open training data requirement

OSAID Definition vs Open Training-Data Requirement

⚠️ Sample. An AI-generated demo of LLM Wiki Newsroom — the "open source AI" topic is just the example corpus.

Opposing Positions

When the OpenSourceInitiative released the Open Source AI Definition (OSAID) 1.0 on 2024-10-28, it made one choice that split the open-source community: a system can be "open source AI" without releasing its raw TrainingData, provided it discloses enough "data information" to recreate a substantially equivalent system. The dispute is not about whether transparency matters but about where the line of "open" falls — and whether a definition that omits open data dilutes a term the community spent 26 years building.

The definition camp — the OpenSourceInitiative, joined by Mozilla — holds that a clear, binary standard is more useful now than a maximalist one that almost nothing could meet. OSI argues that full-data mandates would relegate open-source AI to a niche, because some data (medical records, for example) cannot be legally shared; data information plus code plus parameters is, on this view, the workable definition that lets regulators and civil society distinguish genuine openness from OpenWashing. Mozilla endorses it as "an important step forward" while conceding the data treatment is imperfect and will need refinement.

The open-data camp — the FreeSoftwareFoundation and critics including OSI co-founder Bruce Perens, the Software Freedom Conservancy's Bradley Kuhn, RedMonk's Stephen O'Grady, and OpenUK's Amanda Brock — holds that training data is effectively the source code of a model, so a system is not open unless the data and its processing scripts respect the four freedoms. On this reading the OSAID is "less than Open Source" and erodes the meaning of the term; some warn it threatens the future of "open source" itself. The FSF acknowledges narrow moral exceptions (such as personal data) but concludes these merely yield non-free applications whose use may be ethically excusable.

Representative Evidence

• case-against-osaid — David Cassel rounds up the criticism; the FreeSoftwareFoundation holds a machine-learning application is not free unless its TrainingData and processing scripts respect the four freedoms, and Bradley Kuhn announced a campaign to run for the OSI board on a platform to repeal the OSAID.
• osi-open-source-ai-definition — the OpenSourceInitiative's canonical statement defines openness by four freedoms plus access to data information, code, and parameters, and names combating OpenWashing as the motivation: "Companies are calling AI systems 'Open Source' even though their licenses contain restrictions."
• mozilla-celebrates-osaid — Mozilla (Ayah Bdeir, Imo Udom, Nik Marda) endorses the definition as "an important step forward" and argues a binary standard gives developers, advocates, and regulators needed precision, while conceding the data-information treatment needs refinement.
• open-source-ai-models-how-open — a Hunton legal primer notes OSI's strict standard requires data information, code, and parameters as the preferred form for modification, framing why the data question is the dividing line between full open source and lesser categories.

Derived Tensions & Generational Readings

A procedural grievance runs alongside the substance: the OSAID was approved by OSI's 10-person board rather than a full membership vote, which critics cite as evidence the standard lacks community mandate. This turns a definitional debate into a governance one — Kuhn's board-repeal campaign is the clearest expression of the move from arguing the merits to contesting the institution.

The two camps also read the history of "open source" differently. Perens, an OSI co-founder, argues the 26-year-old Open Source Definition can simply be applied to AI, treating the OSAID as an unnecessary and weaker fork. The definition camp reads the same history as proof that new components (data, weights) demand a new, adapted standard. The disagreement is partly generational within the movement: founders defending an inherited bright line against stewards adapting it to a new technical reality.

A live qualifier on both sides is the legal-data problem. OSI's medical-AI exception is real, and the FSF concedes it; the open question is whether that exception justifies a general data-information standard or only a narrow carve-out. Until that is settled, each camp can claim the hard cases support its position.

Interpretive Direction

In the short term, the practical consequence is two coexisting standards rather than one: the OSAID as the de facto reference, and a stricter open-data criterion under development by the FreeSoftwareFoundation. On the narrow question of adoption, the OSAID has the momentum — at least 20 endorsing organizations and a validated-model list — but on the question of legitimacy, the board-vote grievance and the repeal campaign leave it contested. The fact that fully open-data models such as OLMo and Pythia already exist weakens the "open data is unworkable" defense without settling whether it should be mandatory. The point to monitor is the OSI board election: if the open-data platform gains ground there, the definition itself, not just its reception, is back in play.