Unpacking Open Source AI: Toward a Framework for Openness in Foundation Models (cacm.acm.org)

A new framework for evaluating openness in foundation models (FMs) has been proposed to address the complexities surrounding the concept of "openness" in AI systems. As the implementation of the EU AI Act progresses, discussions about openness are becoming increasingly significant, given that perceptions vary widely—some view it as a security risk while others advocate for transparent development to enhance innovation and participation. The initiative emerged from a workshop organized by Mozilla and the Columbia Institute of Global Politics, gathering insights from over 40 experts. It emphasizes the necessity of a shared nomenclature among stakeholders, including researchers and policymakers, to define and understand how different levels of openness can impact AI’s benefits and risks. The framework suggests that openness must be viewed at both the model and system levels, recognizing the diverse forms it can take, such as access to data, code, and safety safeguards. It refrains from creating a definitive list of requirements but rather offers a flexible tool for evaluating how aspects of openness correspond to various goals, from promoting research to enhancing accountability. By clarifying these trade-offs, the framework aims to facilitate constructive debate on how to harness openness in AI systems effectively, ultimately supporting both innovation and security while ensuring that the discussion moves beyond binary judgments of "open" versus "closed."