The shift in leadership at the Open Source Initiative comes at an interim phase for the Open Source AI Definition, opening new lines of urgency concerning how we should enforce “open” in the age of foundation models. With its founder and executive director, Stefano Maffulli, stepping aside and the board naming Deborah Bryant as an interim leader, the organization is also grappling with whether it should harden OSAID’s guardrails—or recalibrate them—given increasing industry pressure.
A key leadership handoff marks a pivotal moment at OSI
Throughout its history, OSI has continued to promote and protect the open source definition using a variety of programs and initiatives. With Maffulli, the group went from a mainly volunteer empire to one of AI’s most well-known nonprofits and, with it, OSAID 1.0 — a first effort to apply open source stature on AI systems that had been developed over years offline before the word “open source” had taken any shape or meaning in reference to computational mechanisms.
The shift isn’t a retreat from AI. Continuity on licensing, policy, and standards has been a chief priority of the board as it has tasked Bryant — an open-source veteran researcher with public-sector and community credentials — to keep momentum going while a permanent successor is found.
What OSAID 1.0 actually does for defining open AI systems
OSAID 1.0 presumes that an “open” AI system allows users to use, study, modify, and share the system—with expectations that source includes not just code but also model artifacts and weights. On training data, it’s a pragmatic definition: meaningful transparency and documentation vs. strictly mandating publication of full datasets, accounting for legal/privacy/safety constraints.
That settlement garnered as broad—though cautious—an endorsement as he could have ever hoped for from the likes of Mozilla, SUSE, Bloomberg Engineering, and the Cloud Native Computing Foundation. Policy circles started to take notice, as well: Regulators grappling with AI disclosures in the European Union and standards bodies such as NIST have been searching for authoritative, community-driven signals that can help them separate marketing from substance.
The fault lines: data, weights, and real openness
Critics say the definition still leaves too much up in the air. Inverse AI Hackers touched base with legal expert voices from Red Hat and project leaders like Lightning AI to express concern that designating weights as the functional equivalent of “source” is necessary, or else the community will be open to a vulnerability where models are only nominally open but effectively unmodifiable. Others, such as Percona’s leadership, argue that a single definition is too blunt an instrument to suit the manifold contours of AI pipelines and call for a spectrum that differentiates between openness of training data, transparency in model construction, and redistribution rights.
The friction is not theoretical. The advent of services such as code assistants and agentic frameworks has obliterated the distinction between “source code” and a system’s behavior. Earlier platform shifts — such as mobile app stores that restricted users from running modified software — have painted a clear picture of how distribution controls can erode freedoms even when code is technically accessible. AI adds an additional layer: if you don’t have access to the weights, and a trustworthy data lineage, you could be granted modification rights that are largely illusory.
Open-washing and the long-term stakes for OSAID
And there is a thriving market for models advertised as “open” while being encumbered by use restrictions (called copyleft licenses in the open-source world), curtailing weights or limited disclosure about training data. High-profile releases by big tech companies have popularized the notion that open models can mean forbidding commercial reuse and restricting redistribution — practices squarely at odds with decades-old open source norms. Academic-led transparency evaluations have repeatedly shown that leading models share either spotty or little information on how data are collected and processed safely — again highlighting why we need a more lucid definition.
As the sector’s reference point — akin to open-source software before it — OSAID could help reduce open-washing, inform license authors, and offer policymakers a consistent touchstone. On the other hand, if the definition is seen as too much of a diktat for practical considerations or too lax about partial disclosures of otherwise valid efforts, there’s a danger that it becomes irrelevant and orgs can simply build their own bespoke marketing-friendly labeling.
What changes — and what to look out for next
The change in leadership could propel in one of three directions.
- Strengthen the existing definition and fast-track implementation guidelines: model cards with minimum disclosures; standard data provenance reports; clear weights availability requirements.
- Add a graded conformance level — e.g., strict “open AI” versus documented “disclosed AI” — to reflect realistic limitations without watering down what the core term stands for.
- Extend relationships with technical and policy fora and consortia by positioning OSAID within the work of the Linux Foundation AI & Data, OpenJS and Python communities, or standards bodies such as OECD or NIST.
One concrete step for the near term is to have a public conformance program: self-attestation + auditable criteria, and a process for the community to challenge when their “open” claim falls short. Even half-signaled enforcement can serve as a reset back to the kind of incentives that reward genuine openness while putting everybody on notice about what we call “gray area” licensing.
Finally, more than a job search brief will be passed on to the next executive director. They will inherit a branding war around the word “open,” a licensing frontier in which weights and data are as important as code, and a policy moment in which governments are deciding what transparency truly means. The result will decide whether OSAID sinks its fangs into the industry — or is relegated to being a footnote in a war of words.