The scoreboard
One table, every case in this article that has produced a judicial answer, plus the major cases still waiting for one. "Decided" below means a court answered the legal question on the merits; "settled" means the parties paid to end it; "pending" means the question is still live.
| Case | Court | Date | Status | What was decided | What it means for AI training |
|---|---|---|---|---|---|
| Thomson Reuters v. Ross Intelligence | 3d Circuit (No. 25-2153) | Sept 29, 2026 | Decided | Westlaw headnotes are copyrightable; using them to train a competing AI legal-search tool is not fair use. First US appellate fair-use ruling on AI training. | Training-data licensing markets count. Convenience is not a justification for copying. |
| Doe v. GitHub | 9th Circuit (No. 24-7700) | Sept 16, 2026 | Decided | AI-generated code is a new work, not a DMCA "stripped copy." Section 1202(b) reaches removal from existing works, not omission from new ones. | Attribution-stripping claims must prove CMI was removed from a real copy, not merely absent from an output. |
| Bartz v. Anthropic | N.D. Cal. (No. 3:24-cv-05417-AMO) | June 2025 ruling; settled July 20, 2026 | Decided, then settled | Training on lawfully acquired books is fair use; acquiring and keeping pirated-library copies is willful infringement. $1.5B fund, first payments by Nov 15, 2026. | How you got the data matters as much as what you do with it. Piracy carries a "hefty price tag." |
| Kadrey v. Meta | N.D. Cal. | June 2025 | Decided in part; still pending | Summary judgment for Meta on the fair-use theory, on the record the plaintiffs presented, which lacked market-harm evidence. Case continues on BitTorrent distribution claims. | A fair-use win is record-specific, not a blanket rule for AI training. |
| Getty Images v. Stability AI | UK High Court (EWHC 2863) | 2025 | Decided | Primary infringement claims dropped at trial; model weights held not to be an infringing copy. Limited trademark findings over generated watermarks. | Weights as such may not "be" the work, even where outputs echo it. |
| Chegg / Penske Media v. Google | D.D.C. | Sept 30, 2026 | Dismissed (antitrust) | Antitrust claims over AI Overviews dismissed: an expectation of search traffic is not an agreement. No copyright question decided. | Publisher leverage theories must plead a real bargain, not a general expectation. |
| Ted Entertainment v. ByteDance | N.D. Cal. (No. 3:25-cv-10933) | Oct 2, 2026 | Pending | Motion to dismiss denied: YouTube anti-scraping measures plausibly count as DMCA "access controls"; creators may sue over their circumvention. | Defeating scrape blocks to build a training set is legally risky even for public content. |
| New York Times v. OpenAI / Microsoft | S.D.N.Y. (MDL, Judge Stein) | Expected 2027 | Pending | Summary-judgment briefing underway; Sept 17, 2026 unsealing put internal company documents on the record; DOJ filed for the labs. | The generative-AI fair-use question the appeals courts have not yet answered. |
| New author suits v. Meta | N.D. Cal. | Filed Oct 19, 2026 | Pending | Four new class actions (Reno, Sheard, Winspear, Older) by hundreds of authors over book training, some alleging Meta's own book scanning also infringes. | The book-training fight did not end with Bartz. |
| GEMA v. OpenAI | Munich Regional Court I | Nov 11, 2025; on appeal | Decided; on appeal | Song lyrics memorized in model parameters and reproduced in outputs infringe; text-and-data-mining exceptions did not cover the memorization. | Memorized training data inside parameters can be infringement in the EU. |
Ross: the first appellate "no" on AI-training fair use
Thomson Reuters sued Ross Intelligence in 2020, alleging that Ross copied Westlaw headnotes, the short editorial summaries of points of law attached to judicial opinions, and used them to train a competing AI legal-search tool. Ross shut its platform down in 2021, citing litigation costs. In February 2025, Delaware federal judge Stephanos Bibas granted Thomson Reuters partial summary judgment and rejected Ross's fair-use defense. The Third Circuit affirmed both certified questions on September 29, 2026, in an opinion by Judge Montgomery-Reeves: the headnotes are original enough to be copyrighted, and the copying was not fair use.
Three things about this ruling matter, and all three are narrower than the headlines. First, the court held that copying because it was easier is not fair use. Ross already had the underlying judicial opinions and took the headnotes to make building its tool simpler; as Patently-O's Dennis Crouch put it, "unlike necessity, ease is not a justification for copying." Second, the market-harm factor turned on a developing licensing market: Thomson Reuters showed it was marketing headnotes as training data, and Ross's copying undercut that market's value. Third, and most important, the court described the dispute as "no more than an ordinary copyright case" and set generative AI aside. Ross's tool retrieved passages from opinions; it did not generate new text. No US appeals court has yet ruled on fair use for a generative model.
The reasoning itself was reported as sealed, so the public record of the opinion's full text is the later-published version commentators reviewed, and Ross has said it will seek Supreme Court review. What is known is the holding, and it is a first: a federal appeals court has now said that "AI training" on the label does not make copying fair.
GitHub: generated code is not a "stripped copy"
The programmers in Doe v. GitHub sued GitHub, Microsoft, and OpenAI in 2022, alleging that Copilot and Codex sometimes reproduce their open-source code without the attribution notices, copyright statements, and license terms the original files carried. They framed the claim under Section 1202(b) of the DMCA, which bars removing or altering copyright management information (CMI), a provision carrying statutory damages of $2,500 to $25,000 per violation. The plaintiffs' own damages theory ran past $9 billion.
The Ninth Circuit affirmed dismissal on September 16, 2026, in an opinion by Judge Eric Miller, and the reasoning turned on the plaintiffs' own description of how the tool works. A system "that does not look up and reproduce stored work but rather creates new work" cannot remove CMI from a copy that never existed, the court held. Generative output that resembles an existing work without carrying its attribution is not the same as taking an existing copy and stripping the attribution off it. The court rejected a strict identicality requirement, recasting it as an evidentiary inference: near-verbatim copying without CMI can still support an inference that CMI was removed, while material differences suggest a genuinely new work.
What the court did not decide is almost as important. It declined on procedural grounds to reach the "input" theory, whether CMI was stripped during training-data preparation, because the plaintiffs had forfeited that argument below. It also declined to decide whether AI outputs that reproduce copyrighted material can constitute ordinary copyright infringement. And it refused to give AI tools a categorical DMCA exemption: retrieval-based systems, which fetch and disseminate existing works, could face different treatment. Two contract claims remain pending in the district court, so this litigation is not over; only the DMCA theory is dead.
Bartz: piracy has a price tag; licensed training does not
The most expensive sentence in AI copyright law so far came from Judge William Alsup in June 2025: Anthropic's method of training on legitimately acquired books qualifies as fair use, but its acquisition and retention of pirated copies from shadow libraries, Library Genesis and Pirate Library Mirror, was not fair use and was actionable willful infringement. As Alsup wrote, Anthropic could have lawfully purchased the books but chose to "steal" them to circumvent what its own CEO called "legal/practice/business slog."
The parties settled rather than test that finding at trial. After an initial proposal was denied preliminary approval in September 2025, the case was reassigned for settlement administration to Judge Araceli Martinez-Olguin, who granted final approval on July 20, 2026. The numbers: a $1.5 billion non-reversionary fund covering roughly 482,460 works; class counsel's fee cut from $187.5 million requested to about $101.6 million; lead-plaintiff service awards reduced from $50,000 to $15,000 each; a claims rate above 91 percent. On September 15, the court approved a payment plan sending the first round out by November 15, 2026, with the first-round per-work figure estimated around $2,203.56, reflecting that only about 70 percent of the fund will be in escrow by then.
ByteHerald's reading: Bartz split the two questions the industry had been fusing into one. Training itself, on books you acquired lawfully, can be fair use. Building your training pipeline on industrial-scale piracy is a separate act with a separate, and in this case historic, price. Anthropic did not admit liability, and a settlement is not a precedent, but the split fair-use ruling that produced it is.
Kadrey: a fair-use win on a thin record
Kadrey v. Meta is the case fair-use absolutists cite most, and it deserves a careful read. In June 2025, Judge Vince Chhabria granted Meta summary judgment on the plaintiffs' theory that training Llama on their copyrighted books harmed the market for those works. But the opinion stressed that the result followed from the evidentiary record and the arguments actually presented, not from any categorical holding that AI training is always fair use. The plaintiffs, in the court's view, had not developed market-harm evidence that could have succeeded elsewhere.
The case did not end there. In March 2026, Chhabria let the plaintiffs amend their complaint to press contributory-infringement claims based on Meta's use of BitTorrent: while downloading books from shadow libraries, Meta allegedly made portions of those works available to other torrent users, contributing to third-party infringement. Discovery on Meta's torrenting practices has continued through 2026, including privilege fights over internal communications about the legality of the practice. And the book-training fight has now multiplied: on October 19, 2026, four new class actions (Reno, Sheard, Winspear, and Older v. Meta) were filed by hundreds of authors, some alleging that even Meta's scanning of physical books it lawfully owns is infringement. Kadrey proves one thing clearly: a fair-use win on a thin record settles nothing for anyone else.
ByteDance: anti-scraping tools are access controls
The newest front is about how the data was obtained, not what the model does with it. In Ted Entertainment Inc. v. ByteDance, No. 3:25-cv-10933 (N.D. Cal.), YouTube creators Ted Entertainment and Matt Fisher allege that ByteDance circumvented YouTube's technological protections, a rolling-cipher signature system, IP blocking and rate limiting, session-bound streaming URLs, CAPTCHA challenges, and proof-of-origin tokens, to mass-download videos for training its text-to-video systems, including MagicVideo and Seedance.
On October 2, 2026, Judge Jacqueline Scott Corley denied ByteDance's motion to dismiss the DMCA anti-circumvention claim. ByteDance's central argument was that the videos were publicly viewable, so no real restriction was bypassed, and that only YouTube could complain about the circumvention of YouTube's own measures. Corley rejected both: the safeguards plausibly count as "access controls" under the DMCA, and copyright owners whose works those controls protect are within the class Congress meant to protect, even when a third party built the lock. The court did not rule on whether training on YouTube videos is copyright infringement; a motion to dismiss tests only whether the complaint states a claim. The case now proceeds past the pleading stage.
Chegg and Penske: the antitrust door stays shut
The dismissals that require the most careful reading are the ones from September 30, 2026. Judge Amit Mehta of the D.C. federal court granted Google's motions to dismiss the amended complaints in Chegg v. Google (No. 25-cv-543) and Penske Media Corp. v. Google (No. 25-cv-3192), two lawsuits accusing Google of using publisher content in AI Overviews without compensation while making publishers accept that use as the price of staying indexed in Search.
The court held that the publishers' reciprocal-dealing theory, that Google conditions search traffic on receiving content for free, failed because they pleaded no agreement. "Plaintiffs have pleaded only that they have an 'expectation' that Google will send them search traffic if they make their content available for free," Mehta wrote. "But an expectation is not an agreement. It is simply how a general search engine works." The opinion does not address copyright at all, and it expressly declined to rule on the state-law unjust-enrichment claims, dismissing those on jurisdictional grounds. Penske's order is final and appealable.
Do not read this as a ruling that AI Overviews are legal. It is a ruling that these antitrust theories, pleaded this way, cannot proceed. The commercial question of what publishers are owed for AI answers, and the copyright question of whether the answers infringe, are both still live. Mehta himself said he is "not unsympathetic" to publishers whose content Google takes and repurposes without compensation, but that antitrust law does not substitute for legislation.
The data question behind every case
Read the verdicts together and a pattern emerges: courts keep deciding the acquisition question before the training question. Bartz split licensed books from pirated libraries. Kadrey's surviving claims are about BitTorrent distribution. ByteDance is about defeating scrape blocks. Even Ross turned partly on the fact that the underlying opinions were freely available and Ross copied the editorial layer anyway. The industry debate asks "is training fair use"; the courts are increasingly asking "how did you get the data, and who did you bypass to get it."
That is why the parallel contract fight matters. We mapped how Meta's commercial terms, effective October 30, 2026, give it written permission to crawl any domain tied to a business account and use what it collects to train AI, a clause three legal experts told Nieman Lab reads as a training-data license smuggled into terms of service. Contract permission, robots.txt, licensing deals, and lawsuits are becoming one continuous negotiation over who controls the training-data supply, and the courtroom results above are only one part of it.
Fair use in plain English, applied to AI training
The statute lists four factors courts must weigh. Here is what each one means when the use in question is training an AI, with the 2026 rulings as illustrations.
Factor one asks about the purpose and character of the use: was it transformative, adding something new with a different purpose, or just a convenient shortcut? Courts have split here. Training on lawfully acquired books was transformative in Bartz, because the model learns statistical patterns rather than reproducing the books for readers. But Ross's copying was not transformative, because the point was to build a product competing directly with the source, and the court treated the AI-training label as irrelevant to that purpose. ByteHerald's analysis: the more the output competes with the input, the weaker factor one gets.
Factor two asks about the nature of the copyrighted work: creative, expressive material gets stronger protection than factual compilations. Ross answered this one directly: Westlaw headnotes, short editorial summaries written by humans, are original enough to be copyrighted. This factor rarely decides an AI case on its own, but it sets the floor: training on phone-book listings is a very different case from training on novels.
Factor three asks how much was taken: a snippet weighs less than the whole work, and whether the amount taken was reasonable for the purpose. Bartz is the textbook illustration. Training required full copies, and that was fine for the lawfully acquired books; but retaining entire pirated libraries was willful infringement. The factor is not about the model at all, it is about the copies sitting on the defendant's servers.
Factor four asks about the effect on the market for the original: does the use substitute for buying the work? This is usually the most decisive factor, and it cuts both ways. In Ross, the court credited a developing licensing market for headnotes as AI training data, so Ross's copying harmed a market that was forming. In Bartz, the court noted that Anthropic could have purchased the books it downloaded from pirate libraries, bypassing a market that already existed. In Kadrey, the plaintiffs lost precisely because they failed to develop market-harm evidence. ByteHerald's analysis: factor four is where the next decade of AI copyright cases will be won and lost, because training-data licensing markets are forming right now, and every licensing deal signed changes what "the market" looks like for the next case.
July through October 2026: the dated rulings that set the current law. Chart: ByteHerald original.
What no court has decided yet
The honest summary of the state of AI copyright law in October 2026 is a short list of answered questions and a much longer list of open ones. Answered: non-generative training to build a competing product is not fair use (Ross); AI outputs are not DMCA stripped copies (Doe v. GitHub); pirated training data is willful infringement while lawfully acquired training data can be fair use (Bartz); and publisher antitrust theories about AI Overviews fail at the pleading stage (Chegg/Penske). Open: whether training a generative model on copyrighted works is fair use, the question the Third Circuit explicitly reserved; whether removing attribution during training violates the DMCA, the question the Ninth Circuit declined to reach; and what the New York Times case, now in summary-judgment briefing before Judge Stein with a ruling not expected until 2027, will do to the generative-AI fair-use defense after the September 17 unsealing put the labs' own internal documents on the record.
The docket is still growing faster than it is shrinking. Axis Intelligence's AI Copyright Lawsuits Tracker, published under CC BY 4.0 and updated weekly, carried 49 rows covering cases from May 2020 through October 4, 2026. A separate count by copyright professor Edward Lee put the number of copyright suits against AI companies in the United States at 150 in mid-October. An anonymous editorial tracker counted 248 AI-related cases against 76 companies as of October 8, a broader category that includes privacy, antitrust, and employment claims. Against that volume, six answered questions is a beginning, not a body of law.
For anyone building on AI models, the practical reading of the scoreboard is straightforward. Buy your training data or license it; the Bartz price tag is now the public benchmark for piracy. Document how you acquired every dataset; Kadrey shows the record decides the case. Do not defeat scrape blocks to build corpora; ByteDance shows that defeat itself is now actionable. And if your product competes with the sources it trained on, Ross shows that the AI-training label will not save you. The courts have not decided what AI training is. They have decided, case by case, what it is not.
By ByteHerald Staff. Ross details from Patently-O, Jackson Walker, and PYMNTS (opinion reasoning reported as sealed in some coverage). Doe v. GitHub from LexBlog, LexBlog, and Authors Alliance. Bartz timeline and payment figures from AI Lawsuit Tracker and Settlement Insight (figures from the Sept 2, 2026 status report). ByteDance from Thomas James and SLOP TV News. Chegg/Penske from Reuters. Kadrey's ongoing status from JD Supra; new Meta author suits from Bloomberg Law. Tracker counts from Axis Intelligence (49 rows, through Oct 4, 2026) and the ExplainX writeup of the anonymous 248-case tracker. Internal link: Meta's Oct 30 terms and the training-data contract fight.
Comments
Post a Comment