Tremblay v. OpenAI, Inc.Tremblay v. OpenAI, Inc.
- Reporters:
- ,
- Before:
- Araceli Martínez-Olguín
ORDER GRANTING IN PART AND DENYING IN PART THE MOTIONS TO DISMISS
This is a putative class action copyright case. OpenAI, Inc.‘s motions to dismiss were heard before this Court on December 7, 2023. Having read the papers filed by the parties and carefully considered their arguments therein and those made at the hearing, as well as the relevant legal authority, the Court hereby GRANTS in part and DENIES in part the motions to dismiss for the following reasons.
I. BACKGROUND
Before the Court are two nearly identical putative class complaints in Tremblay et al v. OpenAI, Inc. et al, 23-cv-3223 and Silverman et al v. OpenAI, Inc. et al, 23-cv-3416. Plaintiffs are authors of books who allege that their books were used to train OpenAI language models that operate the artificial intelligence (“AI“) software ChatGPT.1 Silverman Compl. ¶¶ 1-4; Tremblay Compl. ¶¶ 1-4. Plaintiffs Paul Tremblay, Sarah Silverman, Christopher Golden, and Richard Kadrey (collectively, “Plaintiffs“) hold registered copyrights in their books. Tremblay Compl. ¶¶ 10, 12, Ex. A (The Cabin at the End of the World (Tremblay))2; Silverman Compl. ¶¶ 10-13,
Defendant OpenAI3 creates and sells certain AI software known as large language models (or “LLM“). Tremblay Compl. ¶ 23. These language models are “trained” by inputting large amounts of texts known as the “training dataset.” Id. The language models copy text from the training dataset and extract “expressive information.” Id. ChatGPT is an OpenAI language model that allows paying users to enter text prompts to which ChatGPT will respond and “simulate human reasoning,” including answering questions or summarizing books. Id. ¶¶ 22, 36-38. ChatGPT generates its output based on “patterns and connections” from the training data. Id. ¶ 39.
OpenAI copied Plaintiffs’ copyrighted books and used them in its training dataset. Id. ¶ 24. When prompted to summarize books written by each of the Plaintiffs, ChatGPT generated accurate summaries of the books’ content and themes. Id. ¶ 41 (citing Ex. B); Silverman Compl. ¶ 42 (citing Ex. B).
Plaintiffs seek to represent a class of all people in the U.S. who own a copyright in any work that was used as training data for OpenAI language models during the class period. Tremblay Compl. ¶ 42; Silverman Compl. ¶ 43. Plaintiffs assert six causes of action against various OpenAI entities: (1) direct copyright infringement (Count I); (2) vicarious infringement (Count II); (3) violation of Section 1202(b) of the Digital Millennium Copyright Act (“DMCA“) (Count III); (4) unfair competition under
OpenAI filed the instant motions to dismiss on August 28, 2023, seeking dismissal of Counts II through VI. ECF 33 (“Motion“).4
II. LEGAL STANDARD
In reviewing the plausibility of a complaint, courts “accept factual allegations in the complaint as true and construe the pleadings in the light most favorable to the nonmoving party.” Manzarek, 519 F.3d at 1031. Nonetheless, courts do not “accept as true allegations that are merely conclusory, unwarranted deductions of fact, or unreasonable
III. ANALYSIS
Defendants seek dismissal of all causes of action except for the claim for direct copyright infringement. Motion at 18. Defendants seek dismissal of Count II for vicarious copyright infringement; Count III for violation of Section 1202(b) of the Digital Millennium Copyright Act (“DMCA); Count IV for Unfair Competition under
A. Vicarious Copyright Infringement (Count II)
The Copyright Act grants the copyright holder exclusive rights to (1) “reproduce the copyrighted work in copies,” (2) “prepare derivate works,” and (3) “distribute copies . . . of the copyrighted work to the public.”
Copyright infringement requires that a plaintiff show (1) “he owns as valid copyright” and (2) the defendant “copied aspects of his work.” Corbello, 974 F.3d at 973. The second prong “contains two separate components: ‘copying’ and ‘unlawful appropriation.‘” Id. at 974. “Copying can be demonstrated either through direct evidence or by showing that the defendant had access to the plaintiff‘s work and that the two works share similarities probative of copying, while the hallmark of ‘unlawful appropriation’ is that the works share substantial similarities.” Id. (citations and internal quotations omitted); see Skidmore as Tr. for Randy Craig Wolfe Tr. v. Led Zeppelin, 952 F.3d 1051, 1064 (9th Cir. 2020) (“the hallmark of ‘unlawful appropriation’ is that the works share substantial similarities“) (emphasis in original).
A claim of vicarious infringement requires a threshold showing of direct infringement, Perfect 10, Inc. v. Amazon.com, Inc., 508 F.3d 1146, 1169 (9th Cir. 2007) (”Amazon.com“); see A&M Recs., Inc. v. Napster, Inc., 239 F.3d 1004, 1013 n.2 (9th Cir. 2001) aff‘d, 284 F.3d 1091 (2002) (“Secondary liability for copyright infringement does not exist in the absence of direct infringement by a third party“). A plaintiff must then show that “the defendant has (1) the right and ability to supervise the infringing conduct and (2) a direct financial interest in the infringing activity.” Perfect 10, Inc. v. Giganews, Inc., 847 F.3d 657, 673 (9th Cir. 2017) (”Giganews“) (citation omitted).
Defendants argue that the vicarious infringement claim fails for three reasons: (1) Plaintiffs have not alleged that direct infringement occurred; (2) Plaintiffs have not alleged that Defendants had the “right and ability to supervise“; and (3) Plaintiffs have not alleged “direct financial interest.” Motion at 19-21. The Court first considers whether Plaintiffs have adequately alleged direct infringement and, ultimately, does not reach Defendants’ latter two arguments.
Plaintiffs suggest that they do not need to allege a “substantial similarity” because they have evidence of “direct copying.” ECF 48 (“Response“) at 15. They argue that because Defendants directly copied the copyrighted books to
Plaintiffs’ allegation that “every output of the OpenAI Language Models is an infringing derivative work” is insufficient. Tremblay Compl. ¶ 59; Silverman Compl. ¶ 60. Plaintiffs fail to explain what the outputs entail or allege that any particular output is substantially similar – or similar at all – to their books. Accordingly, the Court dismisses the vicarious copyright infringement claim with leave to amend.
B. Section 1202(b) of the DMCA (Count III)
In addition to protecting against vicarious and direct infringement, “[c]opyright law restricts the removal or alteration of copyright management information (‘CMI‘) – information such as the title, the author, the copyright owner, the terms and conditions for use of the work, and other identifying information set forth in a copyright notice or conveyed in connection with the work.” Stevens v. Corelogic, Inc., 899 F.3d 666, 671 (9th Cir. 2018).
OpenAI argues that the Section 1202(b)(1) claim fails because Plaintiffs do not plausibly allege that OpenAI intentionally removed CMI during the training process or intended to conceal or induce infringement, and the Section 1202(b)(3) claim fails because Plaintiffs do not allege that Defendant distributed the copyrighted works or copies. Motion at 21-22. The Court begins by analyzing the 1202(b)(1) claim.
1. Section 1202(b)(1) – Intentionally Remove or Alter CMI
Plaintiffs allege that “[b]y design,” Defendants remove CMI from the copyrighted books used during the training process. Tremblay Compl. ¶ 64; Silverman Compl. ¶ 65. However, Plaintiffs provide no facts supporting this assertion. Indeed, the Complaints include excerpts of ChatGPT outputs that include multiple references to Plaintiffs’ names, suggesting that OpenAI did not remove all references to “the name of the author.” See, e.g., Tremblay Compl. Ex. B at 3 (“Throughout these chapters, Tremblay masterfully maintains the suspense and psychological terror“); Silverman Compl. Ex B. at 1 (“Silverman uses her sharp wit to lend a comedic touch“). Moreover, there are no facts to support the assertion that “[b]y design, the training process does not preserve any CMI.” Tremblay Compl. ¶ 64. In Doe 1 v. GitHub, Inc., No. 22-CV-06823-JST, 2023 WL 3449131, at *11 (N.D. Cal. May 11, 2023), plaintiffs alleged where the CMI typically appeared, that defendants were aware that the CMI appeared repeatedly, and that defendants subsequently “trained these programs to ignore or remove CMI and therefore stop producing it.” That was sufficient to support a “reasonable inference that Defendants intentionally designed the programs to remove CMI . . .” Id. By contrast, Plaintiffs here only make conclusory allegations that “[b]y design, the training process does not preserve any CMI.” Tremblay Compl. ¶ 64.
Even if Plaintiffs provided facts showing Defendants’ knowing removal of CMI from the books during the training process, Plaintiffs have not shown how omitting CMI in the copies used in the training set gave Defendants reasonable grounds to know that ChatGPT‘s output would induce, enable, facilitate, or conceal infringement. See Stevens, 899 F.3d at 673 (finding that allegations that “someone might be able to use [the copyrighted work] undetected . . . simply identifies a general possibility that exists whenever CMI is removed,” and fails to show the necessary mental state). Plaintiffs argue that OpenAI‘s failure to state which internet books it uses to train ChatGPT shows that it knowingly enabled infringement, because ChatGPT users will not know if any output is infringing. Response at 21-22. However, Plaintiffs do not point to any caselaw to suggest that failure to reveal such information has any bearing on whether the alleged removal of CMI in an internal database will knowingly enable infringement. Plaintiffs have failed to state a claim under Section 12(b)(1).
2. Section 1202(b)(3) – Distribute Works or Copies
Plaintiffs also allege that Defendants violated
Under the plain language of the statute, liability requires distributing the original “works” or “copies of [the] works.”
Plaintiffs compare their claim to that in Doe 1, however, the plaintiffs in Doe 1 alleged that the defendants “distributed copies of [plaintiff‘s licensed] code knowing that CMI had been removed or altered.” Doe 1, 2023 WL 3449131, at *11. The Doe 1 plaintiffs alleged that defendants knew that the programs “reproduced training data,” such as the licensed code, as output. Id. Plaintiffs here have not alleged that ChatGPT reproduces Plaintiffs copyrighted works without CMI.
Accordingly, the Court dismisses the DMCA claims with leave to amend.
C. UCL (Count IV)
OpenAI also seeks to dismiss Plaintiffs’ UCL claims. “A business act or practice may violate the UCL if it is either ‘unlawful,’ ‘unfair,’ or ‘fraudulent.’ Each of these three adjectives captures ‘a separate and distinct theory of liability.‘” Rubio v. Cap. One Bank, 613 F.3d 1195, 1203 (9th Cir. 2010) (citing Kearns v. Ford Motor Co., 567 F.3d 1120, 1127 (9th Cir. 2009)). “The UCL‘s coverage is sweeping, and its standard for wrongful business conduct intentionally broad.” Moore v. Apple, Inc., 73 F. Supp. 3d 1191, 1204 (N.D. Cal. 2014). Plaintiffs advance UCL claims under each prong, and the Court considers each in turn.
1. Unlawful
Plaintiffs allege that Defendants engaged in “unlawful business practices” by violating the DMCA. Tremblay Compl. ¶¶ 69-70; Silverman Compl. ¶ 70-71. As the Court has dismissed the predicate DMCA claims, the derivate UCL claims cannot remain. See Eidmann v. Walgreen Co., 522 F. Supp. 3d 634, 647 (N.D. Cal. 2021) (“If the ‘plaintiff cannot state a claim under the predicate law . . . [the UCL] claim also fails.‘“) (alterations in original) (quoting Hadley v. Kellogg Sales Co., 243 F. Supp. 3d 1074, 1094 (N.D. Cal. 2017)); see Armstrong-Harris v. Wells Fargo Bank, N.A., No. 21-CV-07637-HSG, 2022 WL 3348426, at *3 (N.D. Cal. Aug. 12, 2022) (citing cases).
Even if Plaintiffs can bring claims under the DMCA, they must show economic injury caused by the unfair business practice. See Davis v. RiverSource Life Ins. Co., 240 F. Supp. 3d 1011, 1017 (N.D. Cal. 2017) (quoting Kwikset Corp. v. Superior Ct., 51 Cal. 4th 310, 322 (2011)). Defendants argue that Plaintiffs have not alleged that they have “lost money or property.” Motion at 29-30; see Kwikset Corp., 51 Cal. 4th at 322-23. Plaintiffs counter that they have lost intellectual property in connection with the DMCA claims because of the “risk of future damage to intellectual property that results the moment a defendant
2. Fraudulent
Plaintiffs also argue that they pleaded UCL violations based on “fraudulent” conduct. Response at 26-27. They point to a paragraph in the complaint that states that “consumers are likely to be deceived” by Defendants’ unlawful practices and that Defendants “deceptively designed ChatGPT to output without any CMI.” Tremblay Compl. ¶ 72. The allegation‘s references to CMI demonstrates that Plaintiffs’ claims rest on a violation of the DMCA, and thus fail as the Court has dismissed the underlying DMCA claim. Supra Sections B, C(1). To the extent that Plaintiffs ground their claim in fraudulent business practices, Plaintiffs fail to indicate where they have pleaded allegations of fraud. Thus, they fail to satisfy the heightened pleading requirements of
3. Unfair
Under the unfair prong of the UCL, Plaintiffs allege that Defendants used Plaintiffs’ copyrighted works to train ChatGPT without Plaintiffs’ authorization. Tremblay Compl. ¶ 71; Silverman Compl. ¶ 72. California courts have defined “unfair” broadly. For example, the California Court of Appeal stated, “California‘s unfair competition law prohibits not only unlawful business practices but also unfair business practices. The statute is intentionally broad to give the court maximum discretion to control whatever new schemes may be contrived, even though they are not yet forbidden by law.” People ex rel. Renne v. Servantes, 86 Cal.App.4th 1081, 1095 (2001). The Court of Appeal explained further, “[o]ne test for determining an ‘unfair’ practice is [whether] the gravity of the harm to the victim outweighs the utility of the defendant‘s conduct.” Id. at 1095.5
Assuming the truth of Plaintiffs’ allegations - that Defendants used Plaintiffs’ copyrighted works to train their language models for commercial profit - the Court concludes that Defendants’ conduct may constitute an unfair practice.6 Therefore, this portion of the UCL claim may proceed.
D. Negligence (Count V)
Negligence claims require that a plaintiff establish (1) duty; (2) breach; (3) causation; and (4) damages. Corales v. Bennett, 567 F.3d 554, 572 (9th Cir. 2009) (citation omitted). There is a “general duty of care on a defendant . . . who has created
Plaintiffs allege that Defendants owed them a duty of care based on the control of Plaintiffs information in their possession and breached their duty by “negligently, carelessly, and recklessly collecting, maintaining, and controlling systems – including ChatGPT – which are trained on Plaintiffs’ [copyrighted] works.” Tremblay Compl. ¶¶ 74-75; Silverman Compl. ¶¶ 75-76. Defendants seek to dismiss the negligence claim, arguing that 1) Plaintiffs have not established that OpenAI owes them a duty and 2) the Complaints only challenge intentional acts. Motion at 30-31.
The Complaints allege that Defendants negligently maintained and controlled information in their possession. Tremblay Compl. ¶¶ 74-75; Silverman Compl. ¶¶ 75-76. Plaintiffs argue without legal support that Defendants owed a duty to safeguard Plaintiffs’ works. Response at 30. Plaintiffs do not identify what duty exists to “maintain[] and control[]” the public information contained in Plaintiffs’ copyrighted books. The negligence claim fails on this basis.
Plaintiffs’ argument that there is a “special relationship” between the parties also fails. See Response at 30. Nowhere in the Complaints do Plaintiffs allege that there is any fiduciary or custodial relationship between the parties. Plaintiffs do not explain how merely possessing their books creates a special relationship, citing only to an inapposite case where defendants were custodians of plaintiffs’ “personal and confidential information.” Witriol v. LexisNexis Grp., No. C05-02392 MJJ, 2006 WL 4725713, at *8 (N.D. Cal. Feb. 10, 2006).7
As Plaintiffs have not alleged that Defendants owed them a legal duty, the Court dismisses this claim with leave to amend.
E. Unjust Enrichment (Count VI)
In the sixth cause of action, Plaintiffs bring an unjust enrichment claim for OpenAI‘s use of the copyrighted books to train ChatGPT. Tremblay Compl. ¶¶ 79-86; Silverman Compl. ¶¶ 80-87. The Ninth Circuit allows an unjust enrichment claim as either an independent cause of action or as a “quasi-contract claim for restitution.” ESG Cap. Partners, LP v. Stratos, 828 F.3d 1023, 1038 (9th Cir. 2016). “To allege unjust enrichment as an independent cause of action, a plaintiff must show that the defendant received and unjustly retained a benefit at the plaintiff‘s expense.” Id. Restitution is not ordinarily available to a plaintiff unless “the benefits were conferred by mistake, fraud, coercion or request; otherwise, though there is enrichment, it is not unjust.” Nibbi Bros., Inc. v. Home Fed. Sav. & Loan Assn., 205 Cal. App. 3d 1415, 1422 (Cal. Ct. App. 1988) (quoting 1 Witkin, Summary
Defendants argue that this claim must be dismissed because Plaintiffs fail to allege what “benefit” they quasi-contractually “conferred” on OpenAI or that Plaintiffs conferred this benefit through “mistake, fraud, or coercion.” Motion at 32 (citing Bittel Tech., Inc. v. Bittel USA, Inc., No. C10-00719 HRL, 2010 WL 3221864, at *5 (N.D. Cal. Aug. 13, 2010) (“Ordinarily, a plaintiff must show that the benefit was conferred on the defendant through mistake, fraud or coercion.“) (citation omitted). Plaintiffs fail to allege that OpenAI “has been unjustly conferred a benefit ‘through mistake, fraud, coercion, or request.‘” See Astiana, 783 F.3d at 762 (citation omitted); LeGrand v. Abbott Lab‘ys, 655 F. Supp. 3d 871, 898 (N.D. Cal. 2023) (same); see, e.g., Russell v. Walmart, Inc., No. 22-CV-02813-JST, 2023 WL 4341460, at *2 (N.D. Cal. July 5, 2023) (“it is not enough that Russell have provided Walmart with a beneficial service; Russell must also allege that Walmart unjustly secured that benefit through qualifying conduct. Absent qualifying mistake, fraud, coercion, or request by Walmart, there is no injustice.“). As Plaintiffs have not alleged that OpenAI unjustly obtained benefits from Plaintiffs’ copyrighted works through fraud, mistake, coercion, or request, this claim fails.8
IV. CONCLUSION
For the foregoing reasons, the Court GRANTS in part and DENIES in part Defendants’ motions to dismiss. Plaintiffs’ amended complaint must be filed by March 13, 2024. No additional parties or claims may be added without leave of Court or stipulation of Defendants. The amended complaint shall consolidate the claims in 23-cv-3223-AMO (Tremblay et al. v. OpenAI, Inc. et al.), 23-cv-04625-AMO (Chabon et al. v. OpenAI, Inc. et al.), 23-cv-03416 (Silverman et al. v. OpenAI, Inc. et al.). The Court will issue a separate order consolidating the three cases. This Order terminates ECF 33 in 23-cv-3223-AMO (Tremblay et al. v. OpenAI, Inc. et al.) and ECF 32 in 23-cv-03416 (Silverman et al. v. OpenAI, Inc. et al.).
IT IS SO ORDERED.
Dated: February 12, 2024
ARACELI MARTÍNEZ-OLGUÍN
United States District Judge