AI Copyright Lawsuit Ruling: Training on a Licensed News Archive Is Fair Use, Scraping Behind Paywalls Is Not

NEW YORK — A federal judge handed each side of the fight over artificial intelligence and copyright half of what it wanted on Monday, ruling that Lumen AI's use of a licensed archive of Harrowgate Media's articles to train its language model was fair use, but that the company's separate scraping of articles behind Harrowgate's paywall was not. The decision on cross-motions for summary judgment in Harrowgate Media v. Lumen AI is the first in the Southern District of New York, where most publisher suits are pending, to rule on the training question, and the first anywhere to distinguish so sharply between how the training data was obtained and what was done with it.

"Fair use protects transformation. It does not protect trespass," the court wrote.

The facts

Lumen licensed Harrowgate's historical archive — several million articles through 2022 — from a data aggregator under a license that permitted "research and analysis." It used the archive to train the model at the center of the case. Separately, Lumen's web crawler collected current articles from Harrowgate's sites, including articles available only to subscribers, by requesting pages in a manner that bypassed the paywall, and added them to the training corpus. Harrowgate sued for infringement on both sets of copies and for the model's outputs, which it said reproduced its articles in response to prompts.

What the court held

Factor	Licensed archive	Scraped paywalled articles
1. Purpose and character	Highly transformative: the articles were used to learn statistical relationships in language, a purpose different from informing readers; commercial, but that is not decisive after Warhol and Google v. Oracle	Same technical purpose, but the acquisition by circumventing the paywall weighs against fairness; the court declined to treat the manner of copying as irrelevant
2. Nature of the work	Factual news reporting, closer to the unprotected end; slight weight for the publisher	Same
3. Amount used	Whole works copied, but no more than the purpose required, as in Authors Guild v. Google (2d Cir. 2015)	Whole works, with the same analysis
4. Market effect	No substitution: the model does not serve as a replacement for the archive, and the license itself shows the market was served	Substitution: the scraped copies displaced the subscriptions and licenses through which the publisher sells current articles, including to AI developers
Result	Fair use	Infringement

The court dismissed the output claims. Harrowgate's evidence that the model reproduced its articles consisted of prompts that instructed the model to recite them, some including the article's opening lines; the court held that regurgitation induced by the plaintiff's own prompts, in a small number of instances, did not show that ordinary use of the model infringed, while leaving Harrowgate free to renew the claim on evidence of real-world outputs.

How the decision fits the case law

The ruling follows the line drawn in Bartz v. Anthropic in June 2025, where a Northern District of California judge held that training on lawfully purchased books was "spectacularly" transformative and fair, but that building a permanent library from pirated copies was not — a distinction that led to a settlement reported at $1.5 billion. Kadrey v. Meta, decided the same month by a different judge, found fair use on the record presented while warning that evidence of market dilution could produce a different result. On the other side, Thomson Reuters v. Ross Intelligence (D. Del. 2025) found no fair use where a non-generative tool was built from a competitor's headnotes to compete with it. The Harrowgate court aligned itself with Bartz: the training use is fair, the acquisition can be the problem, and the fourth factor is where publishers must make their case.

Fair use protects transformation. It does not protect trespass.

From the opinion

What it means

• Licenses are worth more. The decision confirms that licensed training is fair use and that the existence of a license market counts in the user's favor on the fourth factor — a result that strengthens the content-licensing deals publishers and developers have been signing since 2024.
• Paywalls are a legal line. Content a publisher has placed behind a subscription is not available for training without permission, and the court's reliance on the manner of acquisition gives publishers a theory that does not depend on proving output similarity.
• Output claims need real outputs. Plaintiffs will have to show that users, not their own lawyers, obtained infringing reproductions.
• Damages. Statutory damages of up to $150,000 per work for willful infringement, multiplied across tens of thousands of scraped articles, are the pressure point for the appeal and for settlement.

What happens next

Both sides have said they will seek interlocutory review; the court certified the fair use question for appeal under 28 U.S.C. § 1292(b), and the Second Circuit, which decided Authors Guild v. Google, will become the first court of appeals to address generative AI training. The damages trial on the scraped articles is stayed pending the appeal. The consolidated publisher cases in the same district will be shaped by the decision whether or not it is followed.

The Counsel will report the Second Circuit's decision on interlocutory review. Developers assembling training data should document the source and license terms of every corpus; publishers should ensure paywalls and terms of service are enforced technically as well as contractually.

This report is for general information only and is not legal advice. Readers should consult a licensed attorney about their own circumstances. Docket numbers and citations are to the official reporters unless otherwise stated.