On February 11, 2025, a federal judge in Delaware handed Thomson Reuters a win that legal AI vendors had been watching closely. The court granted summary judgment against Ross Intelligence on copyright liability, rejecting the argument that training an AI-powered legal research tool on Westlaw's copyrighted headnotes was protected fair use. Judge Stephanos Bibas, a sitting Third Circuit judge handling the case by designation in the District of Delaware, found that more than 2,200 of the headnotes Ross's data vendor copied were substantially similar to Westlaw's originals, and that none of the four fair use factors saved Ross from liability. Commentators immediately flagged it as the first ruling anywhere to reject a fair use defense specifically for AI training data, and the first one centered on a legal AI product rather than a general-purpose chatbot.
The case had been working through the courts since 2020, long before AI training data was a common courtroom topic. Thomson Reuters sued Ross, a startup building a competing legal research tool, alleging Ross built its product by training on bulk memos that a third-party vendor, LegalEase, compiled largely by lifting Westlaw's editorially written headnotes and Key Number System classifications. This is not an abstract policy question for law firms, especially as they navigate the state of legal AI in 2026. It is a live signal about how the courts will treat any legal AI company that trained, or trains, on proprietary case law summaries, headnotes, or other secondary-source material it did not license.
Evaluating new technology requires looking beyond basic software features. This guide breaks down what the court actually decided, why the judge reversed his own earlier thinking, how the ruling impacts legal AI vendors, and how buyers can run proper due diligence on vendor training data.
What the court actually held
The lawsuit is Thomson Reuters Enterprise Centre GmbH v. Ross Intelligence, Inc., No. 1:20-cv-00613, in the U.S. District Court for the District of Delaware. Judge Stephanos Bibas presided over the matter by designation from the Third Circuit Court of Appeals.
The dispute arose because Ross wanted to build an AI-powered system to challenge Westlaw's dominant position in the market for best AI legal research tools. Because Ross could not secure a license from Thomson Reuters to use Westlaw content, it hired LegalEase, a third-party research vendor. LegalEase produced bulk legal memos for Ross. Thomson Reuters argued that these memos were constructed by copying Westlaw's proprietary headnotes and its Key Number System, which is Westlaw's proprietary classification taxonomy.
On February 11, 2025, Judge Bibas granted partial summary judgment to Thomson Reuters. According to the official court opinion, the court found that 2,243 of 2,830 Westlaw headnotes at issue were copied and substantially similar to the text in the memos Ross used for training. A reasonable jury could only conclude that copyright infringement had occurred on those headnotes.
Before reaching the fair use question, the court addressed a threshold issue. Ross argued that Westlaw's headnotes were not copyrightable because they merely restated public domain judicial opinions. The court rejected this argument. Judge Bibas held that the headnotes cleared the minimal creativity threshold required by copyright law because they introduce creativity by distilling, synthesizing, or explaining parts of a judicial opinion rather than merely restating it. After establishing that the headnotes were copyrightable, the court rejected Ross's fair use defense as a matter of law.
The four-factor fair use analysis
To evaluate whether Ross's use of the Westlaw headnotes was legally permissible, the court conducted a traditional four-factor fair use analysis.
Factor 1: purpose and character of the use
This factor weighed for Thomson Reuters. The court found that Ross's use was commercial, was not transformative, and was designed to compete directly with Westlaw in the same market. The court rejected the argument that intermediate, non-public copying during AI model development is automatically transformative just because the end output looks different from the training input.
Factor 2: nature of the copyrighted work
This factor favored Ross. The court found Westlaw's headnotes and Key Number System have more than the minimal originality needed for copyright protection, but are far less creative than original prose because they stay close to the underlying facts and law. Because the law gives thinner protection to less creative, fact-bound works, the court weighed this factor for Ross, though it noted the factor carries comparatively little weight in the overall balance.
Factor 3: amount and substantiality of the portion used
This factor cut in Ross's favor. Because Ross's finished AI search tool did not surface or display the Westlaw headnotes themselves to end users, the court found the public exposure of the copied work was minimal, even though the headnotes had been used internally to train the system.
Factor 4: effect on the market for the original
This factor weighed for Thomson Reuters. The court ruled that even though Thomson Reuters did not yet operate an active licensing market for its content as AI training data, Ross's unauthorized use could harm or foreclose a potential market for licensing that data to AI developers in the future.
Net result: the four factors split evenly, two for each side. Factors 1 and 4 favored Thomson Reuters; factors 2 and 3 favored Ross. Thomson Reuters won the balance because the first and fourth factors are generally considered the most heavily weighted in fair use analysis, while the court found factors 2 and 3 carried comparatively little weight by comparison. On that balance, the court held that no reasonable jury could find for Ross on fair use.
The unusual self-reversal by the judge
This was not the judge's first word on fair use in this case. According to a client alert by Skadden, Judge Bibas had declined to rule for either side on fair use in 2023, finding genuine disputes of material fact that meant the question belonged to a jury at trial. This left open the possibility that Ross could win on fair use.
Before the scheduled trial, the judge revisited his own earlier ruling and, in the February 2025 opinion, withdrew it. He stated he had reached this new conclusion based on further reflection and a fuller understanding of the law as it applied to the facts, granting summary judgment to Thomson Reuters instead.
Legal commentators described the reversal as unusual and noted the judge was candid that he had changed his mind. The same set of facts, looked at twice, produced opposite conclusions, showing how unsettled this exact area of fair use law is even within a single judge's reasoning. This procedural twist is a major reason why the case is being closely watched on appeal.
What it signals for other legal AI vendors
The decision has immediate implications for the broader legal technology market, though its ultimate impact remains to be seen.
The ruling is specific to Ross's facts: a non-generative AI search tool, trained on a competitor's proprietary editorial content, built by a company that tried and failed to get a license first, for a product designed to compete directly with the rightsholder. As noted by Perkins Coie, this outcome does not resolve fair use questions for generative AI models generally, which involve different products, training methods, and litigation. It does not automatically doom every other legal AI vendor's training choices.
Despite its narrow facts, the ruling is a real warning shot. Legal AI vendors frequently train or fine-tune on proprietary case law databases, court-record aggregations, and secondary sources like headnotes, practice guides, treatises, or annotations that are not in the public domain. The court's holding that editorial synthesis of public-domain case law can itself be copyrightable, and that using it to build a competing product is not protected by fair use, applies directly to those data collection practices. This includes vendors building specialized software such as AI contract review and drafting tools that rely on structured legal examples.
The role of the upcoming Third Circuit appeal
The district court ruling is not the final word. Ross obtained permission to take an interlocutory appeal to the Third Circuit, where the case is docketed as No. 25-2153. The Third Circuit heard oral argument on June 11, 2026, making this the first time a U.S. court of appeals will rule on fair use and AI training data.
A Third Circuit reversal or narrowing would change the calculus industry-wide, while an affirmance would harden this into the first appellate precedent on AI training data fair use. No appellate decision has been issued as of this writing.
The fourth factor's reasoning, that potential, not-yet-existing licensing markets count toward market harm, is likely to be the most consequential and contested part of the ruling on appeal. It could be read broadly to cover almost any rightsholder who could hypothetically license content for AI training in the future, presenting a much larger claim than simple harm to an existing market.
What it means for law firm buyers
For law firm buyers, this litigation shifts how technology must be evaluated. Firms must look beyond basic software features to assess the compliance and liability of their software supply chain.
Historically, buyers evaluated legal AI tools on output quality, such as hallucination rates and citation accuracy. To understand how vendors measure these metrics, firms can read our analysis of how accurate is legal AI.
However, training-data provenance is the question sitting underneath accuracy. A tool trained on properly licensed primary law and original analysis is a different risk profile than one trained on scraped or improperly sourced proprietary secondary sources. Buyers should ask vendors directly what data their models were trained or fine-tuned on, and whether it was licensed. Evaluating technology in this space is notoriously difficult, as explained in our guide on why almost no legal AI tool has reviews, making direct vendor questioning essential.
Firms cannot assume they are protected from copyright liability by default. IP indemnification covering training-data and output infringement claims varies significantly across AI vendor contracts. Many standard agreements only protect the firm if the tool's final output happens to reproduce copyrighted text. They do not cover claims arising from the vendor's unauthorized use of training data. Buyers should specifically ask whether a vendor's indemnification clause covers claims arising from the training data itself, and whether that indemnity is capped or carved out of the contract's general liability cap.
Under professional responsibility standards, lawyers have a continuous duty to understand the technologies they adopt. As detailed in our guide on legal AI ethics rules, ABA Formal Opinion 512 requires lawyers to understand an AI tool's capabilities and limitations before using it. This competence obligation naturally extends to understanding where the underlying model's data came from and what legal exposure that creates for the firm if the vendor loses a copyright suit.
The dispute between Thomson Reuters and Ross is part of a larger wave of copyright lawsuits targeting AI companies. The New York Times has a pending copyright suit against OpenAI and Microsoft over AI training on its journalism. Additionally, a group of authors sued Anthropic for training its models on pirated books. That case, known as Bartz v. Anthropic, resulted in a $1.5 billion settlement to resolve claims regarding approximately 482,000 copyrighted works, which was preliminarily approved by a federal judge in September 2025, according to Wolters Kluwer's legal analysis. While these consumer-facing cases are massive, the Westlaw headnotes lawsuit remains the most relevant precedent for law firms because it directly targets the legal research and analysis tools that firms buy and use daily.
Before signing a contract with any legal AI vendor, procurement teams should ask three specific questions:
- Can the vendor provide a plain, written description of its training data sources and confirm that all proprietary data was properly licensed?
- Does the contract include an IP indemnification clause that explicitly covers claims arising from training-data sourcing, rather than just the final output?
- Is the vendor currently involved in, or threatened by, any active copyright litigation regarding its data collection practices?
If a vendor cannot answer the first question, that response is a critical data point that should factor into the firm's final purchasing decision.
FAQ
What did the court actually decide in Thomson Reuters v. Ross Intelligence?
On February 11, 2025, Judge Stephanos Bibas of the U.S. District Court for the District of Delaware granted summary judgment to Thomson Reuters, finding that Ross Intelligence infringed Thomson Reuters's copyright in more than 2,200 Westlaw headnotes and that Ross's fair use defense failed as a matter of law. The case is now on appeal to the Third Circuit.
Is this the same as a jury verdict or a final ruling?
No. This was a summary judgment ruling on liability and the fair use defense, decided by the judge rather than a jury, and Ross has appealed it to the Third Circuit Court of Appeals, which heard oral argument on June 11, 2026. The outcome of that appeal, not the district court ruling alone, will determine how lasting this precedent is.
Why is it notable that the same judge ruled differently earlier in the case?
In 2023, Judge Bibas had denied summary judgment on fair use, finding genuine factual disputes that meant the question should go to a jury. In the February 2025 opinion, he reversed that earlier conclusion himself and ruled for Thomson Reuters as a matter of law instead. Legal commentators flagged the self-reversal as unusual and as evidence of how unsettled fair use doctrine still is for AI training data, even within one judge's own reasoning on the same facts.
Does this ruling mean every legal AI tool trained on proprietary case law data is now illegal?
No. The ruling is specific to Ross's facts: a non-generative legal search tool trained on a direct competitor's proprietary headnotes and classification system, built after Ross failed to obtain a license, for a product designed to compete directly with Westlaw. It does not automatically resolve fair use questions for generative AI models trained differently, and the broader legal question is still being litigated on appeal.
What should a law firm ask a legal AI vendor before signing a contract, in light of this ruling?
Ask what data the tool was trained or fine-tuned on and whether it was licensed, whether the contract's IP indemnification explicitly covers claims arising from training-data sourcing (not just from the tool's output), and whether the vendor is a party to or otherwise exposed by pending AI copyright litigation. Training-data provenance is now a due-diligence question alongside accuracy and security, not a separate concern only legal departments at AI companies need to think about.
The bottom line
The dispute between Thomson Reuters and Ross Intelligence is a milestone in the development of legal technology. It is the first federal ruling to reject a fair use defense for AI training data, and it focuses entirely on the specialized software products that law firms rely on every day. While consumer tech disputes capture broader headlines, this case directly shapes the legal tech marketplace.
The ruling is narrower than the headlines suggest. It turns on Ross's specific facts, including a failed attempt to license the data and a product built to compete directly with the rightsholder. Furthermore, it is not yet final. The Third Circuit heard oral arguments on June 11, 2026, and a reversal or narrowing of the decision remains a real possibility.
For law firm partners and IT directors, the practical lessons do not depend on the outcome of the appeal. Understanding training data provenance and securing robust contract indemnification are now standard requirements for technology procurement. Taking these steps protects firms from legal liabilities and ensures that their chosen AI solutions are built on a sustainable, legally compliant foundation.