The Third Circuit Court of Appeals has dealt a significant blow to the 'fair use' defense often cited by AI startups, ruling that Thomson Reuters' headnotes in Westlaw are copyright-protected and that ROSS Intelligence's use of them for model training constituted infringement. The decision, affirmed by U.S. Circuit Judge Tamika Montgomery-Reeves, rejects the notion that intermediate steps in AI training automatically qualify as transformative. For developers and infrastructure engineers, this clarifies that scraping proprietary datasets to build competing products carries substantial legal risk.

The 'Creative Spark' Standard

The court focused heavily on the originality of Westlaw's headnotesโ€”short descriptions of legal issues preceding judicial opinions. While ROSS argued these notes were too similar to the uncopyrightable judicial opinions themselves, the panel disagreed. Montgomery-Reeves wrote that the editors' selection and arrangement of these notes demonstrated a 'creative spark,' akin to a sculptor chiseling a raw block of marble. This establishes that curated, structured data derived from public domain sources can still be protected if the human editorial input shows minimal creativity.

Market Competition Undermines Fair Use

A critical factor in the ruling was the commercial intent behind the training data. The court found that ROSS did not use the headnotes for legal research but to 'rush out a competing product.' Because the AI training served the same ultimate purpose as Thomson Reuters' original businessโ€”providing legal research toolsโ€”the use was deemed 'minimally transformative at best.' The panel emphasized that ROSS' use negatively impacted Westlaw's market value, particularly in the potential derivative market of licensing headnotes for AI training.

Key Takeaways

  • Transformation is insufficient: Simply using data to train a model does not make the use transformative if the end goal is direct market competition.
  • Curated data is protected: 'Creative spark' in selection and arrangement grants copyright protection to structured datasets, even if underlying facts are public.
  • Supreme Court appeal likely: ROSS counsel Yar Chaikovsky stated the company will seek Supreme Court review, citing 'continued uncertainty' around AI copyright law.

The Bottom Line

If you are building an AI product, do not assume 'fair use' is a safety net for training on proprietary competitor data. The Third Circuit has signaled that if your model competes with the source of the training data, you likely need a license, not a legal theory.