In a ruling with major implications for AI companies navigating copyright litigation, a federal judge in a copyright action against OpenAI has drawn a bright line between denying wrongdoing and inviting discovery into privileged legal advice. In re OpenAI, Inc., Copyright Infringement Litigation, No. 25-md-3143, D.N. 1776 (S.D.N.Y. Feb. 6, 2026)
The decision offers a roadmap for defendants seeking to protect attorney-client communications while still contesting allegations of willful infringement.
Background
In April 2025, the U.S. Judicial Panel on Multidistrict Litigation centralized twelve copyright infringement actions against OpenAI and Microsoft into a single MDL in the Southern District of New York, assigned to Judge Sidney H. Stein. The cases arise from allegations that defendants used copyrighted works—including books, news articles, and video transcripts—without consent to train their large language models.
Before this litigation commenced, OpenAI had deleted its “Books1” and “Books2” training datasets—datasets created from pirated books downloaded from Library Genesis (“LibGen”). OpenAI stated that the reason for the deletion was the “non-use” of the Books1 and Books2 datasets. But when Class Plaintiffs sought discovery about the reasons for the deletion of datasets, OpenAI asserted attorney-client privilege.
The Magistrate Judge’s November 2025 Ruling
On November 24, 2025, Magistrate Judge Ona T. Wang ruled that OpenAI had waived attorney-client privilege over communications related to its deletion of its “Books1” and “Books2” training datasets. Judge Wang found waiver on three grounds: (1) given OpenAI’s position that all reasons for the deletion were privileged, OpenAI voluntarily disclosed a privileged reason (“non-use”) for the deletion, (2) OpenAI made a “moving target” of its privilege assertions, and (3) OpenAI put its good faith and state of mind at issue by denying willful infringement allegations. Judge Wang did not find that the crime-fraud exception applied here. The scope of that waiver encompassed all communications in 2022 related to the reasons for deleting the Books1 and Books2 datasets and all internal references to LibGen. Judge Wang ordered production of the communications and depositions of OpenAI’s in-house attorneys.
Judge Stein’s Order of February 6, 2026
On February 6, 2026, Judge Stein sustained OpenAI’s objection under Federal Rule of Civil Procedure 72(a) and set aside the magistrate judge’s order in its entirety, finding each basis for waiver to be “clearly erroneous” or “contrary to law.”
No disclosure of privileged material. Judge Stein held that OpenAI’s statements that the datasets were deleted “due to non-use” did not reveal any legal advice from an attorney. Because the statements were not privileged themselves, disclosing them could not trigger a waiver.
No “moving target.” The court found that OpenAI consistently maintained that attorney-client communications about the deletion were privileged. While OpenAI’s later formulations were “inartful”—including an improper instruction to its witness not to answer questions about non-privileged facts—these missteps did not amount to the deliberate re-engineering of privilege assertions that courts have sanctioned with waiver.
No “at-issue” waiver. Most significantly, Judge Stein rejected the conclusion that merely denying willful infringement constituted an “at-issue” waiver. The court drew a clear line between denying willfulness—on which the plaintiff bears the burden—and affirmatively asserting a good faith belief in the lawfulness of one’s conduct. “At-issue” waiver requires a party to rely on privileged advice to support its defense. OpenAI represented that it will not offer state-of-mind evidence at trial and will instead defend on substantive grounds such as fair use.
Crime-fraud exception rejected. Judge Stein affirmed the finding that plaintiffs failed to show probable cause that the privileged communications were made in furtherance of, or intended to conceal, criminal copyright infringement.
Key Takeaways
- Disclosing facts does not waive privilege over related legal advice. A party may describe what it did without forfeiting privilege over communications that informed why it did so.
- Denying willfulness is not the same as asserting good faith. AI defendants can contest willful infringement without triggering “at-issue” waiver, provided they do not introduce evidence of their subjective belief in the legality of their conduct.
- Precision in privilege assertions matters. The court flagged OpenAI’s “inartful” formulations and improper deposition instructions. Practitioners should clearly and consistently state privilege positions from the outset.
- Willfulness strategy carries high stakes in AI copyright cases. Under 17 U.S.C. § 504(c)(2), statutory damages range from $200 for innocent infringement to $150,000 for willful infringement. The line between denial and affirmative defense has real consequences for privilege and exposure.
