What exactly did the New York Times preservation order require OpenAI to retain, and how long did it last?
Executive summary
The May 13, 2025 preservation order directed OpenAI to "preserve and segregate all output log data that would otherwise be deleted on a going‑forward basis" — in plain terms, to stop deleting ChatGPT output logs (including chats users had deleted) and keep them available for potential discovery in The New York Times copyright litigation (Magistrate Judge Ona T. Wang) [1]. That broad pause on routine deletions remained in force through a negotiated narrowing and sunset: OpenAI’s extraordinary ongoing preservation obligations were effectively ended as of September 26, 2025, with later court paperwork in October modifying and lifting the blanket requirement (though some already‑segregated logs and flagged accounts remained subject to retention) [2] [3] [4].
1. What the order required: preserve and segregate “output log data”
The core command on May 13 was precise in legal language: preserve and segregate output log data that would otherwise be deleted — interpreted by courts and counsel to mean ChatGPT chat transcripts and associated output logs across consumer tiers, including conversations users had explicitly deleted or that OpenAI’s 30‑day deletion policy would have removed (the court’s written order and contemporaneous reporting echo this formulation) [1] [5]. Multiple legal analyses and industry write‑ups summarized the order as a mandate to suspend OpenAI’s normal deletion practices for those output logs so the plaintiffs could, if necessary, search them for possible copyrighted material [6] [5].
2. Scope: broad, but with important carve‑outs and later limits
Although the May order was sweeping, both parties and subsequent judicial actions created carve‑outs: ChatGPT Enterprise, ChatGPT Edu, and customers with explicit zero‑retention contracts were treated as excluded from the preservation requirement, and the court later accepted geographic and other limitations in stipulated modifications (OpenAI and commentators emphasized those exclusions) [7] [2] [3]. OpenAI also secured assurances that preserved data would be subject to restricted access protocols while segregated, though the order itself compelled preservation rather than production at that stage [7].
3. Why it mattered: deleted chats, privacy commitments, and industry norms
The order directly conflicted with OpenAI’s public privacy commitments — namely, automatic deletion of ChatGPT conversations and API data after 30 days — because it required retaining chats users believed they had deleted and retaining logs that privacy rules elsewhere would otherwise erase (OpenAI publicly framed the order as an industry‑breaking overreach and promptly appealed) [2] [8]. Lawyers and industry observers called the May preservation mandate a watershed for AI data governance because it converted ephemeral user interactions into litigation‑grade, long‑lived evidence across hundreds of millions of users if applied broadly [6] [5].
4. How long did the preservation last, and what remained after it ended
Practically speaking, the preservation obligation functioned from the May 13, 2025 order until judicial management and a stipulated modification limited ongoing duties — with courts and filings indicating OpenAI’s broad duty to preserve was no longer required for new logs after September 26, 2025, and a later October 9 order formalized the end of the blanket going‑forward preservation while leaving already‑segregated logs and data linked to accounts flagged by the News Plaintiffs accessible for review (OpenAI’s status filings and third‑party reporting converge on that timeline) [2] [3] [4]. Some subsequent orders and rulings then moved the dispute from preservation into production disputes (including later rulings about production of a subset of logs), but the initial indefinite‑retention posture was curtailed by the late‑September cutoff and October modifications [7] [9].
5. Competing narratives and the limits of available reporting
OpenAI framed the order as an overbroad, privacy‑undermining imposition that it appealed and later mitigated through negotiation and judicial modification [2] [8], while court documents and plaintiffs’ filings portrayed preservation as necessary to prevent spoliation given allegations of deleted evidence [1] [6]. Public reporting sometimes compressed preservation and later production orders into a single “indefinite retention” narrative; the record shows a sequence: a sweeping May preservation mandate, litigation and appeal, then narrowing and a September 26/October judicial shift that ended the blanket, going‑forward preservation though leaving previously preserved materials accessible [1] [2] [3]. If further granularity is needed — e.g., the exact fields preserved, the precise mechanisms of segregation, or which flagged accounts remained — those specifics are not fully disclosed in the sources provided here and would require consulting the court docket or sealed technical exhibits.