AI copyright news, graded by source

Meta's interrogatory answers: it downloaded LibGen directly as well as by torrent, and later bought and scanned the plaintiffs' books to train its Muse models

Interrogatory answers Meta served on September 4, 2026, signed by its lawyers and not verified, were filed as exhibits by the plaintiffs. Meta says that besides torrenting, it downloaded LibGen's fiction and science-and-technology collections directly, beginning around spring 2023. It says Books3 data, obtained before 2023, was used to train Llama 1 through 4, the LibGen collections Llama 3 and 4, and Z-Library data from Anna's Archive Llama 4. A book listed under several sources was downloaded separately each time. Meta lists the plaintiffs' books it found in each source, including those in one former employee's partial LibGen torrent. For a substantial majority of the plaintiffs' books used in training, lines with terms such as "Copyright" and "ISBN" were left out of the data. From mid-2025 Meta bought physical books, scanned them and used them to train its Muse models; its lists of the plaintiffs' books used that way run to about 2,300 ISBNs. Meta says its records do not show whether any plaintiff's work was uploaded to others during torrenting, and leaves that to expert analysis. It argues each work supports one statutory damages award however many copies were made.