Diverse Amici Support ROSS at the Third Circuit

On September 29, ten amicus briefs were filed in Thomson Reuters v. ROSS Intelligence, supporting ROSS Intelligence and urging the U.S. Court of Appeals for the Third Circuit to overturn the district court’s controversial ruling. The amici represent a wide range of perspectives from academics and technical experts: (1) Heather Meeker; (2) Brian L. Frye, Jess Miers, & Mateusz Blaszczyk; (3) Edward Lee, Matthew Sag, Pamela Samuelson, Christopher Jon Sprigman, & Rebecca Tushnet; (4) Abraham Kang, Esq. & Kunal Patel; (5) Randy Goebel & Larry Ullman) to nonprofit organizations ((6) Authors Alliance; (7) Electronic Frontier Foundation, American Library Association, Association of Research Libraries, Internet Archive, Public Knowledge, & PublicResource.org; (8) Foundation for American Innovation) to technology companies ((9) Cicerai, Dispute Resolution AI, Free Law Project, Juristai Legal Technology Group, Paxton AI, & Trellis Research; (10) CCIA, Chamber of Progress, & NetChoice). They all share a few core arguments: Westlaw’s headnotes are not entitled to strong copyright protection, and ROSS’s use of them to train an AI-powered legal research tool constitutes fair use. 

In the district court decision, Judge Bibas largely sided with Thomson Reuters and granted most of its motion for partial summary judgment on direct copyright infringement. The court found that ROSS Intelligence infringed 2,243 Westlaw headnotes and that both the headnotes and Westlaw’s Key Number System met the minimal threshold for copyright protection. On fair use, the court held that the first and fourth factors—purpose and character and market effect—favored Thomson Reuters, while the second and third factors—nature and amount used—favored ROSS, but not enough to excuse the infringement. This 2025 ruling revised the court’s 2023 opinion, which had left key factual disputes for a jury and largely denied both parties’ motions for summary judgment. 

Originality and Copyrightability

Several amici, including Meeker, Frye et al., EFF et al., and Cicerai et al., contend that Westlaw’s headnotes fail to meet the originality threshold set forth by the Supreme Court in Feist Publications v. Rural Telephone. They argue that the headnotes are factual statements that closely mirror judicial opinions, which are public domain materials, and therefore lack the creative expression required for copyright protection. The briefs also invoke the “merger doctrine,” which holds that when an idea and its expression are inseparable, copyright protection does not apply. Meeker and Cicerai further criticize the district court’s analogy comparing headnote writing to Michaelangelo sculpting marble, arguing that creating headnotes is more akin to extracting facts from an already “highly sculpted” judicial opinion. 

Fair Use 

All amici maintain that ROSS’s use of Westlaw’s headnotes satisfies the four statutory fair use factors under Section 107 of the Copyright Act. 

Factor 1: Purpose and Character of the Use 

Lee et al., CCIA et al., Authors Alliance, and FAI argue that ROSS’s use was “highly transformative”. They emphasize that ROSS used the headnotes not to reproduce them, but to train its AI system to understand legal language and improve search functionality. This purpose, they argue, differs fundamentally from Westlaw’s use of the headnotes as human-readable summaries. The briefs cite precedents such as Authors Guild v. Google and Sega Enterprises v. Accolade, where courts found that intermediate copying for a new functional purpose constituted fair use. In Authors Guild, the Second Circuit held that “the creation of a full-text searchable database is quintessentially transformative use.” Likewise, in Sega, the Ninth Circuit determined that copying necessary to develop a competing, noninfringing product served a “legitimate, essentially non-exploitative purpose.” 

Factor 2: Nature of the Work 

These amici broadly agree that Westlaw’s headnotes are factual and constrained by the need for accuracy, giving them at most a “thin” layer of copyright protection (works that are composed largely of unprotectable elements are entitled only to limited protection). EFF and Cicerai argue that the district court gave insufficient weight to this factor, especially given that headnotes summarize judicial holdings and lack significant creative expression—far from “the core of intended copyright protection,” as the Supreme Court held in Campbell v. Acuff-Rose. 

Factor 3: Amount and Substantiality of the Portion Used  

Lee notes that ROSS used only 2,243 headnotes, which is a small subset of Westlaw’s overall database, and used them solely to train its AI model. The headnotes were neither included in nor displayed by ROSS’s final product. Amici cite Google v. Oracle, where the Supreme Court held that copying may be permissible when it is “tethered to a valid, and transformative, purpose.”

Factor 4: Effect on the Potential Market

Most amici reject the district court’s conclusion that ROSS’s system harmed Thomson Reuters’s market. They argue that ROSS’s AI does not output or display Westlaw’s headnotes and that competition in the legal research market is not cognizable under copyright law. Citing Sega  and Sony v. Connectix, they maintain that legitimate market competition is not an infringement. CCIA also points out that a use that “simply enables the copier to enter the market for works of the same type as the copied work” is not unfair. Several briefs further criticize the district court’s reliance on speculative “training-license” and “market dilution” theories because such reasoning improperly expands the scope of copyright protection. 

Conclusion: AI and Broader Policy Implications

Several briefs, including those by Kang & Patel, Goebel & Ullman, and FAI, criticize the district court’s analysis of AI training and generative AI. They argue that the court conflated AI training with generative content production. According to these briefs, ROSS’s system is a functional, search-oriented model that “transforms copyrighted text into mathematical vector embeddings” to improve semantic understanding, not a generative system that reproduces content. They warn that misclassifying AI training as infringing reproduction could stifle technological progress and innovation in legal research tools.

Finally, many amici highlight broader public policy implications of the case. They caution that extending copyright protection to short, factual summaries of judicial opinions could privatize access to the law, undermine the public domain, and restrict competition. They argue ROSS’s technology promotes accessibility to judicial opinions and fosters competition in the legal research market, benefiting smaller firms, independent practitioners, and the public. 

Thomson Reuters has since filed an unopposed motion seeking a 30-day extension to file its response brief, requesting that the deadline be extended to November 21, 2025.