A Copyright Suit Became a Fight Over Your ChatGPT Logs.
In December 2023, The New York Times sued OpenAI and Microsoft for copyright infringement, alleging they trained ChatGPT on millions of Times articles without permission. Two years later, that lawsuit spawned a second fight no ChatGPT user signed up for: a battle over whether the private conversations of hundreds of millions of people could be pulled into discovery.
In May 2025 a magistrate judge ordered OpenAI to stop deleting ChatGPT logs and preserve them. By November, the order had narrowed to a demand that OpenAI hand over 20 million de-identified conversation logs to the news plaintiffs. OpenAI called it a “crazy overreach” and an invasion of privacy. The Times called it ordinary civil discovery.
In January 2026, the district judge sided with the Times and affirmed the production order in full. This is a non-partisan dispute between two large institutions, and the facts cut both ways. Here is what each side argued, what the courts decided, and what it means for anyone who has ever typed into ChatGPT.
- 20Mlogs orderedde-identified ChatGPT conversations OpenAI must produce — Nov 7, 2025 order, affirmed Jan 2026
- 1.4Bfirst demandedlogs the NYT originally sought before the court narrowed the sample — OpenAI / Bloomberg Law
- 400M+users affectedOpenAI’s estimate of accounts swept into the preservation order — OpenAI
- 99.99%irrelevantshare of preserved logs OpenAI says have nothing to do with the case — OpenAI
The dispute lives inside a larger case. In December 2023, The New York Times filed suit against OpenAI and Microsoft in the Southern District of New York; that case was later consolidated with related copyright suits into a multidistrict proceeding, MDL 1:25-md-03143, with Judge Sidney Stein presiding and Magistrate Judge Ona Wanghandling discovery. To prove its case, the Times wanted to see how often ChatGPT reproduces Times content in ordinary use — and that meant the conversation logs.
OpenAI normally deletes ChatGPT conversation data on a roughly 30-day cycle, and offers users temporary chats and the ability to delete history. The discovery fight turned that routine deletion into a legal problem: the plaintiffs argued OpenAI could not be allowed to destroy the very evidence at issue while the case was pending.
On May 13, 2025, Magistrate Judge Wang issued a preservation order directing OpenAI to “preserve and segregate all output log data that would otherwise be deleted.” In practice that meant ChatGPT logs that would normally have been wiped after 30 days now had to be kept. The order covered ChatGPT Free, Plus, Pro, and Teamaccounts; it excluded Enterprise and API customers with zero-data-retention agreements. By OpenAI’s count, it reached 400 million-plus users.
OpenAI did not disclose the order to users until June 5, 2025, when it published a blog post and CEO Sam Altman posted on X. OpenAI appealed. Then, on October 9, 2025, Wang terminated the blanket forward-looking preservation requirement — standard 30-day deletion could resume — but logs already preserved, and accounts the Times had specifically flagged, stayed held.
recently the NYT asked a court to force us to not delete any user chats. we think this was an inappropriate request that sets a bad precedent. we are appealing the decision. we will fight any demand that compromises our users' privacy; this is a core principle.
- →Dec 27, 2023: NYT sues OpenAI and Microsoft (SDNY); later consolidated into MDL 1:25-md-03143
- →May 13, 2025: Magistrate Judge Wang orders OpenAI to preserve and segregate all output log data (Free/Plus/Pro/Team; Enterprise + ZDR API excluded)
- →June 5, 2025: OpenAI first discloses the order to users, appeals; Altman posts on X
- →Oct 9, 2025: Wang terminates blanket forward-looking preservation; 30-day deletion resumes; already-preserved + NYT-flagged logs stay held
- →Nov 7, 2025: Wang's production order — OpenAI must produce 20 million de-identified logs (NYT had sought 1.4 billion)
- →Nov 12, 2025: OpenAI publishes 'Fighting the New York Times' invasion of user privacy'
- →Jan 5–6, 2026: Judge Stein affirms Wang's production order in full
- →April 2, 2026: summary-judgment deadline
OpenAI’s position, set out most fully in its November 12, 2025 post by Chief Information Security Officer Dane Stuckey, is that the demand is a privacy violation. People use ChatGPT for intensely personal things — health questions, financial worries, relationship problems — and many of them used temporary or deleted chats precisely because they expected the content to disappear. The company argues that even de-identified data can sometimes be re-identified, and that 99.99%of the logs in question have nothing to do with the Times’ copyright claims.
“anyone in the world who has used ChatGPT in the past three years must now face the possibility that their personal conversations will be handed over.”
Dane Stuckey, OpenAI Chief Information Security Officer · OpenAI blog · Nov 12, 2025
Sam Altman framed the original preservation demand as a “crazy overreach” and floated the idea of an “AI privilege” — a legal protection for conversations with an AI system analogous to attorney-client or doctor-patient privilege. OpenAI also announced it was appealing the rulings and pursuing technical safeguards including client-side encryption for chats.
We're challenging The New York Times' demand for user data. Protecting your private conversations with ChatGPT is a core commitment. We believe this demand disregards long-standing privacy protections, and we're fighting it while pursuing technical safeguards like client-side encryption.
The Times sees the same facts very differently. Its argument is that this is routine civil discovery— the ordinary, unglamorous process by which a plaintiff obtains the evidence it needs to prove its claims. To show that ChatGPT reproduces Times journalism in normal use, and to test OpenAI’s defense that verbatim outputs were the product of manipulated prompts, the Times says it needs to look at real conversations, not a curated sample chosen by OpenAI.
The Times also argues the logs may show users employing ChatGPT to bypass its paywall — reading articles they would otherwise have to pay for — which goes directly to the alleged commercial harm at the heart of the copyright case. And it emphasizes the safeguards built into the order: the sample was narrowed from 1.4 billion logs to 20 million, the logs are de-identified, and they are produced under an attorneys’-eyes-only protective order, meaning the public never sees them and Times reporters never touch them.
“No ChatGPT user's privacy is at risk.”
The New York Times, spokesperson · on the production order
From the Times’ vantage, OpenAI’s “invasion of privacy” framing is a public-relations maneuver that recasts a defendant’s ordinary discovery obligation as an assault on its own users. Litigants routinely produce far more sensitive records — emails, medical files, financial data — under protective orders without it being called a privacy crisis.
Both courts sided with the Times. On November 7, 2025, Magistrate Judge Wang issued the production order: OpenAI must produce 20 millionrandomly sampled, de-identified ChatGPT logs — conversations spanning December 2022 through November 2024 — down from the 1.4 billion the Times had originally sought. OpenAI objected. On January 5–6, 2026, District Judge Sidney Stein affirmed Wang’s order in full and rejected OpenAI’s privacy objections.
Judge Stein found three safeguards adequate to address the privacy concern: the reduction to a 20-million-log sample, the de-identification of the logs, and the attorneys’-eyes-only protective order. He distinguished the situation from a wiretap or government surveillance, reasoning that ChatGPT users had voluntarily submitted their communicationsto OpenAI — a company whose privacy policy already contemplates legal disclosure — rather than having them intercepted without consent.
“[ChatGPT users] voluntarily submitted their communications to OpenAI.”
Judge Sidney Stein, U.S. District Court, SDNY · affirming the production order · January 2026
- →Sample reduction: from 1.4 billion logs originally demanded to 20 million randomly sampled
- →De-identification: personal identifiers stripped from the produced logs
- →Attorneys'-eyes-only protective order: logs seen only by plaintiffs' counsel and experts, never the public or NYT newsroom
- →Voluntary submission: users provided the data to OpenAI directly, unlike a wiretap or covert surveillance
The privacy fight is a sideshow to the main event. The Times alleges that OpenAI and Microsoft trained their models on millions of Times articles without permission or payment, and that ChatGPT can reproduce that journalism — sometimes verbatim, and sometimes by inventing (“hallucinating”) articles falsely attributed to the Times. The Times argues this both infringes its copyrights and competes with its own paywalled product.
OpenAI’s defense is fair use. It contends that training a model on published text is “highly transformative,” that the model learns patterns rather than storing articles, and that the Times manufactured its verbatim examples by feeding the model manipulative prompts designed to force regurgitation. OpenAI’s lead trial counsel is Robert Van Nestof Keker, Van Nest & Peters.
In March 2025, Judge Stein allowed the core copyright claims to proceed, rejecting OpenAI’s bid to dismiss them. The summary-judgment deadline is set for April 2, 2026— the next major checkpoint in a case that could shape how every AI company in the country is allowed to train its models. Whatever the outcome, it has not yet been decided on the merits; both the infringement claims and the fair-use defense remain unresolved.
A copyright lawsuit between The New York Times and OpenAI produced a discovery order that swept in the conversations of hundreds of millions of ChatGPT users. OpenAI says the demand — first 1.4 billion logs, later narrowed to 20 million — is an invasion of user privacy and floated an “AI privilege” to protect such chats. The Times says it is routine civil discovery, narrowed and safeguarded, with no user’s privacy actually at risk.
The courts agreed with the Times: Magistrate Judge Wang ordered the production, and in January 2026 Judge Stein affirmed it, finding that users had voluntarily handed their data to OpenAI and that de-identification plus an attorneys’-eyes-only order were enough. The underlying question — whether training ChatGPT on Times journalism was fair use — remains undecided, with summary judgment due April 2, 2026.
