OSI releases Open Source AI Definition 1.0

Background

Open source software has a long-accepted definition that anyone can hold a license up against. AI systems do not fit that frame easily. A model is not just source code: it also involves training data, the weights produced by training, and the code that trains and runs it, and different projects publish different parts of that stack under different terms. So the phrase open source AI came into wide use without an agreed boundary. The Open Source Initiative (OSI), the organization behind the Open Source Definition, set out to answer the question, describing the work as the result of multiple years of research, an international roadshow of workshops and a year-long co-design process ^[1].

The release

On 28 October 2024, at the All Things Open 2024 conference in Raleigh, North Carolina, OSI announced the release of version 1.0 of the Open Source AI Definition (OSAID) ^[1]. The document is meant to do roughly what the Open Source Definition does for software: state which freedoms a system must grant, and what information must be disclosed, before it can be called open source.

It was not written behind closed doors. Before release, OSI ran a year-long, global, community design process, in which people who build, fund and study AI argued in public about what the criteria should be ^[1].

Training data

The most contested part is data. Ayah Bdeir of Mozilla summarized the new requirement this way: an open source model must provide enough information about its training data that a skilled person can recreate a substantially equivalent system using the same or similar data ^[1]. The emphasis falls on what a skilled person can rebuild, rather than on releasing a dataset byte for byte; the test is whether the information is sufficient, not how much data is published.

Process and confidence

Carlo Piana, chair of the OSI board, said the co-design process had been well developed, thorough, inclusive and fair, and that the board was confident the definition meets the standards of the Open Source Definition and the Four Essential Freedoms ^[1]. Tying the new text to those older standards was a way of keeping open source meaning the same thing in software and in AI, rather than letting each field define it separately.

Impact

Some organizations said on release day that they would build on it. The Digital Public Goods Alliance said its secretariat would build on the definition as it updates the DPG Standard for AI ^[1]. OSI executive director Stefano Maffulli said arriving at version 1.0 had been a difficult journey, and described it as the starting point for continued effort to improve the definition over time ^[1].

Disagreement surfaced the same day. TechCrunch reported that the definition set the stage for a clash with companies such as Meta, whose models do not fit the rules ^[2].

The definition itself has no force of law and regulates no company. What it does is draw a checkable line under a much-used phrase: a model either meets these conditions or cannot call itself open source. For users it is a yardstick; for companies releasing models it is a list of things they have to answer publicly.