On July 10, 2026, Hachette Book Group, Cengage Learning, Elsevier, and author Scott Turow filed a putative class action against Google in the U.S. District Court for the Southern District of New York, accusing the company of copying millions of copyrighted books and journal articles to train its Gemini large language models. The complaint, backed by the Association of American Publishers, alleges Google mined works it obtained for the limited purposes of Google Books, supplemented that trove with web scrapes of pirate sites and paywalled content, and then stripped copyright management information (CMI) to obscure the provenance of what it had taken — a claim that invokes the Digital Millennium Copyright Act, not just the Copyright Act's reproduction right.
Why Publishers Sued Now
The timing is deliberate. Cengage and Hachette had already moved to intervene in In re Google Generative AI Copyright Litigation, the consolidated Northern District of California case before Judge Eumi K. Lee that has been running since 2023 and currently centers on a push by authors and illustrators to certify a broad copyright-owner class. Rather than wait on that court's timeline, the publishers filed a freestanding SDNY suit to preserve claims — including the CMI-stripping allegation — that the AAP says fall outside the scope of the pending class (AAP press release). The complaint's opening line — that Google "abandoned its early motto of 'Don't be evil'" — signals a suit built for headlines as much as for discovery (MediaNama).
The Real Precedent Is Bartz v. Anthropic
The publishers' strongest card isn't a novel legal theory — it's a template that already worked against a different AI company. In Bartz v. Anthropic, Judge William Alsup ruled on June 23, 2025 that training an LLM on legally acquired books was "quintessentially transformative" fair use, but that Anthropic's use of pirated copies downloaded from shadow libraries was not — because unpaid copies "displaced demand for the authors' works, copy for copy" (ArentFox Schiff analysis). Anthropic settled in September 2025 for $1.5 billion — roughly $3,000 per work across an estimated 500,000 pirated titles — while conceding nothing on the fair-use-for-training question it had already won (Norton Rose Fulbright).
That split matters enormously for how the Google case should be read, and how courts should decide it. If Google trained Gemini on books it lawfully licensed or scanned under the terms that survived Authors Guild v. Google — the 2015 Second Circuit ruling that found Google Books' snippet-view scanning to be fair use — that conduct sits on defensible ground. But the new complaint's more serious charge is that Google supplemented that corpus with pirated web scrapes and then stripped CMI to hide it. If true, that isn't a fair-use gray area; it's the exact conduct Bartz punished, plus a DMCA violation Anthropic was never accused of. Courts, and Google's own counsel, should treat these as two separate questions rather than let a valid piracy claim be used to relitigate whether AI training itself is lawful.
Steelmanning the Publishers' Case
The publishers have a fair point that deserves more than a dismissive nod: a functioning market for AI training licenses now exists — OpenAI, Microsoft, and others have signed content deals with publishers including HarperCollins and Axel Springer — and a company the size of Google choosing pirated scrapes over that market, if proven, is a deliberate cost-avoidance decision, not an unavoidable technical necessity. Gemini's ability to generate textbook-style summaries that could substitute for the underlying scholarly works, as Cengage and Elsevier allege, also raises a legitimate market-substitution concern distinct from the training-input question. These are not frivolous arguments, and a court finding pirated acquisition and CMI stripping would be right to impose real damages.
Why Overcorrection Would Backfire
But the policy risk here is conflation. If courts or legislators respond to piracy-driven training-data scandals by treating all AI training on copyrighted text as presumptively infringing — rather than isolating the acquisition method as Bartz did — the effect will be to entrench the handful of firms that can afford nine-figure licensing deals while foreclosing the licensing market itself from developing further. A proportionate outcome treats lawful acquisition and piracy as legally distinct, exactly as Judge Alsup did: punish the shadow-library shortcuts and CMI stripping severely, while leaving room for AI developers who license or scan under existing fair-use doctrine to keep building. Congress and the Copyright Office, which has been studying generative AI training issues since its 2024–2025 report series, should resist the temptation to write new blanket restrictions before the courts finish sorting genuine piracy from lawful transformation — the SDNY and Northern District of California cases now running in parallel are close to doing exactly that.
What Happens Next
Google has not yet filed a response to the SDNY complaint. Expect an early motion practice fight over consolidation with the Northern District of California litigation, and expect Google to lean hard on the Authors Guild precedent for its Google Books-derived training data while trying to isolate — or dispute the scale of — the alleged pirate-scrape component. If the case tracks Bartz, the CMI-stripping and piracy allegations, not the training itself, will determine whether Google faces a nine- or ten-figure settlement.