The Third Circuit’s Decision in Ross Protects Westlaw’s Market for Licensing Headnotes to AI Rivals. Now what?

On September 29, 2026, the Third Circuit held that Ross Intelligence’s use of Westlaw headnotes to train a competing legal-research system was not fair use. The opinion by Judge Tamika Montgomery-Reeves recognizes a potential market for licensing those headnotes as AI training data, even though Thomson Reuters had not licensed them to others. That conclusion gives an incumbent that competes in legal research a protected economic interest in supplying an input to its AI rivals. The competition implications of that arrangement receive no independent examination in the appellate opinion.[1]

The panel describes the dispute as “no more than an ordinary copyright case.”[2] Its reasoning identifies protected expression and an unauthorized commercial use that threatens the owner’s markets. The resulting allocation of power deserves a further inquiry: how much control over entry does ownership of a valuable training resource confer on a firm that also competes with the firms seeking access?

Ross did bring antitrust counterclaims. It withdrew its refusal-to-deal theory and lost its remaining competition claims on summary judgment. The district court also rejected copyright misuse.[3] The problem exposed by the appellate decision is the division of inquiry: copyright protects the incumbent’s potential licensing opportunity while the consequences of withholding that input depend on a separate competition case.

The disputed input combined public law with private editorial work

Thomson Reuters’s refusal preceded the copying. The district court’s 2025 opinion records that Ross asked to license Westlaw’s content and Thomson Reuters refused because Ross was a competitor. Ross then turned to LegalEase Solutions, which supplied approximately 25,000 training memoranda. The memoranda paired legal questions with passages from judicial opinions and rated the passages for responsiveness. Their authors used Westlaw headnotes to formulate the questions.[4]

That editorial contribution was the protected input. The panel upheld copyright protection for the 2,243 headnotes at issue because editors had exercised judgment in selecting significant legal propositions and expressing them concisely. The judicial opinions remained available for others to copy. The court expressly left open the copyrightability of headnotes that reproduce opinion text verbatim.[5]

Ross’s system returned passages from existing judicial opinions. It could not generate original expression. The panel expressly distinguished the generative AI systems involved in other training disputes.[6] The ruling’s significance for AI therefore follows from its reasoning about purpose and markets. Its holding concerns this training use and this competing research product.

The court found that both companies used headnotes to help researchers locate responsive law. Training was an intermediate step toward that shared commercial purpose. Ross’s use was therefore minimally transformative. The headnotes’ factual character favored fair use. The purpose of the copying, its extent and its market effects weighed against it.[7]

The editorial work supplies both the basis for copyright protection and the value of the training input. Editors create useful expression through work that competitors could otherwise appropriate. Ross also accepted for this appeal that the contractors’ copying was attributable to it. The dispute came to the panel with that copying established.[8] A competition analysis must account for the incentive to create the resource as well as the power its ownership supplies.

A potential licensing market protects the option to exclude

The fourth fair-use factor directs attention to effects on the potential market for or value of the copyrighted work. The panel found several kinds of harm. Headnotes attract Westlaw subscribers. Ross used them to develop a substitute research platform that could draw those subscribers away. The court separately recognized harm to the potential market for licensing headnotes as training data.[9]

That final step carries the broader implication. The court relied on evidence that the training-license market was developing rapidly and that Thomson Reuters used its own headnotes to train its own AI search products. The absence of licenses to outsiders did not defeat the claim. Unauthorized training deprived Thomson Reuters of an opportunity to enter that market and license its headnotes.[10]

This reasoning protects an option. Thomson Reuters can use its headnotes to improve its own research service while retaining control over whether a competitor may use them for the same purpose. A recognized licensing market does not require the owner to offer a license. The copyright analysis accordingly gives legal weight to a transaction the owner could have declined to make.

The rule has limits. In Campbell v. Acuff-Rose Music, the Supreme Court confined cognizable derivative markets to uses that creators would generally develop or license. It also explained that a refusal of permission does not defeat an otherwise fair use. The Third Circuit quoted the market limitation and found the evidence sufficient here.[11] Its result rests on its assessment of this record and the limited transformation it found in Ross’s use.

The difficulty arises when the potential license is also permission to compete using an incumbent’s accumulated editorial resource. Under the panel’s analysis, successful substitution for Westlaw supplies evidence of copyright harm. Under a structural competition analysis, an entrant’s capacity to offer such a substitute is a potential public benefit. Both descriptions can be accurate. Deciding the case through the first does not resolve the second.

The antitrust record explains the missing appellate inquiry

Ross’s competition challenge did not reach the Third Circuit through this interlocutory appeal. The certified issues concerned originality and fair use. The panel described its review as generally constrained by those certified questions.[12]

The earlier proceedings ended without an adjudicated finding of dominance. Ross alleged that Thomson Reuters controlled more than 80 percent of the market for legal-search platform products. The district court recited that figure in 2022 under the pleading-stage standard. At summary judgment, however, Ross failed adequately to define the relevant markets and therefore could not reach the issue of market power.[13]

Ross expressly withdrew its refusal-to-deal theory at the 2022 hearing. The surviving antitrust theories concerned tying Westlaw’s public-law database to its search tools. In September 2024, the district court granted Thomson Reuters summary judgment because Ross had not sufficiently established separate products or the relevant product markets. It excluded the market-definition portion of Ross’s expert opinion for inadequate methodology. Those failures independently defeated the remaining claims.[14]

In February 2025, the district court also rejected copyright misuse. Ross argued that Thomson Reuters was using copyright to suppress a competitor. The court relied on the earlier antitrust ruling in concluding that Ross had not shown misuse of copyright to stifle competition.[15]

Refusal-to-deal doctrine imposes a further constraint on examining an incumbent’s control over a rival’s inputs. Under Verizon Communications Inc. v. Law Offices of Curtis V. Trinko, monopoly power alone does not violate the Sherman Act. The circumstances imposing a duty to assist a rival are limited. The Supreme Court emphasized the exceptional character of terminating a voluntary and presumably profitable course of dealing when discussing Aspen Skiing.[16] The refusal of a proposed copyright license would require its own analysis under those demanding standards.

The two market inquiries serve different purposes. Copyright asks whether copying impairs a protected work’s commercial value or a cognizable licensing opportunity. Antitrust market definition examines competitive substitutes to establish the setting in which market power and harm can be assessed.[17] The potential market recognized in the copyright appeal does not cure the evidentiary defects in Ross’s antitrust case. The combined result nevertheless warrants scrutiny: a licensing interest receives protection while the asserted dependence on the rights holder remains unproved.

The interference fallacy directs attention to the resulting power

In The Interference Fallacy, I argued that law repeatedly directs its attention toward discrete acts while leaving the standing relationship of power inadequately examined. The proposed domination test asks whether power is concentrated, arbitrary and continuing. Applied here, that framework directs attention to the conditions under which a research incumbent can determine whether another firm obtains a commercially significant training input.[18]

A structural inquiry turns on practical dependence. Ross already had access to approximately ten million uncopyrighted judicial opinions. The court concluded that it could create training memoranda from those opinions and that using Westlaw’s headnotes offered convenience rather than necessity. It also rejected Ross’s access-to-law argument in light of the availability of the opinions and Ross’s prices, which were comparable to Westlaw’s.[19]

A structural inquiry would examine the cost and time required to produce an effective independent training corpus, the quality of available alternatives and the practical effect of those differences on entry. Formal access to public opinions answers the question of permission to use the law. The feasibility of building a competitive research service requires evidence about the resources needed to make that law useful.

The value of headnotes illustrates the distinction. Their editorial selection and connection to responsive passages supply useful judgments about legal relevance. The panel relied on that work to sustain copyright protection and described how it supported training.[20] A comparison of competitive alternatives should therefore assess access to a usable body of those judgments. Counting available opinions alone cannot establish the commercial equivalence of alternative training resources.

This is the interference fallacy applied to intellectual property. The copying becomes the legally cognizable interference with the owner’s entitlement. The owner’s continuing capacity to condition a rival’s access receives protection through the licensing-market analysis. Whether that capacity creates a relationship of domination requires a further inquiry into dependence and accountable limits.

Protecting editorial investment requires attention to entry

For competition policy, the next step should be an evidence-based examination of access to legal-research inputs. That work should identify where independent annotation is a realistic substitute and where accumulated editorial resources confer durable control over entry. Public investment in accessible opinions and independently reusable research materials could reduce dependence. Any proposal for compulsory access would need to confront both the incentive to produce editorial work and the legal limits on compelled dealing.

The Third Circuit has recognized the value of Westlaw’s opportunity to license its headnotes for AI training. The corresponding public question concerns the conditions under which competitors can develop legal-research tools when the owner elects to withhold that opportunity. A legal framework concerned with domination should make that question answerable on a developed record.

---

[1]   Thomson Reuters Enter. Ctr. GmbH v. Ross Intel. Inc., No. 25-2153, slip op. at 1–2, 26–28 (3d Cir. Sept. 29, 2026) (affirming partial summary judgment concerning 2,243 headnotes and recognizing a potential market for licensing headnotes as AI training data). The court’s public opinion list dates the release September 30. U.S. Ct. of Appeals for the Third Cir., Recent Precedential Opinions (Oct. 1, 2026) (listing No. 25-2153 under September 30, 2026).

[2]   Thomson Reuters, No. 25-2153, slip op. at 2 (describing the questions presented as originality and fair use).

[3]   Thomson Reuters Enter. Ctr. GmbH v. Ross Intel. Inc., No. 20-613, slip op. at 6 n.4 (D. Del. Apr. 26, 2022), ECF No. 170 (recording withdrawal of the refusal-to-deal theory). Thomson Reuters Enter. Ctr. GmbH v. Ross Intel. Inc., No. 20-613, slip op. at 2–4, 11–14 (D. Del. Sept. 27, 2024), ECF No. 669 (granting summary judgment on the remaining antitrust counterclaims). Thomson Reuters Enter. Ctr. GmbH v. Ross Intel. Inc., No. 20-613, slip op. at 15 (D. Del. Feb. 11, 2025) (rejecting copyright misuse).

[4]   Thomson Reuters, No. 20-613, slip op. at 3–4 (D. Del. Feb. 11, 2025) (describing the denied licensing request and the subsequent engagement of LegalEase). Thomson Reuters, No. 25-2153, slip op. at 5–6 (describing the memoranda and their use in training).

[5]   Thomson Reuters, No. 25-2153, slip op. at 10–13 & n.5 (sustaining copyright protection for the headnotes at issue while reserving the question concerning verbatim opinion text).

[6]   Id. at 5, 17 n.7 (describing Ross’s system and distinguishing disputes concerning models that generate original expression).

[7]   Id. at 16–23, 27–28 (finding factors one, three and four against fair use and factor two slightly in its favor).

[8]   Id. at 6 n.2, 7 n.3, 10–11 (identifying the uncontested copying issues and the editorial decisions supporting originality).

[9]   Id. at 24–26 (finding harm to headnote value, the original research-platform market and a potential training-license market).

[10]  Id. at 26 (relying on the developing licensing market and Thomson Reuters’s internal training use despite the absence of licenses to others).

[11]  Campbell v. Acuff-Rose Music, Inc., 510 U.S. 569, 585 n.18, 592–93 (1994) (limiting cognizable derivative-market harm and explaining that denied permission does not defeat an otherwise fair use). Thomson Reuters, No. 25-2153, slip op. at 26 (applying the limitation to the potential headnote licensing market).

[12]  Thomson Reuters, No. 25-2153, slip op. at 7–8 (identifying the certified questions and the scope of interlocutory review).

[13]  Thomson Reuters, No. 20-613, slip op. at 1 & n.2 (D. Del. Apr. 26, 2022) (reciting Ross’s alleged share under the pleading-stage standard). Thomson Reuters, No. 20-613, slip op. at 11–13 (D. Del. Sept. 27, 2024) (finding the market-definition proof insufficient to reach market power).

[14]  Thomson Reuters, No. 20-613, slip op. at 6 n.4 (D. Del. Apr. 26, 2022) (dismissing the withdrawn refusal-to-deal theory). Thomson Reuters, No. 20-613, slip op. at 2–4, 11–14 (D. Del. Sept. 27, 2024) (resolving the surviving tying theories on separate-products and market-definition grounds).

[15]  Thomson Reuters, No. 20-613, slip op. at 15 (D. Del. Feb. 11, 2025) (rejecting the misuse defense by reference to the prior antitrust ruling).

[16]  Verizon Commc’ns Inc. v. Law Offs. of Curtis V. Trinko, LLP, 540 U.S. 398, 407–10 (2004) (explaining that monopoly power alone does not violate § 2 and limiting liability for refusals to assist rivals).

[17]  Thomson Reuters, No. 25-2153, slip op. at 24–26 (examining harm to the value of headnotes and their actual or potential markets). Thomson Reuters, No. 20-613, slip op. at 11–13 (D. Del. Sept. 27, 2024) (requiring adequate definition of relevant markets before assessing market power and the alleged competitive injury).

[18]  J.R. Howell, The Interference Fallacy, The American Counsel (July 27, 2026), https://www.theamericancounsel.com/the-interference-fallacy/ (proposing a common inquiry into concentrated, arbitrary and standing power).

[19]  Thomson Reuters, No. 25-2153, slip op. at 5, 20–21, 23, 27 (identifying available public opinions and rejecting necessity and public-benefit arguments on the record presented).

[20]  Id. at 3–6, 10–11 (describing the editorial work reflected in headnotes and their use in producing training memoranda).