI should be distinguished from open weights models and inform OSI’s plan of action.

For developers, deployers, regulators and other users, the discussions could provide greater clarity about licensing, model weights, provenance and training-data disclosures. OSI has not announced a revised OSAID, but said the salons and its research and policy work will help determine whether and how the definition should evolve.