I have spent most of my adult life telling people to take it to court.
Three countries, two legal traditions, more kitchen-table arguments than I can count — when the powerful will not listen, there is a room where they have to. I believed it every time I said it. I mostly still do.
Here is the assumption buried inside that sentence, the one I never took out and examined. I thought I knew who was in the room. Two sides, a judge, a clerk, and the government as the furniture — the entity that builds the courthouse, pays the judge, prints the forms, and then, in a fight between private parties, has the good manners to sit down and be quiet.
On the evening of September 1, 2026, I opened the docket in the largest copyright fight in the country expecting the usual traffic of a big civil case. Scheduling orders. Sealed exhibits. Letters about letters.
What I found was the United States of America, filed on one side.
Not as a plaintiff. Not as a defendant. Not even as a friend of the court, a role you normally have to ask for. The federal government walked into a private lawsuit between newspapers and an AI company, told the judge how the case should come out, and needed nobody's leave to do it.
Let me be precise about what unsettled me, because it is not the part most people argued about that week. Not that the government took the technology industry's side against writers, though it did. Not even that it did so two months after a proposal to hand the government a piece of the defendant had been reported. It is the door it came through — and the fact that nobody has ever been asked to agree that the door should exist.
First, exactly what happened, in the order it happened
The consolidated cases are before Judge Sidney H. Stein in the Southern District of New York, and the plaintiffs are not only the paper in the headline. As Andrew Deck reported for Nieman Lab, the same proceeding carries daily papers owned by Alden Global Capital, high-profile book authors, and smaller nonprofit newsrooms like The Intercept. When something lands on this docket, it lands on all of them.
On September 1, something landed. Late Tuesday evening, the United States filed a Statement of Interest — twenty pages including the caption, marked "This Document Relates To: All Matters," docket shorthand for every case in the pile at once. The entry itself reads "LETTER addressed to Judge Sidney H. Stein from United States of America dated September 1, 2026 re: STATEMENT OF INTEREST OF UNITED STATES OF AMERICA." A letter. Part of why practitioners kept reaching for the word irregular.
The signature block is not junior: Associate Attorney General Stanley E. Woodward, Jr., Assistant Attorney General Brett Shumate of the Civil Division, and Senior Counsel Michael Weisbuch, over the line "Attorneys for the United States of America."
And the case it landed in is not over. As of this writing, the fair-use question is still undecided — the docket header reads "In Re: OpenAI, Inc. Copyright Infringement Litigation, 1:25-md-03143," assigned to Judge Stein, last known filing September 18, 2026. The parties were still filing summary-judgment papers the week I wrote this. Nothing has been held. Nobody has won anything.
The New York Times answered through a spokesman: "The Administration is siding with a handful of trillion-dollar AI companies at the expense of the countless American creators whose work they stole," Times spokesman Graham James told Deadline that night. A company's statement, not a finding — but it is the reaction of a party who thought the referee was neutral.
The door: one sentence of statute, and nothing else
So how do you walk into someone else's lawsuit?
The authority is 28 U.S.C. § 517, and if you have never heard of it, you are in excellent company. The whole of it is a single sentence about sending a lawyer somewhere: the Attorney General may dispatch an officer of the Department of Justice "to attend to the interests of the United States in a suit pending in a court of the United States."
Read that again for what is missing. No brief. No amicus. Not even the phrase statement of interest. The entire practice of the United States filing merits arguments in private litigation is a gloss grown on top of one sentence about attendance.
The government knows exactly how much room that gives it, and said so in its own footnote. Section 517, the brief notes, quoting a Florida district court, "contains no time limitation and does not require the Court's leave." No deadline. No permission. Compare that to every other way a stranger gets into a case — the motions, the page limits, the judge who can say no.
The open-source and IP lawyer Heather Meeker put it plainly for non-lawyers: the DOJ exercised a right to participate rather than asking for one. She is equally careful about the flip side — statements of interest under § 517 "may carry persuasive weight because they come from the government, but they are not binding, and judges are free to give them as much or as little deference as they choose."
True. Hold onto it. It is also the least interesting true thing about this filing.
Because this door has a history, and it is a history of expansion. The law professor Christine P. Bartholomew — whose empirical work is on antitrust filings, not copyright ones — describes a power that "laid dormant for decades, utilized sparingly in litigation involving foreign sovereigns," was expanded in the 1960s to aid in developing civil rights, and in 2009 was turned toward antitrust. Her description of the mechanism is general, and blunt: "The Attorney General may submit an SOI in any lawsuit, even when the United States is not a party."
Then the line I cannot get out of my head. Congress, she writes, has provided little guidance on the outer bounds for when the DOJ may insert itself, "or how courts should treat the opinions offered." And when someone asked a court to police that discretion, the D.C. Circuit — quoted in her footnotes — said this: "[N]either [Section 517], nor any regulation, nor any administrative practice cabins this discretion or furnishes any standard by which to review the Attorney General's determinations in this area."
No outer bound on when. No standard for how much it counts. Her stated worry, scoped to her own subject, is that this tempts "overreach by the executive branch." Less delicately: a power with a throttle and no speed limit, and we have just watched an administration find fifth gear in a copyright case.
The first three pages are not about copyright
Here is where it gets genuinely strange, whichever team you came in on.
If the United States had walked in and made a copyright argument, this would be a story about copyright. It didn't. It opened by locating its interest somewhere else entirely: "The United States has a strong interest in continuing to develop a robust and competitive artificial intelligence industry that sets the standard for the practice and procedure of AI use globally." That is an industrial-policy sentence, opening a filing about whether copying newspaper articles into a training set is fair use.
Two paragraphs later comes the line that gives the document its stakes: "Of course, foreign adversaries have similar goals. Rules of law that make it significantly more difficult to develop a robust AI industry in the United States therefore threaten national security and give a competitive advantage to foreign adversaries who are not so encumbered."
Note what the government did not write. It never says China. Everyone reading it supplied that word themselves — including, as we will see, its critics on the right. The brief says "foreign adversaries," and if you are going to quote it, quote it.
Before it reaches the Copyright Act, the brief detours through the Government Accountability Office, warning that "[f]ailure to adopt and effectively integrate AI technology could hinder national security," because AI can be used for "analyzing intelligence information," "enhancing weapon systems (such as drones and robotic ships)," and "providing recommendations on the battlefield (such as where to target missile strikes)."
I read that paragraph three times. It is a copyright brief that reaches where to target missile strikes before it reaches the four fair-use factors.
Then it gets to copyright — and the argument is better than its framing
Now the part I have to be fair about, because the piece is worthless if I am not.
On the law, the brief is not a crank document. It argues that training is transformative, borrowing a district judge's phrase — "transformative—spectacularly so." That is the government quoting Judge Alsup in Bartz v. Anthropic, not a coinage of its own and not any appellate holding, but real judicial language from a real opinion.
It is also carefully bounded in a way almost nobody reported. The United States told the court it "focuses on the question whether the use of copyrighted works at the training stage—by copying works in order to feed data into the model as learning material—constitutes fair use." Trade press caught the distinction: the filing left open separate copyright questions involving the acquisition and storage of training data, and outputs that reproduce protected expression. If you read a headline saying the government blessed scraping, the headline was wrong. It did not say acquisition was lawful; it said nothing about acquisition at all. A footnote disclaims any suggestion of its own involvement: "The Government does not contend that any of the activities alleged or described in this litigation were authorized or consented to by the Government." And describing how the technology works, it writes that "The New York Times alleges" the training produces a model predicting likely next words. Alleges.
Then the economics, which is the argument that will outlive the news cycle. Requiring a license to train would hand the largest companies a moat: "It is not in the public's interest for the largest technology companies to have an oligopoly on LLM training due to licensing entry barriers that function primarily as large subsidies for old mainstream media companies." Plainly: large publishers would disproportionately benefit if AI companies had to pay. And so the ask, verbatim — it "would be problematic—and legally incorrect—to impose broad copyright liability that would generally render training of AI models impermissible without licensing."
Then the contradiction nobody has explained. In the same document, the government notes that "the President has encouraged Congress to consider enabling licensing frameworks" — while suggesting, the brief adds, that such legislation "should not address when or whether such licensing is required." Ask the legislature to build the plumbing; ask a judge to make sure nothing flows through it.
The footnote where the government sues its own bookkeeper
One more passage, and it is the one that made me sit up.
The United States Copyright Office — a federal body, housed in the Library of Congress — had already published its own analysis of AI training and fair use. The DOJ's brief deals with it in footnote 17, and the tone is not collegial. "But the Register of Copyrights, who is currently challenging her removal, appeared to endorse a similar theory in a report," it says, before concluding: "Her understanding does not warrant deference." Then it goes after the work directly: "the Register's threadbare reasoning ignored all the caselaw emphasizing the required use-by-use analysis …"
Two facts the coverage kept garbling, both of which I checked rather than assumed. First, the report: the Office has only ever released a pre-publication version of Part 3, issued May 9, 2025, with a final promised "in the future, without any substantive changes expected in the analysis or conclusions." It has never been finished.
Second, the Register. Shira Perlmutter is still in the job. The Office's leadership page still names her Register of Copyrights, and she is there because the D.C. Circuit said so: in September 2025 the court ordered the named officials and their subordinates "enjoined from interfering with appellant's service as Register of Copyrights and Director of the U.S. Copyright Office pending further order of the court." An injunction pending appeal, not a ruling that anything was unlawful. But it means the executive branch went into a federal court, in a private lawsuit, and told a judge to disregard the copyright expert it employs — noting in the same breath that she is fighting her removal.
You can believe the Copyright Office got fair use wrong; plenty of serious people do. It is still a peculiar thing to make that argument by footnote, in someone else's case, about an official you are simultaneously trying to fire.
So this is a scandal. Right?
No. And the version of this article that skips this section is propaganda with citations.
Start with the best objection to everything I have written. James Grimmelmann of Cornell Tech — about as unimpeachable a read on copyright as exists — looked at the filing and said its fair-use reasoning is well within the range of expert opinion. He is not blessing the national-security framing; he is saying the copyright analysis is the kind of thing a competent copyright lawyer might genuinely believe. That has to survive contact with my argument, or my argument is worthless.
Second, the distributive claim is not a Silicon Valley talking point invented for this brief. A year earlier, the Electronic Frontier Foundation — a digital-rights group that spends most of its life opposing this administration — argued that the Copyright Office's draft "puts a thumb on the scale in favor of copyright owners at the expense of creativity and innovation," and warned that prioritizing private licensing markets over user rights "would dramatically expand the market power of major media companies." When EFF and the DOJ land in the same place from opposite ends of the map, the conclusion deserves more than a sneer.
Third, the law genuinely is unsettled — whether AI firms can legally train on copyrighted material has, in TechCrunch's honest summary, "no clear answer," though judges have so far been largely favorable to the fair-use argument. Largely favorable is not settled. And fourth, the strongest defense of all: the filing decides nothing. Judge Stein can read it and put it down.
So why do I still think this matters?
Because persuasive but non-binding is an excellent description of a lobbying document, and we do not normally let the executive branch lobby a judge about two private parties' property. Because what makes an argument legitimate is not only its content but the forum, and how it got there. And because of a question nobody at Justice has answered: if the reason to file was that a plaintiffs' win would threaten national security, what private dispute is outside that logic? Supply chains. Chip patents. A trade-secret fight between two robotics startups. Every one can be narrated as an industrial-policy emergency, and § 517 has no outer bound.
Even the industry's own side found the move odd. Adam Eisgrau of Chamber of Progress — a tech-industry advocacy group, which is to say a friendly witness — said of the filing: "If it's not unprecedented, it's darn unusual."
When your allies describe your procedure as darn unusual, the procedure is the story.
Tokyo had this argument first — and had it in the open
Here is where my own history makes me an irritating person to argue with. Another country answered this exact question years ago, using the other branch of government. Japan did not wait for an AI company to be sued. It legislated.
The provision is Article 30-4 of the Japanese Copyright Act, and the striking thing is that everything the DOJ asked a judge to read into fair use, the Japanese Diet wrote into the statute itself. The text permits exploiting a work where "it is not a person's purpose to personally enjoy or cause another person to enjoy the thoughts or sentiments expressed in that work" — then adds a proviso that does real work: it "does not apply if the action would unreasonably prejudice the interests of the copyright owner." The statute even defines the permitted activity — "data analysis," meaning "the extraction, comparison, classification, or other statistical analysis" of elements from a large number of works — rather than leaving a judge to characterize it later.
Some history, because the popular version is wrong. Japan was, in the words of Waseda's Professor Tatsuhiro Ueno, the first country in the world to introduce a TDM exception — in 2009, not 2018; the 2018 amendment reorganized and broadened what was already there. Ueno, who calls Japan a "Paradise for Machine learning" (his own phrase, not a government slogan), is candid about how far it goes: the Japanese exception "can be considered more robust than those in other jurisdictions," he writes, because "rights holders cannot opt out of any TDM activities, there are no limitations on the content of permissible acts, and there are no lawful access or lawful copying requirements."
Read that carefully, because it torpedoes the lazy version of this comparison. Japan's legislature gave AI developers more than the DOJ is asking an American judge for. Japan is the industry's example, not the creators'.
Which is precisely why it is the right one. The contrast is not what the rule says. It is how the rule got made.
Japan's rule was approved at the 196th ordinary session of the Diet on May 18, 2018, promulgated a week later as Act No. 30 of 2018, and came into effect on January 1, 2019. A bill. A session. A vote. Legislators whose names are attached to the thing forever.
And the Agency for Cultural Affairs said out loud in 2018 — before ChatGPT existed — that the amendment would make it possible to carry out acts such as "recording a work in a database as training data for development of artificial intelligence (AI), without the authorization of the right holder." Nobody has to reverse-engineer that intent from a docket; it was published, in English, eight years ago.
Now the sentence that made me want to write this piece at all. In designing the system, the agency explains, Japan actually studied whether to adopt "highly flexible provisions similar to the U.S. fair use provisions" — and decided against it. Among the reasons it lists, alongside corporate compliance culture and the litigation system, is this one: "how roles are shared between the legislative and judicial branches."
Japan looked at the American model — a broad standard, filled in by judges, case by case — and decided that is not how it wanted the job allocated. Eight years later, the American executive branch walked into a courtroom to ask a judge to fill that standard in a particular way, on a schedule set by two private litigants, in a document most citizens will never see.
Japan is not a utopia here, and its executive branch got impatient too. In 2024 the Agency for Cultural Affairs published a General Understanding on AI and copyright, and was unusually honest about why: "Normally, the provisions of the Copyright Act should be interpreted by the judiciary on a case-by-case basis. However, there are currently very few court precedents that directly address the relationship between AI and copyright." So a council of experts, working since July 2023, set out to clarify things "without waiting for court precedents to accumulate". That document, like the DOJ's filing, is "not legally binding." But the symmetry stops there: one is guidance published to the world, available to a rights holder and a startup on the same morning; the other is a brief filed on one side of one lawsuit, at night, in a case the public cannot join.
Notice, too, what Japan's non-binding guidance did with its freedom: it narrowed the permission. The Article 30-4 proviso, it said, turns on "whether it will compete in the market with the copyrighted work" and "whether it will impede the potential sales channels of the copyrighted work in the future" — and applies, concretely, where a copyrighted database sold for information analysis is reproduced without compensation. It also warned that Article 30-4 "may not apply" to AI training that "picks off" a specific creator's style. May not, not does not; the hedge is theirs and I am keeping it.
The statute itself has not been amended since — the official translation database still gives the Copyright Act's "Last Version" as Act No. 30 of 2018, and the ministry's own list of recent amendments stops at 2023. The answer to the generative-AI era has been guidance and a "Platform for AI and Copyright Stakeholders' Dialogue" convened in June 2025, not a bill.
Which brings me to the part that settles the comparison, because Japanese newspapers are every bit as angry as American ones. The Japan Newspaper Publishers & Editors Association had been saying so publicly since July 2024, warning that AI-powered search engines use news articles without permission and are potentially infringing copyright. Nikkei and the Asahi Shimbun did go to court — a suit filed at the Tokyo District Court on August 26, 2025 against Perplexity, seeking an injunction and ¥2.2 billion each, framed around unauthorized copying, storage and display rather than training. (Live and undecided; that figure is a claim, not an award.)
But when those same publishers wanted the training rule itself changed, look at who they asked. Article 30-4, they complain, "contains no explicit opt-out provision allowing rights holders to refuse use"; what they want is for AI services to "obtain consent from copyright holders"; and the address for that demand is not a judge. They "urge the government and Diet" to expedite the establishment of systems for the appropriate protection of content.
Same grievance. Same industry. Same technology. Different building.
Now run the clock forward five years
Let me do the thing I always do here, and imagine.
It is 2031. The § 517 statement of interest is no longer unusual — it is a product. Every major firm has a practice group for it, with a name like Sovereign Interest Advocacy and a pitch deck explaining that when your litigation touches a designated strategic sector, there is a path to having the United States file paper on your side of the caption.
The process is not corrupt in any way a prosecutor could describe. It is worse: it is routine. You brief a deputy. The deputy briefs a policy shop. Somebody writes twenty pages, nobody outside the building reads a draft, no comment period opens, no vote is taken, and at 9:40 on a Tuesday night the sovereign's view of a contested question of private law appears on a docket — persuasive, non-binding, unreviewable.
The sectors multiply, because the national-security framing generalizes beautifully. Compute. Biotech. Battery chemistry. Port software. A shipbuilding subcontractor's wage arbitration becomes a matter of fleet readiness. Each time the argument is sincere; each time nobody can say where the line is, because Congress never drew one.
Here is the part that should worry you no matter which way you vote. The instrument has no politics. It has an owner, and the owner changes every four or eight years. The same statute that let this administration file for training-as-fair-use will let the next one file against it — or against a company it dislikes, for reasons it never has to publish. If you are cheering today because the filing points your way, you are cheering for a machine, and you do not get to keep the keys.
What the people who have actually read it are saying
The most useful thing about this filing is that it scrambled the usual teams.
From the right. Mike Davis of the Article III Project — a movement conservative, writing an opinion column rather than reporting — landed the sharpest line anyone has managed, in Fox News. The government, he wrote, claims unlimited access is essential to national security and warns that restriction risks losing the AI race to China: "That argument is un-American and absurd." (Note who supplied the word China: the critic, not the brief.) His frame is property: America must win "through markets that respect property rights, not through government-sanctioned expropriation."
From digital-rights advocates on the left. EFF, as above — the Copyright Office got fair use wrong, licensing markets entrench incumbent media, and they got there in May 2025 without any help from this DOJ.
From the academy, industry and the bar, the three voices already quoted above, pulling in three directions: Grimmelmann on the arguments being respectable, Eisgrau on the procedure being darn unusual, Meeker on the whole thing binding nobody.
And from the creators' side, an awkward pair of facts. On July 20, 2026, Judge Araceli Martínez-Olguín granted final approval of a $1.5 billion class settlement in Bartz v. Anthropic — publicized by the Authors Guild, and not a fair-use win, because the training holding in that same case went the AI company's way (it is the opinion the DOJ quotes). The court hedged too, calling the settlement "substantial benefits to the Class in light of the novel claims asserted." Novel claims. Nobody knows yet.
Then, sixteen days after the DOJ filing, the other shoe. Unredacted material unsealed on September 17, 2026 — surfaced by the plaintiffs, in their own framing — quotes Brent Hecht, a Microsoft director of Applied Science, in a January 2023 internal memo calling AI scraping "an astonishing theft of unprecedented proportions" and "the largest theft of labor in human history." One employee's words, never written for a courtroom — landing on the docket where the United States has told a judge that training is spectacularly transformative.
One more item, stated carefully. In July 2026, CNBC reported — sourcing the Financial Times — that OpenAI had proposed handing the U.S. government a 5% stake in the company. Proposed: early-stage talks, not a completed transaction, and I found nothing tying them to this filing. The brief does not mention them, and that silence is the criticism.
What does this mean for you?
You are not a party to this case. You will never read the docket. So here is what to actually do with it.
Learn the phrase "statement of interest," and treat it as a signal. When you read that the government "weighed in" on a private lawsuit, ask three things: was it a party, did it need the court's permission, and which branch decided. Here the answers are no, no, and the executive alone.
Separate the argument from the forum. You can think training is fair use — many thoughtful people do — and still think the way this position was delivered is a problem. Refusing to collapse those two questions is the whole discipline.
When a national-security rationale shows up in a commercial dispute, find the sentence. Here it is short and quotable: rules of law that impede the U.S. AI industry "threaten national security and give a competitive advantage to foreign adversaries." Ask yourself which private lawsuit that sentence could not be written into.
If you write, draw, code, or record for a living, price the uncertainty now. Nothing has been decided, in either direction. Read your platform terms on AI training; ask what your union, guild or professional body has filed. A contract clause you negotiate this year beats a ruling you are hoping for in 2028.
Watch the calendar, not the commentary — the summary-judgment schedule in New York and the pending appeal in Philadelphia, both below, will matter more than any op-ed.
Ask your representative one question. Not "do you support AI." Ask whether the Justice Department should have to publish notice before filing a statement of interest in private litigation, and whether there should be any standard at all for when it may. Congress wrote the one sentence. Congress can write a second one.
The lesson, as I see it
Nothing here is decided, and the honest close says so twice.
In New York, Judge Stein pushed summary-judgment replies to November 6 and set no date for hearing. In Philadelphia, the first federal appellate test of the whole question is already with three judges: the Third Circuit heard oral argument on June 11, 2026 in Thomson Reuters Enterprise Centre GmbH et al. v. Ross Intelligence Inc. — the one case so far to reject fair use for AI training — with a ruling expected in late 2026, which is a law firm's prediction rather than a court's promise. Reporting from the argument has the panel pressing both sides on the first and fourth fair-use factors: transformativeness and market harm. Exactly the two the DOJ chose to argue.
So the government picked a battleground a federal appeals court was already standing on, and arrived through a door with no lock and no visiting hours.
My vote? Keep the door, and build a porch light. I am not arguing the United States should never speak in private litigation — the civil-rights history of this power is a reason to be glad it exists. I am arguing that a power with no outer bound and no standard of review is not a power, it is a habit, and habits are how republics lose track of which branch decided what. Publish notice before you file. Say who asked and who was consulted. Let the other side answer on the record. That is one more sentence from Congress than the single one already on the books.
And if the fair-use answer really is as obvious as the brief says, let it be argued by the parties, decided by the judge, and — if we want it settled rather than merely won — written down by the legislature, in daylight, with names attached. Not a foreign idea: it is what a country with a more generous AI-training exception than ours did eight years ago, in an ordinary session, in May.
Somebody filed twenty pages at night, in a case you are not in, about who owns the sentences you read. That is the beat HAIA works — the paperwork that quietly decides things — and if you would rather hear about it before the ruling than after, this is where it lands.