Judge Pauses Anthropic’s $1.5B AI Book Deal, Demands Clarity

Read Articleadded Sep 9, 2025
Judge Pauses Anthropic’s $1.5B AI Book Deal, Demands Clarity

Judge William Alsup halted preliminary approval of Anthropic’s $1.5 billion authors’ copyright settlement, saying it is incomplete and lacks a clear claims and notice framework. He imposed strict opt-in requirements for all co-owners of a work, insisted on robust notice to class members, and tied attorneys’ fees to actual payouts. A final list of roughly 465,000 covered works is due Sept. 15 as parties rework the proposal.

Key Points

  • Judge Alsup postponed preliminary approval of Anthropic’s $1.5 billion settlement, calling it incomplete and demanding specifics on claims, notice, and a finalized list of works.
  • He ordered a strict opt-in regime: all owners of a work must opt in for coverage; any owner’s opt-out excludes that work, with disputes to be resolved in state court.
  • The court insisted on robust notice to class members to ensure meaningful participation and to shield Anthropic from future duplicative claims.
  • Alsup criticized the large team of settlement administrators and said add-on attorneys won’t be paid from settlement funds; fees will hinge on actual payouts to class members.
  • The deal, pegged at about $3,000 per book and covering roughly 465,000 works, is promoted as a model for other AI copyright cases, but industry groups argue the court’s expectations are impractical.

Sentiment

Mixed and contentious: many support the judge’s tougher scrutiny and agree the case centers on pirated acquisition, while there is sharp disagreement over whether LLM training is fair use and whether $3,000 per book is a fair outcome.

In Agreement

  • The litigation is fundamentally about illegal procurement (e.g., from LibGen/PLM), not the act of training; training on purchased-and-scanned books may qualify as fair use.
  • Judicial scrutiny is warranted: preliminary approval should wait for a clear claims process, robust notice, and a finalized list of works.
  • Per-work opt-in/opt-out and mechanisms to prevent duplicative future claims are appropriate to give defendants a clean release and protect class members.
  • Tying attorney fees to actual payouts and curbing payment to add-on lawyers helps ensure class members aren’t shortchanged.
  • For some authors, roughly $3,000 per book is an acceptable, pragmatic resolution for the piracy acquisition claim.

Opposed

  • Training itself should not be treated as fair use because LLMs copy the expression of works (via memorization or statistical encoding) and can regurgitate protected text.
  • $3,000 per book is too low and risks trading away meaningful rights for a one-time payout, potentially harming authors’ long-term livelihoods.
  • Downloading may not constitute infringement by the recipient (as opposed to the distributor), complicating the theory of liability for acquisition.
  • The envisioned claims and ownership verification process is unwieldy or unworkable, reflecting a judicial misunderstanding of publishing realities.
  • Some argue intellectual property/copyright regimes are net harmful, undermining the case for such settlements altogether.
Judge Pauses Anthropic’s $1.5B AI Book Deal, Demands Clarity