ROSS Technology Lawsuit: Reversal, Fair Use, and AI Copyright

The ROSS Intelligence lawsuit is a federal copyright case filed by Thomson Reuters in December 2020 accusing the AI legal research startup of copying Westlaw’s editorial content to train a competing product. In February 2025, Judge Stephanos Bibas of the U.S. District Court for the District of Delaware rejected ROSS’s fair use defense, producing the first substantive U.S. ruling on whether copyrighted material can be used to train an artificial intelligence system without a license. The case is now on interlocutory appeal at the Third Circuit, which heard oral arguments on June 11, 2026.¹

Who ROSS Intelligence Was

ROSS Intelligence was founded in 2014 by Andrew Arruda, Jimoh Ovbiagele, and Pargles Dall’Oglio out of the University of Toronto.² The company built an AI-powered search engine that let lawyers type legal questions in plain language and receive relevant judicial opinions in response. Critically, the tool was non-generative. It retrieved and ranked existing case law rather than writing new content, a distinction that would matter later in court.³

Thomson Reuters owns Westlaw, one of the two dominant legal research platforms in the United States. The dispute centers on two Westlaw products: headnotes, which are attorney-written summaries of the legal points decided in judicial opinions, and the Key Number System, which classifies those headnotes by area of law.⁴

How ROSS Obtained Its Training Data

ROSS first tried to license Westlaw content directly from Thomson Reuters. It was refused because ROSS was building a competing product.³ After the refusal, ROSS hired a third-party outsourcing firm called LegalEase Solutions to produce roughly 25,000 “Bulk Memos,” each consisting of a legal question paired with ranked answers drawn from case law.⁴

LegalEase gave its contributing lawyers a guide that used Westlaw headnotes as a reference point but prohibited direct copying and pasting. Thomson Reuters alleged that the resulting questions closely tracked headnote language anyway, characterizing them as “essentially headnotes with question marks at the end.”⁴ In its later appellate brief, Thomson Reuters said LegalEase contractors used bots to scrape Westlaw content in bulk and converted headnotes into question-and-answer pairs for ROSS’s training pipeline.⁵

ROSS used about 80 percent of the Bulk Memos for training and 20 percent for validation. Its process converted the text into numerical representations of relationships among legal words and ran them through a machine-learning system called a Featurizer. The headnotes themselves never appeared in output shown to end users. They served solely as training inputs.³

The 2023 Ruling and Its 2025 Reversal

The case was assigned to Third Circuit Judge Stephanos Bibas, sitting by designation. On September 25, 2023, Judge Bibas largely denied both sides’ motions for summary judgment. He found that ROSS and LegalEase had “actually copied” portions of the Bulk Memos from Westlaw headnotes, but concluded that substantial similarity, copyright validity, and fair use all involved factual disputes for a jury.⁴

The trial was postponed. The court invited renewed summary judgment briefing, and on February 11, 2025, Judge Bibas issued a revised opinion overruling much of his own 2023 decision.³

The court held that Westlaw headnotes and the Key Number System are copyrightable. Applying the “extremely low” originality threshold from the Supreme Court’s Feist decision, Judge Bibas concluded that the editorial judgment in creating a headnote is a “creative spark” even when the headnote quotes a judicial opinion word for word. Identifying the key legal point and “chiseling away the surrounding mass” reflects original expression.³

On direct infringement, the court granted summary judgment to Thomson Reuters for 2,243 of the 2,830 headnotes at issue, finding both actual copying and substantial similarity. Judge Bibas concluded that the Bulk Memo questions tracked the headnote language rather than the language of the underlying opinions and that “no reasonable jury could find otherwise.” The court also rejected ROSS’s defenses of innocent infringement, copyright misuse, merger, and scènes à faire.³

Why the Fair Use Defense Failed

The most closely watched part of the ruling rejected ROSS’s fair use defense across all four statutory factors. The court also reversed its earlier position, holding that fair use is a question for the judge rather than a jury.

On the first factor, purpose and character, the court found ROSS’s use commercial and not transformative. Judge Bibas distinguished prior “intermediate copying” precedents involving computer code, such as Google v. Oracle and Sega v. Accolade, where copying functional code was necessary to achieve compatibility. Here the copied material was written text, and the copying was not “reasonably necessary” for innovation. Because ROSS’s tool performed a function similar to Westlaw’s own use of headnotes to return case lists, the use lacked a “further purpose or different character.”⁶

The second factor, the nature of the copyrighted work, favored ROSS. The court acknowledged that headnotes are “not that creative” compared to novels or other artistic works. The third factor, amount and substantiality, also favored ROSS, because the headnotes never appeared in the tool’s final output and so did not serve as a “competing substitute” for the headnotes themselves.⁶

The fourth factor, market effect, was decisive. The court called it “undoubtedly the single most important element” and concluded that ROSS intended to build a market substitute for Westlaw. Even though Thomson Reuters did not have an existing market for licensing its data as AI training material, the potential for such a market was enough to weigh against ROSS. The court also rejected ROSS’s public-benefit argument, stating that while the public has an interest in accessing the law, it has no right to Thomson Reuters’s “parsing of the law.”⁶

Judge Bibas noted explicitly that only non-generative AI was before him and reserved judgment on whether the same reasoning would apply to generative AI systems that create new content.⁷

What Happened to ROSS

The lawsuit ended the company. In December 2020, the same month the complaint was filed, ROSS announced it would stop accepting new customers. By January 31, 2021, it ceased offering its legal research platform entirely, citing the financial burden of defending the litigation.⁸ CEO Andrew Arruda said at the time that ROSS intended to “make a comeback” if it prevailed in court. No bankruptcy filing has been reported, but the company is listed as inactive.⁹

The Third Circuit Appeal

On May 23, 2025, Judge Bibas certified two legal questions for interlocutory appeal under 28 U.S.C. § 1292(b) and stayed the case. The questions are whether the Westlaw headnotes and Key Number System are original as a matter of law, and whether ROSS’s use of the headnotes constitutes fair use. Judge Bibas acknowledged that his own prior opinions had reached different conclusions on these “novel and difficult questions of first impression” and that ROSS faced a potential “existential” financial threat from trial costs.¹⁰

ROSS filed its opening brief on September 22, 2025, arguing that headnotes consist of “verbatim or close-to-verbatim quotes from uncopyrightable judicial opinions” and that granting copyright protection would give Westlaw a monopoly over access to the law. On fair use, ROSS contended that training its AI engine was “quintessential fair use” that “radically promoted scientific progress.”¹¹ Thomson Reuters countered by characterizing ROSS’s actions as “theft, not innovation,” arguing that ROSS knowingly bypassed licensing restrictions after being explicitly refused a license.⁵

The Electronic Frontier Foundation, the American Library Association, the Association of Research Libraries, the Internet Archive, Public Knowledge, and Public.Resource.Org filed amicus briefs supporting ROSS, arguing that headnotes “simply restate individual points from judges’ opinions with no meaningful creative contributions.”¹² The Third Circuit heard oral arguments on June 11, 2026, and no decision has been issued.¹

What the Ruling Means for AI Copyright Cases

The February 2025 ruling is the first substantive federal decision rejecting a fair use defense for AI training data.¹³ Its reach may prove narrow. In Bartz v. Anthropic PBC and Kadrey v. Meta Platforms, California district judges distinguished the ROSS decision by focusing on the generative nature of the AI models at issue. Both courts found that generative AI, which produces new content rather than retrieving existing material, presented a stronger case for transformative use.¹⁴

For now, the ruling signals that when an AI tool acts as a market substitute for the copyrighted source material, and when the training data copying was not functionally necessary for innovation, fair use faces steep headwinds. The Third Circuit’s decision, when it comes, will be the first federal appellate ruling on these questions.¹³

1
CourtListener. Thomson Reuters Enterprise Centre GmbH v. Ross Intelligence Inc.
2
ROSS Intelligence. About Us
3
U.S. District Court for the District of Delaware. Thomson Reuters v. Ross Intelligence, Memorandum Opinion
4
Justia. Thomson Reuters Enterprise Centre GmbH et al v. ROSS Intelligence Inc., No. 1:20-cv-00613-SB
5
LawNext. Thomson Reuters Tells Appeals Court ROSS’s Copying Was Theft Not Innovation
6
James Grimmelmann. Thomson Reuters v. Ross Intelligence, Edited Case
7
Skadden Arps Slate Meagher & Flom LLP. Court Reverses Itself in AI Training Data Case
8
Legal Tech News. ROSS Shuts Down Operations Citing Financial Burden From Thomson Reuters Lawsuit
9
Y Combinator. ROSS Intelligence
10
Justia. Thomson Reuters Enterprise Centre GmbH et al v. ROSS Intelligence Inc., Memorandum Opinion, D.I. 804
11
Courthouse News Service. AI Company Tells Appeals Court Its Decision in Legal Research Copyright Case Will Have Sweeping Consequences for Innovation
12
Electronic Frontier Foundation. Protecting Access to Law and Beneficial Uses of AI
13
Sterne Kessler Goldstein & Fox. AI IP Year in Review: First Federal Ruling Rejects Fair Use Defense for AI Training Data in Copyright Dispute
14
White & Case LLP. Two California District Judges Rule Using Books to Train AI Fair Use