Copyrighted book training legality complicated—2026 lawsuits reshape landscape

If you've been waiting for 2026 to hand down one clean answer on whether AI companies can legally train on pirated books, stop waiting. This is the year the courts split instead — and that split is exactly why new lawsuits keep landing.

In Bartz v. Anthropic, the court ruled that training an AI on copyrighted books is transformative fair use, but that storing pirated copies to do it is a separate, actionable wrong. That distinction is why Anthropic settled for $1.5 billion, working out to roughly $3,000 per stolen work. Weeks later, in Kadrey v. Meta, a different judge reached the opposite conclusion on the piracy question: Meta won partial dismissal, with the court holding that training counts as fair use 'regardless of' how the books were acquired. Same underlying conduct, opposite outcomes.

That contradiction is why the newest lawsuit matters. In May, five major publishers — Elsevier, Cengage, Hachette, Macmillan, and McGraw Hill — joined bestselling author Scott Turow to sue Meta, and Mark Zuckerberg personally. The suit alleges:
- Meta torrented more than 267 terabytes of copyrighted books and journal articles from pirate sites
- The company masked its own IP addresses while downloading
- Copyright notices were stripped to conceal where the material came from
- Meta walked away from legitimate licensing talks on Zuckerberg's own direction

It's a direct attempt to force the piracy question back in front of a court that might rule the other way.

What this means for you: the legal ground under every author lawsuit right now depends on which circuit, which judge, and which fact pattern you land in. A book pulled off a pirate site and a book obtained through a licensing deal are being treated as legally different animals — exactly the argument your own case, if you ever have one, would turn on.

Watch for: the Kadrey plaintiffs are pushing to amend their case around Meta's alleged torrenting activity. If that succeeds, it could narrow the gap between the two rulings and give authors a more consistent standard to stand on.

Update, 23 September 2026. This piece cited roughly $3,000 per work, derived from the $1.5 billion fund divided across about 500,000 titles. The figure for the first round of payments is lower: about $2,203.56 per work, because 440,490 of the 482,460 eligible works were claimed — a 91.3% claim rate, per the Authors Guild's own update — spreading the same pool further than first modelled. The $3,000 arithmetic described the gross fund; $2,203.56 is what is actually being distributed. Today's story has the detail.