Court Rejects Fair Use for AI Training

Dateline: February 16, 2025

A decision out of a Delaware federal court has dealt what many have characterized as a major blow to artificial intelligence companies. In short, the court ruled that using copyrighted material to train AI models does not qualify as fair use. It's an important decision but by no means the end of the AI infringement debate.

The court did not consider the technical aspects of how AI machines are trained and whether there is copying involved in the training. Also, the system at issue is not a generative-AI model. These are areas of inquiry that will be explored in other pending cases against AI companies. Do not underestimate the ability of other AI companies to differentiate this case.

This decision is just one of several legal developments signaling increased scrutiny on how tech giants build their AI systems.

In another case, internal documents from Meta produced in discovery, reveal that employees knowingly used copyrighted material from a pirate website to train its AI models. Meanwhile, OpenAI has been ordered to disclose its training dataset, a move that could expose the full extent of its reliance on copyrighted works.

These cases are narrowing the focus on the question of whether AI companies can freely use creative content without permission—or whether the courts will force them to change their approach.

The Work Being Protected

Before we get to the Court's holding in the Delaware case that fair use is not a defense when using copyrighted material to train AI, let's have some context about the underlying creative work.

The copyright infringement claim was brought by Thomson Reuters, the owners of the legal research database tool Westlaw.

Westlaw gathers court opinions and stores them in a searchable database. As a way of organizing the opinions, Westlaw editors create headnotes. A headnote is a short, key point of law pulled out of a lengthy judicial opinion. There can be dozens of headnotes for any given opinion.

The headnotes are then sorted into a key number system. Key numbers correspond to different areas of law. When you find the key number that covers the question of law being researched, you can get a list of all the cases with that key number. Then you can read the headnotes for those cases and if a decision looks relevant, you can pull it up and read the whole opinion.

The judicial opinions themselves are not protected by copyright, but the headnotes and the key number system developed by Westlaw are.

What the Bad Guys Did

The defendant, Ross Intelligence, is a Westlaw competitor. It built a legal research database using AI. Ross approached Thomson Reuters about licensing Westlaw material for their product, but Thomson declined to work with a competitor.

Because they could not license Westlaw headnotes and key numbers directly, Ross hired a company called LegalEase to draft they called "Bulk Memos." Ross then trained its AI machine on more than 25,000 Bulk Memos.

LegalEase had access to Westlaw and used it to draft the Bulk Memos. Turns out, LegalEase was pretty deft with the old cut and paste when it came to drafting the Bulk Memos.

In short, Ross built its competing product using Bulk Memos, which in turn were heavily copied from Westlaw headnotes. When Thomson Reuters found out, it sued Ross for copyright infringement.

No Fair Use for Ross

This is the first judicial opinion to analyze a fair use defense for a company using copyrighted material to train AI models.

Bear in mind that Ross's product, its legal research database, is not a generative-AI model. It is not creating content based on the copyrighted material it was trained on. It is using the copyrighted material to identify relevant legal resources for the researcher.

The court found infringement (copying and substantial similarity) and then considered whether Ross's conduct is protected by fair use. The copying and substantial similarity the court found was in the Bulk Memos, not in the AI-machine training process. The mechanics of the training process are glossed over in the opinion.

Fair use is a defense to infringement and it is the defendant's job to prove fair use. The fair use analysis requires consideration of four factors. The court analyzed each factor finding:

1. Ross took the headnotes from the Bulk Memos to make it easier to develop a competing legal research tool. So Ross’s use is not transformative. The first factor goes to Westlaw. (Heavy reliance on the Warhol v. Goldsmith decision from the Supreme Court in this part of the opinion.)
2. The headnotes have creative elements but are far from being highly creative works (like novels or artwork). The second factor goes to Ross.
3. Because Ross did not make Westlaw's headnotes available to the public, Ross benefits from factor three -- how much of the underlying work was taken.
4. Ross had every intention of competing with Westlaw by developing a market substitute. In this case, undermining the market for Westlaw is the most important factor in the analysis. This factor goes to Westlaw.

Although each party "won" two of the fair use factors, factor two is less important and factor four is the most important. After weighing all the factors, the judge determined that Ross does not have a fair use defense.

Does Using Copyrighted Material to Train AI Involve Copying?

You're probably thinking that the question of whether training AI on protected material involves copying is silly.

Indulge me for a minute. Here's what I'm thinking about:

Back in 2008, Cablevision introduced a process for consumer digital video recording on demand. They were sued by the networks and Hollywood for copyright infringement. ( Cartoon Network v. Cablevision)

The court took a close look at the technology used for DVR. The protected content, movies and TV shows, was copied and buffered for only 1.2 seconds before it was sent to the subscriber.

The court decided that holding a copy for only 1.2 seconds was transitory and therefore not infringement.

Transitory data use is not copying. Without copying there’s no infringement. When there’s no infringement, there is no question of fair use which is a defense.

This concept, as applied to AI training technology, was not explored in the Westlaw v. Ross Intelligence case.

AI works by using algorithms and computer programs to process and analyze large amounts of data.

Process and analyze data. Not copy.

AI machines consume data, they ingest data, they integrate data, but it is not clear that they copy data in the legal sense of the word “copy.”

The court determined that LegalEase copied Westlaw's headnotes and incorporated them into Bulk Memos. Ross Intelligence used the Bulk Memos to train their AI. Although there was some discussion of "intermediate copying" in the fair use analysis, there was no finding that training = copying.

I think we all need to stay tuned. It may be too late for Ross Intelligence in this case, but there a quite a few cases yet to come.

Authors' Class Action Reveals the Datasets

The Westlaw v. Ross Intelligence case is at a point where it is ready for trial. The discovery phase is finished.

The discovery phase of litigation is the investigation stage of a legal case. It’s when both sides gather evidence before going to trial. Each party gets to ask for documents, emails, and other records, interview witnesses under oath (called depositions), and submit written questions (interrogatories) that the other side must answer.

The class actions brought by Sarah Silverman and Ta-Nehisi Coates, among others, against Meta and OpenAI, are still in the discovery phase. Some interesting things are being discovered.

Meta Does a Faceplant

The morons at Meta downloaded tens of millions of pirated copyrighted works from peer-to-peer networks to train its Llama artificial intelligence product. (For an explanation of how peer-to-peer networks work, see my earlier post How BitTorrent Technology is Used to Pirate Copyrighted Material — A Primer.

The employees at Meta went on to discuss their concerns in emails between themselves and with the company's lawyers:

“I don’t think we should use pirated material. I really need to draw a line there.”

“[U]sing pirated material should be beyond our ethical threshold.”

“[I]f we license one single book, we won’t be able to lean into fair use strategy. So we will have to drop all of the libgen and books datasets[. ]” (NB: This sounds like asking for forgiveness instead of permission.)

“[U]sing torrents would entail ‘seeding’ the files – i.e., sharing the content outside, this could be legally not OK.”

Such hubris. They knew they shouldn't be doing what they did. You can read these and other admissions in the recent court filing.

OpenAI Has to Hand Over Its Dataset

OpenAI used some of the same datasets as Meta did. But OpenAI used another dataset, as well, and the plaintiffs want to see it. OpenAI ostensibly agreed to produce that dataset, but engaged in tech bro-type shenanigans that made the "production" of the dataset nearly useless.

OpenAI only provided the authors with a small portion of the dataset, and only allowed remote access to it which is bogging down the process. Plus, the limited access resulted in searches that have crashed.

The judge ultimately sided with the authors and ordered OpenAI to produce the dataset in a more meaningful manner. But, at the moment, the dataset will be produced under seal. We, the public, will not get to peek.

More to Come

It's February 2025 and the first case was filed back in January 2023. The recent developments in the litigation between creators and the AI companies are key moments, but we are a long way from definitive answers.