The natlawreview.com piece, published without a byline on August 13, 2026, is advocacy dressed as explainer: it argues that if courts narrow fair use for AI training, startups get priced out, subscriptions get more expensive, and knowledge access fragments along paywall lines. Those are real risks. But the essay frames the fight as a binary — full fair use versus a licensing-only regime that forces retraining from scratch — when the actual litigation, per independent reporting, is splitting on a narrower and more consequential question: not whether training is transformative, but how the training data was acquired.
That distinction is where the money already is. Anthropic’s $1.5 billion settlement, reported by bgr.com, wasn’t a loss on fair use — a federal judge found Claude’s training sufficiently transformative in June 2025 — it was a loss on sourcing books from pirate repositories like LibGen. Judge Sidney Stein’s October 27, 2025 refusal to dismiss the Martin-led case against OpenAI turned on the same seam: he found a jury could view ChatGPT outputs as “substantially similar” to the originals, and let piracy claims proceed alongside training claims, according to bgr.com. The natlawreview essay barely engages this fork, even though it’s the one actually deciding cases.
Anthropic’s $1.5 billion settlement is a bet that the real exposure for AI developers was never training on copyrighted text — it was where the text came from.
The numbers accumulating outside this essay make the stakes look different than “innovation versus restriction.” Encyclopedia Britannica and Merriam-Webster’s March 13, 2026 suit alleges OpenAI copied nearly 100,000 articles and dictionary entries verbatim into ChatGPT outputs, per mlq.ai — a direct-substitution claim, not a training-corpus claim. Hachette and Elsevier’s July 2026 suit against Google, reported by Al Jazeera, cites internal Google documents warning that training-data exposure could reach $100 billion in fines. Statutory damages for willful infringement, per bgr.com, range from $750 to $150,000 per work — and these cases involve, in the industry’s own framing, “tens of millions” of works.
Watch for two things: whether Judge Stein’s fair-use ruling in the OpenAI authors’ case, expected after argument in early 2027, treats piracy-sourced training data differently from openly scraped web text, and whether Google’s Hachette suit forces disclosure of the internal risk assessments Al Jazeera says already exist. If courts keep separating “how was it obtained” from “is training transformative,” the fair-use doctrine the natlawreview essay is defending may survive largely intact — while the acquisition practices around it become the industry’s actual liability.
Plaintiffs are asking in many cases for court orders that would force developers to pull their existing gen AI models off the market entirely until they can be retrained using only licensed material.
