Summary: On July 20, Judge Araceli Martinez-Olguin approved Anthropic’s $1.5 billion authors’ copyright settlement, covering about 500,000 works at roughly $3,000 each. The court cut class-counsel fees to 6.8%, while the article stresses that settlement leaves no binding appellate precedent on AI-training copyright limits. It advises developers to treat provenance and licensing as due-diligence issues.
On July 20, Judge Araceli Martinez-Olguin of the Northern District of California entered final approval, closing out Anthropic’s $1.5 billion settlement of the authors’ copyright class action. Nothing in US copyright history has settled for more: about 500,000 works at roughly $3,000 apiece, paid out to the authors and publishers who held the rights. TechCrunch struck the honest note — a milestone it is, but for most writers it will not register as a win.
Looking Back: The Line One Ruling Drew
Rewind to how Bartz v. Anthropic unfolded. In his 2025 ruling, presiding judge William Alsup set down a boundary that is still quoted today: putting legally acquired books to work in AI training is fair use, whereas pulling millions of books off pirate sites such as Library Genesis is not. Anthropic settled, and that kept the piracy question away from a jury. The ruling carried a pragmatic shape — it called the training act transformative while refusing to launder how the material was acquired — and it left the expensive question of what a licensing market for books should look like to the market itself. Alsup has since retired: his signature is on the preliminary approval, and the final one went to Martinez-Olguin — while Reddit, in an unrelated suit, has begun reusing his line almost word for word.
What the Order Changed: Fees Cut to 6.8%
The final approval order is worth reading in its own right. Beyond clearing the payout, the court document says outright that class counsel fees were trimmed to 6.8% — a reduction that shifts how much of the fund reaches the plaintiffs’ lawyers rather than the class itself. For authors and publishers, the roughly $3,000-per-work payments are about to start flowing, and since each work’s compensation is shared among everyone holding rights to it, individual checks will land all over the map. The mood TechCrunch reported is blunt: the money is coming, but “many writers don’t see it as a true win” — per-author amounts are modest, and because the act of training was itself deemed fair use, the central dispute closed with neither a winner nor a loser.
A Settlement Is Not a Verdict
What the $1.5 billion really bought is a deferral of uncertainty, not an ending. Because the case settled, no appellate court will hand down binding precedent on the copyright limits of AI training, and the suits against Google, Meta, Midjourney, and OpenAI remain queued. Reddit’s state-court suit against Anthropic is the fastest glimpse of what comes next. Reddit brought it in June 2025 in San Francisco Superior Court on five state-law claims — breach of contract, unjust enrichment, trespass to chattels, tortious interference, and unfair competition — and alleges that Anthropic’s bots hit or tried to hit the platform more than 100,000 times even after Reddit put it on the crawler block list. In its July 17 opposition brief, Reddit called Anthropic a “freeriding pirate” and invoked the Bartz logic: how data is acquired and what you do with it are two separate questions, so a pirated source does not become lawful merely because the downstream use might be fair. Reddit also reached for a bookstore analogy — a bookstore can stop thieves from breaking in and photocopying its entire inventory, and by the same token a platform can block crawlers from illicitly scraping its content. Reddit adds that Anthropic scraped the User Agreement itself, which Reddit says put the company on actual notice of its terms. Judge Joseph M. Quinn has set a hearing for August 12.
What It Means for Developers and Startups
Three actionable takeaways. First, $3,000 per book is now a public price signal: book training data is no longer a free lunch from a gray zone, and your budget model should carry a line item for data licensing. Second, acquisition channel matters as much as use — provenance audits for training corpora are moving from nice-to-have to standard due diligence; ask where the data came from before you buy a company or sign a data deal, and keep records that can survive discovery. Third, the precedent vacuum is both opportunity and risk: with no appellate boundary fixed yet, the safest engineering default is to follow the Bartz line — lawful acquisition, fair use. The cut to 6.8% also says something about settlement dynamics: judges are watching how these funds get divided, not just whether they get paid, and future defendants should not expect approval to be automatic.