On 9 July 2026 the New York Times, the New York Daily News and other news plaintiffs asked a Manhattan federal court to impose serious sanctions on OpenAI, alleging that the company concealed for more than two years that it could search its own training data and output logs.
This is not a fair-use argument. It is a discovery-conduct argument, and those are decided on a different standard and can end a case without anyone reaching the copyright question at all.
What OpenAI said it could not do
Throughout the case OpenAI's position was that it lacked the ability to search its own training corpus, and that searching or producing its collection of ChatGPT conversations would be technically burdensome and would raise user-privacy concerns.
That position shaped years of discovery. It is the reason the plaintiffs' requests were narrowed, and the reason the fight over chat logs took the shape it did.
What the deposition produced
In a court-ordered deposition in April 2026, OpenAI data privacy engineer Vinnie Monaco testified — per the plaintiffs' motion — that OpenAI had already run internal searches and evaluations of its training corpus looking for copyrighted journalism.
The motion describes two things in particular:
A database of roughly 78 million de-identified ChatGPT conversations, maintained internally and used to assess how much the company was infringing others' works.
"Project Giraffe" — a set of tools including a Bloom filter built to detect and keep a record of regurgitation in model outputs, implemented shortly after the lawsuit was filed.
A Bloom filter is an ordinary data structure — a compact way to test whether an item is probably in a set. There is nothing sinister about the technique. The allegation is about timing and disclosure: a tool built to answer the exact question in the litigation, stood up after the suit was filed, while the company was telling the court it could not answer that question.
The chat log fight
The sequence the plaintiffs describe:
| Step | What happened |
|---|---|
| Request | 120 million chat logs |
| Negotiated down to | 20 million |
| Sample delivered (December) | Redacted so heavily the court deemed it "unusable" |
| Also alleged | Billions of outputs deleted after the suit was filed, and millions of logs substituted in the requested sample |
The deletion and substitution allegations are the serious ones. Spoliation — destroying evidence you are on notice to preserve — is what produces the harshest sanctions available.
What they are asking for
The motion asks the court to:
- Preclude OpenAI from using the 20 million-log sample as evidence
- Deem it established that the logs would have shown substantial regurgitation
- Bar OpenAI from arguing the sample does not show major regurgitation
- Order OpenAI to pay the plaintiffs' legal fees
Read the middle two together. That is an adverse inference request: a ruling that the destroyed or withheld evidence would have proved the plaintiffs' case, so the court should proceed as though it did.
If granted, the central factual question — does ChatGPT reproduce Times journalism at scale — stops being contested. OpenAI would be barred from arguing the negative, and would have to win on fair use with the fact of substantial regurgitation already established against it.
That is an enormous ask, and courts grant it rarely. But it is the ask.
OpenAI's response
Spokesperson Drew Pusateri:
As the Times' case weakens and they've been forced to drop claims against us, they're persisting with their efforts to invade the privacy of people who have nothing to do with this case.
The company says it is defending its users' privacy and long-established principles of fair use.
Note what that answer does and does not do. It contests the plaintiffs' motive and frames the dispute as a privacy question. It does not, in the quote on record, address whether the search tools existed or whether the company's earlier representations were accurate.
Why this matters beyond one case
The pattern across AI copyright litigation so far is that conduct loses cases, not concept.
In Bartz v. Anthropic, Judge Alsup held that training on lawfully acquired books was fair use — and that keeping pirated copies in a central library was not. Only the piracy class proceeded, and it settled for $1.5 billion. The model training was fine. The acquisition was not.
The same shape is visible in the music-industry case against Suno, where the labels' strongest theory is not that training is infringement but that the recordings were obtained by circumventing YouTube's protections.
A discovery-misconduct finding would be the third variant of the same lesson: what decides these cases is how the defendant behaved, not whether machine learning is transformative.
What is not established
Being precise, because this is a motion and not a ruling:
- These are allegations. The characterisation of Monaco's testimony is the plaintiffs'. We have not seen the deposition transcript.
- No ruling has issued. No court has found spoliation, concealment or bad faith.
- The reporting we relied on does not name the presiding judge or give dates beyond the April deposition and the December sample.
- "Deleted billions of outputs" is the plaintiffs' characterisation. Routine log rotation and deliberate destruction look different in a technical record and identical in a press summary.
What to watch
- Whether the court orders the deposition transcript unsealed. That converts allegation into record.
- Whether any adverse-inference relief is granted, even partially. A narrow version — say, precluding the 20m sample — would still be significant.
- Whether OpenAI's earlier representations get characterised by the court, as opposed to by the plaintiffs.
- Whether other AI defendants' burden objections get re-examined. "We cannot search our training data" has been a common position across this litigation. If it fails once on the facts, it becomes harder to assert elsewhere.
