Google is back in court over AI training data. A class action filed by publishers Hachette, Cengage, and Elsevier, alongside author Scott Turow and the group S.C.R.I.B.E., accuses the company of training Gemini on copyrighted books it never had permission to use, according to TechCrunch AI. The plaintiffs go one step further than the usual complaint: they allege Google removed or altered copyright information on those works to “conceal… that its Gemini Models were trained on stolen materials.”
Google did not immediately respond to a request for comment, TechCrunch AI reports.
Why this filing is different
Most AI copyright suits follow the same script. Company scrapes the web, rights holders sue, company says fair use. This one has a wrinkle that matters.
Publishers and authors handed Google their books voluntarily. For years. That was the whole point of Google Books: you give Google the text, Google makes it searchable, readers get snippets and bibliographic data, nobody gets the full book for free. It was a scope-limited deal, and both sides knew the scope.
The lawsuit claims Google took those same copies, plus books uploaded to the Google Play store, and fed them to Gemini. “Google illegally copied works from all these scope-limited programs for AI training, knowing it lacked authorization to do so,” the complaint reads, as detailed in TechCrunch AI.
That reframes the argument. It’s not scraping a public web page. It’s using material you were trusted with, for a purpose nobody agreed to.
Tactical points
- The venue changed. This one landed in the U.S. District Court for the Southern District of New York. Two earlier California decisions went the AI companies’ way, ruling that training on copyrighted work counts as fair use. A different bench now gets a swing.
- There’s an internal document in play. Plaintiffs cite a Google memo allegedly warning that using copyrighted books for AI training could be “highly problematic for Google” and might trigger “$10Bs-$100Bs in potential fines.”
- Anthropic already paid. The company was fined $1.5 billion for pirating training material, the largest payout in U.S. copyright history. Roughly half a million writers qualified for at least $3,000 each.
- Many authors refused the money. They opted out of that settlement specifically to keep suing. The war chest is still forming.
What stands out
The allegation about stripped copyright metadata is the sharpest blade here. Fair use is a defense about purpose. Removing copyright management information is a separate violation under the DMCA, and it carries its own statutory damages. If that claim sticks, the fair use fight becomes only half the battle.
The internal memo quote is the other problem. Fair use arguments lean on good faith. A document allegedly pricing out the fine exposure in tens of billions does not read like good faith. It reads like a risk calculation.
Context: the law is older than the internet
U.S. copyright law hasn’t been meaningfully updated since before the web existed. Judges are applying a 1976 framework to systems that ingest millions of books in a weekend. The two California rulings favored AI companies, but as TechCrunch AI notes, the conflict is too nuanced for those decisions to set an unarguable precedent.
Translation: nobody has won yet. Everyone is still probing the line.
What to expect
If you build on or with foundation models, watch three things:
- Provenance is becoming a product requirement. Enterprise buyers are going to start asking where the training data came from, and “we scraped it” won’t survive procurement.
- Licensing deals will accelerate. Every ruling that leans toward rights holders makes a signed contract cheaper than a lawsuit. Expect more publisher partnerships announced quietly this year.
- Metadata handling matters. The DMCA angle in this case is a template. Other plaintiffs will copy it.
The SDNY judge now holds a decision that could split the circuit and push this toward the Supreme Court. That’s the real stake. Not one company’s bill, but whether training on copyrighted work stays legal by default.
Full details are at the original source.