Thomson Reuters wins an early AI copyright battle for use of headnotes in training of computer models

Score one for the “team media” in the Great AI Wars. On February 11, the US District Court for the District of Delaware issued partial summary judgment in favour of Thomson Reuters in their copyright infringement claim against ROSS Intelligence for ROSS’ use of case headnotes derived from Westlaw. (Thomson Reuters Enterprise Centre GmbH v. Ross Intelligence Inc.)

ROSS Intelligence was a startup company that made legal research products to help users find, summarize, and organize case law. Among ROSS’s features was a tool that let users ask legal questions (such as, say, “What is the threshold for originality under copyright law?”), which would then be interpreted into a search for relevant case law. The Court found that ROSS’s use of Westlaw headnotes to train its AI search tool infringed Thomson’s copyright in the headnotes and rejected all of their defences (innocent infringement, copyright misuse, merger, scenes a faire, and fair use).

I pause to note that ROSS’s tool is perhaps what we would have called a machine-learning-assisted search engine, and is not a Generative AI / Large Language Model tool. Nonetheless, it shares some similar features such as the bulk use of data for training.

Thomson’s infringement case in chief was relatively straightforward – although ROSS did not directly obtain or copy the headnotes from Westlaw (in fact, ROSS tried to get a license from Thomson but the request was refused because they were direct competitors), they obtained a database of legal questions from a third party, LegalEase, who did have a license to Westlaw and had a database of “legal questions” which were more or less Westlaw’s headnotes turned into a question format:

It had long been held under U.S. law that while judicial decisions themselves are not copyrightable, summaries of those cases, including headnotes, were copyrightable so long as they exhibited the requisite originality. (This is more or less the same as under Canadian law.) The Court reiterated that these headnotes were subject to copyright, and it compared each of the headnotes to the impugned “questions” for similarity. Since ROSS had copied the LegalEase database into its own computers for training (a fact that was readily inferable from the record), “actual copying” and “substantial similarity” were both relatively straightforward findings.

As an aside, although it did not play a big role in the judgment (the infringement and defences mainly focused on the headnotes), the court also ruled that the Key Number System used by Westlaw to index cases was also sufficiently original to warrant copyright protection. There remain some questions at law about how much these systems are purely functional (thus not copyrightable) and/or are not sufficiently original or expressive for copyright protection. The analysis by the court was fairly short, relying mostly on the idea that there were many other ways to create such a system. The impact of this aspect of the judgment remains to be seen.

ROSS’s defences were roundly rejected:

1. Innocent infringement: based on the copyright notice on Westlaw’s headnotes, and the fact that this was a summary judgment on liability, the innocent infringement defence against damages did not apply.
2. Copyright misuse: the court had previously ruled that there were no viable antitrust claims and relatedly, found here that Thomson did not misuse copyright to stifle competition.
3. Merger defence: while ROSS claimed that the ideas in the headnotes were so close to their expression that they ‘merged’ (thus not copyrightable), the court rejected this as it found there were many ways to summarize judicial opinions.
4. Scenes a faire: ROSS argued that using “stock elements” from the field was not infringement (this defence historically applied to things like use of certain tropes in historical romance novels, for example), but the court held the defence “did not fit”.

The Court spent most of the judgment on ROSS’s fair use defence. This portion of the judgment is likely to become influential – particularly the discussion relating to how ROSS obtained the information and copied it onto its computers for training of its models, despite the fact that the output of a ROSS query would not reproduce the headnotes (or the questions used in its training models) for the user to read.

On the four fair use factors (commercial nature, transformativeness, amount of use, and effect on the market), ROSS actually succeeded on two grounds (second and third), but the first and fourth grounds dominated the analysis and were in favour of Thomson Reuters.

Some interesting observations:

• Intermediate copying: ROSS argued that their use of the works was transformative – they never showed the headnotes to the users, since they were only used for training and verification. They contended that this was only an intermediate copying like in the case of software reverse engineering. But the court held this was no defence – this was not quite the same as the cases where companies copied copyrighted source code for the purpose of developing a compatible API (its own category of acts exempted from infringement). Indeed, this argument has some superficial attraction, after all, isn’t using case summaries to find cases the same thing as using an API to interact with a program? Well, the analogy broke down because copying headnotes, unlike copying code, isn’t necessary for this purpose. As the judge heavily implied, ROSS could have done their own hard work of summarizing all of the cases in its database to develop their own headnotes and queries rather than take a shortcut.
• Value and market: This was the most determinative factor. Thomson’s case was strong here because ROSS was a direct competitor who intended to take market share from Westlaw. Their legal search tool served the same purpose as Westlaw as a market substitute. Although there is public interest in accessing legal opinions, this was outweighed, in the court’s view, by Thomson’s right to profit from its own copyrights.

Balancing the various fair use factors, the court rejected ROSS’s defence and granted summary judgment in favour of Thomson Reuters. In the result, ROSS was held to have infringed on some 2,243 headnotes.

Practically, though, the result of this case was unlikely to change anything – ROSS had announced its shutdown in 2021 when Thomson Reuters filed its Complaint against ROSS. (ROSS also filed a counterclaim for antitrust, which was dismissed by a separate summary judgment.) It seems the parties were nonetheless content to keep fighting out this case, perhaps in hopes of generating favourable precedent on either side.

Conclusion: In these early days of AI-related judgments, it’s hard to tell how the courts will decide. But, if this decision is any indication, bulk copying of copyrighted works – even for training of computer models – is going to be a problem for the AI industry.