On February 11, a federal judge handed down a ruling in what may turn the tide against artificial intelligence companies claiming a “fair use” defense to the act of copying huge quantities of third-party content to train their technology. The decision in Thomson Reuters Enterprise Centre GmbH et al v. ROSS Intelligence Inc., No. 1:20-cv-613-SB (D. Del. Feb. 11, 2025) includes a thoughtful analysis of the fair use defense to copyright infringement.
The case arose because ROSS Intelligence, a legal research startup, wanted to create an affordable product competitive with Thomson Reuter’s Westlaw platform. Westlaw uses a headnote system to organize legal concepts with short blurbs of text. ROSS asked Westlaw for a license to the headnotes, which Westlaw declined. ROSS then hired a third party to “scrape” the headnotes and used this data to train its AI model. The Westlaw headnotes do not appear in ROSS’s final end-user product.
In the analysis of whether ROSS’s use of Westlaw’s headnotes was “transformative” for the purposes of fair use, the court found that it was not transformative almost entirely on the grounds that the AI was “not generative.” Rather than creating something new, ROSS’s program operated by simply spitting out information in its database in response to a query, much like Westlaw does. The court didn’t say whether the analysis would be different if ROSS’s program was generative AI, potentially leaving the door open for generative AI companies like Stability AI to save their transformative use defense.
In its defense, ROSS argued that copying Westlaw’s headnotes was merely an “intermediate step” to getting to a new and different product, reemphasizing that the Westlaw headnotes did not show up in the final product. This “intermediate step” argument has been used in fair use cases successfully before.
However, the Court took issue with this argument in two ways. Firstly, previous “intermediate step” cases have dealt specifically with software, such as copying the source code of a video game for the sole purpose of making the game compatible with an upgraded console. The Court cites the Google v. Oracle case for the proposition that software is “just different” from other copyrightable material like literary works. (An issue here is that Congress has not made this distinction, and there is therefore no statutory direction for how software should be treated differently in copyright cases.)
Secondly, the Court pointed out that in prior “intermediate step” cases, the “step” in question was necessary to some degree (e.g., you couldn’t solve compatibility issues without the source code of the original software). In ROSS, that was not the case, and the tech startup therefore lost its fair use argument.
This case will certainly chill fair use claims for non-generative AI companies who scrape third-party content for training purposes. It will likely also put a damper on fair use claims from generative AI companies, although as mentioned before, there is perhaps still an opening for these companies. In any event, the ROSS case highlights the continued need for a comprehensive, federal statutory copyright regime for software.
If you have any further questions about artificial intelligence or intellectual property law, please contact Tara Aaron-Stelluto.
Tara Aaron-Stelluto