Authors Guild v. OpenAI Inc.

Opinion

UNITED STATES DISTRICT COURT SOUTHERN DISTRICT OF NEW YORK ---------------------------------------------------------------x IN RE: : 25-md-3143 (SHS) (OTW)

OPENAI, INC., : COPYRIGHT INFRINGEMENT LITIGATION ORDER :

This Document Relates To: :

23-cv-11195 : 23-cv-08292 ---------------------------------------------------------------x SIDNEY H. STEIN, U.S. District Judge. As part of ongoing discovery, Microsoft and the OpenAI defendants1 moved to compel the production of documents by plaintiffs in The New York Times Company v. Microsoft Corporation, et al., No. 23-cv-11195 (the “Times action”) and Authors Guild, et al. v. OpenAI, Inc., et al., No. 23-cv-08292 (the “Authors Guild action”). On November 22, 2024, Magistrate Judge Ona T. Wang issued an order denying OpenAI’s motion to compel the production of documents relating to The New York Times Company’s (“The Times”) creation of, positions on, and use of generative AI generally. (Times, ECF No. 344 (the “November Order”).) On December 2, 2024, Judge Wang issued five additional minute orders denying defendants’ motions to compel the production of documents relating to The Times’s revenue, website traffic, and licensing. (Times, ECF Nos. 351–55 (the “December Orders”).) On December 3, 2024, Judge Wang issued two orders denying defendants’ discovery motions in the Authors Guild action. (Authors Guild, ECF Nos. 289, 290 (the “Authors Guild Orders”).) Defendants have filed timely objections and moved pursuant to Federal Rule of Civil Procedure 72(a) for an order setting aside those Orders and directing plaintiffs to produce the discovery at issue. Because the November Order, December Orders, and Authors Guild Orders are neither clearly erroneous nor contrary to law, the Court denies defendants’ motions.

1 The OpenAI defendants are OpenAI Inc., OpenAI LP, OpenAI GP, LLC, OpenAI, LLC, OpenAI OpCo LLC, OpenAI Global, LLC, OAI Corporation, LLC, and OpenAI Holdings, LLC (collectively “OpenAI”). I. BACKGROUND A. The Times Action The Court assumes familiarity with the facts underlying the Times action, which are detailed in this Court’s Opinion granting in part and denying in part defendants’ motions to dismiss the complaints in the Newspaper Actions.2 (Times, ECF No. 514.) The November Order denied OpenAI’s motion to compel the following discovery: “(1) the Times’s use of nonparties’ [generative AI] tools; (2) the Times’s creation and use of its own [generative AI] products; and (3) the Times’s position regarding [generative AI] (e.g., positions expressed outside of litigation, knowledge about the training of third- party [generative AI] tools using the Time[s]’s works).” (Times, ECF No. 344 at 1.) That same reasoning was subsequently applied in the December Orders to deny defendants’ additional motions to compel discovery involving (1) alleged decreases in The Times’s revenue and the reasons for such decreases; (2) alleged decreases in subscriptions and traffic to The Times’s platforms and the reasons for those decreases; (3) The Times’s licensing efforts and whether there exists a cognizable licensing market for training large-language models (“LLMs”); and (4) The Times’s use of nonparty generative AI. (Times, ECF Nos. 351–55.) Defendants contend that these orders are clearly erroneous and contrary to law because they (1) are contrary to binding fair use precedent; (2) shield documents that are directly relevant to defendants’ fair use and substantial noninfringing use defenses; and (3) ignore the relevance of the requested documents to damages and other remedies. B. The Authors Guild Action The Court also assumes the parties’ familiarity with the facts underlying the Authors Guild action.3 On December 3, 2024, Judge Wang issued two orders in the Authors Guild action. The first order denied Microsoft’s request to compel discovery into the Authors Guild plaintiffs’ use of ChatGPT for any purpose, as well as certain discovery into the valuation of and markets for the Authors Guild plaintiffs’ works. (Authors Guild, ECF No. 289.) The second order denied OpenAI’s request for a pre-motion conference, as well as for the Authors Guild plaintiffs to respond to certain interrogatories regarding harm and

2 For the purposes of this Order, the Newspaper Actions are The New York Times Company v. Microsoft Corporation, et al., No. 23-cv-11195, Daily News LP, et al. v. Microsoft Corporation et al., No. 24-cv-03285, and The Center for Investigative Reporting, Inc. v. OpenAI, Inc. et al., No. 24-cv-04872, which have been consolidated for pre-trial purposes. 3 The Authors Guild action includes Alter, et al. v. OpenAI, Inc., et al., No. 23-cv-10211, which was consolidated with Authors Guild, et al. v. OpenAI, Inc., et al., No. 23-cv-08292 for pre-trial purposes. (See Authors Guild, First Consolidated Class Action Compl., ECF No. 69.) damages. (Authors Guild, ECF No. 290.) Both Authors Guild Orders denied defendants’ requests “for the reasons identified in ECF 344 in the related Newspaper Cases.” Defendants object to the Authors Guild Orders and contend that they are (1) contrary to law, for the reasons defendants articulated in their oppositions to the November Order; and (2) clearly erroneous in their reliance on the November Order, which did not address the specific documents and interrogatories defendants seek from the Authors Guild plaintiffs. * * * For the reasons that follow, the Court finds that the November Order, December Orders, and Authors Guild Orders are neither clearly erroneous nor contrary to law. II. LEGAL STANDARDS A. Discovery & Fair Use Federal Rule of Civil Procedure 26(b)(1) permits “discovery regarding any nonprivileged matter that is relevant to any party’s claim or defense and proportional to the needs of the case.” The party moving to compel discovery “bears the initial burden of demonstrating relevance and proportionality.” Winfield v. City of New York, No. 15-cv- 05236, 2018 WL 840085, at *3 (S.D.N.Y. Feb. 12, 2018); see also Johnson v. J. Walter Thompson U.S.A., LLC, No. 16-cv-1805, 2017 WL 3055098, at *2 (S.D.N.Y. July 18, 2017). Section 107 of the Copyright Act sets forth four factors to consider in determining whether a defendant’s use of a copyrighted work is “fair” and thus “not an infringement of copyright.” 17 U.S.C. § 107. Those factors include: (1) the purpose and character of the use, including whether such use is of a commercial nature or is for nonprofit educational purposes; (2) the nature of the copyrighted work; (3) the amount and substantiality of the portion used in relation to the copyrighted work as a whole; and (4) the effect of the use upon the potential market for or value of the copyrighted work. Id. Section 107 “set[s] forth general principles, the application of which requires judicial balancing, depending upon relevant circumstances.” Google LLC v. Oracle Am., Inc., 593 U.S. 1, 19 (2021). Defendants ground their objections to the November Order and December Orders in the first and fourth statutory fair use factors. “In assessing the first factor, courts consider two sub-factors: (i) the extent to which the secondary use is transformative and (ii) whether the secondary use is commercial in nature.” Hachette Book Grp., Inc. v. Internet Archive, 115 F.4th 163, 179 (2d Cir. 2024). In assessing the fourth factor, courts “focus[] on whether the copy brings to the marketplace a competing substitute for the original, or its derivative, so as to deprive the rights holder of significant revenues because of the likelihood that potential purchasers may opt to acquire the copy in preference to the original.” Authors Guild v. Google, Inc., 804 F.3d 202, 223 (2d Cir. 2015) (“Google Books”). The fair use analysis requires courts to consider not only the “potential loss of revenue,” but also “the source of the loss,” as well as “the public benefits the copying will likely produce.” Oracle, 593 U.S. at 35. B. Federal Rule of Civil Procedure 72(a) Federal Rule of Civil Procedure 72(a) provides that a district judge must “modify or set aside any part” of a magistrate judge’s non-dispositive pretrial order “that is clearly erroneous or is contrary to law.” Fed. R. Civ. P. 72(a). “A decision is ‘clearly erroneous’ when the reviewing [c]ourt is left with the definite and firm conviction that a mistake has been committed.” Smith v. City of New York, 754 F. Supp. 3d 581, 583 (S.D.N.Y. 2024) (citation omitted); see also Fashion Exch. LLC v. Hybrid Promotions, LLC, No. 14-cv-1254 (SHS), 2021 WL 1172265, at *1 (S.D.N.Y. Mar. 29, 2021). “Similarly, a finding is contrary to law ‘when it fails to apply or misapplies relevant statutes, case law, or rules of procedure.’” R.F.M.A.S., Inc. v. So, 748 F. Supp. 2d 244, 248 (S.D.N.Y. 2010) (citation omitted). “[P]ursuant to these standards, a magistrate judge’s determinations on discovery matters are entitled to substantial deference.” Id. III. THE NOVEMBER ORDER (TIMES, ECF NO. 344) The November Order denies OpenAI’s motions to compel discovery into The Times’s creation and use of generative AI, as well as its position on generative AI generally. OpenAI claims the Order is (1) contrary to binding fair use precedent and (2) clearly erroneous, because the evidence at issue is relevant to both the first and fourth statutory fair use factors as well as the determination of damages and other remedies. A. The November Order Accords with Fair Use Precedent. OpenAI challenges the statement in the November Order that “[t]he [fair use] factors do not require a court to examine,” inter alia, “statements or comments a copyright holder may have made about a defendant’s general industry, [or] whether the copyright holder has used tools in the defendant’s general industry.” (ECF No. 344 at 3.) According to OpenAI, this statement is contrary to Google v. Oracle, 593 U.S. 1 (2021) and American Geophysical Union v. Texaco, 60 F.3d 913 (2d Cir. 1994). The Court disagrees. In Oracle, the U.S. Supreme Court analyzed the market effects of Google’s mobile phone (Android) on the market for Oracle’s software (Java SE), when Google had copied part of that software without authorization and incorporated it into its Android product. Oracle, 593 U.S. at 35–40. Importantly, Oracle had attempted to enter the mobile phone market. The Supreme Court therefore found it relevant that Oracle was unsuccessful in its attempts to do so; a jury could conclude from that evidence “that Android did not harm the actual or potential markets for Java SE” and that Oracle “would not have been able to enter those markets successfully” even if Google had not copied part of Oracle’s software. Id. at 36. But Oracle does not stand for the proposition that evidence of the plaintiff’s activities is always relevant. See id. (“We do not say that these questions are always relevant to the application of fair use . . . .”). Indeed, unlike Oracle’s attempt to develop a mobile phone, The Times has never attempted to develop and train its own generative AI LLM.4 It was therefore not contrary to Oracle for Judge Wang to conclude that the fair use factors “do not require a court to examine statements or comments a copyright holder may have made about a defendant’s general industry” or “whether the copyright holder has used tools in the defendant’s general industry.” (ECF No. 344 at 3 (emphasis added).) See also Hachette Book Grp., 115 F.4th at 179 (“[F]air use is a ‘flexible’ concept, whose application varies depending on the context.” (citing Oracle)). The November Order also does not conflict with Texaco. OpenAI argues otherwise, contending that the November Order denies defendants the opportunity to determine whether “potential licensing revenues were impermissibly impaired” under the fourth fair use factor. See Texaco, 60 F.3d at 930 n.17. In Texaco, the defendant copied individual articles from the plaintiffs’ scientific and technical journals. The U.S. Court of Appeals for the Second Circuit explained that, because the copying at issue was of individual articles, the most relevant evidence of a “potential market” for the works at issue—from which the defendant’s copying would impermissibly deprive the plaintiffs of revenue— was the potential market “for the sale of individual copies of articles,” not “the traditional market for journal subscriptions.” Id. at 927 (emphasis added). The November Order is consistent with that reasoning, which explains that the most relevant evidence of harm under the fourth fair use factor involves the effect of the defendant’s copying on the potential market for the thing copied. In Texaco, that meant the relevant potential market was the market for licensing the individual articles, not the journals comprising those articles. Here, the potential market is the market for The Times’ works, not the market

4 Defendants contest this and urge the Court to disregard the affidavit of Karen A. Chesley, which was not before Judge Wang when she issued the November and December Orders, and which states in relevant part that “The Times does not train its own large-language models (“LLMs”) for generative AI purposes.” (Aff. of Karen A. Chesley ¶ 3, Times, ECF No. 429, Ex. 1.) Even disregarding this affidavit, it was not clear error for Judge Wang to conclude that The Times has never attempted to train its own generative AI LLM, notwithstanding that it may have attempted to develop its own generative AI tools utilizing third-party LLMs. for generative AI generally. Denying discovery into The Times’s use of generative AI generally, therefore, is not inconsistent with Texaco. B. The November Order Is Not Clearly Erroneous. 1. The discovery requests are not relevant to assessing the existence of a potential market for licensing training data. As set forth above, the fourth statutory fair use factor considers “the effect of the use upon the potential market for or value of the copyrighted work.” 17 U.S.C. § 107(4). OpenAI submits that the relevant “potential market” in this case is “a functional market for licensing training data,” and concludes that the November Order erred in preventing discovery into evidence of The Times’s “creation or deployment” of an LLM “trained on unlicensed data,” which bears directly on whether there exists a “ready market or means to pay for the use.” (Times, ECF No. 363 at 13 (internal citations omitted).) The Court, again, disagrees. First, The Times has not “created” any of its own LLMs.5 Second, The Times has already agreed to produce documents relevant to the evaluation of whether a “viable market” exists for licensing the rights to its works for the purpose of developing generative AI. See Texaco, 60 F.3d at 930. Those documents include: “(i) licensing agreements related to the asserted works, regardless of whether those licenses relate to generative AI; (ii) documents sufficient to show all attempts to license the asserted works for training or use in generative AI; and (iii) communications related to executed and contemplated agreements to license Times works to train and operate generative AI models.” (Times, ECF No. 429 at 10.) The November Order was not clearly erroneous in its conclusion that the additional documents defendants sought—discovery into The Times’s internal use of, and comments on, nonparty generative AI—were not relevant to the inquiry of whether a ready market exists to license The Times’s works for LLM training. See Texaco, 60 F.3d at 931. 2. The discovery requests are not relevant to assessing transformative use. The first statutory fair use factor asks whether the copier’s use of an original work merely supplants the original, or whether it “adds something new, with a further purpose or different character” such that the use is transformative. Andy Warhol Found. for the Visual Arts, Inc. v. Goldsmith, 598 U.S. 508, 528 (2023) (quoting Campbell v. Acuff- Rose Music, Inc., 510 U.S. 569, 579 (1994)). OpenAI intends to show that its generative AI products are transformative and submits that “[e]vidence that the Times itself has used

5 The Times explains that “[t]o the extent that The Times develops generative AI tools, it has done so using LLMs created by others, and pays the fees those developers—including OpenAI—charge to do so.” (Times, ECF No. 429 at 1.) It was not clearly erroneous for Judge Wang to conclude that The Times’s use of LLMs that The Times did not itself train is not relevant to the fair use inquiry. the same technology to improve and further its journalistic enterprise would directly support OpenAI’s position.” (Times, ECF No. 363 at 16–17.) Judge Wang rejected that argument, concluding that “[w]hether nonparties’ [generative] AI tools confer benefits on the journalism industry is not relevant to a determination of whether Defendant’s acts—i.e., the alleged copying involving Defendant’s [generative] AI tools—constitute fair use.” (Times, ECF No. 344 at 5.) That conclusion is not clearly erroneous. Indeed, the first fair use factor is concerned with “the copier’s use of an original work,” Warhol, 598 U.S. at 528 (emphasis added), and “the extent to which the secondary use is transformative.” Hachette Book Grp., 115 F.4th at 179 (emphasis added). This litigation is focused on defendants’ allegedly infringing use of plaintiffs’ works, and defendants’ fair use defense is likewise grounded in that use. It was not clearly erroneous for the November Order to determine that The Times’s use of nonparty technology does not bear on whether defendants’ use of The Times’s works is transformative. 3. The discovery requests are not relevant to assessing the public benefit of defendants’ products. It was also not clearly erroneous for the November Order to conclude that the requested discovery is not relevant to determining the public benefit conferred by defendants’ copying of The Times’s works, another consideration contemplated by the first fair use factor. See Warhol, 598 U.S. at 530; Oracle, 593 U.S. at 35–36; Texaco, 60 F.3d at 922; see also MCA, Inc. v. Wilson, 677 F.2d 180, 183 (2d Cir. 1981). Again, the use at issue is the use by defendants—not other generative AI companies—of The Times’s works. It is therefore not clearly erroneous to determine that additional discovery into The Times’s use of other generative AI technology is not relevant to the public benefit inquiry.6 4. The discovery requests are not relevant to assessing damages and other remedies. Last, OpenAI’s requests are not relevant to establishing the “objective reasonableness” of OpenAI’s copying and its reliance on the fair use defense for the purpose of determining damages and the availability of other remedies. (See Times, ECF No. 363 at 20 (citing Kirtsaeng v. John Wiley & Sons, Inc., 579 U.S. 197, 207–08 (2016)).) As OpenAI itself explained, whether certain remedies are warranted in the copyright infringement context depends on “the degree to which the defendant reasonably believed his conduct was non-infringing.” (Times, ECF No. 363 at 20–21 (quoting TD Bank N.A. v. Hill, 928 F.3d 259, 283 (3d Cir. 2019)).) As already noted, The Times has agreed to produce documents relating to relevant internal analyses of OpenAI’s use of The Times’s works and whether such use was fair. (Times, ECF No. 429 at 15–16.) It was not clear error for the November Order to conclude that The Times’s internal use of,

6 The Times is producing documents relating to its use of, or comments on, OpenAI’s LLMs. (See Times, ECF No. 429 at 15.) and opinions on, nonparty generative AI is not relevant to the degree to which defendants reasonably believed their conduct—allegedly copying The Times’s works to train their LLMs—was legally non-infringing. IV. THE DECEMBER ORDERS (TIMES, ECF NOS. 351–55) Defendants also challenge Judge Wang’s denial of five motions to compel in the Times action on December 2, 2024. (See Times, ECF Nos. 351–55.) The December Orders each denied defendants’ motions to compel “for the reasons set forth in ECF 344.” (Id.) The discovery requests at issue in the December Orders were Microsoft’s requests for documents relating to The Times’s use of nonparty generative AI, and defendants’ requests for three categories of documents: (1) “[D]ocuments and communications related to decreases in the Times’ revenue, including the reasons for such decreases” (“Revenue Requests”); (2) “[D]ocuments and communications related to decreases in subscriptions and in traffic to the Times’ platforms, including the reasons for such decreases” (“Traffic Requests”); and (3) “[D]ocuments and communications related to the Times’ licensing efforts, including regarding the existence (or non- existence) of a market for LLM training” (“Licensing Requests”). (Times, ECF No. 372 at 5–6.) As set forth below, the Court finds that Judge Wang’s denials of defendants’ motions to compel were neither clearly erroneous nor contrary to law. A. The December Orders Did Not Clearly Err in Denying Defendants’ Broad Revenue, Licensing, and Traffic Requests. Defendants contend that the Revenue Requests are critical to the issue of causation, which is relevant for determining both market harm under the fair use defense and damages. See Harper & Row Publishers, Inc. v. Nation Enters., 471 U.S. 539, 567 (1985); Fitzgerald Publ’g Co. v. Baylor Publ’g Co., 807 F.2d 1110, 1118 (2d Cir. 1986). OpenAI argues that “[i]f the Times’s alleged losses derive from factors having nothing to do with OpenAI, then OpenAI is entitled to know.” (Times, ECF No. 372 at 13); Microsoft contends that “[b]y asserting that Microsoft Copilot is causing it harm, The Times is hypothesizing that but-for the effect of Copilot, The Times would have maintained or improved its position.” (Times, ECF No. 446 at 3.) These arguments are misguided. The Times is not alleging that defendants’ conduct is the only cause of harm to their works, or that they would suffer no harm but for defendants’ conduct. Rather, they contend that defendants are causing them a particular harm that would not occur but for defendants’ conduct. To measure that particular harm, relevant evidence examines whether defendants’ use, as opposed to some other factor, caused the particular harm. To that end, The Times has agreed to produce “analyses and reports regarding whether decreases in traffic to Times websites, if any, are attributable to GPT services versus other market forces and factors.” (Times, ECF No. 439 at 9.) With respect to licensing, The Times has agreed to produce “(i) licensing agreements related to the asserted works, regardless of whether those licenses relate to generative AI; (ii) documents sufficient to show all attempts to license the asserted works for training or use in generative AI; and (iii) communications related to executed and contemplated agreements to license Times works to train and operate generative AI models.” (Times, ECF No. 429 at 10 (citing ECF No. 310-1 at 45–46).) And with respect to subscriptions, The Times has agreed to produce surveys that “allow[] users to select, as one of the reasons for the cancellation, that ‘AI makes my subscription unnecessary’ or that otherwise takes Generative AI into consideration.” (Times, ECF No. 439 at 12.) As the November Order explained, the fourth fair use factor evaluates “the effect of the use upon the potential market for or value of the copyrighted work.” 17 U.S.C. § 107(4). That is, the fourth factor focuses on “whether the copy brings to the marketplace a competing substitute for the original, or its derivative, so as to deprive the rights holder of significant revenues because of the likelihood that potential purchasers may opt to acquire the copy in preference to the original.” Google Books, 804 F.3d at 223 (emphasis added). The December Orders were not clearly erroneous in denying defendants’ motions to compel discovery into any change in The Times’s revenue whatsoever, which would fail to shed light on whether or not those metrics were impacted by defendants’ conduct. Similarly, the December Orders did not clearly err in denying Microsoft’s motion to compel the production of all surveys regarding subscription cancellations. Subscription cancellation surveys are relevant insofar as they illumine whether subscribers canceled their subscriptions because of generative AI or for some other reason. But The Times is already producing all surveys that allow the user to give an AI-related reason for his or her subscription cancellation, regardless of whether the user actually selected the AI- related option as the basis for their decision to cancel their subscription. The additional surveys defendants request, which do not include any AI-related answer choice, would be of little use in determining the impact of AI on subscriptions. In sum, it was not clearly erroneous for the December Orders to exclude this discovery. B. The December Orders Do Not Conflict with Oracle. The denial of broad discovery into every source of revenue loss to The Times also does not conflict with Oracle. The Supreme Court in Oracle analyzed Google’s fair use defense by considering the harm, if any, (1) caused by Google, and (2) to Oracle’s attempt to enter the mobile phone market. Oracle, 593 U.S. at 36. In doing so, the Supreme Court considered Oracle’s potential loss in revenue from the mobile phone market due to Google’s activity in that market; it did not consider every potential source of revenue loss to Oracle. Indeed, the Supreme Court wrote that when Oracle’s former CEO was asked “whether [Oracle’s] failure to build a smartphone was attributable to Google's development of Android, he answered that it was not.” Id. And the Court further stated that even before Google copied Oracle’s software, Oracle’s executives “projected declining revenue for mobile phones.” Id. Again, The Times does not allege that defendants are the cause of every subscription loss, traffic decline, or decrease in revenue. The Times’s allegations are focused on the alleged harm caused by defendants’ specific unauthorized use of The Times’s works. It was therefore not contrary to Oracle nor clear error for the December Orders to conclude that discovery into any decrease in subscriptions, traffic, or revenue, divorced from defendants’ conduct and products, was not relevant to the parties’ claims or defenses. C. The December Orders Do Not Erroneously Deprive Microsoft of Evidence of Substantial Noninfringing Uses. Microsoft contends that the December Orders improperly denied it evidence relevant to whether defendants’ products are capable of substantial noninfringing uses, a defense Microsoft intends to bring in accordance with the Supreme Court’s reasoning in Sony Corp. of America v. Universal City Studios, 464 U.S. 417 (1984). Microsoft urges that The Times’s internal use of generative AI is directly relevant to this inquiry. But the question of substantial noninfringing uses asks whether “the product is widely used for legitimate, unobjectionable purposes.” Sony, 464 U.S. at 442; see also EMI Christian Music Grp., Inc. v. MP3Tunes, LLC, 844 F.3d 79, 100 (2d Cir. 2016) (“[W]hen a product is capable of ‘substantial noninfringing uses,’ its manufacturer cannot be liable simply for knowing that the product can be used in a way that would constitute infringement.”). The “product” here is defendants’ generative AI; it is not other generative AI products created by nonparties. The Times has already agreed to produce documents relating to its use of defendants’ products. (Times, ECF No. 439 at 18.) Judge Wang’s denial of forays into The Times’s use of nonparty generative AI does not erroneously deprive Microsoft of relevant evidence pertaining to the existence vel non of substantial noninfringing uses of defendants’ products.7 V. THE AUTHORS GUILD ORDERS (AUTHORS GUILD, ECF NOS. 289, 290) A. The Authors Guild Order Pertaining to OpenAI In the Authors Guild Order pertaining to OpenAI (Authors Guild, ECF No. 290), Judge Wang denied OpenAI’s request for the identification of documents related to its fair use defense, stating in full: The discovery dispute relates to Defendants’ fair use defense and is denied for the reasons identified in ECF 344 in the related Newspaper Cases (23-cv-11195). Additionally, the request is denied as premature: this dispute relates to interrogatories requesting that Plaintiffs identify certain documents that may relate to Defendants’ fair use defense. The parties should be guided by the scope of agreed-upon document production that the Plaintiffs in these cases and the Newspaper Cases have been and will be producing. If Defendants still seek identification of documents produced, they may renew this motion after reviewing the produced documents and meeting and conferring with their adversaries, but would need to explain why and how they still need identification of documents after they have reviewed the documents themselves. (Id.) That Order is neither clearly erroneous nor contrary to law. First, the Order is not clearly erroneous in its reliance on the reasoning in the November Order. The November Order related to OpenAI’s fair use defense and the question of market harm under the fourth statutory fair use factor; OpenAI’s interrogatories in Authors Guild also went to the question of market harm under that factor. Second, the Order specifically grants OpenAI the opportunity to renew its request in the future. The Court denies OpenAI’s request to set aside the Authors Guild Order pertaining to OpenAI.

7 Microsoft also argues that the December Orders are clearly erroneous because The Times’s use of nonparty generative AI is relevant to (1) the market harm inquiry, and (2) transformative use. However, as the Court has reiterated throughout, fair use factors one and four contemplate the defendant’s use of the plaintiff’s work, and whether or not that use is transformative and/or a threat to the market for the plaintiff’s work. See 17. U.S.C. §§ 107(1), (4). Accordingly, the December Orders were not clearly erroneous in denying discovery into The Times’s (not defendants’) use of nonparty (not defendants’) products. B. The Authors Guild Order Pertaining to Microsoft Finally, Judge Wang did not clearly err in denying Microsoft's request for documents relating to (1) the Authors Guild plaintiffs’ use of ChatGPT, and (2) the valuation of and markets for the Authors Guild plaintiffs’ works. (See Authors Guild, ECF No, 289.) Regarding the first category of requests—ChatGPT use—the Authors Guild plaintiffs have already agreed to produce “all non-privileged documents in our client’s emails that contain the term OpenAI or ChatGPT,” and “all ChatGPT outputs cited in the Complaint.” (Authors Guild, ECF No. 315 at 2.) The Order’s reliance on the November Order—which involved similar discovery —to deny Microsoft's request for documents relating to all ChatGPT use by the Authors Guild plaintiffs-including uses entirely unrelated to the litigation or works at issue —is not clearly erroneous. Nor is it contrary to the case law on substantial noninfringing uses, which discusses the noninfringing uses of products whose manufacturers have been accused of contributory copyright infringement. See Sony, 464 U.S. at 442. The substantial noninfringing use analysis applies when the contributory copyright infringement claim involves “the sale of an article of commerce that is used by the purchaser to infringe.” See id. at 440, Here, however, the Authors Guild plaintiffs’ contributory copyright infringement claim against Microsoft alleges that Microsoft contributed to OpenAl’s primary infringement, not the primary infringement of third-party end-users. (See Authors Guild, First Consolidated Class Action Compl., ECF No. 69 {J 423-27.) Regarding the second category of requests—the valuation of and markets for the Authors Guild plaintiffs’ works— Microsoft's requests to Authors Guild were not ripe at the time of the Authors Guild Order pertaining to Microsoft (see Authors Guild, ECF No. 315 at 3), and its requests to the individual author plaintiffs were resolved by agreement of the parties (see Authors Guild, ECF No. 315, Ex. 1 at 1). Accordingly, the Authors Guild Order pertaining to Microsoft was neither clearly erroneous nor contrary to law. VE. CONCLUSION For the foregoing reasons, the Court denies defendants’ motions to set aside the November Order, December Orders, and Authors Guild Orders. Dated: New York, New York May 30, 2025 SO ORDERED: □

Sidney H/Stein, U.S.DJ. 12