Westlaw Headnotes Meet AI Training: Third Circuit Applies Traditional Fair Use Rules

Key Takeaway

• The Third Circuit held that Westlaw headnotes are copyrightable and ROSS’s use of those headnotes to train a competing legal-research platform was not fair use. The court applied traditional copyright principles and left unresolved whether training generative AI models presents a different fair-use analysis.

Why It Matters

• The decision shows that AI training does not displace the ordinary fair-use framework., In applying that framework, the court focused on ROSS’s competing purpose, the availability of uncopyrighted source material, the amount copied from each headnote, and the harm to Westlaw and a developing AI-training license market.

Next Steps

• Companies developing AI systems should evaluate whether copyrighted training material is needed, whether licensed or public-domain alternatives are available, and whether the use could affect the copyright owners’ existing or developing markets. Content owners should identify protectable human-created elements, consider registration, and assess whether AI-training rights present licensing opportunities.

The recent wave of companies using copyrighted material owned by others to train artificial intelligence models has reached the legal research market. In Thomson Reuters Enterprise Centre GmbH v. ROSS Intelligence Inc., the Third Circuit affirmed a ruling that Westlaw headnotes are eligible for copyright registration and that ROSS’s use of those headnotes to train its AI legal-research platform was not fair use.[1] The decision is an important addition to the developing law governing copyrighted material used for AI training, but its reach has limits: ROSS’s system was not a generative AI model, and the court distinguished cases involving large language models that can generate new copyright expressions.

What follows is a review of the Third Circuit’s decision and an outline for what companies should understand about protecting content and using materials subject to copyrights to train AI models as the law continues to develop.

Thomson Reuters v. ROSS Intelligence: How the Dispute Reached the Third Circuit

For decades, Westlaw has organized its judicial opinions through its Key Number System and headnotes. The underlying judicial opinions are not subject to copyright protection by Thomson Reuters, but its editors create headnotes that identify points of law and summarize the corresponding portions of opinions. ROSS, a new entrant in the legal-research market, sought to build an AI-powered search platform that would respond to plain-language legal questions with relevant passages from judicial opinions.

To train its system, ROSS hired a third party to create approximately 25,000 training memoranda. The memo writers used thousands of Westlaw headnotes to formulate legal questions and paired those questions with passages from judicial opinions. Thomson Reuters sued, alleging copyright infringement. ROSS argued that the headnotes lacked sufficient originality for copyright protection and, in any event, that its use of them was fair use.

The district court granted partial summary judgment to Thomson Reuters as to 2,243 headnotes. On interlocutory appeal, the issues before the Third Circuit were whether those headnotes were original works entitled to copyright protection and whether ROSS’s use was fair use. The court answered both questions in Thomson Reuters’s favor.

Originality: An Extremely Low Bar, but a Bar Nonetheless

The Third Circuit made short work of the originality issue. Copyright’s originality requirement is an “extremely low” bar[2]: a work must be independently created and possess a “modicum of creativity.”[3] The court held that the 2,243 headnotes met that standard because Westlaw editors made creative choices about which points of law to include and how to express those points while providing enough context for each headnote to stand on its own.[4]

The court also rejected ROSS’s merger argument. The merger doctrine can prevent copyright protection when an idea can be expressed in only one or a few ways. That was not the case with Westlaw’s headnotes because, as the court explained, “headnotes [can] have differing expressions.”[5] Copyright protection for the headnotes therefore did not give Thomson Reuters a monopoly over the law itself.

Fair Use: Training an AI Does Not End the Inquiry

The harder question the court grappled with was fair use. The court applied the familiar four-factor test: (1) the purpose and character of the use; (2) the nature of the copyrighted work; (3) the amount and substantiality of the portion used; and (4) the effect of the use on the potential market for or value of the copyrighted work.[6] Three of the four factors weighed against ROSS.

Under the first factor, the court held that ROSS’s use was commercial. ROSS intended to compete with Westlaw, charged comparable prices and hoped its platform would serve as a substitute. The court acknowledged that ROSS used the headnotes through an intermediate training step rather than displaying them to users in the same way Westlaw did. But that difference did not carry the day. Thomson Reuters used the headnotes to help researchers find and understand relevant judicial opinions; ROSS used the headnotes to train a platform designed to help users find responsive judicial-opinion passages. Because the uses shared the same ultimate purpose, the court found ROSS’s use “minimally transformative, at best.”[7]

The court also rejected ROSS’s reliance on cases involving intermediate copying of computer code. In those cases, copying was necessary to access unprotected functional elements.[8] ROSS, by contrast, had access to millions of judicial opinions that were not subject to copyright protection and could have used them to create training material. It chose the headnotes because they made that task easier. As the court put it, “[u]nlike necessity, ease is not a justification for copying.”[9]

The second factor, nature of the work, favored ROSS. Although the headnotes contained sufficient creativity to qualify for copyright protection, they were published and were more factual than fictional because they summarized points of law drawn from judicial opinions that were not subject to copyright protection.[10]

The amount and substantiality factor weighed against ROSS. The training memoranda copied the entire text of 25,000 Westlaw-written headnotes.[11] ROSS argued that this represented only a small fraction of Westlaw’s millions of headnotes. The court focused instead on what ROSS took from each individual copyrighted work: for each copied headnote, ROSS took the whole work. As the court held, “ROSS took more than necessary when it copied these headnotes without justifying transformative purposes.”[12]

ROSS fared no better with the fourth factor, market effect. The court rejected the premise that market harm should be limited to a standalone market for headnotes. The headnotes had value as a feature that attracted users to Westlaw, and ROSS used them to build a competing legal-research platform.[13] The court also recognized a developing derivative market for licensing copyright-protected text as AI training data. Thomson Reuters had used its headnotes to train its own AI search products, and its decision not to license the headnotes to others did not eliminate the potential licensing market. ROSS’s unlicensed use usurped Thomson Reuters’s opportunity to enter that market.[14]

What the Ruling Means for Companies

ROSS is not a blanket rule for generative AI. The court stressed that ROSS’s system did not generate new expression and distinguished cases involving large language models. The decision makes AI training subject to ordinary copyright rules, but it does not resolve whether training a generative AI model is fair use.

Content owners should identify and protect human-created value. The originality bar remains low. Companies that invest in summaries, annotations, taxonomies and other editorial content should document the creative choices behind those assets, consider copyright registration and assess whether AI-training rights may support a licensing market.

AI-training licensing markets can matter before they are mature. The court recognized a developing market for licensing headnotes as AI training data even though Thomson Reuters had not licensed the headnotes to third parties. The court’s analysis there is not that any hypothetical licensing market defeats fair use; on the contrary, the record showed that the market was developing and that Thomson Reuters already used its headnotes to train its own AI products. Therefore, content owners should consider whether training rights are a market worth developing or preserving, while AI developers should account for credible licensing markets when evaluating training data.

AI developers should focus on purpose and market substitution. Calling a use “training” does not make it transformative. Risk increases when material subject to copyright registration is used to build a commercial product that serves the same purpose as, or competes with, the copyright owner’s product.

Training-data choices matter. ROSS could have used judicial opinions not subject to copyright protection but chose Westlaw headnotes because they were easier to use. Companies should ask whether copyrighted material is necessary, whether licensed or public domain alternatives are available, and how much of each individual work is being copied.

Input rights and output rights are separate questions. Companies using AI should assess both whether they have the right to use training material and whether resulting content contains enough human authorship to receive copyright protection.