Using copyrighted content to train AI without permission is not ‘fair use’, US court rules in Thomson Reuters case

WHAT’S HAPPENED?

For the first time, a US court has ruled on whether using copyrighted material without permission to train AI amounts to “fair use” under copyright law, and the news is mostly good for copyright holders: The judge in the case has ruled against the AI company.

However, there is one big wrinkle: The judge noted in the ruling that it does not necessarily apply to generative AI, the kind of AI over which music companies are suing AI developers.

The case in question pitted media and news conglomerate Thomson Reuters, owner of the Reuters news service, against Ross Intelligence, a now-defunct service that offered users access to a database of court cases compiled through machine learning (AI) technology.

In 2020, Thomson Reuters sued Ross Intelligence for copyright infringement, alleging that Ross had scraped Thomson Reuters’ own law database, known as Westlaw, to create its database, and that this amounted to copyright infringement.

Much like the AI companies defending themselves against lawsuits by music companies, Ross Intelligence argued – among other things – that its use of Thomson Reuters’ copyrighted content should be given a “fair use” exemption.

In a ruling on Tuesday (February 11), Judge Stephanos Bibas of the US District Court for the District of Delaware rejected that argument. The judge overturned an earlier order from 2023 in which he sent the matter of copyright infringement and the issue of the “fair use” defense to be heard in front of a jury.

Judge Bibas said that, after taking a closer look at the matter, he was overturning his decision on the “fair use” aspect of Ross’ defense, and issued a summary judgment.

“A smart man knows when he is right; a wise man knows when he is wrong,” the judge wrote in his new ruling, which can be read in full here. “Wisdom does not always find me, so I try to embrace it when it does – even if it comes late, as it did here.”

The case will still go before a jury, as there are still a number of issues to be resolved, such as the validity of Thomson Reuters’ copyrights and whether or not Ross copied a tracking system used by Westlaw. But the fair-use argument has been thrown out of court.

On its face, this looks like unmitigated good news for copyright holders: The fair use defense has been rejected in a case involving scraping of copyrighted content to train AI, seemingly exactly the same sort of thing that AI developers like Anthropic, Suno, and Udio have been accused of in cases brought by music companies.

Claiming that scraping copyrighted content is “fair use” has become common for AI developers facing copyright lawsuits. Anthropic has made the claim in a lawsuit brought by Universal Music Publishing, Concord, and ABKCO, which alleges that Anthropic’s Claude chatbot was trained on copyrighted lyrics and even regurgitated those lyrics when prompted.

Suno and Udio – sued last year by record companies owned by Sony Music Entertainment, Universal Music Group, and Warner Music Group – have made fair use the crux of their defense. They’ve gone so far as to basically admit that their AI tools – which allow users to create complete music tracks in minutes with just a text prompt – may have been trained on copyrighted recordings, but, in their view, there’s nothing wrong with that.

In light of the Thomson Reuters decision, there’s little doubt that the road to victory has just become harder for AI companies and easier for copyright holders. However, when digging into the details of the judge’s decision, the issue becomes less clear-cut.

To understand why, we need to understand just how the “fair use” concept works in US law.

HOW ‘FAIR USE’ WORKS

The fair use doctrine exists to safeguard freedom of expression and to create an environment that encourages learning and innovation.

To that end, in many cases, fair use protects parody and commentary on copyrighted works and some uses of copyrighted works for educational purposes and (primarily non-commercial) research.

When deciding whether unauthorized use of copyrighted content can be allowed, US courts look at four factors:

• Factor 1: The purpose and character of the use – whether the use of the copyrighted work is “transformative,” that is, whether it would have a “further purpose or character” that’s different from the original work.
• Factor 2: The nature of the copyrighted work – works that are more creative, such as written fiction, original music, and visual art, enjoy more protection than less creative works, such as news articles or compilations of publicly available information.
• Factor 3: The amount and substantiality of the portion taken – using a small amount of a copyrighted work is less likely to be seen as infringement than using a large part or all of that work.
• Factor 4: The effect of the use on the potential market for, or value of, the copyrighted work – in essence, does the new work created using the copyrighted work compete in the same market as the original work?

Importantly, in the Thomson Reuters case, Judge Bibas ruled in favor of Thomson Reuters on factors one and four, but against Thomson Reuters on factors two and three. Still, according to judicial precedent, factors one and four weigh more heavily than the other two, so the judge gave the win to Thomson Reuters.

THE GOOD NEWS FOR COPYRIGHT HOLDERS

On the first factor – the purpose and character of the use of the copyrighted works – Judge Bibas handed the victory to Thomson Reuters.

The way Ross Intelligence used the Thomson Reuters data was “not transformative… it does not have a ‘further purpose or different character’ from Thomson Reuters’s. Ross was using Thomson Reuters’s headnotes as AI data to create a legal research tool to compete with [Thomson Reuters’] Westlaw,” the judge wrote.

“Ross’s use is commercial,” the judge noted. “Ross admits as much.”

That works in favor of music companies, who can argue before the courts that training AI on copyrighted music is not “transformative” because it results in the same product – music.

On the fourth factor – the effect of the use on the potential market for, or value of, the copyrighted work – Judge Bibas again handed the victory to Thomson Reuters.

“Ross… meant to compete with Westlaw by developing a market substitute,” the judge concluded.

“There is nothing that Thomson Reuters created that Ross could not have created for itself… without infringing Thomson Reuters’s copyrights.”

“There is nothing that Thomson Reuters created that Ross could not have created for itself… without infringing Thomson Reuters’s copyrights.”

Judge Stephanos Bibas, US District Court for the District of Delaware

Again, this works in music companies’ favor. They can argue that AI tools that create music result in a product that directly competes with the music companies whose material was taken. They can also argue that the AI developers could have developed their generative engines without infringing copyright — for instance, by licensing the music they use, as some AI companies do.

On the second factor – the nature of the copyrighted work – the judge ruled against Thomson Reuters and in favor of Ross, but the rationale for this doesn’t really apply to music.

Judge Bibas concluded that Thomson Reuters’ legal database Westlaw isn’t very creative – it’s a searchable database of publicly available court documents. This would likely not apply to music, which is seen by the courts as being highly creative.

On the third factor – the amount of a work taken – again the judge ruled in favor of Ross Intelligence and against Thomson Reuters. But he did so on the argument that Ross had taken only a small part of Westlaw’s database.

Again, this likely won’t apply in the cases brought by the music companies because, in those instances, each work of music has its own copyright (and copyright owners), and in this instance, the music companies will likely argue that the AI developers took all of each of those recordings.

THE BAD NEWS FOR COPYRIGHT HOLDERS

So, if all four factors of “fair use” favor the record companies, why would there be bad news for copyright holders here? Because in his ruling, Judge Bibas made a point of stressing that his ruling does not apply to generative AI – which is the kind of AI that’s being sued over by the record companies and other rights holders.

“It is undisputed that Ross’s AI is not generative AI (AI that writes new content itself). Rather, when a user enters a legal question, Ross spits back relevant judicial opinions that have already been written,” the judge wrote.

“Because the AI landscape is changing rapidly, I note for readers that only non-generative AI is before me today.”

However, companies like Anthropic, Suno, and Udio use generative AI. This means the AI creates something new, if not necessarily entirely original. And that means the courts in the cases brought against generative AI could assess fair use differently from how it was assessed in the Thomson Reuters case.

A FINAL THOUGHT…

If the courts in the music companies’ copyright cases follow precedent as Judge Bibas did in the Thomson Reuters case, the issue of fair use will be determined primarily by two questions: Is the AI companies’ use of the copyrighted works “transformative,” and does the output lessen the value and/or compete with the original works?

Undoubtedly, music companies will argue that it’s not transformative. In simplest terms, these AI companies (allegedly) took recorded music in order to create recorded music. However, AI music generation companies will likely respond by noting that the music their AI creates is not the same as the music they trained on, and therefore, it’s transformative.

“I note for readers that only non-generative AI is before me today.”

Judge Stephanos Bibas, US District Court for the District of Delaware

On the second question – does AI-generated music lessen the value of copyrighted human-made music and compete in the same market – it might look like an open-and-shut case from the perspective of rights holders. Of course they do.

But AI music-making tools like Suno and Udio allow people who’ve never created music – people who would never have signed with a record label, who would never have performed in a live venue – to generate tracks. And these companies make money by selling subscriptions to their music-making services, and not by selling music. So is that still the same market as the one recording companies operate in?

The Thomson Reuters case has brought the rights of copyright owners one step closer to being recognized by courts in the age of generative AI. But before either side can declare victory, some important questions will have to be answered.Music Business Worldwide