Training data copyright in the United States turns on two questions: whether copying works into a training set reaches the owner's exclusive rights, and whether that copying is a fair use under the four factors of 17 U.S.C. 107 [1]. On 29 September 2026 the Third Circuit held that a legal search company's use of 2,243 Westlaw headnotes to train a non-generative tool was not fair use [8]. In June 2025 two federal judges in California ruled for model developers on fair use in training language models on books, each on its own record [13, 14]. The Copyright Office's report on training was still listed as a pre-publication version on 8 November 2026 [24].
This article states each source as written, with its date. It is research, never legal advice, and gives no view on whether any reader's use is fair.
What the statute and the Supreme Court set out
Section 106 of the Copyright Act gives an owner the exclusive rights "to reproduce the copyrighted work in copies or phonorecords", to prepare derivative works and to distribute copies to the public [1]. Section 107 provides that fair use "is not an infringement of copyright" and lists four factors: the purpose and character of the use, the nature of the work, the amount used and "the effect of the use upon the potential market for or value of the copyrighted work" [1]. Statutory damages are counted per work, from $750 to $30,000, and up to $150,000 where infringement was willful [2].
Section 105 puts works of the U.S. Government outside copyright [1]. A federal record can still carry another party's terms: older federal award records include fields licensed from a commercial data provider whose terms prohibit systematic access, as where federal contract data comes from shows.
Feist (1991) holds that facts are not copyrightable [4]. Campbell (1994) asks whether a new work merely "supersede[s] the objects" of the original creation or "adds something new", that is, "whether and to what extent the new work is 'transformative'" [3], and Warhol (2023) whether a use "has a further purpose or different character" [5].
Thomson Reuters v. ROSS: headnotes and a non-generative tool
ROSS Intelligence built a search tool that answered legal questions with passages from judicial opinions and "would not create any new expression" [8]. It was trained on about 25,000 memos written by LegalEase, each pairing a legal question with passages [8]. On 11 February 2025 the District of Delaware found actual copying of 2,243 Westlaw headnotes and ruled against fair use [8, 9].
The Third Circuit affirmed on 29 September 2026 [8]. Judicial opinions "remain 'free for publication to all'", the court wrote, while the headnotes written by Thomson Reuters's editors showed "some creative spark" [8]. It weighed the factors this way [8]:
- Purpose: training was "an intermediate step" that "arguably presents a slight degree of difference in use", and the shared purpose left ROSS's use "minimally transformative, at best".
- Nature: the headnotes "are more factual than fictional", which favored fair use.
- Amount: each headnote is its own work, so for each one taken "ROSS copied an entire work".
- Market: "the market for licensing headnotes as text to train AI is rapidly developing", and ROSS took Thomson Reuters's opportunity to license into it.
It set a book-scanning search case apart because ROSS's use, as ROSS admitted, "aims to replace Westlaw" [8]. Noting a Justice Department statement in a generative case, it said those concerns "do not apply here" because ROSS's platform "cannot generate original expression" [8].
The ruling binds federal courts in Pennsylvania, New Jersey, Delaware and the Virgin Islands, notes Authors Alliance, which filed an amicus brief supporting ROSS [12]. On 3 November 2026 ROSS petitioned for rehearing [11]. Tools that answer with passages from sources can be checked claim by claim, with the method in how to check LLM citations against their sources.
Is training on copyrighted data fair use?
On 23 June 2025, in the Northern District of California, in a suit by authors against a model developer, Judge Alsup held that training on the books was "exceedingly transformative" and a fair use [13]. The order states that the developer "pirated over seven million copies of books", and it held that keeping them in a "permanent, general-purpose library" was not a fair use, leaving those copies and the damages for trial [13]. It also said a market for licensing books for training "is not one the Copyright Act entitles Authors to exploit" [13].
Two days later, in the same district, in a suit by thirteen authors against another model developer, Judge Chhabria granted the developer partial summary judgment on fair use [14]. He called the authors' theories of regurgitation and a lost licensing market "clear losers" and named market dilution as the "potentially winning argument", which the authors had barely argued [14]. His ruling "only affects the rights of these thirteen authors", he wrote [14]. A separate claim that the developer distributed the works while torrenting them remained, and on 25 March 2026 the authors were allowed to add a contributory infringement claim [14, 15].
The class settlement and what it released
The first case settled as a class action. The notice says the suit claimed infringement "by downloading works in two allegedly pirated online datasets", and describes a $1.5 billion fund and about $3,000 per work before costs, fees and expenses [17]. As of 16 April 2026, 440,490 of the 482,460 works on the list had been claimed, and the court granted final approval on 20 July 2026 [16]. The release covers past conduct only, with nothing after 25 August 2025, leaves claims about outputs out, and "does not constitute, in any respect, a license to torrent, scan, or train AI models on any copyrighted works" [16, 17].
Training data copyright: pleading-stage and appeal rulings
The image and news orders below decide only whether claims may go forward [19, 20].
| Date | Court | Parties, by role | What the order decided |
|---|---|---|---|
| 12 August 2024 | N.D. Cal. | Visual artists against image model developers | Copyright Act claims go forward, Digital Millennium Copyright Act (DMCA) claims dismissed [19] |
| 4 April 2025 | S.D.N.Y. | News publishers against a model developer and its commercial partner | Contributory infringement claims go forward, fair use not raised [20] |
| 16 September 2026 | Ninth Circuit | Programmers against a code-hosting platform and the developers of its code tools | DMCA dismissal affirmed: the tools "create new works that never contained" copyright management information [22] |
On 1 September 2026 the Justice Department told the Southern District of New York that training language models on copyrighted texts is "extraordinarily transformative", and called the second California ruling's application of the market factor "deeply flawed" [21].
What the Copyright Office's training report says
After launching an initiative in early 2023 and holding listening sessions, the Office published a notice of inquiry on 30 August 2023 [24, 25] and received more than 10,000 comments by December 2023 [24]. Part 3, released on 9 May 2025, covers training [24]. It finds that the steps to build a training dataset of copyrighted works "clearly implicate the right of reproduction" [23]. On the factors [23]:
- Training a foundation model on a large and diverse dataset is often transformative, in the Office's view, and training on specific types of works "to produce content that shares the purpose of appealing to a particular audience" is "at best, modestly transformative".
- "the knowing use of a dataset that consists of pirated or illegally accessed works should weigh against fair use without being determinative".
- The fourth factor "should not be read so narrowly" as to leave out the dilution of markets for works of the same kind.
- Where licensing markets are available, unlicensed uses are disfavored under the fourth factor.
Given "the robust growth of voluntary licensing", the Office called government intervention "premature at this time" [23]. Most data licensing deals for model training are announced without a price, as training data licensing deals finds.
Where the courts and the Office part ways
The licensing market splits the sources. The Third Circuit counted a market for headnotes as training data against ROSS [8]. Both California judges said book authors are not entitled to a market for licensing their works for training [13, 14]. The Office disfavors unlicensed uses where licensing is available [23].
Market dilution splits them too. The Office counts it under the fourth factor [23] and Judge Chhabria called it the potentially winning argument [14], while the Justice Department called his reasoning on the market "deeply flawed" [21].
On how works were obtained, the Office and Judge Alsup point the same way, against pirated copies [13, 23].
Facts a training data record can hold
Each ruling turned on facts a record can hold: which works were used, how they were obtained, what the use was for and whether a market to license them existed [8, 13, 14]. The Office's conclusion turns on what works were used, from what source, for what purpose and with what controls on outputs [23].
A license settles some of them in writing, and what a training data license grants sets out how one can name training, fine-tuning and retrieval as separate grants. Rillor's training data acquisition and licensing finds the publishers and data holders that hold the data a team needs, agrees use, term and onward sharing in writing before delivery, and delivers each record with the license it came under.
Read the full report
The full report sets out ten findings, from the statute to each ruling's factors, the settlement record, the Office's report and where the sources disagree, with all 27 sources.