An Ordinary Copyright Case: Westlaw Headnotes and AI Training in the Third Circuit

An AI copyright defendant would usually rather its case be about the future. ROSS Intelligence put it that way to the Third Circuit, and on September 29 the court declined the invitation in the second paragraph of its opinion affirming partial summary judgment for Thomson Reuters: “Under ROSS’s framing, this case appears to concern the future of AI legal technology. But appearances can be deceiving. In truth, this is no more than an ordinary copyright case.”¹

The court meant it. The opinion announces no special rule for machine learning and treats the training step as one more fact to be weighed under the four statutory factors. That makes it easy to read as a narrow decision about an old-fashioned product: an AI that could not write anything, trained to compete with the publisher whose work it copied. In one respect that reading is right. In two others it undersells what the court decided.

IWhat was actually before the court

The posture comes first, because it does much of the limiting. The district court — Circuit Judge Stephanos Bibas, sitting by designation — granted Thomson Reuters partial summary judgment, holding as a matter of law that ROSS’s use of 2,243 Westlaw headnotes infringed Thomson Reuters’s copyright and was not fair use. At the parties’ request it then certified two questions for interlocutory appeal under 28 U.S.C. § 1292(b): whether the headnotes and the West Key Number System are original as a matter of law, and whether ROSS’s use of the headnotes was fair.²

The appeal then narrowed further. For purposes of the appeal, ROSS did not dispute that the copying done by its vendors was attributable to it, or that the vendors copied Westlaw headnotes. It never briefed the originality of the Key Number System, and the court treated that issue as forfeited. And the 2,243 headnotes were a filtered set. A ROSS expert had identified a batch of 2,830 headnotes whose training questions closely resembled the headnote text; the district court kept only those for which no reasonable juror could conclude the headnote had not been copied. None of the 2,243 copies opinion text verbatim.³

The facts were not in dispute in any way that mattered. Westlaw’s editors write each headnote to state a point of law, under guidelines that call for concision, a limit of 800 characters where possible, and enough context to “stand on its own.” ROSS built a search engine that answered plain-language legal questions with passages from roughly ten million judicial opinions. Its AI “was not a generative AI, meaning it would not create any new expression; it would only return text passages from preexisting judicial opinions.” To train it, a vendor, LegalEase, wrote some 25,000 memos, each posing a legal question and labeling four to six opinion passages as great, good, topical or irrelevant. The memo-writers used thousands of Westlaw headnotes to frame the questions, because the headnotes provided “an easy way [to] fram[e] questions.” ROSS ran advertisements comparing itself directly to Westlaw at prices “in line with” Westlaw’s, and some law firms switched.⁴

IIOriginality, and the question the court left open

The originality holding is the least surprising part of the opinion. Under Feist the bar is “extremely low”: independent creation and a modicum of creativity, “some creative spark.” The court found that spark in the editors’ decisions about which points of law were important enough to include and how to word them, so that each headnote would stand alone while accurately reflecting the passage it summarized. It found support in Callaghan v. Myers, where the Supreme Court recognized in 1888 that a reporter may hold a copyright covering “the matter which is the result of his intellectual labor,” headnotes included, and in Georgia v. Public.Resource.Org, which read Callaghan the same way.⁵

ROSS’s three answers each failed in a sentence or two. The monopoly argument: “Headnotes are not law; judicial opinions are.” The merger doctrine: there are many ways to express a point of law, just as — in the court’s own earlier example — there are many ways to design a banana costume. And Matthew Bender & Co. v. West Publishing Co., the Second Circuit decision denying West copyright in its arrangement of party names and its parallel citations, turned on industry conventions that dictated those choices; the same opinion described headnotes as “independently composed.”⁶

The more interesting sentence is in footnote 5. The district court had “mused” that even headnotes quoting opinions verbatim might be original enough for protection. The Third Circuit called that dictum, said it “raises an interesting question,” and declined to answer it, because the summary judgment covered only headnotes that do not copy opinion text verbatim. The reservation matters more than its placement suggests. The editorial guidelines quoted in the opinion tell Westlaw’s editors that they “[g]enerally . . . follow the court’s language.” A headnote written that way sits close to the line the court declined to draw. In the Editor’s view, that is where the next dispute over headnotes is likely to be fought: how much of a headnote must be the editor’s own words before it is protected at all.⁷

IIIPurpose: the training step does not change the answer

On fair use, the first factor carried most of the argument. ROSS’s use was commercial: it set out to build a legal-research platform charging comparable prices to compete with Westlaw for customers. Commercial use is not dispositive, so the question became how far ROSS’s purpose differed from Thomson Reuters’s, with that difference balanced, under Andy Warhol Foundation v. Goldsmith, against the commercial nature of the use.⁸

ROSS’s best point was that it did not use the headnotes the way Westlaw does; it used them to train a model. The court accepted that ROSS took “an intermediate step” that “arguably presents a slight degree of difference in use,” and then looked through it: “both Thomson Reuters and ROSS use the headnotes to create and optimize a legal-research platform that helps users find responsive legal material.” Same ultimate purpose, and so a use “minimally transformative, at best.”

That method, more than the result, is what other courts are likely to take from the first-factor analysis. The court did not ask what the training did to the headnotes. It asked what the trained product was for, and compared that purpose with the purpose of the original. On the court’s approach, a model trained to do what the original already does shares the original’s purpose, whatever happens in between.

Two lines of authority could have pointed the other way, and the court distinguished both. Authors Guild v. Google involved a search tool that served a different function from the books it scanned and might send readers to buy them; ROSS’s platform does what Westlaw does, and “does not lead a user to Westlaw; instead, as ROSS admits, it aims to replace Westlaw.” The intermediate-copying cases — Google v. Oracle and the Ninth Circuit’s decisions in Sega and Connectix — involved copying computer code that was necessary to reach unprotected functional elements and make software work with an existing system. ROSS did not need the headnotes to reach anything unprotected. It had the opinions, which it was free to copy, and it chose the headnotes because they offered an “easy” way to build its memos.⁹

A footnote adds that the undisputed evidence showed ROSS “at times acted in bad faith,” including attempts to access Westlaw with law-firm investor credentials despite notice that Westlaw’s terms of service prohibited it, and that to the extent good faith still matters after Oracle, it weighed against ROSS.¹¹

IVNature and amount: the unit is the headnote

The second factor went to ROSS. The headnotes were published, and they are more factual than fictional because they must accurately convey the law. The court noted, quoting the Second Circuit, that the second factor “has rarely played a significant role,” and in the end it weighed only “slightly” in ROSS’s favor.¹²

The third factor is where the opinion does something a generative-AI defendant should notice. ROSS argued that it took a trivial share of Westlaw: 0.08 percent of Thomson Reuters’s 28 million headnotes. The court answered in two steps. First, even an insubstantial portion is not fair if an important part of the work is taken, and “[t]he crux of the inquiry is whether ‘no more was taken than necessary’” for the copier’s purpose. Because the opinions were freely available, copying the headnotes was not necessary at all. Second, in a footnote, the court affirmed that each headnote is itself a copyrightable work, so that “for each headnote taken, ROSS copied an entire work.”¹³

The second step changes the arithmetic. Measured against a corpus of millions, any training set looks small. Measured against the individual work, each item copied whole is a whole work taken. The court did not limit that reasoning to non-generative systems, and nothing in it depends on what the trained model later produces.

VMarket harm, and a market for training data

The fourth factor produced the opinion’s broadest holding. ROSS argued that the relevant market was a market for headnotes as a standalone product, which, ROSS said, does not exist. The court assumed as much and held that it did not matter: under its own decision in Video Pipeline, the statute directs attention to the effect on the “value of the copyrighted work, not only the effect upon the market, however narrowly that term is defined.” Thomson Reuters promotes the headnotes as a reason to buy a Westlaw subscription, and ROSS diminished their value as a draw. As to the market for legal-research platforms, which the court also considered, ROSS offered “no evidence to rebut the alleged harm.”¹⁴

Then the derivative market. Thomson Reuters argued that ROSS had harmed its position in a market for licensing headnotes as AI training data, and the court agreed. The evidence showed “that the market for licensing headnotes as text to train AI is rapidly developing.” Thomson Reuters presented evidence that it uses its headnotes as training data for its own AI search products, and ROSS offered none to the contrary. That it had not licensed them to anyone else “does not disprove that a market exists to do so.” And by copying without authorization, ROSS “usurped Thomson Reuters’s opportunity to enter that derivative market and license its headnotes for that purpose.”¹⁵

That holding is not tied to the kind of model ROSS built. It recognizes a cognizable market in training data itself, established by evidence that the market is developing and that the owner uses its own material that way, without any showing that the owner has licensed to others. Rights holders in other training-data cases can be expected to cite it, and defendants will need evidence to answer it. ROSS’s public-benefit arguments fared no better. The opinions were already free and ROSS charged prices comparable to Westlaw’s, so it was unclear how its copying would greatly increase access to the law; and ROSS “presents no evidence” that the ruling would halt AI development, and none connecting its platform to national security. Some AI technology may implicate national security, the court allowed, but that “does not give ROSS carte blanche to violate copyright law merely because it incorporates AI.”¹⁶

VIWhat the court said about generative AI

The court addressed generative AI directly once, in footnote 7. The Department of Justice had filed a statement of interest in In re OpenAI, Inc. Copyright Infringement Litigation, pending in the Southern District of New York, relying on Bartz v. Anthropic PBC to contend that training a large language model, which can “generate original responses,” is a transformative use, and arguing that the training there did not result in “substitutive competition.” The Third Circuit’s response: “The concerns raised in that separate case do not apply here.” ROSS’s platform “cannot generate original expression,” and ROSS trained its AI for the purpose of creating “a commercial substitute for Westlaw.” The court added that the government is tracking these issues and knows how to assert its interests, “but the DOJ notably did not do so here.”¹⁷

Read carefully, the footnote distinguishes; it does not decide. The court did not say that training a generative model is transformative, or that it is not. It said that the evidence about ROSS “supports the opposite conclusion about transformativeness” from the one urged for models that generate new expression, and that ROSS built a substitute. Both distinctions go to the first factor.

So the answer to the question this case raises for generative-AI litigation is a split one, and the Editor offers it as commentary. On purpose, the narrow framing does limit the opinion’s weight: a court deciding whether training a generative model is transformative will find that the Third Circuit expressly set that question aside. On amount and on market harm, nothing is confined. The individual work as the unit of measurement, the cognizable market for licensing training data, and the rule that ease is not necessity are all stated as ordinary copyright law, and the opinion insists that ordinary copyright law is all it applies. Under Warhol, moreover, even a real difference in purpose must still be balanced against the commercial nature of the use. A generative-AI defendant that persuades a court on transformativeness will still have to meet the third and fourth factors as this opinion states them.

For the case the court actually decided, the rule is durable: copying a competitor’s editorial work to build a product that replaces it is not fair use, and routing the copy through a training set does not change that.

VIIFor the litigator

Several practical points are on the face of the opinion.

Fair use can be lost on summary judgment. Fair use is a mixed question of law and fact, but where the undisputed facts are sufficient to evaluate each statutory factor, a court may decide it as a matter of law. ROSS, as the proponent of the affirmative defense, bore the burden of showing that its copying was justified.¹⁸

Policy arguments need a record. ROSS’s claims about the future of AI development and national security failed for want of evidence, and so did its answer on the market for legal-research platforms. Its access-to-law argument ran into the record instead: the opinions were already free, and its prices were comparable to Westlaw’s.

Concessions shape the appeal. ROSS’s decision not to contest copying and attribution on appeal, and its failure to brief the Key Number System, left the court with a record it could treat as undisputed and one issue fewer than the district court had certified.

Provenance is discoverable, and it matters. The bad-faith evidence in footnote 9 came from the record of how ROSS sought access to Westlaw. Even with the role of good faith in doubt after Oracle, the court put that evidence in its opinion.

Necessity must be shown, not asserted. The intermediate-copying cases did not help a defendant with a lawful route to the unprotected material. Where the public-domain source was available — here, the opinions themselves — a defendant that copied the protected version instead had no necessity to plead.

---

The court called this an ordinary copyright case, and on the merits it is one: no new doctrine, the four factors applied in order, an affirmance. The ordinariness is the point. Because the court refused to treat AI as special, the parts of its reasoning that did the most work against ROSS — the work-by-work measure of what was taken and the market for training data — are not confined to the kind of AI that ROSS built. §