Responsible innovation in AI for science: copyright governance, attribution preservation, and public interest

Large language models now routinely draw on scholarly publications as training data, although neither EU nor US copyright law was drafted with this use in mind. This presentation, delivered at the Eu-SPRI Annual Conference 2026 (Valencia, 10–12 June 2026), reports a qualitative content analysis of 36 primary governance documents (legislation, court decisions, and soft-law instruments) examined through the Responsible Research and Innovation (RRI) framework. Using the AIRR dimensions (Anticipation, Inclusion, Reflexivity, Responsiveness) together with four cross-cutting categories tailored to the copyright–AI interface, I coded 659 segments and generated 1,080 category assignments (mean 1.64 codes per segment) across the EU, US, UK, and China for the period 2019 to May 2026. The dataset underpinning this analysis is openly available on Zenodo (https://doi.org/10.5281/zenodo.20555375).