A New Approach to Open Source AI: A Framework for Openness
Mozilla and Columbia University convened experts to define 'open' in AI foundation models, culminating in a peer-reviewed framework published in *Communications of the ACM* that replaces binary labels with a layered, gradient approach to openness.
The debate over open versus closed AI has intensified in policy circles, prompting questions about who controls AI development and oversight. In 2024, Mozilla and the Columbia Institute of Global Politics gathered over 40 researchers, builders, and policymakers to establish a shared definition of openness for foundational AI models, moving beyond traditional software paradigms. This effort aimed to create a durable framework that could guide developers, regulators, and civil society in navigating the complexities of AI openness.
The resulting paper, published in *Communications of the ACM*, introduces a framework that dissects the AI stack into components like data, code, model weights, and documentation. It argues that openness is not a binary choice but a gradient, where different layers may vary in transparency. This approach allows for precise descriptions of what is shared and how, addressing the limitations of current policy debates that treat open source AI as a monolithic concept.
A key contribution of the framework is its emphasis on safety as a system-wide concern, not just a model-level attribute. It highlights that a model’s real-world risks depend on its deployment environment, safeguards, and governance structures, rather than isolated technical details. By doing so, the framework provides a vocabulary for evaluating tradeoffs between openness and safety on a case-by-case basis, rather than prescribing a one-size-fits-all solution.
The paper reflects contributions from a diverse coalition of academics, industry professionals, and civil society representatives, underscoring a collaborative effort to advance open AI systems. This milestone aligns with Mozilla’s mission to democratize AI, ensuring that development remains accessible and accountable to a broader community rather than concentrated in the hands of a few.