Open or closed AI isn’t a fringe question anymore; it’s a live policy debate, with real consequences for who gets to build, audit, and compete in the AI ecosystem. But as that debate plays out in Washington, Brussels, and beyond, the harder questions are still very much up for discussion. What does meaningful openness actually require? Where should the guardrails sit? And who gets a say in setting them?
That’s exactly the gap a group of us set out to close two years ago.
Back in 2024, Mozilla and the Columbia Institute of Global Politics brought together more than 40 researchers, builders, and policy experts for the Columbia Convening on Openness and AI. The goal was simple to state, but hard to do: figure out what “open” actually means once you’re talking about foundational models rather than traditional software, and build a shared vocabulary that researchers, developers, policymakers, and civil society could all work from.
The convening was the start of a years-long effort to turn a room full of perspectives on openness into something durable and useful for developers, builders, and policymakers alike. Recently, that effort reached a new milestone: Communications of the ACM published “Unpacking Open Source AI: Toward a Framework for Openness in Foundation Models,” which I co-authored alongside a host of other allies spanning academia, industry, and civil society.
The paper outlines a framework that does something the open source AI debate badly needs. It stops treating “open” as a single switch you flip on or off. Instead, it breaks the AI stack down into its component parts – data, code, model weights, documentation, and more – and looks at how openness at each layer shapes who actually gets to inspect, use, and build on a system. That kind of specificity matters. A lot of the current policy debate treats open source AI as a monolith, when in reality the details of what gets shared (and how) are what determine whether openness delivers on its promise of safety, competition, and accountability.
A few of the framework’s ideas are worth pulling out. First, it treats openness as a gradient rather than a binary. A system can be open on weights while remaining closed on training data, or open on code while offering little documentation on how it was evaluated. The framework offers language to describe that mix precisely, instead of lumping it under one vague label. Second, it argues that safety can’t be assessed at the model level alone: what matters just as much is the surrounding system – the deployment environment, the safeguards, the moderation layers, and governance structures. That’s because a model’s real-world risk profile is shaped by all of it together, not by the weights in isolation. Critically, the framework deliberately stops short of prescribing one “correct” level of openness for every system. Instead, it gives developers, researchers, and regulators a common vocabulary to make and evaluate those tradeoffs on a case-by-case basis.
It’s worth pausing on who’s behind this work. Scroll through the author list, and you’ll find a genuine cross-section of the individuals and organizations contributing to moving the development of open systems forward. It’s a collection of people who’ve spent years arguing, in good faith and often across real disagreements, that open approaches can be a foundation for safer, more democratic, more accountable AI. That’s the same spirit that drove the original convening, and it’s what makes this paper a genuinely important milestone for our mission to make AI systems accessible to the many, rather than the few.
You can read the full paper here.