Judge Rules Against Fair Use Defense for AI Company

“A smart man knows when he is right; a wise man knows when he is wrong. Wisdom does not always find me, so I try to embrace it when it does––even if it comes late, as it did here.”

These are the opening words in a 23-page summary judgment written by Judge Stephanos Bibas of Delaware in the case Thomson Reuters v. ROSS Intelligence.

The case pits Thomson Reuters, owners of the Westlaw legal search tool, against the AI company ROSS Intelligence. ROSS had sought to create an AI-powered competitor to Westlaw but was accused of training its system on Westlaw content.

We first discussed the case here in September 2023. At that time, Judge Bibas had issued a different summary judgment. That judgment sent most of the case to a jury rather than directly deciding on key issues.

A year and a half later, Bibas has changed his mind. After further examining the case and studying more legal rulings, he has granted Reuters summary judgment on several key issues. Most notably, Bibas rejected ROSS’ fair use defense, holding the company liable for thousands of the works it is accused of using.

So, what changed his mind, and what happens next? To find out, we first need to examine the background of the case.

Background of the Case

In the 2010s, ROSS Intelligence sought to create a new legal search engine powered by AI systems. However, to train that system, they needed a large number of legal questions and answers.

ROSS approached Reuters about licensing the Westlaw database. However, Reuters declined to do so. Instead, ROSS partnered with the legal access company LegalEASE. ROSS asked LegalEASE to create some 25,000 “bulk memos” that would be used to train their system.

However, the lawyers who created those memos were instructed to use Westlaw headnotes. Headnotes are short summaries written by Westlaw employees that make it easier for searchers to find relevant cases.

Though LegalEASE instructed the lawyers not to copy and paste content from Westlaw, Reuters alleges that thousands of the headnotes were infringed.

Reuters sued ROSS in 2020. ROSS shut down operations shortly afterward, but the case has moved forward.

In September 2023, Bibas ruled on dueling motions for summary judgment. He declined to dismiss the case but also didn’t side with Reuters. Most of the key issues of the case were slated to go to trial.

Though the trial was slated for August 2024, Bibas’ review of the documents and legal issues indicated that he might not have gone far enough. He continued the trial, and both sides filed new motions for summary judgment.

Now, Bibas has ruled on those motions. In the ruling, he revises his previous decision to grant summary judgment to Reuters on several key issues, most notably fair use.

The Judge’s Analysis

When Reuters filed its motion, it sought summary judgment on 8,197 headnotes (it alleges infringement of 21,787 headnotes). However, the judge set 5,367 of those aside, saying that there were open questions about whether they or LegalEASE’s versions were too close to the original judgment.

This left 2,830 headnotes that an expert for ROSS admitted were copied. The judge then went through all 2,830 headnotes and ruled that 2,243 had been infringed. These were the notes that were actually copied in the bulk memos that ROSS used.

With those 2,243 notes, the judge dismantled other legal defenses. He said, “None of Ross’s possible defenses holds water. I reject them all.”

This included the defenses of innocent infringement, copyright misuse, the merger defense and scenes à faire. In each case, he found that ROSS’s arguments were left wanting.

However, the fair use analysis makes up the plurality of the judgment. There, Bibas examined the four fair use factors:

1. The purpose and character of your use
2. The nature of the copyrighted work
3. The amount and substantiality of the portion taken, and
4. The effect of the use upon the potential market.

In his analysis, Judge Bibas determined that the first and fourth factors favored Reuters. The second and third, on the other hand, favored ROSS. However, since the first and fourth factors are considered the most important, he granted summary judgment in favor of Reuters.

As such, the judge ruled that ROSS had infringed those 2,243 works, leaving the question of whether the copyright had expired.

The Fair Use Analysis

The judge’s fair use analysis is likely the most critical part of this case. There, the judge ruled in favor of Reuters, ruling that ROSS’s use of the content to train their system was an infringement.

Regarding the first factor, the judge found that ROSS’ use was commercial and not transformative. Heavily citing the recent Warhol ruling, he found that ROSS’ use was in the same purpose and character as the original work.

However, the analysis was complicated by the fact the headnotes were not presented to end users. They were simply used to train the system. However, the cases that ROSS cited (such as the Google v. Oracle case) dealt with computer code, not literary works. Those rulings were meant to be limited to software.

As such, Bibas found the first factor favored Reuters.

With the second and third factors, Bibas found that they both favored ROSS. Headnotes are not highly creative works (barely meeting the threshold for copyrightability), and ROSS did not output the headnotes to end users.

On the fourth factor, which Bibas describes as the most important, he ruled in favor of Reuters. He found that ROSS intended to create a competitor to Westlaw and that it would harm the market for Westlaw both directly and when seeking to license content for AI training.

Though two factors favored each side, the first and the fourth factors are more heavily weighed. As such, Bibas granted summary judgment for Reuters.

What This Means for AI Companies

In this case, the judgment means that, unless another decision or an appeal overturns it, ROSS is liable for infringing 2,243 headnotes. What damages that entails will be determined later.

Many issues, including thousands of other headnotes ROSS allegedly infringed, could still go to trial. But these will not, and ROSS will face at least some liability (barring it being overturned).

This is bad news for ROSS and other AI companies. The path to holding ROSS liable was longer than with other AI companies. While other AI companies are known to have trained directly on copyright-protected content, ROSS had an intermediary, LegalEASE.

That said, other AI companies may be able to shift the fair use calculus some. They could argue that their output is less of a direct competitor and/or more transformative.

However, those will be difficult arguments as AI systems are often tasked with producing work “in the style of” various writers and artists. They’ve also been used to replace journalists and publish books on Amazon.

The biggest solace AI companies can enjoy is that this is just one district court judge. Other judges (and juries) are still free to reach different conclusions. But if this is upheld on appeal, it could set a president in the Third Circuit. However, even then, other circuits could disagree.

In short, there is a long way to go before this opinion becomes the standard. But, as an indication of which way the winds are blowing, it’s an awful sign for AI companies.

Bottom Line

When I discussed the case in September 2023, I compared it to the RIAA case against Napster. Though the Metallica lawsuit against Napster dominated the news coverage, the RIAA one ended up being the most important.

We are in a similar situation here. Though celebrity lawsuits dominate the news headlines, this case is farthest along, starting two years before the launch of ChatGPT. Though it isn’t likely to be the final word on the matter, it will still severely impact other AI decisions.

Simply put, there’s no correlation between important cases and those that get media attention. The Blurred Lines case generated endless headlines but had almost no long-term impact.

This puts me in an awkward position. I am trying to simultaneously draw attention to an important case and decision while acknowledging that it’s just the first of many. It’s a difficult balance.

While we are still early in what will likely be a lengthy legal saga, the early indicators do not bode well for AI companies. This is especially true considering how much time and energy the judge put into his analysis in this case, which led to a rare self-reversal.