Delaware court finds fair use defence unsuccessful in AI training copyright motion

The main three questions addressed by the Court in its decision were as follows: 

1. Were the Westlaw headnotes original and therefore protected by copyright?  
2. If so, was the copyright in such headnotes infringed by Ross? 
3. If so, could Ross rely on the fair use defence?

In short, the Court found that the headnotes were indeed original, on the basis that the threshold for originality is “extremely low“, requiring only “some minimal degree of creativity … some creative spark” (per Feist, 499 U.S. 340). 

As a headnote is a short, key point taken from a lengthy judicial opinion, there is editorial judgment exercised in what to cut away and what to leave in place. Further, the Court found that a headnote can introduce creativity by “distilling, synthesizing, or explaining part of an opinion”. This editorial expression has enough of a “creative spark” to be original. 

Having found the Westlaw headnotes to be original, the Court turned to the question of whether there had been “copying of constituent elements of the work that are original” (Feist), which would involve a determination of both (a) actual copying; and (b) substantial similarity (per Dam Things, 290 F.3d 548).

The Court did not apply its analysis to all c. 21,000 headnotes alleged by Reuters to be infringing, which will be a matter for the jury, but to the batch of 2,830 identified by Reuters for the purposes of the summary judgment motion. 

Addressing first actual copying, the Court noted that this can be proven directly or indirectly, by showing that the defendant had access to it and produced something similar (Tanksley v Daniels, 902 F.3d 165). The parties were agreed that LegalEase had access to Westlaw and used it when creating the Bulk Memos. 

Where a Bulk Memo question looks more like a Westlaw headnote than it does the underlying judicial opinion, this is strong circumstantial evidence of actual copying. On this basis, the Court went through all 2,830 headnotes individually, and found that there has been actual copying of 2,243 of them. 

Turning to substantial similarity, the question was whether an ordinary user of a product would find it substantially similar to the copyrighted work (Dam Things, at 562). However, more similarity is required when less protectable matter is at issue (per Nimmer on Copyright §13D.32[A]). Even applying this standard, the Court was willing to grant summary judgment on the 2,243 headnotes on the basis that substantial similarity was “so obvious that no reasonable jury could find otherwise”. This was because the language very closely tracked the language of the Bulk Memo question but not the case opinion. 

Having dismissed Ross’ various other defences, including innocent infringement and copyright misuse, the Court turned to the key question of whether Ross could rely on the fair use defence. The Court noted that this is an affirmative defence, for which Ross bore the burden of proof (Video Pipeline, 342 F.3d 191) and which is to be considered on the basis of at least four fair use factors, namely: 

1. the use’s purpose and character, including whether it is commercial or non-profit; 
2. the copyrighted work’s nature; 
3. how much of the work was used and how substantial a part it was relative to the copyrighted works’ whole; and
4. how Ross’ use affected the copyright work’s value or potential market. 

The Court decided the second and third factors in Ross’ favour. Firstly, although the headnotes could be original, they were not that creative, and secondly the output to Ross’ end user was a judicial opinion, and not a headnote, so it communicated little sense of the original work. 

As for the fourth factor, which is “undoubtedly the single most important element of fair use” (Harper & Row, 471 U.S. 539), the Court focused on the fact that Ross had intended to compete with Westlaw by creating a market substitute. Even if Reuters had not yet used its data to train its own legal search tools, the likely effect of Ross’ use on a potential market for AI training was sufficient. 

Balancing its findings on the four factors, the Court granted summary judgment for Reuters on fair use, and on direct infringement for the 2,243 headnotes. The remaining factual questions in the case will proceed to trial. 

Although the Court was very keen to emphasise that this was not a case regarding generative AI, clearly it would be highly significant for the spate of GAI-related copyright infringement cases pending before the US courts, if those courts were to apply a similar reasoning, particularly in respect of intermediate copying within the context of the first factor of fair use analysis for GAI-training. Although it is likely too early to tell, given the different technologies involved, the claimants in such cases are likely to be feeling buoyed by the Court’s decision, and it will be interesting to see whether it precipitates a greater number of settlements or position changes by defendants.