Analyzing One of the First AI Copyright Cases: the Thomson Reuters v. Ross Intelligence Copyright Infringement Case

Want More of a Deep Dive? View or Download the Podcast:

Analyzing the Thomson Reuters v. Ross Intelligence Copyright Infringement Case

Today, we’re diving into a fascinating and complex copyright case that sits at the intersection of artificial intelligence (AI) and intellectual property law: Thomson Reuters Enterprise Centre GmbH v. Ross Intelligence Inc. I’m thrilled to unpack this recent court opinion from Judge Stephanos Bibas in the United States District Court for the District of Delaware, issued on February 11, 2025. This case not only pits two legal research companies against each other but also raises critical questions about the use of copyrighted material to train AI systems—a topic increasingly relevant in our tech-driven world.

In this blog post, I’ll summarize the key rulings and findings of the opinion, analyze whether the judge’s decisions are logical and fair to both parties, and explore a specific point of contention: whether Westlaw’s headnotes should be considered transformative (spoiler: the court didn’t rule them as such, but I’ll clarify this misunderstanding from the query). Most importantly, I’ll provide insights and predictions about how this ruling might influence future copyright cases, particularly those involving AI and large language models (LLMs), and assess whether the commercial nature of Ross’s use disproportionately swayed the outcome. This post is designed to be comprehensive yet accessible, so whether you’re a legal professional, an AI developer, or simply curious, there’s something here for you. Let’s get started!

Background of the Case

Thomson Reuters, through its subsidiaries Thomson Reuters Enterprise Centre GmbH and West Publishing Corp., owns Westlaw, a powerhouse in legal research platforms. Westlaw provides subscribers with access to an extensive database of legal materials—case law, statutes, regulations, journals, and treatises—enhanced by editorial content like headnotes and the Key Number System. Headnotes are concise summaries of key legal points distilled from judicial opinions by Westlaw’s editors, while the Key Number System is a proprietary taxonomy organizing legal topics numerically. Both are copyrighted components of Westlaw’s offerings.

Enter Ross Intelligence Inc., a startup aiming to disrupt the legal research space with an AI-powered search engine.

To train its AI, Ross needed a robust dataset of legal questions and answers. It approached Thomson Reuters to license Westlaw’s content but was rebuffed due to their competitive relationship. Undeterred, Ross partnered with LegalEase, a third-party service, which provided “Bulk Memos”—compilations of legal questions and answers crafted by lawyers using Westlaw headnotes as a foundation. LegalEase instructed these lawyers to draw from headnotes without copying them verbatim, and Ross ultimately used around 25,000 Bulk Memos to train its AI tool.

When Thomson Reuters discovered this, it sued Ross for copyright infringement, alleging that the Bulk Memos infringed on its copyrighted headnotes. The case progressed through the District of Delaware, with an initial 2023 ruling largely denying summary judgment. However, after deeper scrutiny, Judge Bibas revised his stance in 2025, issuing a new opinion that shifted the tide significantly in Thomson Reuters’s favor.

Summary of the Court’s Rulings

Judge Bibas’s revised opinion, issued under Fed. R. Civ. P. 54(b), reflects a thoughtful reassessment of the case. Here are the key rulings:

1. Copyright Validity and Originality:

   Westlaw’s headnotes and Key Number System are original works eligible for copyright protection. The court found that the editorial selection and arrangement in headnotes, even when quoting judicial opinions, and the organizational choices in the Key Number System meet the “minimal degree of creativity” required under Feist Publications, Inc. v. Rural Telephone Service Co. (499 U.S. 340, 345 (1991)).

   
   

2. Direct Copyright Infringement:

   The court granted partial summary judgment to Thomson Reuters, finding that Ross infringed 2,243 headnotes. For these, the Bulk Memo questions were substantially similar to the headnotes, not the underlying judicial opinions, evidencing actual copying. Factual disputes remain about copyright expiration for some headnotes, reserved for trial. Summary judgment was denied on other headnotes (e.g., a batch of 5,367) and the Key Number System due to unresolved factual issues.

   
   

3. Defenses to Infringement:

   Ross’s defenses—innocent infringement, copyright misuse, merger, and scenes à faire—were rejected. Notably, innocent infringement didn’t apply because Westlaw’s headnotes bear copyright notices, and no evidence supported misuse or the other doctrines.

   
   

4. Fair Use Defense:

   The court rejected Ross’s fair use defense, granting summary judgment to Thomson Reuters. Analyzing the four statutory factors under 17 U.S.C. § 107:

   • Purpose and Character of Use: Ross’s commercial, non-transformative use favored Thomson Reuters.

   • Nature of the Copyrighted Work: The headnotes’ moderate creativity slightly favored Ross.

   • Amount and Substantiality: Ross’s non-public use of headnotes favored Ross.

   • Effect on the Market: Potential harm to Westlaw’s market, including AI training data licensing, favored Thomson Reuters.

     
     

     Factors one and four outweighed two and three, defeating fair use.

     
     

This revised ruling marks a significant pivot from 2023, reflecting Judge Bibas’s deeper engagement with the evidence and legal principles.

Analysis of Copyright Validity and Originality

The court’s finding that Westlaw’s headnotes and Key Number System are copyrightable hinges on originality—a bedrock of copyright law requiring independent creation and a “minimal degree of creativity” (Feist, 499 U.S. at 345). Judicial opinions, as public domain works, aren’t copyrightable, but Thomson Reuters argued that its editorial enhancements are.

Judge Bibas employs a striking analogy:

“A block of raw marble, like a judicial opinion, is not copyrightable. Yet a sculptor creates a sculpture by choosing what to cut away and what to leave in place. That sculpture is copyrightable. So too, even a headnote taken verbatim from an opinion is a carefully chosen fraction of the whole.”

This reasoning is spot-on. Crafting a headnote involves distilling a lengthy opinion into a concise, focused summary—a creative act of selection and expression, even if the words mirror the original. The Key Number System, too, clears the originality bar through Thomson Reuters’s unique organizational choices, despite overlapping with common legal categories.

Novel Insight: The sculptor analogy elevates the discussion beyond mere overlap with source material, emphasizing the editorial judgment as a protectable expression. This could extend copyright protection to other curated outputs—like AI-generated summaries—if they involve similar creative distillation.

Actual Copying and Substantial Similarity

To prove infringement, Thomson Reuters had to show (1) actual copying and (2) substantial similarity between the Bulk Memo questions and the headnotes. The court’s analysis here is meticulous, comparing 2,830 headnotes individually and finding 2,243 infringed.

• Actual Copying: Evidence included LegalEase’s access to Westlaw and Ross’s expert conceding similarity between Bulk Memo questions and headnotes, not judicial opinions. Judge Bibas’s hands-on comparison confirmed this for 2,243 headnotes.

• Substantial Similarity: As an “ordinary user” of Westlaw (a lawyer and judge himself), Bibas assessed whether the questions materially appropriated the headnotes’ original expression. He granted summary judgment only where similarity was undeniable, leaving closer calls for trial.

  
  

Insight: This granular approach—reviewing thousands of headnotes—sets a high bar for precision in copyright cases involving large datasets. It’s a model for future disputes where AI training data comprises numerous discrete works, ensuring fairness through detailed scrutiny.

Fair Use Defense: A Detailed Examination

Fair use, under 17 U.S.C. § 107, balances four factors. The court’s rejection of Ross’s defense is a cornerstone of the opinion.

Factor 1: Purpose and Character of Use

• Commerciality: Ross admitted its use was commercial, aiming to profit without licensing costs.

• Transformativeness: The court found Ross’s use non-transformative, as its AI tool competed with Westlaw without altering the headnotes’ purpose. Ross argued intermediate copying (for training, not output) justified fair use, citing Google v. Oracle (593 U.S. 1 (2021)). Bibas distinguished this case: unlike code copied for interoperability, Ross’s copying wasn’t necessary, as it could have created its own data.

  
  

Factor 2: Nature of the Copyrighted Work

• The headnotes’ moderate creativity (less than a novel, more than a phonebook) favored Ross slightly, though this factor rarely sways outcomes.

  
  

Factor 3: Amount and Substantiality

• Since Ross used headnotes internally without public disclosure, this favored Ross. The focus was on substitution, not quantity copied.

  
  

Factor 4: Effect on the Market

• The court highlighted harm to Westlaw’s existing platform and potential AI training data market—a derivative use Thomson Reuters could pursue. This outweighed public benefit arguments, as judicial opinions are already accessible.

  
  

Balancing: Factors one (commercial, non-transformative) and four (market harm) trumped two and three, rejecting fair use.

Insight: The distinction from code cases underscores a functional divide—AI training data disputes may hinge on necessity, not just intent.

Does the Opinion Make Sense and Is It Fair?

The opinion is legally sound and equitable:

• Logical Consistency: The sculptor analogy justifies headnote originality, and the fair use analysis aligns with Warhol (598 U.S. 508 (2023)), prioritizing purpose and market effects. Distinguishing code cases is principled, given the lack of necessity here.

• Fairness to Thomson Reuters: Protecting headnotes rewards their investment, preventing free-riding by competitors.

• Fairness to Ross: The court’s cautious summary judgment (only 2,243 of 2,830 headnotes) and reservation of factual disputes for trial ensure Ross isn’t unduly prejudiced.

Critique: Some might argue intermediate copying for training warrants more leniency, akin to reverse-engineering cases. However, here the commercial intent and viable alternatives (creating original data) justify the ruling in this case.

Clarifying the Transformative Misunderstanding

Several clients have asked whether Judge Bibas found Westlaw’s headnotes transformative—a misreading of the opinion. The court didn’t assess the headnotes’ creation as transformative; it ruled them original for copyright purposes. Transformativeness arose in the fair use analysis of Ross’s use, which was deemed non-transformative because it mirrored Westlaw’s purpose without adding new character.

Why Not Transformative for Ross? Ross’s AI didn’t repurpose headnotes creatively (e.g., for art or commentary) but used them to build a competing tool. This aligns with Warhol’s focus on differing purposes, not just differing forms.

Insight: Conflating originality with transformativeness is a common pitfall. Here, headnotes are original works, but Ross’s use didn’t transform them, reinforcing the distinction’s importance in AI contexts.

Implications for Future AI and Copyright Rulings

This ruling reverberates beyond legal research into AI and LLM development:

1. Training Data Scrutiny:

   Using copyrighted material for AI training risks infringement if it’s commercial and non-transformative. Companies must secure licenses or use public domain data, especially as courts recognize potential markets like training data licensing.

   
   

2. Generative vs. Non-Generative AI:

   Ross’s non-generative AI (retrieving opinions) contrasts with generative LLMs (creating new content). Generative AI might argue transformativeness if outputs don’t replicate inputs, potentially shifting fair use outcomes. This case flags this divide for future exploration.

   
   

3. Commerciality’s Weight:

   Commercial use amplifies infringement risk absent transformation. Non-commercial uses (e.g., academic research) might fare better, though market harm remains key.

   
   

4. Necessity as a Factor:

   Unlike Google v. Oracle, where copying enabled interoperability, Ross’s avoidable copying weakened its defense. Future cases may hinge on whether copyrighted data is the only viable training source.

Predictions:

• LLM Cases: Generative AI lawsuits (e.g., against ChatGPT) may see plaintiffs arguing input similarity, while defendants push output transformation. Courts might split outcomes based on replication versus innovation.

• Legislative Push: Persistent disputes could prompt Congress to clarify fair use for AI training, balancing innovation and IP rights.

• Industry Shift: AI firms may prioritize synthetic or licensed data, reducing legal exposure.

The Impact of Commercial Use on the Ruling

Ross’s commercial intent was pivotal. Under factor one, the court cited Warhol: when original and secondary uses align in purpose and the latter is commercial, fair use falters. Ross’s aim to profit without licensing costs, directly competing with Westlaw, tipped this factor decisively.

Oversized Effect? Not quite. Commerciality was balanced against transformativeness and market harm. A non-commercial use (e.g., research) might have softened factor one, but factor four’s market impact—crucial per Harper & Row (471 U.S. 539 (1985))—would likely still favor Thomson Reuters. The commercial lens amplified, but didn’t solely dictate, the outcome.

Insight: Commerciality’s role reflects copyright’s economic roots—protecting creators’ markets. In AI disputes, it’ll remain a litmus test, though not the sole arbiter.

Conclusion

The Thomson Reuters v. Ross Intelligence opinion is a landmark in copyright and AI law. By affirming headnote originality, finding infringement on 2,243 instances, and rejecting fair use, Judge Bibas crafts a precedent that safeguards creative works while challenging AI developers to innovate responsibly. The ruling makes sense legally, fairly balances the parties, and sets a cautious tone for AI training data use.

For future cases, it signals rigorous scrutiny of purpose, transformation, and market effects—especially for commercial AI. As LLMs proliferate, this case hints at a bifurcated future: generative AI may find fair use footholds where non-generative tools stumble. Meanwhile, the commerciality factor underscores copyright’s protective ethos, urging ethical data practices.

Please let me know what you think!  Am I off base?  Should the court have provided more leniency?  Let me know.