The NYT OpenAI Preservation Order: 136 Days, 20M Logs and Who It Skipped

For 136 days in 2025, deleting a ChatGPT conversation did not delete it. On 13 May 2025 a federal magistrate judge ordered OpenAI to keep output logs it would normally erase within 30 days, as evidence in New York Times v. OpenAI. The NYT OpenAI preservation order ended on 26 September 2025, but 20 million of those logs were later ordered handed to the plaintiffs.

What the NYT OpenAI preservation order actually did

A preservation order tells a party to a lawsuit to stop destroying material that might be evidence. It does not hand anything to the other side. It freezes the material so that a later fight about disclosure has something to fight over.

On 13 May 2025, Magistrate Judge Ona T. Wang of the Southern District of New York directed OpenAI to "preserve and segregate all output log data that would otherwise be deleted on a going forward basis". The case is In re OpenAI, Inc. Copyright Infringement Litigation, No. 23 Civ. 11195, which consolidates the Times' suit with other news publishers.

The phrase that mattered was "would otherwise be deleted". It reached chats users had deleted themselves, Temporary Chat sessions designed to vanish, and API outputs on a standard 30-day retention schedule. Duane Morris summarised it as covering data subject to user deletion requests or statutory erasure rights.

This is written for the head of operations or delivery who owns vendor risk and answers the security questionnaire. That person signs off on where customer data goes, and that person is the one who has to explain afterwards why a deletion promise did not hold.

The Times' theory was straightforward. Users who prompt ChatGPT to reproduce paywalled articles might delete those chats, and routine deletion would destroy evidence of infringement. Whether that theory is right is for the copyright trial. The overview of where the AI copyright cases on training data stand covers the merits. This post covers what the discovery fight did to the data of everyone else.

The dated timeline of the ChatGPT logs retention fight

Most coverage compresses this dispute into "a judge made OpenAI keep everything". The record is more useful than that. It shows a sequence of 10 dated steps, each one a point where a customer could have changed what they sent.

ECF numbers come from the law firm summaries cited, not from the docket itself. Duane Morris gives the 9 October modification as ECF No. 922 in its text and links a file labelled 992, so treat that number as unconfirmed.

Which ChatGPT and API tiers the retention order covered

OpenAI's notice of 5 June 2025 answered the question every buyer asked first. Was my data in scope? The answer depended entirely on which contract you were on, not on how much you paid or how large your company was.

According to the notice, as quoted by Simon Willison, the order applied if you had "a ChatGPT Free, Plus, Pro, and Teams subscription" or used the API without a zero data retention agreement. It stated that it did "not impact ChatGPT Enterprise or ChatGPT Edu customers".

The Team plan is the uncomfortable row

Team is sold to businesses. Many small companies bought it precisely because it promised not to train on their data. That promise held. The deletion promise did not, because Team sat on the covered side of the line.

My view is that this row is where most of the real exposure sat. A 15-person agency on Team had business data in scope. A 5,000-person bank on Enterprise did not. Size did not decide it. The contract name did.

Zero data retention worked because it is architecture

OpenAI's notice described zero data retention traffic plainly: "we never retain the prompts you send or the answers we return". A court can order a company to keep what it has. It cannot order it to keep what it never wrote down.

That distinction is the most useful thing a buyer can take from the whole episode. A deletion schedule is a promise about the future. Zero retention is a fact about the present. Courts override promises routinely.

Where the order collided with OpenAI data retention promises

OpenAI's standing policy was that deleted ChatGPT conversations and standard API content were removed within 30 days. The order suspended that for covered tiers. OpenAI's own reconsideration filing said the order would "disregard legal, contractual, regulatory, and ethical commitments to hundreds of millions of people".

OpenAI's chief operating officer, Brad Lightcap, put it more bluntly. He said the demand "fundamentally conflicts with the privacy commitments we have made to our users". Chief executive Sam Altman argued publicly that talking to an AI should be treated like talking to a lawyer or a doctor.

The privacy policy already allowed it

Here is the part most commentary skipped. Judge Stein did not find that OpenAI was breaking its promises. According to Nelson Mullins, he highlighted that OpenAI's terms of use allowed preservation for legal requirements. The carve-out was in the policy all along.

That is the real collision. It was not between a court and a promise. It was between what customers believed the promise said and what the promise actually said. Almost every vendor retention clause carries the same "unless legally required" exception. The broader comparison of how model provider terms handle data use and retention shows how common the pattern is.

Stein reasoned, as reported by Bloomberg Law, that it was a permissible inference that some users who tried to extract news content would want to delete those chats. OpenAI countered that the plaintiffs had no substantial evidence anyone did this, and that the plaintiffs themselves estimated only 0.006% of the data might even be relevant.

I find OpenAI's burden argument stronger than its privacy argument. A tiny relevant fraction does not mean the data is irrelevant. It means the order was a very wide net cast for a very small fish, and every covered customer was inside the net.

How the OpenAI data retention court order was narrowed

The order never had an end date. It ran "until further order of the Court". What ended it was a negotiated stipulation, which Judge Wang signed on 9 October 2025.

Per Duane Morris, the modification ended the ongoing preservation duty "as of September 26, 2025" and limited retention to logs already segregated. It also excluded requests originating from the European Economic Area, Switzerland, and the United Kingdom.

Three things survived the narrowing

First, every log already preserved stayed preserved. Engadget reported that chat logs saved under the earlier order would still be accessible. Ending the duty stopped new collection. It did not release the old collection.

Second, preservation continued for accounts the Times flagged. Duane Morris describes "targeted, domain-based preservation for select accounts listed in an appendix", and Engadget reported the Times could expand the flagged list. Going-forward retention became targeted rather than universal.

Third, the European carve-out arrived at the end, not the start. For 136 days, data from users with statutory erasure rights sat under a US hold. That is a direct illustration of the cross-border problem covered in the analysis of AI data residency rules in the EU and India.

From preserving ChatGPT logs to producing 20 million of them

Preservation is the quiet stage. Production is when data leaves the vendor. In July 2025 the plaintiffs moved to compel a sample of 120 million chat logs. OpenAI offered 20 million instead, scrubbed of personally identifiable information, and the plaintiffs accepted while reserving the right to ask for more.

OpenAI then proposed running searches to find conversations involving the plaintiffs' content, rather than producing the full sample. Judge Wang rejected that in November 2025. On a Monday in January 2026, Judge Stein affirmed. The ABA Journal quoted his order: Wang's rulings were "neither clearly erroneous nor contrary to law".

Stein found that the smaller sample, de-identification and the protective order adequately protected users, according to the ABA Journal. Bloomberg Law reported he also distinguished wiretap precedent because users had submitted their prompts voluntarily.

That reasoning should change how you think about chat logs. A court treated conversations with an AI vendor as business records of the vendor, subject to ordinary discovery. Ilia Kolochenko of Platt Law, quoted in the same ABA Journal piece, called it a reminder that interactions with AI systems may one day be produced in court regardless of privacy settings.

Where this argument is weakest

I have framed this as a lesson for every buyer. The honest counter-case deserves the same space.

The practical exposure was probably small

The preserved data sat with a small, audited legal and security team, according to CyberInsider's account of OpenAI's notice. It was not shared wholesale with plaintiffs. The 20 million sample was de-identified and covered by a protective order. No public report I found describes a single customer harmed by this hold.

If you weigh realised harm rather than theoretical exposure, the episode looks more like a scare than a breach. That is a fair reading, and I do not think it is the right one for planning, because the next hold may come with weaker safeguards.

This was one case with unusual facts

The Times alleged that users deleted chats to hide infringement. That fact pattern drove the breadth. A routine commercial dispute might never justify an order this wide, and the court itself said at the 27 May hearing that it did not intend wholesale preservation.

Several details rest on secondary sources

I could not open OpenAI's notice or the court orders directly during this research. OpenAI's page refused automated access and the October order is a scanned image. Tier coverage, dates and quotations come from law firm summaries and press reports that quote the documents. They agree with each other, but agreement among summaries is not the same as reading the order.

The contract lesson for zero data retention and legal holds

The lesson is not "avoid OpenAI". Any vendor that stores your prompts can be ordered to keep them, and any vendor in litigation is a candidate. The question is what your contract does when that happens. The checklist in the clauses that matter in an AI data processing agreement is the right starting template.

Notification is the clause that was missing

Nelson Mullins advised companies to add terms requiring notification when their data becomes subject to a hold, with an opportunity to object. I agree, and I would make it the first redline. Customers learned of this order from a public blog post 23 days after it was signed.

Route by sensitivity, not by vendor

Split your traffic. Send regulated or confidential workloads through zero retention endpoints or an enterprise contract with admin-controlled retention. Leave low-risk drafting on standard plans. The question of personal AI accounts sitting outside the company perimeter matters here, because staff on personal Plus accounts were inside this order with your data.

Then check what each vendor actually guarantees on delete. The ranking of ten vendors' AI data retention guarantees shows that retention terms now vary by model as well as by plan, which makes the routing decision a recurring one.

Frequently asked questions

How to check your own exposure this week

Pull a list of every AI tool your team uses and mark the plan name next to each one. Any row that reads Free, Plus, Pro or Team, or API without a written zero retention amendment, was the kind of account this order reached. That list is your exposure map.

Then send one question to each vendor on it. Ask what they do, and how fast they tell you, when a court orders them to keep data they would otherwise delete. Write down who answered in writing and who sent a link to a policy page. The difference is your shortlist for the next renewal, and the clauses finance teams are adding to AI contracts covers how to write it in.

Related reading