AI copyright news, graded by source

Anthropic says it halted a 2022 training run to remove LibGen and PiLiMi books, and asks a court to declare its copying of an opt-out author's book fair use

Answering Daniel Gilbert's suit, Anthropic admits it downloaded books from LibGen in 2021 and from PiLiMi in 2022 and says at least some portion of his book was in the versions it downloaded. In its counterclaim, Anthropic says it used those books in research training runs from 2021 to late 2022, then halted a run in October 2022, removed the LibGen and PiLiMi data and started over at a cost of about $5 million. It says the resulting model became Claude 1 and that no publicly released model was trained on LibGen-sourced text. It says it later bought a print copy of the book in a bulk purchase for training, and that it still uses LibGen-sourced text as reference data in a filter that blocks outputs matching existing text. Anthropic calls the Bartz court's view that it kept a general-purpose library "erroneous" and seeks a declaration of non-infringement. Gilbert moved on Aug. 27 to dismiss the counterclaim in part; Anthropic opposed on Sept. 10. These are Anthropic's own statements, not court findings.