IP monitor: Training AI / machine learning models and copyrighted materials: A Canadian perspective on recent US decision

For the first time, a US court has found that copyright was infringed by copying works for the purpose of training an AI legal research tool. The US District Court for the District of Delaware granted summary judgment in favour of Thomson Reuters Enterprise Centre GmbH and West Publishing Corp (Thomson Reuters) finding that ROSS Intelligence Inc.’s copying of Westlaw headnotes to train its AI legal research tool constituted copyright infringement and none of the fair use, innocent infringement, copyright misuse, merger, or scènes à faire defences absolved the infringement.

Factual background

Thomson Reuters owns one of the largest legal research platforms (Westlaw). As part of its offering, it creates headnotes of cases that summarize key points of law and case holdings. 

ROSS initially sought to license Westlaw's content to train its AI tool, but Thomson Reuters refused. ROSS commissioned "Bulk Memos" from a third party, LegalEase. These Bulk Memos were lawyers’ compilations of legal questions created using Westlaw’s headnotes. The Bulk Memos were used to train ROSS’ AI search tool; however, the tool responds to queries by providing full cases and the headnotes are not part of its output.

Procedural history

Thomson Reuters moved for summary judgment, which was largely denied in 2023. Interestingly, the court invited the parties to renew their briefings and both parties moved again for summary judgment. 

Revisiting its 2023 decision, the court granted summary judgment on Thomson Reuter’s claims for copyright infringement in 2,243 headnotes, and further that none of ROSS’ defences apply.

Copyright infringement

Originality: The court found the headnotes in issue met the low hurdle of originality. Also at issue was Westlaw’s key number system, a taxonomy used to organize materials in Westlaw. The court found this number system also met the threshold for originality, even though a “rote computer program did the work.” It was sufficient there were many possible ways to organize legal topics and Thomson Reuters chose this particular one. 

Copying: The court inferred that 2,243 headnotes had been copied on the basis that the Bulk Memos were so similar to the headnotes and dissimilar to the underlying opinion, that a reasonable jury had to find that actual copying had occurred.

Substantial Similarity: The court acknowledged the less protectable expression a work contains, the greater the similarity required to find infringement. On this basis, it found the Bulk Memos in question were substantially similar to 2,243 headnotes. 

Accordingly, the court found infringement by Ross of those headnotes. While not explicit, the court notes in its assessment of the fair use defence, that the level of originality of the headnotes was low, suggesting that in assessing substantial similarity, the court was looking for (and found) significant overlap between the Bulk Memos and these 2,243 headnotes. While additional headnotes were in issue, the court found there were open factual questions that had to be resolved at trial and declined to grant summary judgment on the balance of the headnotes in issue.

Fair use defence

The court also granted summary judgment and rejected ROSS' fair use defence. 

The court noted the fact that ROSS did not make the headnotes available in its output, and that the level of creativity in the headnotes was low weighed in favour of fair use; however, on balance, the court found fair use did not apply.

The most significant factors were that ROSS’ use of the headnotes was commercial and not transformative. ROSS used the headnotes to create a competing legal research tool and did not alter the purpose or character of the original work. In determining that ROSS’ use was not transformative, the court distinguished cases where copying occurred as an intermediate step, explaining the work copied in the cases relied on by Ross related to computer code, which almost always serves functional purposes, and because the copying in those cases was necessary for the infringers to innovate. The court found that copying the headnotes was not reasonably necessary to achieve ROSS’ purpose.

Other defences

Finally, the court also rejected all of ROSS' other defences.

Innocent Infringement: This defence could not be used to limit damages because the headnotes had copyright notices.

Copyright Misuse: ROSS failed to demonstrate that Thomson Reuters misused its copyrights to stifle competition. 

Merger: The court found there were many ways to express points of law, so the ideas did not merge with the expression. 

Scènes à Faire: The court ruled nothing about a judicial opinion required it to be slimmed down to headnotes or categorized by key numbers. 

Conclusion

This is the first case in which copyright infringement has been found in the context of training an AI tool. While the copying and training occurred through manual processes and not data scraping or crawling, the case has set goalposts on key issues of copyright in the context of AI, most notably on the limitations of certain defences in this area.

For more information, please contact your IP professional at Norton Rose Fulbright Canada LLP.

For a complete list of our IP team, click here.