Albert Lane · albertlane.net

Gotcha Wiki · Frontier AI training-data provenance · Nov 9, 2026

Fair Use, Unfair Library

The robot may read the book. The robot may not keep a pirated library of half a million of them. The difference was about $1.5 billion.

Satire & commentary. Lines marked Source summarize the linked documents published on this site (AI-assisted research reports by Albert Lane; check their cited references before relying on them). Lines marked Gotcha are opinion and satire, not statements of fact. No private individuals; no invented quotes. Corrections: lane.albert@pm.me.

What the record says

  • SourceIn Bartz v. Anthropic, authors alleged that pirated books from shadow libraries (LibGen, Books3, Pirate Library Mirror) were downloaded to train models. (Inquiry #93, Part III)
  • SourceThe court held that training itself was an “exceedingly transformative” fair use, but that acquiring and storing pirated books in a permanent central library was infringement. (Inquiry #93, Part III)
  • SourceThe case settled for $1.5 billion, about $3,000 per class work, with the pirated copies to be destroyed. The inquiry calls it the largest copyright class-action settlement in US history. (Inquiry #93, Part III)
  • SourceThe inquiry also lists class actions against Google over scraping copyrighted works for generative AI training (e.g. In re Google Generative AI Copyright Litigation). (Inquiry #93, Part II)
  • SourceIts recommendation: mandatory C2PA provenance manifests (signed assertions, SHA-256 hard-binding hashes, COSE signatures and a manifest store) so regulators can check where training data and outputs came from. (Inquiry #93, Recommendations)

Gotchu

GotchaReading Is Fine, Hoarding Is Not

The court’s line, roughly: learning from books may be fair use; keeping a stolen library is still a stolen library. A $1.5 billion library card.

GotchaHashes Cost Microseconds

Tamper-evident provenance (SHA-256, signed manifests, C2PA) has existed for years. The missing ingredient was never technology.

GotchaMeanwhile, on albertlane.net

Every published file carries a SHA-256, and every page carries a public provenance canary listed in a manifest. Provenance is not hard. It is a choice.

Sources on this site