One of the biggest issues that has arisen in the copyright litigation against OpenAI is whether OpenAI waived its attorney-client privilege as to communications in-house OpenAI lawyers made regarding the deletion of Books1 and Books2 datasets in mid-2022 after OpenAI torrented them in 2018 and then used them to train an earlier AI model(s) of OpenAI at least by 2020.
Willful Civil Infringement v. Willful Criminal Infringement
In her 28-page opinion, Judge Wang accepted 2 of the plaintiffs’ arguments for waiver ((1) selective voluntary disclosure by OpenAI and (2) at-issue waiver based on its relevance to OpenAI’s alleged willful civil infringement and the issue of OpenAI’s willfulness). But Judge Wang rejected a 3rd reason for waiver, the crime-fraud exception for the deletion of the datasets.
This is an important but controversial set of complex issues worthy of closer scrutiny.
Some of the issues have come up in other AI copyright litigation. Judges Chhabria and Hixson rejected the assertion of the crime-fraud exception for Meta’s counsel’s communications in Kadrey v. Meta; Judge Hixson ruled that a Meta document created after the torrenting of pirated copies by Meta cannot be in furtherance of an ongoing unlawful scheme. Earlier, Judge Chhabria had rejected a similar attempt by Kadrey to invoke the crime-fraud exception. But Judge Alsup, in Bartz v. Anthropic, questioned whether Anthropic waived its attorney-client privilege as to communications regarding the torrenting of pirated books by Anthropic’s assertion of innocent (civil) infringement. Anthropic settled before Judge Alsup made any ruling on waiver.
Criminal Infringement: Magistrate Judge Wang ruled on both waiver issues in the OpenAI MDL Litigation. Yet her opinion is divided. Judge Wang rejected the plaintiffs’ assertion of the crime-fraud exception. On p. 26, she explained: “The crime-fraud exception strips privilege from communications made during and in furtherance of a suspected crime, not those that take place after. See In re John Doe, Inc., 13 F.3d at 636. For the same reasons outlined in Judge Hixson’s opinion, Class Plaintiffs have failed to sustain their burden under this theory. See Kadrey (N.D. Cal., 23-CV-3417, ECF 627 at 3) …”
Civil Infringement: But Magistrate Judge Wang found that OpenAI waived its attorney-client privilege by both (1) “Disclosing A “Privileged Reason” for the Deletion of the Books1 and Books2 Datasets” (p. 15) and (2) “Putting Its ‘Good Faith’ At Issue,” by OpenAI failing to concede it won’t pursue innocent infringement or “advance factual arguments opposing Class Plaintiffs’ claims of willfulness” in the case (p. 19).
Here’s the Contradiction in the Decision:
As to the second ground for waiver (willful infringement at issue), it’s not apparent from Magistrate Judge Wang’s opinion why activity and communications of OpenAI lawyers that cannot be evidence of willful criminal copyright infringement can nonetheless be evidence of willful civil copyright infringement.
After all, the standard for willfulness for criminal and civil copyright infringement are largely the same and require the defendant’s knowing infringement. See United States v. Moran, at 1050 (“Rather, since Congress used ‘willful’ in the civil damage copyright context to mean that the infringement must take place with the defendant being knowledgeable that his/her conduct constituted copyright infringement, there is no compelling reason to adopt a less stringent requirement in the criminal copyright context.”). See also Lydia Pallas Loren, Digitization, Commodification, Criminalization: The Evolution of Criminal Copyright Infringement and the Importance of the Willfulness Requirement, 77 Wash. U. L. Q. 835, 871-85 (1999) (discussing civil and criminal willfulness).
Here, OpenAI downloaded the datasets in 2018. It then used them to train earlier AI model(s), at least by 2020. It was not until 2022 when the disputed attorney-client communications by OpenAI attorneys occurred regarding the deletion of the Books1 and 2 datasets.
Magistrate Judge Wang ruled that these 2022 communications “are likely probative of willfulness”: “it is not a stretch for Class Plaintiffs to posit that communications regarding the reasons for deleting Books1 and Books2 could be probative of OpenAI’s willfulness. Indeed, the Court’s in camera review of Log Nos. 14, 15, 17, and 18, discussed above, confirm that such communications are likely probative of willfulness.”
Without having access to those communications, it’s impossible to fully assess that conclusion. But the timing seems to cut sharply against the relevance of the 2022 OpenAI attorney communications as to the state of mind or alleged willfulness of OpenAI employees in torrenting in 2018. Unless OpenAI attorneys were also involved in giving advice for the 2018 torrenting, the 2022 communications seem completely irrelevant to the state of mind of the OpenAI employees in 2018.
Some of the key OpenAI attorneys mentioned in Magistrate Judge Wang’s opinion, Jason Kwon (April 2021 joined OpenAI), Che Chang (2021 joined OpenAI), and Michael Trinh (2023 joined OpenAI) were not even working at OpenAI in 2018, the time of the alleged infringing torrenting in 2018. What they said in 2022 is not relevant to establishing the willfulness of conduct by other OpenAI employees in 2018. Their hires also occurred after the 2020 OpenAI paper discussing the training of its model with those Books1 and 2 datasets.
And it seems contradictory to conclude (1) the OpenAI attorney communication in 2022 was not in furtherance of willful criminal infringement because it occurred after the alleged criminal infringement was already completed in the 2018 torrenting activity but (2) the same OpenAI attorney communication 2022 can be used to establish willful civil infringement in the 2018 torrenting activity.
If legal advice by counsel that occurred after the completed torrenting (and AI training) cannot be used to establish willful criminal copyright infringement, that same approach should apply as well to willful civil copyright infringement. In either context, criminal or civil, the act of alleged infringement was completed well before, by several years, the legal advice given.
Contradiction also shown in Magistrate Judge’s rejection of alleged OpenAI’s “misconduct”
The internal contradiction in the Magistrate Judge’s opinion is further shown in her rejection of the plaintiffs’ argument that “the deletion of Books1 and Books2 “while facing ‘substantial legal uncertainty’ regarding the legality of its actions” constitutes misconduct at odds with the basic premises of the adversarial system.” (p. 27).
In rejecting the plaintiffs’ argument, Magistrate Judge ruled:
Magistrate Judge Wang in OpenAI Order rejecting misconduct
Class Plaintiffs fail to allege a sufficient factual basis to find probable cause to believe that that the [OpenAI in-house counsel’s 2022] communications at issue were in furtherance of the types of misconduct that are recognized by Second Circuit precedent and/or that such communications were intended to facilitate or conceal such activities.
Just as with the rejection of plaintiffs’ argument that OpenAI in-house counsel’s 2022 communications regarding deletion of Books 1, 2 fell within the crime-fraud exception, Magistrate Judge Wang’s rejection of plaintiffs’ argument that OpenAI in-house counsel’s 2022 communications undermines the judge’s own finding that the same 2022 communications are relevant to willful civil infringement by other OpenAI employees in 2018 in torrenting from shadow libraries.
As Magistrate Judge Wang concludes, the plaintiffs failed to even establish probable cause that OpenAI in-house counsel’s 2022 communications were in furtherance of OpenAI’s alleged misconduct and/or “and/or that such communications were intended to facilitate or conceal such activities.”
Because OpenAI in-house counsel’s 2022 communications were neither in furtherance of alleged misconduct by OpenA nor intended to facilitate or conceal such activities, those same 2022 communications are not relevant to OpenAI’s alleged willful infringement in 2018 (torrenting) or 2020 (training). That conclusion is especially warranted given that Jason Kwon (April 2021 joined OpenAI), Che Chang (2021 joined OpenAI), and Michael Trinh (2023 joined OpenAI) were not even working at OpenAI in 2018 and 2020, when OpenAI’s alleged willful infringement with Books1 and 2 occurred. The OpenAI in-house counsel’s state of mind in 2022 cannot retroactively or magically alter the state of mind of the OpenAI researchers who engaged in the torrenting in 2018 or the AI training in 2020.
When the defendant’s mental state in committing an offense is at issue, it is axiomatic that the law examines the mental state of the defendant at the time of the offense. Cf. Larkin v. Brown, 41 F.3d 387, 389 (8th Cir. 1994) (holding that defendant’s prior letters about the plaintiff “do not demonstrate [defendant] Brown’s intent at the time of the battery”); United States v. Cooper, 473 F.2d 95, 97 (D.C. Cir. 1972) (for offense of burglary, there is a “requirement of a criminal intent at the time of entry”); Wilson v. People of the State of New York, 39 N.Y. 459, 461 (1868) (larceny requires defendant had intent to steal at the time of taking the property, not afterwards); id. (“Lord COKE lays down the rule as drawn from the year books, that the taking must be felonious, ‘cum animo furandi,‘ and ‘the intent to steal must be when it cometh to his hands or possession; for, if he hath the possession of it unlawfully, though he hath ‘animum furandi’ afterward, and carrieth it away, it is no larceny.’” (3 Coke’s Inst. 107.) Such is the rule established by an unbroken current of decisions in England and this country.”); Lorenzo Dow v. The State, 12 Tex. App. 343, 345 (1882) (“The intent is the gist of the offense, and such intent must exist at the time of the taking; for if the intent did not exist at the time of taking no subsequent felonious intention will render the previous taking felonious.”) (emphasis added).
This requirement of the defendant’s mental state also applies to questions of willful infringement of intellectual property. As the Supreme Court explained in Halo Electronics, Inc. v. Pusl Electronics, Inc., a patent case:
Id. at 105-06. See also Kewazinga Corp. v. Microsoft Corp., 558 F. Supp. 3d 90, 119 (S.D.N.Y. 2021) (“[C]ulpability is generally measured against the knowledge of the actor at the time of the challenged conduct. Knowledge is defined for [the purpose of willful [patent] infringement] as knowledge of the allegedly infringed patent and its claims.”) (cleaned up) (emphasis added); Therabody, Inc. v. Tzumi Electronics LLC, 2022 WL 17826642, at *3 (S.D.N.Y. Dec. 19, 2022) (“After Halo Electronics, ‘[t]o establish willfulness, the patentee must show the accused infringer had a specific intent to infringe at the time of the challenged conduct.’”) (emphasis added); Fitzgerald Pub. Co., Inc. v. Baylor Pub. Co., Inc., 807 F.2d 1110, 1115 (2d Cir. 1986) (“a defendant’s knowledge that its actions constitute an infringement establishes that the defendant acted willfully within the meaning of § 504(c)(2) for purposes of enhancing statutory damages.”); Curet-Velasquez v. Acemla De Puerto Rico, 2010 WL 1150913, at *12 (“As the court already found, the copyright infringements were not ‘willful’ under the statute because, at the time of the infringements, the defendants were appealing the court’s ruling on the ownership of Pueblo Latino.“) (emphasis added).
And subsequent acts by the defendant after the offense is completed, especially starting several years later, typically are not probative of the defendant’s intent at the time of the offense. Cf. In re Hanson, 432 B.R. 758, 772-73 (Bankr. N.D. Ill. 2010) (“Proof of intent to deceive is measured by the debtor’s subjective intention at the time the representation was made. Therefore, subsequent acts of fraud or omission do not demonstrate that the debtor had the requisite intent at the time the representations were made.”) (internal citations omitted and emphasis added); In re Trevisan, 300 B.R. 708, 717 (Bankr. E.D. Wis. 2003) (finding defendant’s subsequent conduct after the alleged fraudulent checks were “inadequate proof of the Debtor’s fraudulent intent” at the time of writing the checks).
Courts typically require evidence of the defendant’s state of mind contemporaneous to the alleged offense, while exercising great caution in allowing the plaintiff to prove the defendant’s state of mind based on later evidence of the defendant’s conduct subsequent to the offense. For example, in United States ex rel. Edward O’Donnell v. Countrywide Home Loans, Inc., 822 F.3d 650 (2d Cir. 2016), the Second Circuit explained the longstanding principle that a subsequent breach of contract by the defendant cannot establish the defendant’s earlier intent at the time of making the allegedly fraudulent contract:
What fraud in these instances turns on, however, is when the representations were made and the intent of the promisor at that time. As explained below, where allegedly fraudulent misrepresentations are promises made in a contract, a party claiming fraud must prove fraudulent intent at the time of contract execution; evidence of a subsequent, willful breach cannot sustain the claim. * * *
The SEcond Circuit in Edward O’Donnell v. Countrywide Home Loans, Inc., 822 F.3d at 662.
In sum, a contractual promise can only support a claim for fraud upon proof of fraudulent intent not to perform the promise at the time of contract execution. Absent such proof, a subsequent breach of that promise—even where willful and intentional—cannot in itself transform the promise into a fraud. Far from being an arcane limitation, the principle of contemporaneous intent is, like materiality, one without which “the common law could not have conceived of ‘fraud.’”
See also Akinsade v. Holder, 678 F.3d 138, 146 (2d Cir. 2012) (respondent’s later acceptance of plea for embezzlement was not evidence “that he knew the checks were fraudulent at the time he committed the offense or that he had knowingly processed fraudulent checks”).
In some cases involving an ongoing “overarching scheme to defraud” by the defendant, “subsequent conduct [of the defendant] may be relevant to an earlier state of mind.” Palmacci v. Umpierrez, 121 F.3d 781, 792-85 (1st Cir. 1997) (citing In re Haining, 119 B.R. 460, 464 (Bankr. D. Del. 1990); Comerica Bank v. Weinhardt (In re Weinhardt), 156 B.R. 677, 670 (Bank. M.D. Fla. 1993)). But, absent such an overarching scheme to defraud, the defendant’s subsequent conduct is not probative of the defendant’s intent or state of mind at the time of the alleged offense.
Indeed, in a bankruptcy case, the First Circuit upheld the exclusion of “proffered testimony related to events almost two years after [the alleged offense when] Palmacci’s investment was induced.” Palmacci, 121 F.3d at 793 (emphasis added). Two years later is not contemporaneous.
Like Palmacci, the communication by OpenAI’s in-house counsel in 2022 occurred well after the alleged offense occurred: 2 years after the training of OpenAI’s model in 2020 (as discussed by OpenAI’s published paper) and 4 years after OpenAI’s torrenting from shadow libraries in 2018.
Like Palmaccci, there is no evidence that OpenAI’s in-house counsel were engaged in an “overarching scheme to defraud.” Indeed, the 3 OpenAI in-house counsel were not even working for OpenAI in 2018 or 2020, when the alleged infringement occurred. And, as Magistrate Judge Wang concluded, what the lawyers said regarding the deletion of Books 1, 2 were not in furtherance of a crime or fraud. (p. 26.) Plus, “Class Plaintiffs fail to allege a sufficient factual basis to find probable cause to believe that that the communications at issue were in furtherance of the types of misconduct that are recognized by Second Circuit precedent and/or that such communications were intended to facilitate or conceal such activities.” (p. 27).
Under these circumstances, the 2022 communications of OpenAI’s in-house counsel are irrelevant to the willfulness or state of mind of OpenAI researchers who engaged in torrenting in 2018 and AI training in 2020, the allegedly willful infringing activity. What OpenAI’s in-house counsel thought in 2022 cannot magically change what OpenAI’s researchers thought in 2018 and 2020.
Related Stories