A memorandum of law filed September 4, 2026 by Orrick, Herrington & Sutcliffe LLP for defendant Microsoft Corporation, signed by Annette L. Hurst, in In re: OpenAI, Inc. Copyright Infringement Litigation, No. 1:25-md-3143-SHS-OTW (S.D.N.Y.), Doc. No. 1690, 42 pages, filed as a redacted public version. It relates to Case Nos. 1:23-cv-08292 and 1:23-cv-10211 — the Books Plaintiffs' consolidated cases tracked at Authors Guild v. OpenAI — and appears in the member docket as entry 1211. Microsoft filed a separate brief the same day in the News Cases, which this brief refers to but does not contain.
The requested relief is summary judgment for Microsoft on all claims the Books Plaintiffs assert against it, on the ground that "the use of copyrighted works to train a general-purpose LLM is fair use as a matter of law."
The claims at issue
The brief characterizes the allegations against Microsoft as narrow: the only LLMs addressed in the consolidated complaint are OpenAI's models, and the complaint "contains no allegations concerning the operation, use, or output of any Microsoft product." The claims pleaded are direct infringement for "reproducing the works owned by Plaintiffs … to create and develop datasets used to train … artificial intelligence models" (Compl. ¶ 311), plus vicarious and contributory liability for OpenAI's training of the GPT models. A footnote states that the contributory claim is foreclosed by Cox Communications, Inc. v. Sony Music Entertainment, 146 S. Ct. 959 (2026), and is the subject of a contemporaneous motion for judgment on the pleadings.
The fair-use argument
Factor one — purpose and character. The brief argues training is transformative in both character and purpose, leaning on two district decisions it says have already so held: Kadrey v. Meta Platforms, Inc., 788 F. Supp. 3d 1026 (N.D. Cal. 2025) ("highly" transformative, no "serious question"), and Bartz v. Anthropic PBC, 787 F. Supp. 3d 1007 (N.D. Cal. 2025) ("spectacularly" and "exceedingly" transformative). On character, it describes the mechanism — the transformer breaking text into tokens assigned unique integers, which combined with trillions of others influence embeddings in high-dimensional space, producing "a vast 'mathematical operation' comprised of hundreds of billions of 'parameters' used to predict the next token" — and argues this "goes vastly beyond a book, a library, or, say, a searchable digital database of books," analogizing to reverse engineering under Sony Computer Entertainment v. Connectix. On purpose, it argues that books are made "to be read" while training uses them "for a technological purpose," naming John Grisham's The Street Lawyer and Stacy Schiff's Cleopatra as illustrations.
Factors two and three — nature and amount. The brief argues these "carry little weight" given the transformativeness, citing Authors Guild v. Google, 804 F.3d 202, 220 (2d Cir. 2015) that the second factor "has rarely played a significant role." In the alternative it argues both favor the defendants: training is not an unchanged use because it breaks works into tokens and translates patterns into model weights, and any copying is "necessary to gain access to … unprotected functional elements."
Factor four — market effect. The brief argues the GPT models operate in "a completely 'distinct market[]': the market for generative AI technology," and that the discovery record shows no traditional-market harm. It cites plaintiff testimony, quoting David Henry Hwang that he "do[es not] believe" the defendants' LLMs are "hurting the market for [his] plays," and states that sales analysis shows no drop attributable to the models. A quotation from Jonathan Franzen's deposition is redacted from the public version.
Responses to the plaintiffs' three theories
Shadow libraries. On OpenAI's downloading of books from Library Genesis, the brief makes two arguments. Against direct infringement it relies on volitional conduct under Cartoon Network v. CSC Holdings ("Cablevision"), asserting that Microsoft "did not download—and indeed had no involvement at all in acquiring—the datasets from LibGen." Against the theory generally it argues provenance is immaterial to fair use, since "any use for which a defendant asserts a fair use defense is unauthorized," citing Kadrey that downloading from a shadow library "must still be considered in light of its ultimate, highly transformative purpose." It addresses the bad-faith line of cases directly — Harper & Row weighing the "propriety of the defendant's conduct" against the Supreme Court's later "skepticism about whether bad faith has any role in a fair use analysis" in Google v. Oracle — and quotes Judge Jacobs's concurrence in NXIVM that fair use is not a privilege "earned by good works and clean morals," and Pierre Leval that "[c]opyright is not a privilege reserved for the well-behaved." The brief states that Microsoft "works hard to combat piracy every day," citing an injunction granted against Library Genesis in Cengage Learning v. Doe (S.D.N.Y. Sept. 24, 2024).
Memorization. The brief argues the term is "something of a misnomer" because it describes outputs resembling training data rather than an observed model property, and that regurgitation "is not an intended feature" — engineers "actively work to avoid" it. It then gives the discovery figures, which are the brief's principal factual contribution:
| Measure | Figure as stated |
|---|---|
| Extraction attempts by the plaintiffs' expert Dr. Shawn Shan, feeding verbatim book passages hundreds of words long | approximately 5.3 million |
| Share of those attempts producing any 30-word match | fewer than 1% |
| Copilot conversations reviewed | 8.2 million |
| Copilot responses containing 30 matching words | 24 |
| Resulting rate | .00029% |
| Books Plaintiffs' asserted works with no regurgitation found in the logs | 202 of 212 |
The brief characterizes Shan's method as an "adversarial extraction protocol" in which he "already had the text, decided what passage to try to extract, then used surrounding text to bombard the LLM until it emitted what he was looking for," and argues that the low out-of-lab rate is itself evidence that regurgitation is "neither the purpose nor the effect of LLM training."
Bing Index. On Microsoft's provision to OpenAI of portions of its Bing search index, the brief says the evidence is inconclusive as to whether the transferred portions contained any plaintiff works, and that at best the plaintiffs' expert identified "between 8 and 24 of Plaintiffs' works" in an earlier and a later transfer respectively. It argues the transfer was part of training in any event, quoting the plaintiffs' own expert conceding that "if you define 'training' broadly as contributing to a final product model," the provision was a step in the process, and that early portions were used for "computational analysis" to determine whether web data is useful — "a reasonable first step toward training an LLM" under Kadrey.
The two alternative market theories
A training-data licensing market. The brief argues the loss of licensing revenue is not cognizable because the market for "licensing … works for the narrow purpose of training LLMs" is "not one the Copyright Act entitles Authors to exploit" (Bartz), and that under HathiTrust a lost-sale theory tied to a transformative use "does not work" because "economic 'harm' caused by transformative uses does not count." It argues that existing licensing deals are legally irrelevant under Castle Rock, that inferring market harm from a "willingness to pay" is a fallacy under Whyte Monkee Productions v. Netflix, 174 F.4th 761, 805 (10th Cir. 2026) — parties "may have done so simply to avoid the disruption, expense, and risk associated with lawsuits like this one" — and that sporadic licensing does not establish a "reasonably efficient, reasonably priced, [and] convenient [to] access" market, since a single author's work supplies "only an infinitesimal fraction of the tokens" and rights are dispersed across millions of authors with no settled means of valuation.
Market dilution. On the theory that downstream non-infringing works written with LLM tools might compete with the plaintiffs' books, the brief argues the framing "has it precisely backwards" because "[n]ew non-infringing creative works are a copyright good—they are the whole point of fair use," and notes that Kadrey, which raised the theory sua sponte, declined to weigh it for want of evidence. It cites the defendants' economic expert Tucker that "increasing the number of available books does not mechanically reduce demand for existing works" and that "[t]he overwhelming majority of consumers would not purchase an AI-generated book—even if it were drastically less expensive," and states that several Books Plaintiffs conceded in deposition that they do not expect consumers to prefer AI-generated works.
The scope Microsoft concedes
The closing section limits the holding sought: "such a holding does not mean that every downstream use of an LLM—every LLM-based product, every end-user contrivance—is lawful. That the foundational technology underlying the internet is lawful does not mean every website and app is too. Likewise here, every LLM-based tool must be considered in light of its own design and function." The brief points to Microsoft's separate News Cases brief as addressing the specific tool the News Plaintiffs challenge, and says "[t]ools challenged in future cases may not be" lawful.
Provenance
Retrieved from CourtListener RECAP at storage.courtlistener.com; the docket page is courtlistener.com. The document carries its own ECF header and page stamps on all 42 pages and Hurst's signature block. The Verge reported the same filing date and the same fair-use argument on September 4, 2026 (Source: theverge.com).
Relationships
- supports: Authors Guild v. OpenAI
- related: NYT v. Microsoft, OpenAI et al.
- depends-on: Bartz v. Anthropic
- depends-on: Kadrey v. Meta
- related: AI Copyright Litigation — Analysis
- related: AI Copyright
- related: Microsoft
- related: OpenAI