[This post is authored by Akshat Agrawal and Sneha Jain. Akshat Agrawal is a Senior Associate at Saikrishna and Associates and has previously written on the blog here. Sneha is a Partner at Saikrishna and Associates and leads their Artificial Intelligence and Law Practice. Views expressed here are those of the authors alone. Long post ahead.]
The US District Court for the District of Delaware’s recent opinion in Thomson Reuters and West Publishing v. Ross Intelligence Inc. carries significant implications for Generative AI and pending Copyright cases worldwide. The Court has granted summary judgment in respect of 2,830 “headnotes” belonging to Thomson Reuters and admittedly used by Ross Intelligence to train its Natural Language Processing and Artificial Enabled Legal Research tool, finding Direct Copyright Infringement and rejecting fair use.
But, first, Some History!
On September 25, 2023, Judge Bibas rendered his initial opinion on summary judgment motions concerning direct infringement and fair use. The core dispute centered on Thomson Reuters’ Westlaw “headnotes”—concise summaries of legal points in judicial opinions, hyperlinked to corresponding passages.
During its development, Ross Intelligence engaged “Legal Ease” to create memos with questions lawyers might ask, alongside answers featuring verbatim quotes from legal opinions. These “Legal Ease” memos were converted into machine learning training data to enable Ross’s AI to retrieve relevant judicial opinions for user search queries matching the questions.
Thomson Reuters argued that Legal Ease’s questions, incorporated by Ross in machine-readable form, essentially copied its creative headnotes. They contended that extracting relevant key phrases from judicial opinions to create summarized headnotes demanded significant skill, labor, and creativity. Ross undisputedly did not display the headnotes in its output using them solely for training its software through linguistic pattern recognition for accurate search functionality. The dispute involved the use of 2,830 allegedly original Thomson Reuters headnotes.
This post will analyze the Court’s revised 2025 opinion, highlighting:
- The flawed originality analysis;
- Misapplication of the infringement standard;
- Inadequate consideration of merger and fair use defenses;
- Contrast with the more nuanced Indian law approach;
- Negative implications for AI innovation.
The post will also look at how contention of parties would fare under Indian law.
The 2025 Reversal: Judge Bibas’ Change of Heart
In a moment of judicial introspection, Judge Bibas presented an illustration central to his reconsideration and revised opinion, juxtaposing a Legal Ease question crafted for Ross’s software training, a Thomson Reuters headnote, and the corresponding judicial opinion text. The illustration forming part of the actual issue at hand is:
Judge Bibas candidly states that his previous perspective on originality had been fundamentally flawed. Initially believing originality “depended on how much the headnotes overlap with the uncopyrightable text of opinions,” his evolved understanding now recognizes that extracting a concise legal point from an extensive judicial opinion represents an act of intellectual distillation—a creative endeavor synthesizing and explaining complex legal reasoning.
The Court’s new perspective asserted that a headnote, even if verbatim from the judicial opinion, embodies a carefully curated fraction of a lengthy opinion. By identifying the most consequential words and principles, such a headnote reflects a significant creative exercise meriting copyright protection, transcending mere textual reproduction to recognize the intellectual labor of legal synthesis.
Critical Analysis of the Court’s Reasoning
This approach mirrors India’s “brain work” doctrine of originality from Eastern Book Company v. D.B. Modak [Para 41]. However, the key reflection of this “brain work” lies in displaying these headnotes—potentially incorporating verbatim quotes or trivial paraphrases—alongside the judgment for navigation, not in the headnotes or their use ipso facto.
While granting copyright to headnotes, the Court overlooked that the key “brain work” wasn’t in the headnotes’ expression itself (being trivial variations of judgment quotes) but in their creative display alongside judicial opinions for easier navigation. The headnote’s ‘expressive’ essence lay in its summarizing nature, enabling users to grasp the judgment’s crux without reading the entire text, combined with the minimal creativity in sequentially displaying it along opinions for easier navigation.
The Copyright Infringement Analysis
In holding summary judgment for infringement, the Court fundamentally erred by ignoring that even if Westlaw headnotes were copied, mere copying would not ipso facto constitute infringement if not used for any ‘expressive’ objective/intent or to portray aesthetic and creative expression. The Court recognized this principle when considering the “amount and substantiality” fair use factor, stating “What matters is not the amount and substantiality of the portion used in making a copy, but rather the amount and substantiality of what is thereby made accessible to a public for which it may serve as a competing substitute.” However, this point was completely missed when evaluating prima facie infringement.
The Court’s analysis reveals a misunderstanding of how technological innovation intersects with copyright law. Undisputedly, the Westlaw headnotes were transformed into a numerical, machine-readable format to enable Ross Intelligence’s AI software to learn linguistic syntax and patterns. The fundamental purpose was to create an effective search functionality—a process that necessarily requires extracting key linguistic patterns from judicial opinions.
What the Court fundamentally misses is the dual nature of these headnotes. Beyond their expressive content—which helps readers quickly comprehend judicial opinions—headnotes serve a critical functional significance in enabling accurate search engine results. The Court fell into what Professor Oren Bracha describes as “copy fetishism”—the fallacious notion that “a copy is a copy is a copy”—without recognizing the nuanced, functional purpose of the use of headnotes.
Critically, not a single human accessing Ross’s platform would be exposed to Westlaw headnotes. The copying was purely instrumental, aimed at training an AI system to recognize and retrieve relevant legal information efficiently. By failing to distinguish between ‘expressive’ reproduction and ‘functional’ utilization, the Court risks creating a precedent that could significantly impede functional uses of creative works in process of technological innovation–something beyond the purpose, scope and conceptualization of copyright law.
Analysis of Legal Defenses
The Merger Doctrine
The Court’s dismissal of the merger doctrine was a missed opportunity to address an issue intertwined with legal and technological colors. Effective legal search engines require incorporating significant phraseology and core holdings from judicial opinions.
The Westlaw headnotes substantially incorporated essential phraseology from the judicial opinions themselves—a fact that demands a more nuanced examination of the merger doctrine. For any search engine to provide accurate responses, incorporating significant legal phrases and fundamental holdings is not just beneficial, but essential. The Court’s cursory treatment failed to recognize the intrinsic challenge of distilling legal principles without relying on their original textual expression.
This analysis ignored the fundamental tension between copyright protection and the functional uses in the process of technological innovation. The substantial overlap between headnote expression and the original judicial opinion warranted a more sophisticated and detailed exploration of the merger doctrine’s applicability.
The Fair Use Analysis
- Purpose and Character of Use
On the first factor concerning purpose and character, the Court held that Ross’ use was for creating a competing legal research tool that functioned similarly to how Westlaw uses headnotes to return fitting cases. While noting that headnotes didn’t appear in Ross’s final product and copying occurred at an intermediate step, the Court rejected the intermediate copying defense from software cases as inapplicable, claiming computer programs to be different from other literary works. The Court also restricted the intermediate copying defense to cases where copying was “necessary” for innovation.
This analysis:
1. Relies too heavily on similar functionality of headnotes to improve search tools, ignoring that such functional aspects fall outside copyright’s scope and within patent law’s domain;
2. Ignores its own recognition that headnotes, like computer software, served both aesthetic and functional purposes;
3. Interprets the ‘necessity’ requirement too restrictively.
The Court’s undue emphasis on “necessity” for intermediate copying has been consistently diluted in progressive legal interpretations. The 9th Circuit in Sony Computer Entertainment v. Connectix Corporation cautioned against “wasted effort” and “artificial hurdles” to accessing ideas in copyrighted software. The Court held:
“This is precisely the kind of “wasted effort that the proscription against the copyright of ideas and facts is designed to prevent.” Such an approach would erect an artificial hurdle in the way of the public’s access to the ideas contained within copyrighted software programs. These are “aspects that were expressly denied copyright protection by Congress.” We decline to erect such a barrier in this case. If Sony wishes to obtain a lawful monopoly on the functional concepts in its software, it must satisfy the more stringent standards of the patent laws. “
In Consumers Union of U.S., Inc. v. General Signal Corp., the Court held that copying exact words may be the only valid way to precisely report evaluations. Lexmark International v. Static Control Components also, while holding that a modification chip designed to make third party ink cartridges compatible with Lexmark printers was not primarily designed to circumvent copyright, emphasized focusing on practical, not theoretical, alternatives within real-world constraints while evaluating the “necessity” requirement.
The Court also missed a similar situation in the Google Books (Authors Guild v. Hathi Trust) case (which it, in fact, refers to while evaluating the “Amount and Substantiality of Use Factor”) where use of copyrighted works and even its expressive display (albeit partially) in the quest of enabling search functionality on the Google Books search engine was held to have a transformative function and purpose that was non-infringing under Copyright law.
Crucially, the Court ignored that headnotes, like computer software, serve both aesthetic and functional purposes. By creating an arbitrary distinction between literary works and technological innovations, the judgment risked stifling the very innovation copyright law should encourage.
- Market Effect Analysis
The Court’s analysis of market effect created additional concerns by introducing a “potential derivative market” for AI training data. Without recognizing this market’s functional rather than expressive nature, the Court held this factor against Ross based on potential competition with Westlaw.
Creating a market in expressive works for AI training assumes that training AI using copyrighted works constitutes expressive use amounting to infringement. This premise and approach misunderstand copyright’s core purpose: protecting ‘expressive’ use for human consumption, not regulating ‘functional’ technological development.
The Indian Law Perspective
The Indian legal framework offers a particularly sophisticated approach to understanding copyright in technological contexts. The Supreme Court’s landmark judgment in R.G. Anand v. Deluxe Films provides critical guidance on what constitutes an infringing copy. The Court recognized that infringement is not merely about copying ideas or embedded information, but about adopting the manner, arrangement, and specific expressive elements with minimal alterations in a subsequent expressive work.
Section 2(m) of the Indian Copyright Act presents a nuanced definition of an ‘infringing copy’ that fundamentally differentiates between reproduction and mere copying. This distinction is crucial in technological contexts like AI training. The law suggests that reproduction becomes problematic only when it involves creating/or intending to create an expressive work that substantially imitates the original’s form of expression.
The Madras High Court’s judgment in M/s Blackwood and Sons Ltd. v. A.N. Parasuraman further refined this understanding. The Madras High Court interpreted infringing reproduction as reproduction that amounts to publication, as opposed to reproduction for private use or without publication intent. Section 3 of the Act defines publication as “making a work available to the public by issue of copies or by communicating the work to the public.”
This interpretation creates a critical distinction: the mere act of making a copy is not inherently infringing if such copy is not exposed or intended to be exposed for human consumption. The exclusive reproduction right of a copyright owner is arguably intrinsically linked to an element of publication / intent of communication. The work need not be published to be protected, but for an infringement claim under the bracket of the “reproduction right”, it necessarily requires publication or the intention of publication of a subsequent work that incorporates expressive aspects of the primary original work (which may be published or unpublished). For the “communication to the public” bracket, the communication of the said unpublished work of the copyright owner would be infringing. For e.g., reproduction or communication right in an unpublished work is arguably not infringed if the original expressive work is not exposed to a human by a subsequent work, but the communication right (and not the reproduction right) would be infringed even if one human unauthorizedly copies the work as the original expressive work then stands published to that one human.
Transformative Use Under Indian Law
While transformative use is not explicitly part of Section 52 defenses in Indian copyright law, courts have increasingly used the concept to restrict the scope of copyright protection itself. The Calcutta High Court’s approach in Barbara Taylor Bradford v. Sahara Media Entertainment significantly limited reproduction and adaptation rights for transformed purposes.
In the context of AI training, Ross Intelligence’s use of Westlaw Headnotes presents a compelling case for transformative use:
a) Fundamental Purpose Transformation: The use of copyright-protected content for AI training represents a fundamentally different purpose from the original publication. The objective is to train an artificial intelligence tool without exposing the expressive form to human readers—a stark departure from the original communicative intent.
b) Objective and Intent Differentiation: The reproduction serves a substantially different objective. While the functional purpose might overlap, the intent is distinctly technological—to enable machine learning and pattern recognition rather than human comprehension.
c) Output Distinctiveness: The subsequent output—judicial opinions extracted through AI processing—bears no substantial similarity to the original headnotes’ expressive form.
d) Deliberate Non-Reproduction: Ross Intelligence’s careful effort to prevent headnote regurgitation demonstrates a clear intent to transform, rather than merely copy, the original content.
This approach aligns with the fundamental principle that copyright protection should not impede technological innovation and extended to functional uses of creative expression. By focusing on the transformative nature of AI training, the law can balance the interests of content creators with the imperative of technological advancement.
Conclusion
While the Court carefully caveated that the judgment does not specifically address Generative AI, the implications are profound and potentially far-reaching. The findings about potential AI training markets and intermediate copying could create significant impediments to technological innovation. The ruling fundamentally miscomprehends the essential distinction between functional and expressive uses of “expression” in copyright law.
Most critically, the decision fails to recognize that copyright law’s primary purpose is protecting expressive human consumption—not regulating functional human or machine learning processes. This fundamental misunderstanding threatens to create artificial barriers in the AI industry, potentially stifling the very innovation that drives technological progress.
Although the modus in the facts at hand is significantly different from “scraping the whole internet” type Generative AI models, however, this decision serves as a cautionary tale—a reminder of the need for a thoughtful, purpose-driven understanding of copyright protection. Only by aligning legal doctrine with the realities of technological transformation can we build a framework that fosters both creative expression and groundbreaking innovation.