AI Startup Suffers Defeat in Copyright Case
Earlier today, the United States District Court for the District of Delaware issued a memorandum opinion in the case of Thomson Reuters v. Ross Intelligence, Inc. You’ll find a brief background on the case and a summary of the opinion in this blog.
What was this case about?
Thomson Reuters is one of the largest corporate entities in legal research. Its primary product — Westlaw — is a familiar tool for law students and practicing attorneys. It’s a platform to search caselaw and statutes, read judicial opinions, and perform advanced research tasks that include querying the Westlaw AI Assistant.
One element of Westlaw is headnotes. Headnotes appear at the top of each judicial opinion. They are brief summaries of the major elements of each opinion. They are listed by subject matter, along with a numerical designation. In addition to providing links to the relevant portion of the opinion, the system of headnotes also creates a comprehensive library of subject matter across all published judicial decisions. For example, there are five headnotes that appear at the top of the Supreme Court’s decision in Brown v. Board of Education (ending segregation in public schools). The first is about “Public elementary and secondary education.” Users can read a brief summary of the relevant portion of Brown v. Board in the headnote, click a link beside the headnote to navigate to the same portion of the opinion, or click the label to find the other federal cases that also relate to that headnote.
This case arose because Ross Intelligence, an AI startup that has since shut down because of the costs associated with this lawsuit, trained its model on those headnotes without paying Thomson Reuters. Thomson Reuters decided that it wasn’t comfortable with that. It sued Ross Intelligence for copyright infringement.
In 2023, Circuit Judge Bibas issued an opinion that left most of the fundamental issues in the case for the jury’s determination. He refused to conclude, as a matter of law, whether the headnotes were protected by copyright law. He also chose not to determine whether Ross Intelligence’s ingestion of the headnotes into its AI training regimen was protected by the defense of “fair use.” The case would go to a jury, who would closely consider these questions.
Then, Judge Bibas changed his mind. He has revised his earlier judgment. His latest opinion sings a very different tune.
What’s in the opinion?
The new opinion says that, although the text of judicial opinions are not copyrightable, the headnotes are an original creation by Thomson Reuters that is entitled to copyright protection. Here’s the heart of his explanation from page 7 of the memorandum order:
[E]ach headnote is an individual, copyrightable work. That became clear to me once I analogized the lawyer’s editorial judgment to that of a sculptor. A block of raw marble, like a judicial opinion, is not copyrightable. Yet a sculptor creates a sculpture by choosing what to cut away and what to leave in place… 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.
Judge Bibas then sorts through more granular details about which headnotes have been copied as a matter of law, but the essential point here is that Thomson Reuters’ headnotes are protected by copyright, and Ross Intelligence infringed that copyright by training their AI model on that material.
That isn’t the end of the story. The looming question in the pending legal battles over generative AI is less about what’s protected by copyright and more so about whether training AI models on that material is protected by fair use.
The rough-and-dirty explanation of fair use is that a would-be infringer of copyrighted materials can defend against a lawsuit by showing that they only borrowed enough material to create something new and different. There are four factors that courts weigh:
- The purpose and character of the use, including whether it is for financial gain.
- The nature of the copyrighted work.
- How much of the copyrighted work has been used.
- How the infringing use will affect the value of the original work.
Fair use protects the artist who borrows the tune of a famous song to make a parody. It protects the YouTuber that includes clips from a movie in their latest review. Does it also protect a AI model that ingests copyrighted materials, learns from them, and then performs a separate function?
No, says Judge Bibas, it does not. Not here.
The first factor is probably the most important, and it favors Thomson Reuters. Defenders of AI tools argue that their ingestion of copyrighted data is “transformative”: it analyzes the copyrighted material, draws statistical relationships, and creates new outputs. Often, they analogize to a person who gains creative inspiration from a book they read and then writes their own story that might share themes and structure with the original. Here, writes Judge Bibas, the use of Westlaw’s headnotes is “not transformative because it does not have a ‘further purpose or different character’ from Thomson Reuter’s.” Here, he points out that Ross Intelligence is not generative AI. It doesn’t generate its own case summaries. It just outputs relevant judicial opinions based on a user’s query. Analysis of the headnotes is “intermediate step” that helps facilitate the process of connecting a user’s query to the relevant opinions. The headnotes have been copied to create a legal research tool, and that’s the same function that Westlaw itself serves. No transformation, then.
The second and third factors favor Ross Intelligence because, respectively, the headnotes are not an especially creative endeavor (compared to, say, writing a lengthy book) and the output of Ross Intelligence’s AI model doesn’t output the text of the headnotes to its users.
The fourth factor is important, too, and it also favors Thomson Reuters. Ross Intelligence may not perform precisely the same function as Thomson Reuter’s Westlaw, but it “meant to compete with Westlaw by developing a market substitute.”
Judge Bibas weighs the factors, and rejects the fair use defense. A huge win for Thomson Reuter.
Why does it matter?
The initial consequences are for this particular case. Technically, the memorandum opinion has only decided some issues at the summary judgment stage and the case could proceed to trial on the remaining issues. It probably won’t, though. Ross Intelligence has lost on the fundamental issues, and they’re out of money. The case will probably settle before trial.
More broadly, this is a major pronouncement about how United States federal courts will approach the still-evolving question of how AI training relates to copyright law. Judge Bibas writes carefully to point out that this is not a case about generative artificial intelligence, but that isn’t quite right. Ross Intelligence may not have been creating a chat bot in the mold of ChatGPT or Claude, but its brute force approach to AI training and its theory of transformative-ness is nonetheless an early staging of the legal theory that is beginning to play out across cases, including the noteworthy New York Times v. OpenAI case.
I’m not persuaded that the court gets it right. In its analysis of the first factor of fair use, the opinion draws some distinctions from other cases that have accepted a fair use defense in the context of digital technologies and, instead, veers towards the most recent (if confusing) pronouncement from the Supreme Court. I’m not sure that’s a good fit. I’ll leave this here as a summary of today’s new, but message me if you’d interested in chatting about the opinion.
