IP Law Daily, COPYRIGHT—S.D.N.Y.: OpenAI did not waive attorney-client privilege, magistrate judge’s discovery order vacated, (Feb 9, 2026)

By Ravindra Kumar Singh, B.L.

A magistrate judge erred in finding a waiver based on non-privileged disclosures and denial of willfulness.

The federal district court in Manhattan, New York, has set aside a magistrate judge’s discovery order that compelled the production of attorney-client communications relating to the deletion of two book-based training datasets used by OpenAI and related entities. The court held that references to “non-use” as a reason for deletion did not disclose privileged legal advice, that OpenAI did not take inconsistent privilege positions warranting waiver, and that merely denying allegations of willful copyright infringement did not place its state of mind at issue. The court further agreed that the crime–fraud exception did not apply because the challenged communications were not made in furtherance of criminal conduct or spoliation. Accordingly, the court vacated the magistrate judge’s November 24, 2025, order as clearly erroneous or contrary to law (In re: OpenAI, Inc., Copyright Infringement Litigation, No. 1:23-cv-08292-SHS-OTW (S.D.N.Y. Feb. 6, 2026)).

Background. OpenAI operates as an artificial intelligence research and deployment organization that develops large language models, including GPT-3 and GPT-3.5. In the consolidated putative class actions, authors and copyright holders alleged that the company unlawfully copied and used copyrighted books to train its models.

The plaintiffs asserted copyrights in literary works that were allegedly reproduced without authorization and incorporated into training datasets. The challenged materials related to books obtained from Library Genesis, which the plaintiffs described as a “shadow library” hosting pirated digital copies.

According to the record, an employee downloaded books from Library Genesis in 2018, and the company used those copies to create two datasets, “Books1” and “Books2” that were used to train GPT-3 and GPT-3.5. The company deleted the datasets in mid-2022. During discovery, outside counsel stated in March and April 2024 letters that use of the datasets had been discontinued after training and that the datasets were deleted due to “non-use.”

In January 2025, a corporate designee testified in a Rule 30(b)(6) deposition that legal counsel was involved in the decision to delete the datasets, after which counsel instructed him not to answer further questions on privilege grounds. The plaintiffs moved to compel discovery, arguing that the company waived privilege by previously citing “non-use” or, in the alternative, that the crime–fraud exception applied. The magistrate judge agreed and found a waiver, prompting the defendant’s objection under Federal Rule of Civil Procedure 72(a).

Disclosure of non-privileged facts did not waive privilege. The district court first held that statements attributing deletion to non-use did not reveal privileged legal advice. Citing United States v. Mejia, 655 F.3d 126 (2d Cir. 2011), the court reiterated that attorney-client privilege protects confidential communications made for the purpose of obtaining or providing legal advice. The court emphasized that waiver arises only when a party discloses privileged communications or selectively relies on them, referencing In re Grand Jury Proceedings, 219 F.3d 175 (2d Cir. 2000), and United States v. Bilzerian, 926 F.2d 1285 (2d Cir. 1991). Because non-use was a factual explanation that did not convey legal advice, its disclosure could not lead to a waiver. The court further relied on In re FirstEnergy Corp., 154 F.4th 431 (6th Cir. 2025), for the proposition that disclosure of non-privileged material never waives privilege.

No “moving target” of privilege assertions. The court next rejected the magistrate judge’s conclusion that the defendant made a “moving target” of its privilege positions. It found that the defendant consistently asserted privilege over communications seeking or providing legal advice about deletion, even if later deposition conduct and phrasing were imprecise. Comparing the record to cases such as In re Honeywell International Income Securities Litigation, 230 F.R.D. 293 (S.D.N.Y. 2003), and SEC v. Yorkville Advisors, LLC, 300 F.R.D. 152 (S.D.N.Y. 2014), the court concluded that the conduct here did not resemble strategic or shifting assertions warranting a waiver sanction.

Denial of willfulness. Addressing at-issue waiver, the court held that the defendant did not place its state of mind in issue by merely denying allegations of willful infringement. Relying on In re County of Erie, 546 F.3d 222 (2d Cir. 2008), the court explained that an at-issue waiver requires an affirmative reliance on privileged advice, such as asserting a good-faith belief in legality. The court distinguished between denying willfulness, which remains the plaintiff’s burden, and affirmatively asserting good faith. Citing Fitzgerald Publishing Co. v. Baylor Publishing Co., 807 F.2d 1110 (2d Cir. 1986), the court emphasized that lack of willfulness is not the converse of innocent intent. Because the defendant expressly disclaimed any good-faith defense and stated it would rely on fair use and lack of liability, no waiver occurred.

Crime–fraud exception. Finally, the court agreed that the crime–fraud exception did not apply. Applying In re Grand Jury Subpoenas Dated September 13, 2023, 128 F.4th 127 (2d Cir. 2025), the court held that the plaintiffs failed to show probable cause that the communications were made in furtherance of criminal infringement or intended to conceal wrongdoing. The court also rejected the argument that deletion amid legal uncertainty constituted spoliation sufficient to trigger the exception, noting that recovered versions of the datasets were later produced.

The Case is No. 1:23-cv-08292-SHS-OTW.

Judge: Stein, S.

Attorneys: Anna Josefine Freymann (Lieff Cabraser Heimann & Bernstein) for Authors Guild, David Baldacci and Mary Bly. Allison S. Blanco (Latham & Watkins LLP) for OpenAI Inc., OpenAI GP LLC and OpenAI LLC.

Companies: Authors Guild; OpenAI Inc.; OpenAI GP LLC; OpenAI LLC

MainStory: TopStory AINews Copyright NewYorkNews GCNNews