Judge Stephanos Bibas changed his mind. That is the part lawyers keep circling back to. In 2023 he had looked at Thomson Reuters versus a scrappy legal-research startup called Ross Intelligence and decided a jury should sort out whether copying Westlaw’s editorial work to build a rival search tool was fair use. He set it for trial. Then he sat with the record over a long stretch, reread the briefs, and did something federal judges rarely do out loud.<\/p>\n
He reversed himself.<\/p>\n
On February 11, 2025, in a written opinion out of the District of Delaware, Bibas granted summary judgment to Thomson Reuters on direct infringement and threw out Ross’s fair-use defense (Thomson Reuters Enterprise Centre GmbH v. Ross Intelligence Inc.<\/em>, D. Del., 2025). “I continue to think that summary judgment is inappropriate on some of these issues,” he wrote of other questions, before landing the hammer on the one that mattered most. Ross had used Westlaw headnotes and the Key Number System, filtered through a set of legal question-and-answer pairs it commissioned from a company called LegalEase, to train a search engine that competed with Westlaw itself. That, the judge held, was not transformative. It was substitution wearing a technical costume.<\/p>\n Everyone who advises an AI company read that opinion the week it dropped. Everyone who writes for a living and suspects their words are sitting inside some model’s training corpus read it too, or should have. Because for two years the copyright war over AI had been fought almost entirely in motions to dismiss, in preservation orders, in discovery skirmishes over what exactly went into the machine. Bibas gave us the first real merits ruling. And the plaintiff won.<\/p>\n Ross was not a generative model. Bibas was careful about that, almost insistent. His opinion drew a line around itself: this case involved a company that took copyrighted material to build a competing information product, and the output was search results, not synthetic prose. He explicitly set aside the harder questions that generative systems raise.<\/p>\n Still, the reasoning traveled. Bibas ran the four statutory fair-use factors under Section 107 and handed two of them, the first and the fourth, to Thomson Reuters. The first factor asks about the purpose and character of the use. The fourth asks about the effect on the market for the original. Those are the two factors that decide most fair-use cases, and they are the two that the AI defendants most need to win.<\/p>\n On purpose, Bibas leaned on the Supreme Court’s most recent word on the subject. In Andy Warhol Foundation for the Visual Arts, Inc. v. Goldsmith<\/em>, 598 U.S. 508 (2023), Justice Sotomayor wrote for the Court that a use is not transformative just because it adds new expression or serves some fresh aesthetic. What matters is whether the new work shares the same commercial purpose as the original and whether that shared purpose lets it substitute in the marketplace. Warhol’s silkscreen of Prince, licensed to a magazine, competed with Lynn Goldsmith’s photograph of Prince, licensed to magazines. Same lane. Same buyers. The Foundation lost.<\/p>\n Bibas took that framework and set it down on top of Ross. Westlaw sells legal research. Ross was selling legal research. The headnotes went in one end and a competing product came out the other. “Ross’s use is not transformative,” he concluded, “because it does not have a further purpose or different character” from Westlaw’s. The market harm followed almost automatically. Thomson Reuters had a market for its editorial content, including a potential market for licensing that content to train AI tools, and Ross had helped itself without paying.<\/p>\n That phrase about a potential licensing market is worth pausing on. It hands copyright owners a circular but powerful argument. The market for AI training data exists, they can now say, precisely because companies keep trying to pay for it, and the ones who did not pay have damaged that market by taking for free what others would license. Warhol supplied the doctrine. Bibas supplied the first application to anything AI-adjacent. And the plaintiffs’ bar noticed.<\/p>\n The largest of the fights sits in the Southern District of New York, where The New York Times Company v. Microsoft Corporation and OpenAI<\/em> has been grinding forward since the paper filed in December 2023. The complaint is a document written to be read aloud. It reproduces, in side-by-side columns, prompts fed to ChatGPT and the near-verbatim Times articles that came back out. Paragraph after paragraph of the Times’s own reporting, regurgitated by a system the paper says was trained on that reporting without a license.<\/p>\n “Defendants seek to free-ride on The Times’s massive investment in its journalism,” the complaint charged, “by using it to build substitutive products without permission or payment.” OpenAI answered that the memorization examples were a bug, a rare and manipulated outcome, and that training on publicly available text is quintessential fair use because a model learns statistical patterns rather than copying expression in any legally meaningful sense.<\/p>\n Judge Sidney Stein was not persuaded that the theory should die at the pleading stage. In an opinion issued in late March 2025, and expanded in a written order that April, he let the core copyright claims proceed, along with claims from the consolidated cases brought by other news organizations. He trimmed some counts. He kept the spine of the case alive. The parties are now deep in discovery, which is where these cases get interesting and where they get ugly.<\/p>\n The discovery fights in the Times case have been trench warfare. OpenAI has resisted turning over the full composition of its training data. The Times has fought to protect its reporters’ notes and internal materials. And Magistrate Judge Ona Wang entered a preservation order that put OpenAI under an obligation to retain output log data, including deleted user conversations, over the company’s loud objection that it was being forced to keep private chats it had promised users it would purge. By the fall of 2025 the dispute had swelled to an order touching on tens of millions of conversation logs, a scale that made privacy advocates nervous and made OpenAI’s lawyers louder.<\/p>\n Here is what the fights are really about. Nobody outside these companies knew, with precision, what went into the models. Discovery is the mechanism by which the public and the courts find out. And what discovery has turned up, across several of these cases, is that a lot of the training material did not come from tidy licensed feeds. It came from the open sea of the internet, and in some cases from places that look a great deal like digital piracy.<\/p>\n The name that keeps surfacing is LibGen. Library Genesis. A shadow library of pirated books, millions of them, long a target of publisher lawsuits and long a convenient bucket of clean text for anyone building a language model.<\/p>\n In the Northern District of California, Judge Vince Chhabria has been overseeing Kadrey v. Meta Platforms, Inc.<\/em>, the case brought by Sarah Silverman, Ta-Nehisi Coates, and other authors who say Meta trained its Llama models on their books without permission. Discovery in that case pried loose internal Meta communications, and the communications did not read well. Employees discussed sourcing books from LibGen. One message, quoted in the plaintiffs’ filings, captured an engineer worrying that torrenting from a known pirate source “doesn’t feel right.” Another exchange described concern that using the material could carry “medium-high legal risk.” The plaintiffs put those words in front of the judge and asked him to draw the obvious conclusion.<\/p>\n He did not, at least not the way they wanted. In June 2025, Chhabria granted summary judgment to Meta on the fair-use question. But he did it grudgingly, and he wrote an opinion that read less like a win for AI companies than a warning shot fired over their bows. The authors had lost, he explained, largely because they made the wrong arguments and failed to develop a record on the theory most likely to work, which he called market dilution: the idea that a model trained on an author’s work can flood the market with competing books and undercut the author’s livelihood. “This ruling does not stand for the proposition that Meta’s use of copyrighted materials to train its language models is lawful,” he wrote. “It stands only for the proposition that these plaintiffs made the wrong arguments and failed to develop a record in support of the right one.”<\/p>\n Translation for the next set of plaintiffs: sharpen the market-harm theory and come back.<\/p>\n The companion ruling, issued days earlier by Judge William Alsup in Bartz v. Anthropic PBC<\/em> (N.D. Cal., 2025), drew the sharpest line of the year. Alsup split the baby with a cleaver. Training a large language model on books, he held, is transformative and can be fair use, because the model learns from the works the way a human reader learns and does not reproduce them wholesale as output. That part cheered the AI industry. But then Alsup turned to how Anthropic had obtained the books in the first place, and here he was merciless. The company had downloaded millions of pirated titles from sources including LibGen and had kept a permanent library of them. Building a central library of pirated books, he ruled, was not fair use no matter what you later did with it. The acquisition itself was the infringement.<\/p>\n “Anthropic had no entitlement to use pirated copies for its central library,” Alsup wrote. The theft was not cured by a later transformative purpose.<\/p>\n What happened next got everyone’s attention. Facing a trial on statutory damages for potentially hundreds of thousands of pirated works, with statutory awards that can reach $150,000 per willful infringement, Anthropic settled. In September 2025 the company agreed to pay roughly $1.5 billion to resolve the class claims, a figure that penciled out to something on the order of $3,000 per book across a class of around half a million titles. It was, by a wide margin, the largest copyright recovery of its kind. And it sent a message that the AI companies heard clearly: how you got the data may matter more than what you did with it.<\/p>\n The visual-art front has moved slower and messier. In the Northern District of California, Andersen v. Stability AI Ltd.<\/em> gathered a group of artists, including Sarah Andersen, Kelly McKernan, and Karla Ortiz, who sued Stability AI, Midjourney, and DeviantArt over image generators trained on the LAION dataset, a scraped web archive of billions of image-text pairs. Judge William Orrick took an axe to the first complaint in October 2023, dismissing most of it while leaving the door open to amendment. The artists came back with more detail about how the models worked and how images could be regurgitated, and in an August 2024 order Orrick let a reworked set of claims proceed, including direct infringement and an inducement theory. The case survived. It did not win. But surviving is its own kind of leverage when the discovery bill starts running.<\/p>\n Getty Images picked a different weapon. Rather than a class of individuals, Getty brought the resources of a stock-photography giant that meticulously tracks its own catalog. It sued Stability AI in the District of Delaware and, in parallel, in the United Kingdom. Getty’s evidence was vivid: generated images that carried mangled remnants of the Getty Images watermark, a ghostly signature suggesting the model had ingested enormous quantities of Getty’s watermarked stock.<\/p>\n The UK case, though, delivered a bruise. As the London trial unfolded through 2025, Getty narrowed its claims dramatically, dropping the central training-related copyright allegations for jurisdictional and evidentiary reasons, and in a November 2025 judgment the High Court handed Stability a largely favorable result on what remained. The takeaway was not that training on copyrighted images is fine. It was that proving where and how a model was trained, under a particular country’s law, is brutally hard when the defendant controls the servers and the record is opaque. The US Delaware case pressed on separately, on American copyright law, where the watermark evidence carries more weight.<\/p>\n Then there is the Authors Guild. Its consolidated suit against OpenAI in the Southern District of New York, filed in September 2023, put marquee names on the caption: George R.R. Martin, John Grisham, Jonathan Franzen, David Baldacci, and a class of working novelists behind them. The theory tracks the Times case. Books went in without a license; the model can produce derivative summaries, sequels, and stylistic imitations; the market for human-written books suffers. That case, too, cleared the pleading stage in substantial part and moved into discovery, riding the same doctrinal currents as everything else.<\/p>\n Strip away the docket numbers and a shape emerges. By late 2025 the fair-use fight over AI training had sorted itself into two questions that the courts are answering separately, and the answers are pulling in opposite directions.<\/p>\n The first question is whether training a model is transformative at all. On this, the AI defendants have been doing better than the headlines suggest. Alsup in Bartz<\/em> said training itself can be transformative. Chhabria in Kadrey<\/em> assumed as much while faulting the plaintiffs’ proof. The industry’s core argument, that a model learns patterns rather than copies expression, has real traction with judges who understand the technology as something closer to reading than to reproduction.<\/p>\n The second question is about the source and the output, and here the defendants keep losing ground. Piracy in the acquisition, as in Bartz<\/em>, is not saved by a transformative purpose downstream. Verbatim regurgitation in the output, as the Times alleges, undercuts the claim that no expression was copied. And market harm, particularly the licensing market and Chhabria’s market-dilution theory, gives plaintiffs a factor-four argument that Warhol made stronger and that Bibas already accepted once.<\/p>\n Warhol is the hinge. Before Goldsmith<\/em>, defendants could argue that any meaningful change of purpose made a use transformative, and courts often nodded along. After Goldsmith<\/em>, the question narrowed to whether the new use competes in the same market for the same customers. A model that can write news articles competes with news publishers. A model that can generate images in an artist’s style competes with that artist. A legal-research tool competes with Westlaw. The moment a court frames the market that way, factor one and factor four start tilting toward the copyright owner, and the transformative-training argument has to carry more weight than it may be able to bear.<\/p>\n That is the doctrinal tension that will define the next round. Judges believe training is more like learning than copying. Judges also believe, after Warhol, that fair use dies when the output substitutes in the same market. Those two beliefs cannot both control every case. Which one governs will depend on facts that are only now surfacing in discovery: how the data was obtained, how often the model spits out protected expression, and whether a licensing market existed that the defendant chose to bypass.<\/p>\n I have covered enough copyright litigation to distrust confident predictions, so I will keep this to what the record actually supports.<\/p>\n If you advise AI developers, the lesson of 2025 is not written in the fair-use holdings. It is written in the settlement figure. Anthropic did not pay roughly $1.5 billion because training is unlawful. It paid because it built a permanent library out of pirate sites and could not defend the acquisition, and it faced statutory damages that scaled into the stratosphere with every willful count. Provenance is now a discoverable, bet-the-company fact. The internal Slack message that says a source “doesn’t feel right” is going to be read aloud someday, in a courtroom, by a plaintiff’s lawyer who found it in your document production. Firms counseling these clients should be treating data sourcing the way they treat securities disclosures: as an area where the paper trail either saves you or sinks you, and where clean licensing is cheaper than it looks against a Section 504 damages exposure.<\/p>\n If you represent creators, Warhol and Bibas handed you a template, and Chhabria handed you the missing piece. The winning case is not the one that shouts about theft. It is the one that builds a granular record: a defined market for licensing, evidence that licenses were available and refused, proof of output substitution, and a market-dilution theory developed with an economist rather than asserted in a footnote. The plaintiffs who lost in 2025 mostly lost on proof, not principle. That is a fixable problem, and the plaintiffs’ bar is fixing it.<\/p>\n And then there is the lawyer reading this who is neither of those things. The practitioner who has published articles, briefs, treatise chapters, blog posts, and CLE materials for twenty years, and who now suspects, correctly, that some of that writing was scraped into a training corpus without anyone asking. I will not pretend the law offers you a clean remedy today. Individual authors have found the courthouse doors heavy. Class certification is contested. Registration requirements under the Copyright Act trip up people who never registered the article they dashed off for a bar journal in 2014. The Andersen<\/em> plaintiffs are still fighting years in. But the ground is shifting under the assumption that scraping is free, and the Anthropic settlement proved that a large enough class of unlicensed works can command a real number.<\/p>\n What bothers me, watching all of this, is the asymmetry that Thomson Reuters v. Ross<\/em> exposed and that nobody has resolved. Thomson Reuters won because it is a sophisticated owner of a well-documented editorial product with a proven licensing market and the resources to litigate for years in Delaware. The individual writer whose sentences trained the same category of tool has none of those advantages, and the fair-use doctrine, as currently applied, rewards exactly that gap. The company with the license department and the market-harm expert prevails. The freelancer with a byline and no registration does not get through the door.<\/p>\n Bibas changed his mind once, in the open, because the record persuaded him. That is the encouraging part of this whole saga: the judges are actually reading, actually reasoning, actually letting the facts move them. The discouraging part is that the facts they are being handed still come mostly from the parties who can afford to develop them. Until that changes, the copyright war over AI training data will keep being decided by who has the better discovery budget, and the people whose words built these machines will keep watching from the cheap seats, waiting to see whether the law that protects Westlaw ever gets around to protecting them.<\/p>\n","protected":false},"excerpt":{"rendered":" From Thomson Reuters v. Ross to the Times suit and a $1.5 billion Anthropic settlement, the fair-use fight over AI training data hardened into real law by late 2025.<\/p>\n","protected":false},"author":1,"featured_media":69,"comment_status":"open","ping_status":"open","sticky":false,"template":"","format":"standard","meta":{"footnotes":""},"categories":[16],"tags":[],"class_list":["post-35","post","type-post","status-publish","format-standard","has-post-thumbnail","hentry","category-legal-tech-ai"],"yoast_head":"\nWhy a search engine, not a chatbot, drew first blood<\/h2>\n
The Times case and the fight over what the machine memorized<\/h2>\n
The pirate libraries in the record<\/h2>\n
Images, artists, and a defense that keeps surviving dismissal<\/h2>\n
What the battle lines actually are now<\/h2>\n
What I would tell a firm on either side of this<\/h2>\n