The Hidden Copyright Risk in AI Training

Copyright analysis of AI training often begins too late. By the time a model generates an output, potentially significant decisions about acquiring, copying, transferring, and retaining copyrighted material have already been made.I explored these issues in my keynote, “The Hidden Copyright Risk in AI Training,” at the IFRRO World Congress 2026 in Bogotá, Colombia. I proposed a four-stage framework, Source, Share, Store, and Stack, for examining the copyright implications of AI data supply chains.

The central argument is that reliable provenance is essential to understanding what uses are authorized, what exceptions may apply, and how licensing and remuneration can operate effectively. The challenge is to identify the relevant acts and apply copyright principles more precisely.

You can find an edited and abridged version of my keynote below.

Thank you to IFRRO, particularly Anita Huss-Ekerhult, Antonios Baris, and Veraliah Bueno, and to our Colombian hosts at CDR for the invitation and hospitality.

I am also grateful to my research collaborators Saige Lim, Jeongwon Jo, and Jihoon Choi, to LG AI Research for supporting the research, and to Dean Danielle Conway for her encouragement and support.