The lie that anchored the defense

OpenAI told a federal court it could not search its own training data. Its engineers swore under oath that combing the massive corpus for copyrighted journalism was technically infeasible. That claim shaped two years of litigation. It let OpenAI stonewall the New York Times and The Daily News when they asked for proof of what the models had ingested and regurgitated.

Then a deposition cracked the story open.

Vinnie Monaco, an OpenAI data privacy engineer, testified in April that the company had already run internal searches of its training corpus for copyrighted journalism works. The searches existed. The capability existed. The court heard that OpenAI had not just the ability but the actual practice of auditing its own ingestion.

A secret audit, a hidden database

The revelation goes deeper. Before the Times filed suit, OpenAI had already assembled a database of roughly 78 million de-identified ChatGPT conversations. It used that trove internally to measure how much its models were infringing on others' work. After the complaint landed, OpenAI layered a "Bloom" filter — part of a toolset dubbed "Project Giraffe" — onto the pipeline to detect and log regurgitation in outputs.

Think about that sequence. The company built a private infringement dashboard while telling the court and the public that such measurement was impossible. It negotiated the plaintiffs' discovery request from 120 million conversations down to 20 million. Then it delivered a sample so heavily redacted the court itself called it "unusable."

The destruction that followed

The plaintiffs allege something worse. After the lawsuit was filed, OpenAI deleted billions of ChatGPT outputs in direct violation of a court preservation order. It then substituted millions of logs in the very sample it had agreed to produce. The result: a discovery record that cannot be trusted, curated by a party that had every incentive to scrub it clean.

This is not a dispute over technical nuance. It is a pattern of concealment. OpenAI claimed privacy burdens made production impossible. Its own engineers had already solved the de-identification problem at scale. It claimed search was infeasible. Its own tools had already indexed the corpus for copyright hits. It claimed good faith. Its own preservation violations suggest the opposite.

What the plaintiffs now demand

The Times and The Daily News are asking the judge to treat the 20-million-log sample as inadmissible. They want the court to accept as established fact that ChatGPT logs would have shown major regurgitation and grounding of their journalism. They want sanctions for spoliation.

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

The credibility gap

OpenAI's public posture rests on a foundation of responsible innovation. Its lobbyists argue that training on publicly available text is fair use. Its researchers publish papers on alignment and safety. Its executives testify before Congress about transparency.

None of that survives the deposition record. A company that secretly measures its own infringement while denying the capacity to measure it has forfeited the benefit of the doubt. A company that deletes evidence under a preservation order has forfeited the right to call its opponents litigious.

The stakes extend beyond one case

Every major AI developer trains on vast scrapes of human knowledge. Every one of them faces the same copyright exposure. The OpenAI discovery fight will set the template for how those battles play out. If courts tolerate hidden databases, negotiated-down samples, and post-filing deletions, the discovery process becomes a theater where the best-resourced defendant writes the script.

The Times and The Daily News are not asking for a novel legal theory. They are asking for the ordinary machinery of civil litigation to function: preserve evidence, produce what exists, stop lying about what you can do. OpenAI's alleged conduct suggests it fears what a fair process would reveal.

A moment of reckoning

The judge now decides whether the 20-million-log sample stays in the record. The judge decides whether spoliation sanctions land. The judge decides whether a company that built a private infringement dashboard gets to keep pretending it cannot see what it has already measured.

That decision will echo in every newsroom, every publisher, every creator who has watched their work disappear into a model's weights without consent or compensation. The technology is impressive. The conduct, as alleged, is indefensible.