AI Copyright: Third Circuit Rejects Fair Use as DOJ Backs OpenAI

Two signals from Washington and Philadelphia, a month apart

The biggest open question in AI copyright law is whether copying protected works to train a model counts as fair use. In September it got two answers that point in different directions. One came from a federal appeals court. The other came from the executive branch. They do not conflict as much as the reactions suggest, but read side by side they show where the fight is going and how much is still undecided.

On September 29, 2026, the U.S. Court of Appeals for the Third Circuit upheld Judge Stephanos Bibas's summary-judgment ruling for Thomson Reuters against Ross Intelligence. It agreed that Westlaw's headnotes are protected by copyright and that Ross's use of them to train an AI tool was not fair use.11 Coverage generally calls it the first time a U.S. appeals court has decided fair use in AI training.313 The panel first issued a one-line order reading "AFFIRMED," then released a sealed opinion that was made public the next day.19

Four weeks earlier, on September 1, the Department of Justice filed a Statement of Interest in the consolidated OpenAI copyright litigation in New York. It argued that training AI models on copyrighted material does not, by itself, violate copyright law.41 Several outlets describe this as the first time the federal government has formally taken a side in the wave of AI training cases.44

What the Third Circuit actually decided

The facts in Thomson Reuters v. Ross are narrow. The infringement finding covered 2,243 Westlaw headnotes. Ross used them to build a legal research tool meant to compete directly with Westlaw.1320 The opinion, by Judge Tamika Montgomery-Reeves, says the case may look like it concerns the future of AI legal technology but is really "no more than an ordinary copyright case."15

The court found that three of the four fair-use factors weighed against Ross.19 It called the use only minimally transformative because Ross used the headnotes for the same purpose Westlaw did: helping lawyers find relevant law. Using them only during training did not change that.3 The court also rejected Ross's comparison to software cases that allowed "intermediate copying." Ross did not need the headnotes to reach the unprotected court opinions underneath them. Copying was simply the easy route.313

The market-harm reasoning is likely to matter most later. The court found harm in two places: the existing legal research market, and a developing market for licensing headnotes as AI training data. Thomson Reuters was already using headnotes to train its own AI products.3 That second finding treats a market for AI training licenses as real, legally recognized harm. Rights holders have been trying to establish exactly that.

Everyone claimed victory, and both sides have a point

Reactions split quickly. Music industry groups RIAA and NMPA had filed a joint brief arguing that training a model to compete with copyrighted works "cannot qualify as fair use." They pointed to label licensing deals with Stability AI, Udio, KLAY and ElevenLabs as proof that a training-data market exists.4 On the other side, Adam Eisgrau of the tech group Chamber of Progress said the ruling implicitly confirms that highly transformative generative AI training is likely fair use. Aaron Moss of Copyright Lately joked that the victory laps made it look as if everyone had read a different opinion.1

The opinion supports both readings, and that is why its effect on the larger cases is limited. The court separated Ross's tool, which returned existing judicial passages, from systems that generate new expression. It expressly left the generative AI question unresolved.319 Nieman Lab's reading is similar: the ruling makes clear that a tool copying text to compete head-on will struggle, but it leaves room for generative tools to be found more transformative.15 Commentator James Astrachan argues the decision should apply narrowly to cases where protected material is used to build a direct competitor.17

My reading is that the case matters more for how it reasons than for who won. A court refused to treat AI training as a new category of law and applied the ordinary four factors, with market substitution doing the decisive work. That standard is harder to argue around than special rules for AI would be.20 For generative AI defendants, the warning is factor four: if plaintiffs can show a licensing market and direct competition, "transformative" may not carry the day.

The government's opposite bet

The DOJ filing goes further in the other direction. It calls LLM training "extraordinarily transformative." It warns that requiring licenses could leave model development to only the largest tech companies.42 It also rejects as "deeply flawed" the market-dilution theory Judge Vince Chhabria raised in Kadrey v. Meta: the idea that floods of AI-generated content can harm authors even when no output copies their work.4248 The brief describes licensing fees as barriers that would mainly act as subsidies for "old mainstream media companies."48

The filing has clear limits. It is not binding, and Judge Sidney Stein can ignore it.4446 It deals only with the training stage and leaves aside whether training on pirated material is fair use.4243 That gap is significant. In Bartz v. Anthropic, Judge William Alsup found training on lawfully acquired books "spectacularly" transformative but refused to excuse a central library of pirated copies.5 That case ended in a roughly $1.5 billion class settlement. It received final approval on July 20, 2026, covering 482,460 books, with all 53 objections overruled.1213

The filing also fits a consistent administration position. In its March 20, 2026 National Policy Framework, the White House said training on copyrighted material does not violate the law, acknowledged that "reasonable arguments to the contrary exist," and said the courts should decide.23 The same framework suggested Congress consider collective licensing similar to the ASCAP and BMI system for music, without deciding when licenses would be required.28 That puts the administration at odds with the Copyright Office. The Office's May 2025 report concluded that training which produces outputs competing with the originals is unlikely to be fair use.26 The administration later removed Register Shira Perlmutter, and the fight over that removal reached the Supreme Court.43

A docket that keeps growing

The lawsuits are still multiplying while the doctrine stays unsettled. One tracker counts 144 AI copyright cases.12 Recent filings include USA Today's suit against OpenAI on October 8, a suit by 26 local publishers seeking coordination with the OpenAI case, and a second Universal Music suit against Suno covering 60,202 recordings.11 Many related cases are paused until Judge Stein rules on summary judgment in the consolidated OpenAI litigation.12 That ruling will be the next major test.

Some side issues are getting narrower. On September 16, the Ninth Circuit upheld dismissal of the DMCA §1202(b) claim in Doe v. GitHub. It reasoned that an AI system producing new code without copyright notices does not remove or alter anything.12 That weakens a claim found in many author and publisher complaints. The Supreme Court's Cox v. Sony decision has also changed contributory-infringement claims. OpenAI used it to challenge news plaintiffs, and in August the court refused to let the New York Times and Daily News add new contributory theories against Microsoft.1314 Appellate guidance on generative AI is not coming soon. In July, Judge Chhabria refused to send Kadrey to the Ninth Circuit early, so the Third Circuit remains the only appellate ruling on the question.13

One related question is largely settled. The Supreme Court declined on March 2, 2026 to hear Thaler v. Perlmutter, leaving in place the rule that copyright requires a human author.24 Congress has bills pending, including the Schiff-Curtis CLEAR Act, which would require companies to file notices summarizing the copyrighted works in their training data. None has become law.2230

The legal profession's own AI test

The first appellate AI copyright ruling involves legal research software, and the legal profession is now one of the most aggressive adopters of the technology. Measured adoption varies a lot by survey. Clio reports 79% of legal professionals using AI in 2026.39 Thomson Reuters puts generative AI use at 41% of law firms.36 The ABA's own technology survey found 30% of lawyers using AI tools.32 Part of the gap comes from what each survey counts: individual use versus firm-wide deployment.32

Oversight has not kept pace. One 2026 survey found only 9% of firms have a written AI policy that is actually enforced.31 A database maintained by researcher Damien Charlotin has logged 2,185 court filings worldwide containing AI-hallucinated material. Monetary sanctions appear in 380 of them.33 The penalties are getting bigger. They include $15,000 per attorney from the Sixth Circuit in Whiting v. City of Athens34, and a reported $110,000 sanction in Oregon.39

For the companies building legal AI tools, the Ross decision adds a concrete risk: training a competing product on a rival's editorial content without a license has now lost on appeal. Generative models trained on books and news may still win on fair use. But the courts have made it clear that every ruling will depend on what was copied, how it was obtained, and whose market the product takes.