AI Copyright: DOJ Backs Fair Use While Calling Distillation Theft

Two positions on copying

In September, the U.S. government took two positions on AI copying that are hard to reconcile. Early in the month, the Justice Department asked a judge to side with AI companies in more than a dozen copyright lawsuits. Its filing argued that any ruling for creators would be "erroneous" and would rest on a misreading of fair use.11 A few days later, on September 8, the FBI, NSA and Cybersecurity and Infrastructure Security Agency named six Chinese developers: DeepSeek, Moonshot, Alibaba, MiniMax, StepFun and Z.ai. The agencies accused them of running industrial-scale campaigns to pull capabilities out of models built by OpenAI, Anthropic and Google.21

Put together, Washington says it is fine for American AI labs to learn from the work of every author, musician and newspaper without paying. When Chinese labs learn from the outputs of American AI models, Washington calls it theft. The Atlantic's Alex Reisner described this as a hypocritical stance on intellectual property.11 That reading holds up. The more useful question is what the double standard means for the 130-plus lawsuits now moving through the courts, and for the first appeals ruling on AI training, which came down at the end of the same month.

What the DOJ argued

The Justice Department filed a statement of interest, a tool the executive branch uses to tell a court how it reads the law. Its main argument was that American "AI dominance" is a national-security matter and that no legal doctrine should slow it down.11 The filing pointed to AI's uses in intelligence analysis, weapons such as drones and robotic ships, and battlefield targeting advice.11 Reisner noted that the filing offers no evidence for why short-story collections or music videos would matter to choosing missile targets.11 Its more modest point was that language models help officials work through facts reported in outlets like The New York Times. That is also how millions of ordinary users already use chatbots, and it is a thin basis for setting copyright aside.11

The filing fits the administration's track record. When it released its AI Action Plan in 2025, the plan said nothing about copyright. Trump then said in a speech that paying for every piece of training data was "not doable" and that China was "not doing it."23 The administration later fired the head of the U.S. Copyright Office after her office published a report suggesting AI training might not always be fair use.11 Former AI czar David Sacks has argued that models would be "crippled" if developers had to license training data.11 Earlier this year, the administration's AI framework said it believed training did not violate copyright but that the courts should decide.11 The Fairness and Innovation Foundation also described the White House as advising Congress to stay out of the fair-use question.20 The DOJ filing drops that restraint. The administration is no longer waiting for the courts to decide. It is trying to steer them.

Not every conservative agrees. Mike Davis, an attorney and political strategist allied with the administration, called on the DOJ to withdraw the filing. He argued that in a free market, AI companies pay for their inputs like any other business.11 Senate Republicans have gone further. At a 2025 hearing, Sen. Josh Hawley called mass ingestion of copyrighted works "the largest intellectual property theft in American history." He dismissed the companies' argument about beating China as a request for permission to take creators' work.14 Sen. Marsha Blackburn's TRUMP AMERICA AI Act would amend the Copyright Act to say that unauthorized AI training "shall not constitute fair use."20 Policy groups that back the industry call that provision a disaster that would make virtually every major model illegal.20

The distillation side

The administration has escalated its campaign against Chinese distillation all year. In April, White House science adviser Michael Kratsios accused China-based entities of using tens of thousands of fake accounts to extract capabilities from American models.27 In July, Moonshot released Kimi K3. Treasury Secretary Scott Bessent then said Chinese models found to be "stealing from our great companies" could face sanctions, and he called it "IP theft."22 Officials have drawn a line between legitimate, small-scale distillation and "covert, industrial-scale" extraction done through fraud and violations of companies' terms of service. They have said sanctions and Entity List designations are possible responses to the second kind.24 The September advisory went further. It said stolen capabilities are the core of China's AI strategy, not just a supplement to it.28

The government's own language undercuts its legal footing here. Asked about distillation, Georgetown law professor Anupam Chander pointed out that the AI companies themselves have argued that learning from other people's copyrighted work is fair use.26 A legal analysis published by IP.com goes further. It argues that the "IP theft" label has no clear legal claim behind it. Trade-secret law covers stolen model weights, but distillation doesn't take the weights. A copyright claim would cut against the arguments the labs have largely won. Breach of contract may be the best option, but Stanford's Mark Lemley doubts that non-compete-style terms in user agreements would hold up in court.15 The same analysis notes that no frontier lab has sued a distiller. It also points to Elon Musk's sworn testimony in April that xAI had "partly" trained on OpenAI models, which he called common practice in the industry. Nothing came of it.15 Beijing has made the same point from the other side. Its foreign ministry said many U.S. AI companies have used Chinese models for distillation.28

There is a real distinction available to Washington. The fraud, the fake accounts and the use of proxies to get around access controls are different from scraping books and websites.2124 But the administration has not chosen to argue on those narrower grounds. Its officials keep calling it theft of intellectual property, and that language matches the creators' complaint against American labs almost word for word.

The first appeals ruling

On September 29, the Third Circuit gave the government's fair-use argument its first real setback. The court upheld a ruling that Ross Intelligence, a now-defunct legal-research startup, infringed Thomson Reuters' copyrights by training its tool on Westlaw headnotes.9 Judge Tamika Montgomery-Reeves wrote that the headnotes had the "creative spark" needed for copyright protection. She found that Ross used them for a highly similar purpose because it wanted to compete directly with Westlaw.1 It is the first federal appeals decision to reject fair use for training a commercial AI product on publisher content without permission or payment.3

Coverage agrees on the facts but not on how much the ruling matters. Tech Times called it unambiguous at the riskiest end of the spectrum.7 Music Business Worldwide emphasized that the RIAA and NMPA backed Thomson Reuters, and that the ruling does not decide the Suno and Udio music cases.4 Nieman Lab and the law firm Jackson Walker focused on how carefully the court limited its holding. Ross's tool could only retrieve existing opinions and could not produce new expression. The court said that made it unlike the generative model in Bartz v. Anthropic.39 Tech Times also noted that the opinion discusses the DOJ's description of large-language-model training as "extraordinarily transformative."7

The narrower reading is the right one. The ruling binds federal courts only in Delaware, New Jersey and Pennsylvania, and does not bind the Ninth Circuit, where most generative-AI cases are being heard.7 It establishes something more limited but still important: training is not automatically protected. When a product competes with the work it was trained on, and a licensing market for that work exists, the defense gets much weaker.7 That fits the district-court picture. In Bartz, Judge Alsup found that training on lawfully bought books was fair use but that keeping pirated copies was not.5 That split produced a $1.5 billion settlement, which received final approval in July.6 In Kadrey v. Meta, the court called training transformative but stressed that harm to the market remains the most important factor.2

The Ross case also matters for legal technology. Many legal AI tools are built to do much the same job as the research databases they draw on. The Third Circuit has now said that is the riskiest position to be in.7

What comes next

More rulings are close. Summary-judgment hearings are scheduled in October for Concord v. Anthropic, the music publishers' lyrics case, and for the Mosaic LLM litigation.6 The music-label cases against Suno have been pushed back, so a U.S. fair-use ruling on them is not expected until 2027.8 Courts elsewhere are moving faster. A Munich court ruled for the German music-rights society GEMA against Suno in July, after an earlier ruling for GEMA against OpenAI over memorized lyrics.8 Meanwhile, The Times's lawsuit against OpenAI turned up documents showing OpenAI developers got around the paper's paywall. A Microsoft executive described the operation as "an astonishing theft of unprecedented proportions."11

The administration has built its China policy on the idea that copying an AI model's output without permission is theft. It has built its domestic litigation strategy on the idea that copying everyone else's work is fair use. Judges deciding the coming cases will be able to see both positions. The Third Circuit has already shown that a court will look closely at what was copied and for what purpose, regardless of how strongly Washington invokes national security.