Deposition Confirms Stephen Roller Torrented Books3 While at Meta 2017-2023

In the joint status report filed 22 October in Kadrey v. Meta, plaintiffs report Roller’s 17 October deposition admissions. He confirmed acquiring Books3 via BitTorrent, acknowledged its 2020 creation date, and stated he held no other employment during the Meta period. Meta counters that the download may have been personal and that Roller could not confirm additional copies or their use for Llama training.

Court records already establish that training Llama 1 and 2 on Books3 constitutes fair use; the remaining issue is the provenance of the data. Roller also identified two other FAIR researchers who likely maintained separate copies and referenced an internal “ML Books” Google Drive folder plus a company book club that circulated shadow-library files. These details expand the scope beyond the single torrent.

The testimony reveals systemic handling of pirated corpora inside Meta rather than isolated researcher action. Deletion of the dataset more than two years after departure, explicitly tied to litigation, raises spoliation questions under the same federal rules applied to other AI training cases. Plaintiffs now seek compelled production of folder contents and book-club records.

Judge Chhabria’s next order is expected to address the new discovery requests; failure to produce will trigger sanctions motions already previewed in the status report.