US litigation over AI training and copyright has been waiting on appellate guidance since district courts split in 2025. On September 1, 2026, the US Department of Justice (DOJ) filed a 20-page Statement of Interest in the consolidated proceeding (MDL) in the Southern District of New York that combines The New York Times' (NYT) suit against OpenAI and Microsoft. The brief divides large language model (LLM) development into three stages, "acquisition," "training," and "output," and takes up only the copying done at the training stage, arguing that it is fair use. The conclusion matters less than where the line is drawn. Which stage the brief isolates, and which it avoids, determines where it can help in the litigation.
A 20-page brief that splits acquisition, training, and output, and the four fair use factors
The legal basis for the brief is 28 U.S.C. §517, which authorizes the Justice Department to "attend to the interests of the United States" in pending suits. As footnote 1 explains, no court permission or filing deadline is required. It differs procedurally from an amicus brief, which a third party files with the court's leave. It is a document in which the government states its views on the strength of its own interest, and it has no binding force on the court. It was signed by three people: Stanley E. Woodward, Jr., Associate Attorney General and the department's third-ranking official; Brett Shumate, Assistant Attorney General heading the Civil Division; and Michael Weisbuch, Woodward's senior counsel. The case is In re OpenAI, Inc. Copyright Infringement Litigation (25-md-3143, Judge Stein).
A quick recap of the case. The NYT sued on December 27, 2023. In a March 26, 2025 ruling on the motion to dismiss, the direct and contributory infringement claims survived, while the DMCA §1202 (copyright management information) claim was dismissed in part. In April 2025 the case was consolidated into the MDL with suits by the Authors Guild, the Daily News, The Intercept, and others, and it is now in the middle of summary judgment briefing. The NYT has continued to update its complaint into 2026, and the brief cites the complaint dated August 21, 2026.
The brief has four parts. Part I lists the interests of the United States: national security, the competitiveness of the AI industry, and the benefits to independent media. Section II.A divides LLM development into acquisition, training, and output. Section II.B sets out the fair use framework, Section II.C argues that training satisfies the "transformativeness" inquiry of the first factor, and Section II.D argues that the fourth factor, "market effect," is not met.
The first line is drawn in Section A. On page 5 the brief states that "Each stage may present distinct questions of copyright law," and says explicitly that the United States addresses only whether copying at the training stage is fair use. Footnote 2 also rejects application of §1498, which immunizes certain government uses, making clear that OpenAI's conduct is not attributable to the government.
Section 107 of the US Copyright Act weighs four factors to decide infringement: the purpose and character of the use, the nature of the work, the amount used, and the effect on the market. As a rule the analysis is done "use by use." Copying the same book to build a search index and copying it to sell an audio reading are evaluated separately under the four factors. The brief leans on this principle to the fullest, framing training as a use separate from output. What is at issue at the training stage is the purpose of copying the text to create the model's weights and whether that act displaces the market for the original. What the model later emits, the brief says, is a question for the output stage.
On the first factor, the brief calls training "extraordinarily transformative." It cites Judge Alsup in Bartz v. Anthropic, who called training transformative "spectacularly so," and Judge Chhabria in Kadrey v. Meta, who found it "highly transformative." It argues that the purpose of training is to build a conversational model, which is "different in kind" from the original purpose of entertaining and teaching readers. On the commercial nature of the use, it cites Campbell and says that does not "move the needle." On the fourth factor, it builds on Authors Guild v. Google, arguing that the only cognizable harm is "substantial substitutive competition," and that training does not produce substitution because it "reveals nothing to the public." Footnote 16 handles the second and third factors by saying they "play no major role" but both support fair use.
Rejecting Kadrey's market dilution theory and Copyright Office p. 64 as "conflating stages"
The strongest language in the brief is aimed at Kadrey, a decision that would seem to be on its side. On June 25, 2025, Judge Chhabria held that Meta's training was fair use on the record before him, but in dicta set out a theory of "indirect substitution," or "market dilution." If a trained model can generate large volumes of works that compete with the originals, he reasoned, the harm of diluting the market could be considered under the fourth factor even if no individual output resembles an original. The brief characterizes this as dicta adopted "without the parties' having argued it," and writes that "the Kadrey court's application of the fourth factor is deeply flawed." Its reason is that the court improperly collapsed training and output into a single continuous use.
To rebut it, the brief invokes an anecdote about Joan Didion. As a teenager, Didion typed out Hemingway's prose to learn sentence structure, as she told the Paris Review in 1978. Under Kadrey's logic, the brief says, Didion would owe Hemingway liability for that copying every time she later published a book. It also quotes Bartz's line that it is inconceivable to make someone pay every time a book is later used in a new way. The argument is that competition produced in the future by abilities learned through training is not harm from the act of copying in training.
The same reasoning is turned on the US Copyright Office. In the pre-publication version of its Part 3 report on generative AI training, released May 9, 2025, the Office presented market dilution as a new fourth-factor harm on page 64 and recommended voluntary licensing and collective rights management. In footnote 17, the brief notes that the report came from "the Register, whose dismissal is currently being challenged," and, citing Loper Bright, dismisses her understanding as not deserving deference, calling it "poor reasoning" that ignores precedent. Register Shira Perlmutter was dismissed in May 2025, shortly after the report's release, but remained in her post following later court rulings, and as of September 3, 2026 she is serving as Register. As the brief itself refers to her in the present tense, she is not a "former Register."
The brief is not silent on the output stage. Footnote 15 states that even if infringement were found, relief would be limited to what is "necessary to provide complete relief" under Trump v. CASA, and that "a small fraction of aberrant reconstructive outputs" would not justify broad liability for training as a whole. It also notes that OpenAI described measures to prevent reproduction in its June 27, 2025 MDL filing. In the brief's structure, then, any problem with output should be handled by output remedies, and should not be used to push liability back onto the training stage. On page 18, the brief says NYT reporters themselves use LLMs for "ideation and editing" of articles and cites a Futurism article dated May 12, 2026 in footnote 19. The Decoder points out, however, that the article reports the NYT strictly warning freelancers about AI use, and that the source in fact supports the opposite conclusion.
Five rulings on the same axes
The rulings since 2025 addressing the legality of AI training come down to three district court decisions and two agency views. We sorted them on four axes: the transformativeness of training, whether the market dilution theory was adopted, treatment of the acquisition stage (piracy), and procedural status as of September 3, 2026. Because court rulings and agency views carry different legal weight, the type of each is shown separately.
| Ruling | Type | Transformativeness of training | Market dilution theory | Treatment of acquisition stage | Procedural status (as of Sept. 3, 2026) |
|---|---|---|---|---|---|
| Bartz v. Anthropic (N.D. Cal., Judge Alsup, June 23, 2025) | District court ruling | Yes (spectacularly transformative) | Not adopted | Keeping a pirated library is not fair use | $1.5 billion class settlement given final approval July 20, 2026 |
| Kadrey v. Meta (N.D. Cal., Judge Chhabria, June 25, 2025) | District court ruling | Yes (highly transformative) | Raised in dicta | Torrent distribution claim continues | Interlocutory appeal denied July 8, 2026; SJ hearing on distribution claim reportedly set for Feb. 25, 2027 |
| Thomson Reuters v. Ross (D. Del., Judge Bibas, Feb. 11, 2025) | District court ruling | No (substitute for same purpose) | Not addressed | Not at issue | Third Circuit heard oral argument June 11, 2026; no decision yet |
| Copyright Office Part 3 report (May 9, 2025, pre-publication version) | Agency view | Depends on use | Raised on p. 64 | Pirated sources weigh against under first factor | Final version unpublished |
| DOJ brief (S.D.N.Y. MDL 3143, Sept. 1, 2026) | Government statement | Yes (extraordinarily transformative) | Rejected as "deeply flawed" | Not addressed | Summary judgment briefing under way |
Three parties, Bartz, Kadrey, and the DOJ, accepted the transformativeness of training without qualification, while the Copyright Office reserved judgment as depending on use. On market dilution, Kadrey and the Copyright Office adopt it and Bartz and the DOJ reject it, so both the government and the courts are internally split. On the acquisition stage, Bartz found liability for keeping pirated copies, Kadrey left the distribution claim standing, and the Copyright Office wrote that pirated sources weigh against fair use under the first factor. Only the DOJ brief has nothing in that column. Ross involves a legal research tool rather than generative AI, and the brief does not cite it, but a Third Circuit decision would be the first appellate ruling in this area.
Note also that Kadrey's market dilution theory is dicta, not part of the holding, as the brief itself says. Judge Alsup decided Bartz at the district level, but final approval of the settlement was given by Judge Martínez-Olguín, who took over. The Copyright Office report and the DOJ brief are both non-binding views, and the table does not treat them as equivalent to precedent.
What the brief does not say about acquisition
Nowhere in the 20 pages does the brief address the legality of the acquisition stage. The only references to acquisition are the definition of the three stages on page 5 and the footnote 2 rejection of §1498 immunity. Whether collecting material from piracy sites or shadow libraries is permissible is not discussed. Having said itself that "each stage may present distinct questions," the brief left one of them untouched. That is a fact about the scope of the brief, not a statement that the DOJ has no position on acquisition.
The price of the avoided issue has already been set once. In Bartz v. Anthropic, on June 23, 2025, Judge Alsup held that training was fair use but that keeping books collected from pirate sites as a central library was not. Anthropic agreed to a $1.5 billion class settlement over liability at this acquisition stage (about ¥237.3 billion at the September 2, 2026 New York close of ¥158.22 to the dollar), and Judge Martínez-Olguín gave final approval on July 20, 2026. It comes to about $3,000 per work, and the payment start date has not been set. As TechCrunch has also noted, the amount is payment for use of pirated shadow libraries, not compensation for training itself.
In other words, the issue on which money has actually moved since 2025 sits in the stage the brief did not address. The facts in the NYT case differ from Bartz: the dispute is over collection of articles from the web, not piracy, so acquisition would not be argued the same way. But the MDL also includes book authors and publishers, and footnote 11 of the brief says its argument applies to those parties as well. A position that isolates training does not directly collide with a Bartz-style outcome that finds training fair use but imposes liability for acquisition. What the brief seeks to protect appears to be freedom to train, not how the data is gathered.
The policy argument points the same way. On page 4, the brief argues that requiring licenses for training could harm competition in the LLM market, since only the largest tech companies might be able to pay, creating an entry barrier that would work as "a massive subsidy to legacy mainstream media companies." Footnote 13, however, reserves any position on whether a licensing regime would be financially or practically feasible. The conclusion reads: "Fair use determinations turn on the specific facts and uses in each case. But imposing broad copyright liability that would make training AI models without a license generally impermissible is problematic and legally wrong."
The plaintiffs' side reacted to the framing itself. NYT spokesperson Graham James told the AP that AI companies should simply pay fairly for the content that makes their products possible, as copyright law requires. Matt Topic, counsel for The Intercept, criticized the brief, saying that accepting its argument would amount to an "unprecedented transfer of IP rights." News/Media Alliance CEO Danielle Coffey said courts are called on to apply the law strictly and that she believes the NYT's claims will prevail, the Hollywood Reporter reported. OpenAI did not respond to The Intercept's request for comment.
The US government's path from January 2025 to September 2026
The brief is not a one-off. Lining up the executive orders and legislative recommendations it cites shows the administration's position has been consistent since January 2025, with only the Copyright Office pointing the other way. Dates are taken from primary sources or from articles directly reporting the events.
| Date | Event | Direction |
|---|---|---|
| January 23, 2025 | Executive order "Removing Barriers to American Leadership in AI" (cited in the brief) | Administration |
| May 9, 2025 | Copyright Office releases pre-publication Part 3 report on generative AI training; presents market dilution | Copyright Office |
| May 2025 (just after report release) | Administration dismisses Register Perlmutter | Administration |
| May 22, 2025 | Perlmutter v. Blanche filed in D.C. federal district court | Court |
| July 23, 2025 | At AI Action Plan speech, Trump says a successful AI program cannot be expected if it must pay for every book and article it reads | Administration |
| September 10, 2025 | D.C. Circuit grants, 2-1, an injunction protecting Perlmutter's continued service | Court |
| March 20, 2026 | White House releases "National Policy Framework for AI" legislative recommendations | Administration |
| June 2, 2026 | Executive order "Promoting Advanced AI Innovation and Security" | Administration |
| June 30, 2026 | Supreme Court denies administration's emergency stay application, leaving the injunction in place | Court |
| September 1, 2026 | DOJ files Statement of Interest in S.D.N.Y. MDL 3143 | Administration |
The March 20, 2026 legislative recommendations had already put the brief's position into words. They state the administration's view that "training of AI models on copyrighted material does not violate copyright laws," while acknowledging that opposing arguments exist and leaving resolution to the courts. They also ask Congress to consider licensing frameworks and collective rights management, but say that whether licensing is required at all should not be addressed through legislation. The brief cites these recommendations and the two executive orders in Part I. In his July 23, 2025 speech, President Trump also said "China's not doing it," and the brief's national security argument runs along the same line.
One caution on the Copyright Office. It is part of the Library of Congress and sits in a different chain from the White House. It is therefore more accurate to say that the executive branch and an agency on the legislative side hold different views, with the courts protecting the Register's position, than to call it a "conflict within the government." As of September 3, 2026, the Part 3 report remains in its pre-publication version.
Compared with Japan's Article 30-4 and the EU's opt-out
Japan and the EU also allow copying for training in principle, but the way they allow it and where they draw the boundary of liability differ from the US. The comparison is limited to copying text works for training, and uses the legal position as of September 3, 2026. Infringement analysis of outputs (generated content) is treated separately.
| Jurisdiction | Copying at the training stage | Where liability arises | Relationship to output stage |
|---|---|---|---|
| US (Copyright Act §107, DOJ brief's position) | Four-factor analysis case by case. DOJ argues training is an independent, extraordinarily transformative use and fair use | Substantial similarity of output (brief treats the scope of output remedies separately in footnote 15) | Training and output separated as distinct uses |
| Japan (Copyright Act Art. 30-4) | Information analysis for non-enjoyment purposes generally requires no permission | Proviso (unreasonably prejudicing the copyright holder's interests), coexisting enjoyment purpose | For infringing outputs, the business operator can become the "normative actor" |
| EU (DSM Directive Art. 4, AI Act Art. 53) | Possible under the TDM exception only for lawfully accessible works; can be excluded by machine-readable opt-out | Expression of opt-out, memorization (Munich court in GEMA v. OpenAI, first instance, found reproduction; on appeal) | General-purpose AI model providers must publish a copyright policy and a training data summary (applicable from Aug. 2, 2025) |
The three jurisdictions agree on the side of "copying at the training stage is permitted in principle." They differ in where the line of liability sits. The DOJ brief draws it at output similarity, Japan at the purpose of training and the manner of acquisition, and the EU at the rightsholder's declared intent, while the Munich court at first instance also drew it at memorization inside the model.
Article 30-4 of Japan's Copyright Act permits information analysis without permission where the purpose is not "to personally enjoy or cause another person to enjoy the thoughts or sentiments expressed in the work," and its proviso excludes cases that "unreasonably prejudice the interests of the copyright holder." The Agency for Cultural Affairs' March 15, 2024 document "General Understanding on AI and Copyright in Japan" places additional training with a coexisting enjoyment purpose, such as intentional overfitting, outside the provision, and gives as an example of the proviso the case where a database of works organized for information analysis is sold. On collecting material knowing it to be pirated, it says this should be "strictly refrained from," and treats it as an unfavorable consideration when deciding whether to recognize a business operator as a "normative actor" for infringing outputs. The acquisition stage that the DOJ brief did not touch is built into Japan's liability analysis as a factor.
The EU, under Article 4 of the DSM Directive (2019/790), allows training on lawfully accessible works through the TDM (text and data mining) exception, but rightsholders can exclude their works by opting out in a machine-readable way. Article 53(1)(c) and (d) of the AI Act require providers of general-purpose AI models to maintain a copyright policy that respects opt-outs and to publish a summary of training data, applicable from August 2, 2025. In GEMA v. OpenAI on November 11, 2025, the Munich Regional Court found that nine songs' lyrics were memorized in the model and output, treated that as reproduction, and denied the TDM exception (OpenAI has appealed, and the case is pending before the Munich Higher Regional Court). Where the DOJ brief says training "reveals nothing to the public," an EU court has called lyrics remaining inside a model a reproduction.
The comparison has limits, however. US fair use is case law, and the DOJ brief is the opinion of a non-party government that does not bind the court. The Agency for Cultural Affairs' document is a non-binding subcommittee summary, Japan has no final judgment directly addressing generative AI training, and the interpretation of "unreasonably prejudice" in Article 30-4 will be decided case by case. In the EU, member states differ in how they have implemented the DSM Directive, and the GEMA ruling is being contested on appeal.
Whether the brief's line will hold in the US depends on two rulings. One is the Third Circuit's Thomson Reuters v. Ross (oral argument June 11, 2026, before Judges Restrepo, Montgomery-Reeves, and Bove), which will be the first appellate answer on fair use in AI training. The other is the MDL's summary judgment ruling, where the reply deadline is reportedly November 6 and no hearing date has been set. The Kadrey SJ hearing on torrent distribution is reportedly set for February 25, 2027, and liability at the acquisition stage will be contested there too. If the court adopts the brief's framework of carving out training as an independent use, AI companies' liability would concentrate on output similarity and the manner of acquisition, and the US would line up with Japan's Article 30-4 and the EU's opt-out in a structure of "training is free in principle, liability attaches at other stages."