institutional-leaders/osi-maffulli-binary-threshold-open-source-ai

Quote / Position

“While some people describe openness as a ‘spectrum,’ OSI sees Open Source as a binary gate threshold. Just as Linux and BSD represent different licensing models but both qualify as Open Source, AI must meet a minimum set of criteria to be considered genuinely Open Source. Many models claiming to be ‘Open Source’ actually impose usage limitations based on user numbers, commercial applications, or other criteria.”

Context

Maffulli’s framing is the load-bearing definitional substrate for distinguishing open-weights from open-source AI. Three structurally important moves:

  1. Binary threshold vs spectrum. OSI rejects the “more or less open” framing; open-source AI is either meets-OSAID or doesn’t. Same binary-threshold logic the OSI applies to free software.
  2. The four-element test. OSAID 1.0 requires: (i) model weights and parameters; (ii) code used to build and train the system; (iii) code for dataset creation; (iv) complete training data or detailed information enabling reproduction. By this test, Llama 3, Qwen, Mistral fine-tunes, and most “open” frontier models are not open-source — they are open-weights.
  3. The policy-definitional coupling. OSAID’s primary use case is policy: the EU AI Act’s exemptions, India’s national AI strategy, Canada’s procurement language all hinge on which definition a jurisdiction recognises. Mozilla’s “5 freedoms” framing builds on top of OSAID; OSAID’s “4 freedoms + data information” framing builds on top of the four-freedoms tradition.

This entry is the OSI-side complement to Mozilla’s Mozilla-krikorian-open-source-turning-point.md and the FAO-side fao-vincent-martin-open-source-innovation.md — together they triangulate the three institutional voices (Mozilla industry / OSI definitional / FAO multilateral) for any open-source AI in agrifood talk.

Application to agrifood

The agrifood translation: most open-source agrifood AI projects (FarmVibes.AI, FarmOS, GAIA, AgML, CGIAR tools) release code; only a subset release model weights + build code + dataset code + data-or-data-information together. The four-element test is rarely published against any agrifood deployment. Gap G-357 in the substrate scan tracks this as an outstanding OSAID-conformance scoring pass.

Source

Stefano Maffulli, Open Source Initiative Executive Director. Report from OSS EU 2025 and AI_dev: What’s next for OSAID. September 10, 2025. https://opensource.org/blog/report-from-oss-eu-2025-and-ai_dev-whats-next-for-osaid