Lawyers: First AI fair use case may prove to be outlier

Listen to this article

BOSTON — They are far from the last words a judge will write on the subject, nor do they have any precedential value. But that has not stopped attorneys from poring over a recent Delaware federal court decision that represents the first substantive application of copyright law and fair use doctrine to artificial intelligence.

The case, Thomson Reuters Enterprise Centre GMBH, et al. v. Ross Intelligence Inc., involves an AI-powered legal research search engine that the defendant created to compete with Westlaw, a product of plaintiff Thomson Reuters.

Rebuffed when it asked to license Westlaw’s content, Ross Intelligence instead struck a deal with the company LegalEase to get access to LegalEase’s “Bulk Memos” to use as training data for its AI. Bulk Memos are lawyers’ compilations of legal questions with good and bad answers. They were created using Westlaw headnotes, which summarize key points of law and case holdings. The lawyers hired by LegalEase were told they should use Westlaw’s headnotes, but they should do more than just copy and paste them directly into the questions they were writing.

Back in 2023, 3rd U.S. Circuit Court of Appeals Judge Stephanos Bibas, sitting by designation in Delaware, largely denied Thomson Reuters’ motions for summary judgment on the issues of copyright infringement and Ross’ fair use defense.

But as the case was moving toward trial last August, Bibas “studied the case materials more closely and realized that my prior summary-judgment ruling had not gone far enough.” He continued the trial and invited the parties to renew their summary judgment briefing. On Feb. 11, Bibas issued an opinion reversing himself on both fronts.

He also made it a point to highlight the fact that “only non-generative AI is before me today.”

That fact may limit the persuasive effect of Bibas’ ruling in cases like the one brought by the New York Times against OpenAI and Microsoft, alleging that the newspaper’s copyrights were infringed in training ChatGPT and Copilot, attorneys say.

“This is a decision that’s fairly protective of copyright owners of existing content, and doesn’t necessarily spell doom for large language models and generative AI but certainly is not in their favor and perhaps tips the scales a little more in the legacy content owners’ favor in terms of their bargaining position to try to get compensated for use of their work in training models,” said Boston attorney Adam J. Kessel.

But Kessel and others were quick to note the unique facts in Thomson Reuters, which should provide ample opportunity to distinguish it from future cases.

Thomson Reuters could get cited a lot and “lead the path forward,” or it could quickly fall into obscurity, said New York attorney Karen K. Won.

“I think there are almost 40 pending cases involving AI and copyright issues right now across the various courts, and I think most of them feature a fundamental fact pattern that is distinct from this one,” she said.

Boston attorney John N. Anastasi agreed.

“[Thomson Reuters] gives both plaintiff content owners and defendant AI companies some ammunition in the realm of infringement and fair use. But every case comes down to its own unique set of facts, so they’ll all have to be tackled on the merits,” he said.

There may be a day when we can look back and say that the unique nature of AI technology has spawned its own body of law, Boston attorney Patrick J. Concannon said.

“But what we’re seeing thus far is applying traditional copyright principles,” he said. “In this case, it could have just been a case about Ross putting content on its website or something much more plain that doesn’t include the sort of quote-unquote ‘sexy’ topic of AI.”

Like a sculptor

As a threshold matter, Bibas had to decide whether Westlaw’s headnotes were copyrightable at all, given that they are drawn from an underlying judicial opinion they are summarizing, often word for word.

In 2023, Bibas had used that characteristic to deny Thomson Reuters’ motion for summary judgment.

His change of heart on whether the headnotes were copyrightable became clear once he analogized the editorial judgment of a lawyer writing a headnote to that of a sculptor, Bibas explained.

“A block of raw marble, like a judicial opinion, is not copyrightable,” Bibas wrote. “Yet a sculptor creates a sculpture by choosing what to cut away and what to leave in place. That sculpture is copyrightable. So too, even a headnote taken verbatim from an opinion is a carefully chosen fraction of the whole. Identifying which words matter and chiseling away the surrounding mass expresses the editor’s idea about what the important point of law from the opinion is.”

In other AI cases, in which the content at issue is original writing, news articles or songs, this will be a non-issue, Boston attorney Daniel Rudoy noted.

Not like computer code

Having established that Westlaw’s headnotes were copyrightable, Bibas moved on to reconsider whether any defenses to infringement applied, including fair use.

Once again, Bibas reversed himself. Looking at the four-factor fair use test, Bibas gave two wins to each side. But the more important factors — the purpose and character of Ross’ use of Thomson Reuters’ copyrighted material and how Ross’ use affected the copyrighted work’s value or potential market — tipped in Thomson Reuters’ favor.

Initially, Bibas had bought into Ross’ argument that it had merely engaged in a permissible act of “intermediate copying,” which a series of cases analyzing computer programs had found to be fair use.

Upon further reflection, Bibas realized that copyright law recognizes that “computer programs differ from books, films, and many other literary works in that such programs almost always serve functional purposes.”

In addition, the computer programming cases about intermediate copying rely on a factor not present in Thomson Reuters: “The copying was necessary for competitors to innovate,” Bibas wrote.

For example, in the 2021 U.S. Supreme Court case Google LLC v. Oracle Am., Inc., Google had copied part of a computer programming language “necessary for different programs to speak to each other,” Bibas wrote.

“Here, though, there is no computer code whose underlying ideas can be reached only by copying their expression,” Bibas found. “The ‘copying is [not] reasonably necessary to achieve the user’s new purpose.’”

What Bibas seemed to be trying to do in delving into the computer programming cases is address the issue of whether the copying involved in Thomson Reuters was reasonably necessary to get at the unprotectable content — the judicial opinions — that is embedded into the copied work, Won said.

In Thomson Reuters, the copyright is “thin,” in the sense that the material barely clears the low bar for originality, and the nature of the copying is “purely intermediate,” in the sense that the copied material is not publicly rendered, Won noted.

At first blush, Bibas’ decision could set off alarm bells for AI companies wanting to make use of material benefiting from more robust copyrights and producing outputs that more closely resemble the inputs, she said.

“On the other hand, this was so niche,” Won said. “There are a number of things that the judge did not consider or did not look at.”

That means the other pending cases may well be distinguishable on other grounds, she said.

Market effects

With respect to the test’s fourth factor, which he noted was of paramount importance, Bibas said he was required to consider not only current markets but potential derivative ones developed either by content creators or their licensees. The Supreme Court in Google had also directed him to consider the public benefits the copying was likely to produce, he added.

Bibas’ previous decision hinged on the possibility that Ross might create “a brand-new research platform that serves a different purpose than Westlaw.”

“If that were true, then Ross would not be a market substitute for Westlaw,” he wrote.

In addition, he thought there was a relevant, genuine issue of material fact about whether Thomson Reuters would use its data to train AI tools or sell its headnotes as training data. If not, the public’s interest might be better served by protecting a “copier” rather than a creator.

However, Bibas had come to believe those concerns were “unpersuasive.”

Ross meant to compete with Westlaw by developing a market substitute, he wrote.

“And it does not matter whether Thomson Reuters has used the data to train its own legal search tools; the effect on a potential market for AI training data is enough,” he wrote.

While there might be a public interest in accessing the law, “the public has no right to Thomson Reuters’s parsing of the law,” Bibas wrote.

Copyrights encourage people to develop things that help society, like good legal research tools, and their builders should be paid accordingly, he said.

With generative AI, there are two big looming questions, Kessel said.

One is whether training a model with third-party, copyrighted, unlicensed content is the sort of copying that could be pursued as copyright infringement.

The second is whether the outputs of the models could be infringing, even if they do not reproduce the inputs exactly or even to a copyright infringement standard.

The latter question is of more concern to the larger world that is already using AI models extensively, Kessel said.

As others have analyzed the Thomson Reuters decision, Won said she had seen some people suggest that companies need to start considering licensing the data that they use to train their AI models, which she believes to be an overreaction.

“The fact of the matter is, as a commercial reality, that’s not going to be feasible for the vast majority of AI companies, which are startups,” she said. “Nor should it really be required, depending on the manner of the usage, the nature of the AI technology, and the output.”

Fair use law has evolved to focus increasingly on the question of transformativeness, to the point where it trumps or at least encapsulates all four prongs of the fair use test, Kessel said.

In the Thomson Reuters case, what was notable was that Bibas found that Ross’ use was not transformative as a matter of law.

“If that ruling holds, that could raise a lot of risks, at least for the LLM providers,” Kessel said. “I think they would take the position that what they do is incredibly transformative because they are creating technologies, opportunities, systems that didn’t exist. You just have to spend five minutes with an LLM to see that it’s doing something different from everything that came before it.”

Whether the Thomson Reuters decision itself is appealed, there will continue to be a lot of focus on whether generative AI and systems are sufficiently transformative to protect them under the fair use doctrine, Kessel said.

But Boston attorney Joseph Rutkowski noted that the output of Ross’ AI system — a list of uncopyrightable judicial opinions — was specifically not something created by a large language model like the ones at the heart of other ongoing AI copyright infringement cases.

“That’s one thing that could set this case off from, or reduce its effect on, the generative AI cases that are the most significant or at least the most attention-grabbing and making their way through the courts,” he said.