The original deal
On September 27, 1983, Richard Stallman, then at MIT's Artificial Intelligence Lab, announced a plan to write GNU, a complete Unix-compatible operating system that would be given away free. 3
The movement that grew from it defines freedom as four abilities: run a program, study how it works, change it, and share it with or without changes. Studying or changing a program requires its source code, the human-readable instructions it is built from. "Free" refers to those abilities, not to price; selling copies and services is allowed. 4
Why AI complicates it
An AI model has more parts than a program. The code that trains it and runs it is one part. The trained numbers that determine its answers, called weights, are another. The collection of examples it learned from is a third. A person can download the weights and still be unable to see what went into them, or to afford the computing needed to retrain them. The OECD's 2025 primer on AI openness says openness is a spectrum across these components, not a switch, and that licenses designed for source code do not map neatly onto weights. 5
Meta's Llama models show the gap. The weights can be downloaded, and the models have been described as open source. But the OECD notes that Meta drew criticism, including from the Open Source Initiative (OSI), because the license places conditions on use and distribution. 5
The split
| Free Software Foundation | Open Source Initiative | |
|---|---|---|
| Code | Must give users the four freedoms | Complete code to train and run the system, under OSI-approved licenses |
| Training data | Must be free too; the FSF says it cannot call a machine-learning application free otherwise | Not required; a detailed description of the data, its sources and how it was filtered is |
| Stage | Statement from October 2024, with its definition still being drafted | Open Source AI Definition, version 1.0 |
The Free Software Foundation, which Stallman founded, said in October 2024 that an application is not free unless its training data and the scripts for processing it respect users' freedoms as well. 1 The Open Source Initiative's definition, which some in the debate use as the reference point, asks for the freedoms to use, study, modify and share a system, and requires the code and the weights. It stops short of the data itself: companies must describe what they trained on, including data they cannot share, but need not hand it over. 25
The difference matters because a model's behavior comes largely from its data. Under the OSI definition, an outsider can see what the data was and where it came from, but may not be able to rebuild the model. Under the FSF's stricter test, a model whose training data is not released would not count. The question is who may call a system "open", and whether the term is a promise that users can verify.
Others in the argument
Other groups say the code-and-weights frame is too narrow. Creative Commons argued in March 2026 that openness "cannot stop at model weights", and called for open datasets and shared infrastructure, including for languages and communities poorly represented online. It favors stewardship of data over both exploitation and enclosure. 6 Mozilla's January 2026 strategy frames the stakes as owning versus renting intelligence, and puts the weak point on usability: closed systems win today on developer experience, it says, not on values. Its plan covers tools, licensed and clearly sourced data, and partnerships. 7
These are advocacy positions. None of them shows that a particular model is safer or fairer, and none addresses the creators whose work went into the training data and who may want a say in how it is used.
For anyone choosing an AI tool, the practical test is to ask which of the parts you could actually inspect, change and run without the vendor's permission: the code, the weights, the data, and the computers to retrain them.
References
- FSF is working on freedom in machine learning applications. (n.d.). Retrieved November 6, 2026, from https://www.fsf.org/news/fsf-is-working-on-freedom-in-machine-learning-applications
- The Open Source AI Definition. (n.d.). Retrieved November 6, 2026, from https://opensource.org/ai/open-source-ai-definition
- Initial Announcement. (n.d.). Retrieved November 6, 2026, from https://www.gnu.org/gnu/initial-announcement.html
- What is free software? (n.d.). Retrieved November 6, 2026, from https://www.gnu.org/philosophy/free-sw.html
- AI openness. (n.d.). Retrieved November 6, 2026, from https://www.oecd.org/content/dam/oecd/en/publications/reports/2025/08/ai-openness_958d292b/02f73362-en.pdf
- AI’s infrastructure era. (n.d.). Retrieved November 6, 2026, from https://creativecommons.org/2026/03/04/ais-infrastructure-era/
- Owners, not renters: Mozilla’s open source AI strategy. (n.d.). Retrieved November 6, 2026, from https://blog.mozilla.org/en/mozilla/mozilla-open-source-ai-strategy/