OpenAI engineers erased search programs, search-result data, folder structures, and some file names from a virtual machine used in copyright discovery on November 14, 2024. The New York Times and Daily News said the loss undermined their ability to determine how their articles appeared in OpenAI’s training datasets. OpenAI said a requested configuration change affected a temporary cache drive, that the underlying file contents were not shown to be lost, and that the searches could be rerun.
The public record supports a narrower conclusion than the headline “OpenAI deleted evidence” suggests: potentially important discovery work product and metadata were erased, apparently unintentionally, but the cited filings do not establish that OpenAI destroyed its underlying training datasets or that the incident resulted in legally sanctionable spoliation.
As an Amazon Associate I earn from qualifying purchases.
What lawsuit was involved?
The dispute arose inside copyright cases brought by The New York Times and Daily News against OpenAI and Microsoft. The publishers allege that their copyrighted articles were used to develop artificial-intelligence models without permission. OpenAI has maintained that using publicly available material for model training can be protected by fair use.
The November 2024 event was not a separate lawsuit about data deletion. It was a discovery dispute within that litigation, involving the parties’ efforts to search OpenAI’s training data.
#1 Best Overall
What OpenAI erased
In a November 20, 2024 court filing, the publishers said OpenAI engineers erased material from one of two virtual machines provided for their searches. They identified:
- Search programs used by the publishers’ lawyers and experts.
- Search-result data.
- Folder structures.
- File names and other organizational metadata.
The filing said OpenAI recovered much of the data but could not restore the original folder structure and file names. The publishers argued that this made the recovered material unreliable for identifying where their articles appeared in the relevant training datasets.
That distinction matters. The filing does not establish that OpenAI erased the training datasets themselves. It describes the loss of search work product and the metadata that gave those search results context.
Quick wins for a faster PC:
Clear out junk files and repair common Windows errorsFree Scan →Fix the driver behind crashes, sound loss and screen glitchesFind Drivers →Repair Windows errors before they cause bigger problemsFix Now →Read the publishers’ November 20 filing.
The two sides’ accounts
| Issue | Publishers’ account | OpenAI’s account |
|---|---|---|
| Search programs | OpenAI engineers erased the programs. | A configuration change affected the search environment. |
| Search-result data | It was erased and the work had to be recreated. | The file contents were not shown to be lost and the searches could be rerun. |
| Folder structure | It could not be restored, making the recovered work unreliable. | The configuration change removed folder structures from a temporary cache drive. |
| File names | Original names were lost along with organizational context. | Some file names were removed, but the underlying contents were not believed to be lost. |
| Intent | The publishers said they had no reason to believe the erasure was intentional. | OpenAI characterized it as an accidental technical problem following a requested setup change. |
OpenAI’s November 22 response said the plaintiffs had requested a configuration change, that the drive was intended as a temporary cache for OpenAI data, and that the plaintiffs had stored search results there without backups. OpenAI said the searches could be rerun with additional computing time.
Read OpenAI’s November 22 response. OpenAI later gave a similar public account on its New York Times litigation page.
Rank #2
How much work was affected?
The publishers told the court that their lawyers and experts had spent more than 150 person-hours searching OpenAI’s training data after November 1, 2024, in addition to substantial computing time. They said an entire week of work had to be recreated.
OpenAI’s position was that the searches could be rerun. Both statements can be true: rerunning a search may reproduce raw results, but it does not necessarily restore the analysts’ annotations, validation steps, file relationships, or record of how earlier conclusions were reached.
Why metadata can be evidence
Discovery evidence is not limited to the text inside a file. In a large technical investigation, the surrounding information may determine whether a result can be understood and authenticated.
- Underlying data: The files or datasets being searched.
- Search artifacts: Scripts, indexes, result files, logs, and analyst annotations.
- Metadata: File names, paths, timestamps, directory relationships, and other information that explains a result’s origin.
- Evidence of absence: A failed search does not necessarily prove that a work was absent, especially if the search process or index was compromised.
For the publishers, the issue was therefore not simply whether some bytes survived on a hard drive. They said the missing organization prevented them from reliably connecting search results to locations in OpenAI’s training data.
Was this legally spoliation?
It should not be described as proven spoliation. In general, spoliation concerns the destruction or alteration of information that should have been preserved. Courts typically consider questions such as:
- Whether the information was relevant to the litigation.
- Whether a preservation obligation applied.
- Whether the loss was intentional or negligent.
- Whether the opposing party suffered meaningful prejudice.
- Whether the loss can be cured through replacement searches or other remedies.
The cited November filings establish a dispute over lost search artifacts and metadata. They do not, by themselves, establish that OpenAI intentionally destroyed evidence or that a court imposed sanctions specifically for this incident.
The most accurate description is that the publishers characterized the erased material as potentially relevant discovery, while OpenAI denied destroying evidence and said the work could be recreated. The publishers themselves said they had no reason to believe the erasure was intentional.
Why the 2024 incident became part of a larger preservation story
Later proceedings raised separate questions about OpenAI’s preservation and production practices. These events add context, but they should not be collapsed into the November 2024 virtual-machine incident.
May 2025: preservation of output logs
On May 13, 2025, the court ordered OpenAI to preserve and segregate output-log data that otherwise would have been deleted. The order addressed future preservation obligations; it was not a ruling that the November 2024 virtual-machine event was intentional.
November 2025: Books1 and Books2
A November 24, 2025 opinion addressed a different episode: OpenAI had deleted the Books1 and Books2 datasets in 2022, before the litigation began. The opinion says OpenAI initially told opposing counsel that the datasets were deleted because they were not in use. It also describes later disputes over privilege and changing explanations.
The Tool Desk
Outbyte Driver Updater FREEScan for outdated or missing drivers - takes under a minuteDriver Scan →Outbyte PC Repair FREERepair Windows errors before they cause bigger problemsFix Now →The same opinion says copies of Books1 and Books2 were later recovered. This is separate from the 2024 erasure of search work and metadata.
Read the November 2025 opinion.
ChatGPT logs and deleted data
Later discovery proceedings addressed production of approximately 20 million chat logs, including how user-deleted data would be handled, how personally identifiable information would be redacted, and why litigation-related conversations were excluded. The court denied the publishers’ request for an evidentiary hearing without prejudice. That means the request was not granted at that stage, but it was not necessarily foreclosed permanently.
OpenAI has said that data covered by the preservation order was stored separately and was not automatically handed to The New York Times. Production was subject to court-supervised procedures rather than public release of ordinary users’ conversations.
Read the chat-log discovery order and OpenAI’s explanation of its data-retention position.
Recommended Free Tools
March 2026: larger log reservoirs
On March 9, 2026, another order required production of reservoirs containing 78 million and 10 million logs, subject to privacy-protective de-identification protocols. The court also found that an OpenAI witness had not been sufficiently prepared for a Rule 30(b)(6) deposition and that objections and answers had impeded the examination. The court reserved its decision on sanctions.
Best Value
Those logs were part of court-supervised discovery, not a public database available to ordinary ChatGPT users.
Independent reader supportYour contribution helps us test, update, and keep practical guides available for everyone.What remains unresolved?
- Whether the lost metadata materially affected the publishers’ ability to prove their claims.
- Whether all of the file contents were fully recoverable and usable.
- Whether a court imposed a remedy specifically for the November 2024 incident.
- What preservation and backup procedures were in place before the erasure.
- How privacy obligations should be balanced against discovery demands for user logs.
- How much of the later log production remains contested.
What this means for readers concerned about privacy
The dispute does not show that ordinary ChatGPT conversations were automatically exposed to The New York Times or made public. The later proceedings concerned preservation and production under litigation procedures, including segregation and de-identification requirements.
It does show why large technology litigation creates competing obligations. A company may have to preserve information for a legal hold while also protecting user privacy, controlling access, documenting deletion, and maintaining systems that can reproduce search results. Losing metadata can create a serious discovery problem even when the underlying files remain available.
Windows Errors? Fix Them Before They Spread
Repair common Windows errors and clear accumulated junk for a smoother, more stable PC - no reinstall needed.Free scan · no reinstallOutdated Drivers Are Slowing You Down
One free scan finds every outdated or missing driver and matches the right update for your exact hardware.Free scan · exact hardware matchBottom line
OpenAI’s November 14, 2024 system change erased potentially important search work product and organizational metadata from a virtual machine used in the publishers’ copyright discovery. The publishers said that loss forced them to redo work and made recovered results unreliable. OpenAI said a requested configuration change affected a temporary cache, that the underlying file contents were not shown to be lost, and that the searches could be rerun.
The cited public record supports calling the incident an accidental loss of potential discovery evidence—not proof that OpenAI deleted its training corpus, tampered with evidence, or was found liable for sanctionable spoliation. The later disputes over Books1 and Books2, output logs, and ChatGPT conversations are important context, but they are distinct episodes.
Quick Recap
Product prices and availability are accurate as of the date/time indicated and are subject to change. Any price and availability information displayed on Amazon at the time of purchase will apply.
