Judge said Meta illegally used books to build its AI

Original Article ↗ Hacker News Discussion ↗

A US copyright lawsuit against Meta over training its Llama AI on pirated books is prompting broader questions about whether using copyrighted works for AI training is “fair use” and what kind of economic harm authors must prove. Commenters debate whether the real infringement lies in copying works into training sets or only when models reproduce or substitute for those works, and contrast this with past cases against individual file‑sharers. Many see the outcome as precedent‑setting for the future of large language models, compensation for creators, and US competitiveness against countries like China that train models on unrestricted data.

Status of the Case & Headline Dispute

  • Commenters stress the judge has not ruled; this was a pretrial hearing on summary judgment.
  • The original Wired title framed the case as about “the next Taylor Swift,” while the HN title implies a ruling that hasn’t happened.
  • The judge appears focused on economic harm: plaintiffs must show Meta’s models plausibly reduce sales or market value of their works, not just that training “feels wrong.”

Harm, Markets, and Substitution

  • Many see the plaintiffs’ focus on “lost book sales” as a weak theory of harm, especially given current AI fiction quality.
  • Others argue the real harm is long‑term erosion of creative livelihoods and diversion of readers’ attention toward platform content (feeds, chats, AI outputs), which is hard to quantify.
  • Debate over whether LLM summaries or “books in the style of X” materially substitute for reading or buying the originals. Analogies invoked: Reader’s Digest, Wikipedia, book‑summary sites.

Fair Use, Copying, and Training vs Outputs

  • One camp: training is a transformative, intermediate use akin to search indexing or humans learning; infringement, if any, arises only when outputs substantially reproduce a protected work.
  • Opposing camp: copying entire books (especially from pirate sources) to train a commercial model is itself infringement, regardless of what the model later emits. Napster/DVD‑copying comparisons recur.
  • Long subthread on whether LLMs are “just tools” or effectively lossy compression of the corpus, capable of verbatim regurgitation; this matters for whether model weights are “copies.”
  • Human‑learning analogies are attacked as legally irrelevant, since brains aren’t regulated as fixed media under copyright.

Piracy Sources and Double Standards

  • Strong criticism that Meta allegedly used LibGen/Books3–type corpora: “ordinary” downloaders were punished for similar behavior, yet big firms seek fair‑use shelter.
  • Others counter that the primary infringers are the uploaders/hosts, and that merely downloading (without seeding) was rarely prosecuted.

Policy, Power, and Global Competition

  • Some want drastic copyright reform or abolition; others are furious that corporations may get exceptions while individuals remain exposed.
  • Several predict stricter Western rules will advantage Chinese models trained on “everything,” pushing innovation offshore.

Proposed Directions & Workarounds

  • Suggestions include:
    • Training only on licensed, open, or public‑domain corpora; emerging “legal” foundation models.
    • Content‑ID‑like systems for AI outputs, with revenue sharing.
    • Per‑book or catalog licenses for training, priced by the market.
    • Tools to trace outputs back to training data and enforce attribution or opt‑out.