Artists, writers, and musicians are building their own curated datasets to train AI models, reclaiming control from tech giants that scrape the web for content without permission. This matters because major AI systems like GPT-4 and Stable Diffusion train on vast, poorly curated datasets—the LAION-5B image set, for example, includes work collected without creators’ consent. Current data commons remain inaccessible to independent organisations, and open-source governance models for creative data are fragmented. The project develops two frameworks: one helping artists structure and share datasets while keeping local control, and another setting rules for community-managed data commons. If successful, this could shift how creative industries—worth $750 billion in the UK and US alone—interact with AI. Instead of feeding corporate models, small, curator-made datasets would let creators steer outputs, reduce bias, and produce culturally representative content. The frameworks would also support low-resource “small tech” machine learning, cutting the environmental cost of training. Policy recommendations aim to influence copyright regulation, giving right holders fair compensation and accountability from platforms.
View original technical description
As AI increasingly integrates into creative practices, we need data governance approaches prioritising communities and addressing representation, ownership, and accountability issues. This project will develop a governance framework enabling UK and US creative communities to share and manage federated data commons for responsible machine learning applications. A federated data commons is a distributed network of datasets governed by creators and stakeholders, allowing them to maintain local control while participating in a shared ecosystem with common standards. Balancing individual agency with collective benefit enables creators to manage their work while fostering innovation through shared resources across visual arts, music, literature, and social science scholarship. While large AI companies suggest enormous datasets produce better results, models like GPT-4 and Stable Diffusion train on vast, poorly curated datasets, raising concerns about representation, copyright, ownership, and authorial integrity. The LAION-5B dataset includes images collected without the creators' permission. In contrast, small, curator-made datasets empower creators with enhanced creative agency through iterative model steering, bias reduction, and precise control over outputs while fostering culturally significant and representative content honouring source material. This project will conduct participatory co-design workshops to develop two complementary frameworks: (1) a Creator-Centric Federated Dataset Framework that empowers artists to structure and share their creative datasets while maintaining control, and (2) a Federated Data Commons Governance Framework that establishes the rules, permissions, and protocols for community-managed data commons. Both frameworks will specifically address the needs of creative communities working with open-source machine learning technologies. The deliverables will include these two frameworks in a comprehensive repository, with implementation case studies, and policy recommendations—all hosted online as an accessible resource for international collaboration. The UK and the USA generate $750 billion in media, culture, entertainment, and the arts, leading global discussions on creativity, critical issues of sustainability, and intellectual property. Collaborating with researchers from both countries will enhance creative communities and allow us to explore how cultural capital and technology differ, especially regarding intellectual property and machine learning. As both governments and the commercial sectors push for AI development, it is essential for the Creative Industries that creative works and community data benefit creators and technology firms. The UK Government's recent consultation on Copyright and AI emphasises the need for "supporting right holders' control over their content and fair compensation." Still, the ongoing debate illustrates the challenges in regulating generative AI. Furthermore, open-source governance models for creative data remain fragmented and underutilised. We aim to develop community-governed small-scale datasets for local training on low-resource models, empowering creators with greater control while reducing environmental impact. This research addresses three key areas: Access: Empowering creative communities with resources and governance structures Sustainability: Implementing low-resource "small tech" machine learning solutions Regulation: Addressing copyright, authorship, and intellectual property across UK/USA frameworks While existing online data commons provide some infrastructure, they remain inaccessible to independent organisations. Analysis of platforms like Google's datacommons.org, EU's Open Future, and Serpentine Galleries' Future Art Ecosystems (FAE) publications, particularly FAE4: Art x Public AI, reveals a need for improved accessibility, user relevance, and stronger governance regarding privacy and data rights. This research addresses public media challenges by implementing transparent data governance for content authenticity and platform accountability. This approach promotes the United Nations Sustainable Development through resource-efficient AI development while facilitating international partnerships to interrogate key differences in intellectual property around machine learning.
Plain English summaries and category classifications on this site are generated by AI and may not perfectly reflect the original research.
Is something wrong? Let us know