Revising research practices for singing data collection

As AI voice synthesis enables increasingly sophisticated vocal deepfakes and non-consensual voice cloning, the governance, licensing and access of singing datasets has become an urgent concern for data-contributors, who face significant harms from downstream and non-consensual usage of their singing data. Singing datasets are foundational to the development of high fidelity voice AI synthesis, yet current data collection practices pose challenges: data-contributors have an event-centric contribution to datasets which constrains the scope of the power they have in influencing dataset licensing and access decisions; and they face greater potential harms from non-consensual downstream use of their data. This is in contrast to data-collectors, whose involvement with the dataset directly concerns licensing and access decisions, and who do not risk the same harms as data-contributors. Current singing data collection practices suggest an imbalance in the power-to-interest stakes held by data-contributors relative to data-collectors. To investigate the power-to-interest differences between data-contributors and data-collectors, we apply the Ethically Aligned Stakeholder Elicitation (EASE) framework to three singing datasets, mapping the power-to-interest stakes of data-contributors and considering these stakes against their exposure to potential harms from downstream data usage. Our analysis consistently demonstrates that data-contributors occupy lower-power positions compared with data-collectors; they hold minimal decision-making authority over how their data may be accessed; and they face greater risk of harm from non-consensual downstream use of their data. A persistent temporal symmetry emerges: while data-contributors’ involvement in datasets is event-centric, their vulnerability to harm is temporally unbounded, extending beyond the moment of contribution as new AI capabilities develop. Drawing from a cross-section of data ethics, decolonial data perspectives and artificial intelligence legislation, we propose revised research guidelines which reposition data-contributors more centrally in dataset governance and licensing decisions. We contribute with three consideration factors to better consider data-contributors’ interests: stakeholder authority in dataset licensing, use and access decisions; the temporal scope of stakeholder involvement with the dataset; and vulnerability to harm from downstream data usage. Our findings highlight an urgent need for researchers to consider what licensing-related steps could be taken to protect the personality or identity rights of human participants within an AI climate in which singing data is increasingly treated as a freely available commodity for downstream use.