Glonce

OSI's open source AI definition splits the FOSS community

policyThe Glonce editorsThe Register🔥52Glonce rating52Impact50Novelty35Relevance80Surprise25Players45Credibility75
OSI's open source AI definition splits the FOSS community

The AI industry keeps calling "open weight" model releases "open source," but the two are not the same thing, according to a Register column by Steven J. Vaughan-Nichols. Publishing a model's weights lets anyone download it, run it locally, fine-tune it on their own documents and skip a vendor's API. It does not reveal the training data, documentation or methods behind the model, so outsiders cannot check what it was trained on, whether copyrighted or private material was used, which languages or communities were underrepresented, whether benchmark data leaked into training, or what safety and alignment work was applied afterward.

The Open Source Initiative (OSI), which stewards the Open Source Definition, draws the same line: open weights, it says, are "the final weights and biases of a trained neural network," and on their own they expose only "a fraction of the information required for full accountability." James Landay, director of the Stanford Institute for Human-Centered AI (HAI), put it more bluntly: "Open weights are progress. You can download the model, run it on your own machine, keep it out of someone else's data pipeline. But you still can't see how the thing was built, what it was trained on, or why it behaves the way it does. That's not an open model. That's open distribution." He says a model earns the open source label only once its developers disclose the training data or provide a thoroughly documented, auditable account of it, and argues that right now almost every lab, American and Chinese alike, is answering only the easier question of whether the model can be run, not whether it can be trusted, improved and built on.

OSI's own Open Source AI Definition (OSAID 1.0), released in October 2024, requires model parameters including weights to be made available under OSI-approved terms, but it does not prescribe a specific legal mechanism for doing so, and critics say that omission is the core flaw. Luca Antiga, CTO of Lightning AI and a prominent PyTorch contributor, calls it "a gaping hole that will make licenses less effective in determining whether OSI-licensed AI systems can be adopted in real-world contexts." Bruce Perens, author of the original Open Source Definition, denounced OSAID back in 2024 and has since said, "It's not Open Source! ... It's unfortunate that the Open Source Initiative itself is now involved in Openwashing." Bradley Kuhn, policy fellow and hacker-in-residence at the Software Freedom Conservancy, and Red Hat Senior Commercial Counsel Richard Fontana have gone further and called for OSAID to be repealed, arguing that OSI "acted too quickly to impose an overly ambitious policy compromise on the community," that the resulting rift damaged OSI's reputation, authority and influence, and that OSAID "shows no signs of having any positive policy influence on machine learning practitioners, the FOSS community, or regulators." OSI has said since OSAID 1.0 shipped that the definition would keep evolving, but critics say the shortcomings they flagged remain unresolved.

A rival approach is now moving through OSI's own process. The Linux Foundation's Mike Dolan submitted the Open Model, Data, and Weights (OpenMDW) license, which has existed since 2025 and already lists contributors from Amazon, Meta, IBM, Microsoft and Nvidia. Instead of treating a model as one indivisible release, OpenMDW sets separate terms for a model's architecture, its training data and its weights, then brings all three under a single licensing agreement. The submission has run into objections on OSI's license review mailing list. Stefano Maffulli, OSI's former executive director, who led the organization while OSAID was being formulated, says he suspects the pushback is "tainted by an ideological bias" against big tech rather than a substantive objection to the license itself. Whether OSI will adopt OpenMDW remains, in the article's own words, "an open question."

Key facts

  • OSAID 1.0, released by the Open Source Initiative in October 2024, requires weights to be released under OSI-approved terms but sets no legal mechanism for doing so, which critics call its central flaw.
  • Bruce Perens (author of the original Open Source Definition), Bradley Kuhn of the Software Freedom Conservancy and Red Hat's Richard Fontana have all called for OSAID to be repealed, with Perens accusing OSI of "Openwashing."
  • The Linux Foundation's OpenMDW license, submitted by Mike Dolan and in existence since 2025 with contributors from Amazon, Meta, IBM, Microsoft and Nvidia, proposes separate terms for a model's architecture, training data and weights.
  • Stanford HAI director James Landay argues open weights only answer "can I run this?" while open source must answer "can I trust this, improve it, and build the next thing on top of it?"
  • OpenMDW's review inside OSI has hit objections that former OSI executive director Stefano Maffulli attributes to ideological bias against big tech, and OSI has not decided whether to adopt the license.

Why it matters

When a lab calls a model "open source" while withholding its training data, it lets people run and modify the model without letting anyone verify what shaped its behavior, check it for copyrighted or private material, or audit it for bias. That conflation blurs a distinction that used to be clear-cut for ordinary software. The piece argues that unless the industry settles what "open" actually requires for an AI model, not just whether the weights can be downloaded, the label risks becoming, in its own words, "another hollow tech marketing term."

Who it affects

AI labs that want the credibility of an "open source" label for a release that is really open-weight; developers and companies deciding how much to trust a model they cannot fully audit; the Open Source Initiative, whose OSAID definition is the closest thing to an official standard and is now facing open revolt from senior figures in the free-software movement; and the Linux Foundation, which is trying to get its OpenMDW license adopted as an alternative inside OSI's own review process.

How to use it

No product ships from this piece, so the practical takeaway is how to read a license claim. Check whether a release actually names its training data sources and provides an auditable account of how the model was built, not just whether the weights can be downloaded. Treat an "open source AI" claim as answering, per Landay, whether you can trust the model, improve it and build on it, not merely whether you can run it. And note that OSAID compliance alone does not guarantee that disclosure, since OSAID does not prescribe how weights, let alone training data, must be released.

How solid is it

This is a signed opinion and analysis column by Register columnist Steven J. Vaughan-Nichols, not a report on a single dated event. It synthesizes on-the-record statements from OSI itself, Stanford's James Landay, Lightning AI's Luca Antiga, original Open Source Definition author Bruce Perens, the Software Freedom Conservancy's Bradley Kuhn, Red Hat's Richard Fontana and OSI's former executive director Stefano Maffulli, with verbatim quotes attributed to each. It carries no independent statistics on how many models are open-weight versus open source and no timeline for when OSI might rule on OpenMDW.

Risks and caveats

The dispute is explicitly unresolved: OSI said OSAID would keep evolving after its October 2024 release but has not closed the gap critics describe, and whether OSI adopts OpenMDW is, in the article's own words, "an open question." As an opinion column, it presents named critics who want OSAID repealed alongside a defender, Maffulli, who suspects ideological bias behind the pushback on OpenMDW, without adjudicating between them on the repeal question, though the author does take a side on OpenMDW itself, calling its licensing approach "reasonable" and crediting its backers for at least trying, and it names no specific model or company caught mislabeling itself.

“Open weights are progress. You can download the model, run it on your own machine, keep it out of someone else's data pipeline. But you still can't see how the thing was built, what it was trained on, or why it behaves the way it does. That's not an open model. That's open distribution.”

— James Landay, director of the Stanford Institute for Human-Centered AI (HAI)