Toward Open-Source AI Systems as Digital Public Goods: Definitions, Hopes and Challenges | SpringerLink

Abstract "The recent advancements in artificial intelligence (AI), especially generative AI, have ignited a vivid debate about the potential of AI for positively impacting societies at large and attaining the United Nations Sustainable Development Goals (SDGs), but also about its risks and challenges. For the global majority, this debate is also linked to pressing questions about the distribution of (economic) power and influence, representation, diversity, access, and ownership. To address these issues, the democratization of artificial intelligence’s development, use, benefits, and governance is at the core of the debate, and with it, the importance of open-source AI to provide widespread access to these technologies, including the freedom of using, studying, sharing, and modifying the underlying AI models. This text examines the risks and opportunities associated with disseminating AI models and related components openly and unpacks the current debate about the definition of open-source AI. It finds that the term “open-source” AI is not well-defined and is used in misleading ways. It then explores the intersection of open-source AI systems and digital public goods (DPGs). Contrary to open-source software, DPGs have SDG relevance at their core, adding a layer of purpose to an otherwise technical and legal definition. In a nutshell, DPGs are tools that are open and accessible and can be adapted to local contexts to contribute to inclusive development. AI DPGs will be essential in addressing urgent global challenges and attaining the SDGs by 2030. However, apart from making AI technologies available under an open-source paradigm, additional aspects must be addressed to help democratize AI in all its facets. This includes looking beyond the hype of generative AI and systematically exploring other AI technologies for under-resourced contexts, leveraging AI research innovations to improve training efficiency and safety, creating open, high-value datasets for under-represented contexts, novel data governance approaches and licenses that benefit data subjects and communities over corporations, and policies and regulations addressing market concentration and enable access to essential resources. These efforts must be complemented by enhancing human, organizational, and societal capacity to shape AI development locally and support under-represented groups to partake in AI governance decisions globally."