Article by Camille François et al: “…It has been challenging to define openness in the context of FMs, as definitions of “open source” software do not easily translate to AI systems. However, shared nomenclature is essential to developing shared understandings, norms, benchmarks, and best practices. By shared nomenclature, we mean establishing common terms across researchers, developers, policymakers, and civil society organizations working to unlock the benefits and mitigate the risks of FMs. We use AI systems here as shorthand for systems built with FMs, while recognizing the limitations and critiques of the term AI. This work is also necessary to illuminate the range of potential design choices around openness throughout the AI stack, and to ensure that this conversation moves beyond a narrow focus on model weights.
In this article, we survey existing approaches to defining openness in AI models and systems. We also propose a descriptive framework to evaluate how each component across the FM stack contributes to openness, enabling normative definitions of openness in AI. We intentionally do not present a definitive list of requirements for openness. This work builds on a February 2024 workshop convened by Mozilla and the Columbia Institute of Global Politics, which brought together more than 40 leading scholars and practitioners working on openness and AI. These individuals—spanning open source AI startups and companies, nonprofit AI labs, and civil society organizations—focused on exploring what open should mean in the current era of foundation models…(More)”.
