NYT Accuses OpenAI of Hiding Searchable Evidence in Copyright Fight

The New York Times and The Daily News are accusing OpenAI of concealing evidence in their copyright lawsuit, saying the company had already built internal tools to search chat logs and training data for copied journalism. The dispute matters because it could shape how courts evaluate whether AI companies trained on news content and whether they later reproduced it in ChatGPT responses.

In new filings tied to the two-year case, the publishers say OpenAI’s public claims about what it could search, what it preserved, and what it was able to produce in discovery were misleading at best. OpenAI denies wrongdoing and says the newspapers are trying to pry into private user conversations while the core legal claims against the company are weakening.

What the publishers are alleging

The Times and The Daily News are asking a federal judge to punish OpenAI for what they describe as discovery misconduct. Their central claim is straightforward: the company told the court it could not easily search its own materials for evidence of copying, but internally it had already done exactly that.

According to the plaintiffs, an April deposition of OpenAI privacy engineer Vinnie Monaco revealed that the company had performed searches and evaluations of its training data to look for copyrighted journalism. They say that testimony undermines OpenAI’s repeated argument that identifying such material would be technically impractical.

The publishers also contend that OpenAI had already accumulated a large internal archive of de-identified ChatGPT conversations and used it to assess how often the system might be reproducing or relying on protected works. They argue that this dataset became one of the most important sources of evidence in the case, yet OpenAI fought over how much of it to produce and in what form.

The plaintiffs’ lawyers argue that if OpenAI truly believed its use of their journalism was lawful, it would not have hidden evidence that could show otherwise.

Why the chat log dispute matters

The fight over chat logs is not just a procedural skirmish. It goes to the heart of how copyright cases against AI companies will be litigated going forward. If plaintiffs can show that a model outputs copied or closely paraphrased news stories, that may help establish harm, frequency and the scope of alleged infringement.

OpenAI has said it lacks a workable way to search its full corpus of training data and that searching user conversations creates privacy risks because the logs must be retrieved, processed and de-identified. The plaintiffs say those arguments were overstated because OpenAI had already created search tools for precisely that purpose.

That tension reflects a broader issue in AI litigation: companies often control the systems, records and logs most relevant to claims against them, while plaintiffs must persuade courts to force production of internal data that may be expensive, sensitive or heavily redacted.

How many chat logs are at issue?

OpenAI’s internal archive is at the center of one of the case’s biggest fights. The plaintiffs say the company had assembled roughly 78 million de-identified ChatGPT conversations for internal review well before the lawsuit was filed.

They also say the original demand was much larger: the newspapers had sought a sample of 120 million chat logs. After negotiation, that request was cut to 20 million. OpenAI eventually turned over the smaller sample in December, but the plaintiffs argue that the production was so heavily redacted that the records could not be meaningfully analyzed.

The companies’ clash over the sample matters because it may determine whether the court believes the produced logs are reliable enough to support OpenAI’s defense. If the sample is incomplete or altered, the plaintiffs want the judge to treat it as unreliable and to draw adverse conclusions from the missing material.

What is Project Giraffe and the Bloom filter?

Project Giraffe is the name the plaintiffs say OpenAI used for a set of internal tools that tracked output regurgitation. They say the company deployed a Bloom filter, a data structure commonly used to quickly check whether a piece of information might match a known item, to detect and log instances where ChatGPT appeared to repeat protected text.

The newspapers say that system was put in place shortly after the lawsuit began and that it recorded possible regurgitation in model outputs. If true, that would mean OpenAI had a mechanism to identify the very type of evidence the plaintiffs are seeking to prove their case.

The company has not publicly detailed Project Giraffe in the same way the plaintiffs describe it. But the allegation is significant because it suggests OpenAI may have had an internal pathway to measure repetition and potential copying long before it acknowledged such evidence in court.

Timeline of the evidence fight

What exactly do the newspapers want the judge to do?

The plaintiffs are not merely asking for more disclosure. They want the court to impose sanctions that would affect how OpenAI can defend itself at trial.

According to their request, the judge should:

• Block OpenAI from relying on the 20 million chat log sample as evidence.
• Accept as established that the logs would have shown substantial regurgitation and grounding of the newspapers’ content.
• Prevent OpenAI from arguing that its produced logs fail to prove meaningful copying.
• Order OpenAI to pay the plaintiffs’ legal fees tied to obtaining the evidence.

Those are severe requests. In civil litigation, asking a court to draw facts against the other side is a powerful remedy, one generally reserved for situations where a party is believed to have destroyed, hidden or manipulated evidence.

The publishers’ position is that OpenAI made it unnecessarily difficult to uncover information it had already gathered internally, forcing them to spend time and money chasing data that should have been disclosed in a more complete way.

How is OpenAI responding?

OpenAI is rejecting the allegations and framing the dispute as an attempt to invade users’ privacy. A company spokesperson said the Times is pursuing private conversations from people unrelated to the lawsuit, and argued that the paper’s claims are false.

OpenAI says the newspapers’ case is weakening and that the latest filings are an effort to reach into private user data while the company continues to defend fair use.

The company also maintains that its use of copyrighted material falls within the legal doctrine of fair use, a central defense in many generative AI lawsuits. Fair use arguments generally focus on whether the use of the material was transformative, the amount copied, the effect on the market and the purpose of the use.

That defense is far from settled in the context of large language models. Courts in multiple cases are still working through whether training on copyrighted works without permission is lawful and, if it is not, what damages or remedies should follow.

Why this case is bigger than one dispute

The lawsuit is part of a wider wave of copyright challenges against AI developers, including claims from authors, publishers, artists and media companies. News organizations are especially focused on how generative systems ingest their work and whether the systems reproduce it in ways that substitute for the original reporting.

For the media industry, the case is about more than compensation. It is about controlling how journalism is reused, whether AI products can summarize or quote reporting without permission, and what obligations AI companies have to preserve evidence once litigation begins.

For AI firms, the case highlights an equally important concern: the challenge of documenting how models are trained and how outputs are generated without exposing private data or proprietary information. Courts will increasingly have to decide how much transparency is required when the companies themselves control most of the evidence.

Why discovery fights are central in AI lawsuits

Discovery fights are central because AI systems are difficult to inspect from the outside. Unlike a standard software dispute, plaintiffs often cannot tell from public-facing outputs what training records were used, how often a model repeats protected text, or whether internal monitoring tools captured those events.

That makes internal databases, logs and engineering notes highly valuable. It also means a company’s document retention practices can become just as important as the underlying copyright issue.

In this case, the plaintiffs say OpenAI not only resisted disclosure but may also have deleted billions of ChatGPT outputs after suit was filed, even though a preservation order was in place. OpenAI has not publicly accepted that characterization, and the court has not yet ruled on the substance of those claims.

How the evidence dispute could affect the broader AI legal landscape

If the judge agrees with the newspapers, the decision could strengthen plaintiffs in future AI cases by showing that courts will penalize companies that overstate technical limitations or fail to preserve relevant data. That could lead to stricter preservation rules and more aggressive discovery demands in later suits.

If OpenAI prevails, the company could reinforce a defense strategy used by many AI developers: argue that internal logs are difficult to search, that privacy concerns limit disclosure, and that plaintiffs are seeking to turn ordinary user data into a litigation fishing expedition.

Either way, the outcome will likely influence how courts balance copyright enforcement against user privacy in the AI era. That balance is becoming one of the most consequential legal questions in technology.

What happens next?

The judge will decide whether the plaintiffs have shown enough to warrant sanctions or other corrective measures. That could include limiting what OpenAI can present at trial, requiring additional disclosures or allowing the court to infer that missing evidence would have favored the newspapers.

Even before any ruling, the latest filing raises the pressure on OpenAI by suggesting that the company had deeper internal visibility into copying than it previously acknowledged. It also gives the plaintiffs a stronger narrative: that the company was not unable to find evidence, but unwilling to turn it over.

For now, the case remains one of the most important legal tests of generative AI’s relationship with news content. Its next phase may determine not only what OpenAI did, but how much a court is willing to trust the company’s account of it.

Key facts at a glance

The legal battle now hinges on a familiar but difficult question: when AI companies say they cannot produce evidence, what happens if internal testimony suggests they already found it?