OpenAI told the court it couldn't search its own training data. It was a central pillar of their defense. Now, The New York Times and The Daily News claim that defense was a lie.

In a court filing this week, the publishers alleged that OpenAI possessed the technical capability to track copyrighted content all along. They argue the company actively concealed internal tools designed to monitor how often ChatGPT regurgitates protected journalism. The revelation, if proven, could fundamentally alter the trajectory of a two-year legal battle over AI training practices.

The Smoking Gun in the Deposition

The allegations center on a deposition from Vinnie Monaco, an OpenAI data privacy engineer. During his testimony, Monaco reportedly revealed that the company had already conducted internal evaluations of its training corpus to identify copyrighted works.

Even more damaging is the claim that OpenAI had quietly amassed a database of 78 million de-identified ChatGPT conversations. This dataset was allegedly used to assess the company’s own infringement risks long before the lawsuit even began.

There was more. The plaintiffs claim OpenAI implemented a "Bloom" filter under a project codenamed "Giraffe." This tool was designed to detect and record instances where the model spit out verbatim copyrighted text. It was active shortly after the litigation started. OpenAI had the data. They just didn't want to share it.

A Pattern of Obfuscation

The discovery process has been contentious from the start. Originally, the plaintiffs requested a sample of 120 million chat logs. OpenAI negotiated that number down to 20 million, citing privacy concerns and technical burdens.

When the data finally arrived last December, it was effectively useless. The publishers claim the sample was heavily redacted and, worse, that millions of logs had been substituted. They also allege that OpenAI deleted billions of outputs after the lawsuit was filed, potentially violating a court-ordered preservation mandate.

"If OpenAI genuinely believed that copying our clients’ journalism was fair and legal, it wouldn’t have hid the truth about having done it," said Ian B. Crosby, lead counsel for the plaintiffs.

What This Means for OpenAI

The publishers are now asking the judge to drop the hammer. They want the court to exclude the 20 million chat log sample from evidence entirely, arguing it is unreliable. They are also seeking a ruling that would force the court to accept as fact that ChatGPT logs contain significant regurgitation of their content.

OpenAI is fighting back. Spokesperson Drew Pusateri dismissed the claims as "blatantly false" and accused the Times of attempting to invade user privacy as their legal position weakens. "We’ll continue defending our users’ privacy and the long-established principles of fair use," Pusateri said in a statement.

Key Takeaways

  • Internal Tracking: The NYT alleges OpenAI used a tool called "Project Giraffe" to track how often ChatGPT regurgitated copyrighted journalism.
  • Hidden Datasets: Plaintiffs claim OpenAI held a 78-million-conversation database that it used to monitor its own infringement risks.
  • Discovery Sanctions: The publishers are asking the court to punish OpenAI for allegedly redacting and substituting evidence, and for deleting logs after the suit began.

The court must now decide whether these discovery failures warrant severe sanctions. If the judge agrees that OpenAI intentionally misled the court, the company’s "fair use" defense could be severely compromised. The next hearing will determine if the publishers get access to the internal data they have been chasing for months. The stakes are high. The outcome will set a precedent for every AI firm currently facing copyright claims.