Open source licenses have quietly become the default for AI model releases, and that’s a problem. Developers routinely slap Apache 2.0 or Creative Commons onto model weights, treating machine learning artifacts as if they were source code. But models aren’t code — and the mismatch is creating what researchers call socio-ethical distortions in how these systems get used.
The core issue is Freedom 0: the open source principle that anyone can use software for any purpose. That works for code. For a model capable of generating disinformation at scale or being repurposed for military targeting, it’s a different story. As Widder et al. (2022) and Moran (2021) have documented, licenses designed for software don’t account for the technical nature and capabilities of ML models. Anduril Industries’ general counsel Babak Siavoshy put the dilemma bluntly when asking what license terms should govern a computer-vision algorithm adapted for military threat evaluation — neither commercial software licenses nor standard government data rights clauses adequately answer that question.
Enter OpenRAIL — Open & Responsible AI Licenses. The approach, championed by the RAIL Initiative and supported by Hugging Face, keeps the good parts of open source: royalty-free access, flexible downstream use, redistribution rights. But it adds something new: use-based restrictions. If a model is released under an OpenRAIL license, you can use it freely — except in specifically identified critical scenarios where the developers have drawn a line based on evidence of potential harm.
The bet is that openness and responsibility aren’t mutually exclusive. Open source brought enormous benefits to AI development, but the field has matured past the point where unrestricted access is always the right call. OpenRAIL represents a middle path: not closed systems, which bury problems under corporate opacity, but not naive openness either. The question is whether the broader ML community — which has internalized open source norms for decades — will accept restrictions that many see as a betrayal of the movement’s founding principles. If OpenRAIL gains traction, it could redefine what “open” means in AI. If it doesn’t, the industry may need to find another way to reconcile access with accountability.
