The Tool They Said Didn't Exist: Publishers Ask a Court to Sanction OpenAI

A discovery fight that stopped being about discovery

For two and a half years, the copyright war between the news industry and OpenAI has been fought on the terrain everyone expected: fair use. Did training a model on millions of articles transform them into something new, or did it simply copy them at industrial scale? That is the question the industry has been waiting for a court to answer.

On July 9, the plaintiffs stopped arguing about fair use and started arguing about honesty. The New York Times, the New York Daily News, the Chicago Tribune, the Santa Rosa Press Democrat, the Authors Guild and a group of bestselling writers filed a motion for sanctions in the consolidated In re OpenAI Copyright Infringement Litigation in Manhattan federal court, asking Magistrate Judge Ona Wang to punish OpenAI for what they describe as a sustained campaign of misrepresentation and evidence destruction. The core claim, as reported by TechCrunch, is blunt: OpenAI repeatedly told the court it lacked the technical ability to search its own training data and output logs for the publishers' articles — while, internally, it had already built tools to do exactly that and had already used them.

This is a category shift. A fair-use ruling is a legal judgment about what the law permits. A sanctions ruling is a judgment about whether a party can be trusted, and courts have far blunter instruments for the second problem than the first.

What the motion alleges

The pivot point is a deposition. According to the Associated Press, John Vincent "Vinnie" Monaco, who works on privacy engineering at OpenAI, was designated as the company's witness on these questions. Judge Wang found in a January 2026 order that he had come to his first deposition unprepared and ordered him back. It was at the second sitting, in April 2026, that the plaintiffs say the picture finally came apart: Monaco allegedly acknowledged that OpenAI had in fact searched its training corpus for copyrighted journalism, had built searchable datasets of de-identified ChatGPT conversations, and had run searches for publishers' content.

Some of those searches, the publishers say, predate the Times' original complaint in December 2023. TechCrunch reports the motion describes an internal database of roughly 78 million de-identified ChatGPT conversations assembled before the suit was filed. If that holds up, it means the capability OpenAI told the court did not exist was not merely theoretical — it was already in production, and had been for years.

OpenAI rejects the framing. Spokesperson Drew Pusateri said the Times is "persisting with their efforts to invade the privacy of people who have nothing to do with this case, including by making these blatantly false allegations," and said the company will keep "defending our users' privacy and the long-established principles of fair use." Nothing in the motion has been ruled on. These are allegations by an adversary in a high-stakes case, and OpenAI has not yet filed its full response.

Project Giraffe and the Bloom filter

The most technically interesting allegation is an internal effort the motion calls Project Giraffe. Per TechCrunch's account of the filing, shortly after the Times sued, OpenAI deployed a Bloom filter — a probabilistic data structure that answers "have I seen this before?" cheaply and at scale — to detect when ChatGPT's outputs reproduced copyrighted text.

Think about what that implies. A Bloom filter for regurgitation is, functionally, a regurgitation detector. To build one you must have a corpus of protected text to check against, a pipeline that inspects outputs, and a mechanism that logs or suppresses hits. Every one of those is the thing OpenAI allegedly told the court it could not do. The publishers' argument is not that building such a filter was wrong — arguably it was the responsible engineering move — but that having built it, the company could not credibly claim to be blind.

There is a real and uncomfortable tension here that the industry should sit with rather than cheer past. The same infrastructure that lets a lab suppress verbatim reproduction is the infrastructure that lets a plaintiff prove verbatim reproduction happened. A lab that builds no such tooling has less to hide and less to fix. A lab that builds it becomes, in effect, its own best witness. That is a genuinely perverse incentive, and it is now sitting in the middle of the most important AI copyright case in the United States.

The preservation order and the deleted logs

The second half of the motion is the more legally dangerous one: spoliation. The publishers allege that OpenAI deleted chat data — the AP reports the filing describes millions of chat histories and billions of ChatGPT responses — after a court preservation order was in force. Preservation orders are the load-bearing wall of American civil discovery. Violating one knowingly is the kind of thing judges take personally.

The discovery record around the surviving logs is itself contested. TechCrunch reports plaintiffs originally sought some 120 million chat logs, negotiated down to a 20 million-log sample, and that when OpenAI produced it in December 2025 it arrived so heavily redacted as to be of limited use. The motion's language is unambiguous about intent: "Serious sanctions are especially appropriate here because OpenAI's conduct, including the violation of the Court's preservation Order, was knowing and intentional."

The remedies requested track that theory. The publishers want OpenAI barred from relying on the 20-million-log sample; they want the court to treat as established fact that the logs would have shown substantial and systematic reproduction of their work; they want OpenAI prevented from arguing otherwise; they want adverse jury instructions about the destroyed data; and they want their fees. That last cluster is what lawyers call an issue sanction, and it is close to a case-ending weapon. It would hand the plaintiffs, by judicial fiat, the factual finding they would otherwise have to prove at trial.

Hype versus what is actually established

It is worth being precise about what is and is not known. What is known: a sanctions motion was filed, it is public, it names Project Giraffe and Monaco's depositions, and it seeks issue-preclusive relief. What is not known: whether any of it is true. OpenAI calls the allegations blatantly false. Judge Wang has not ruled. There is no finding of spoliation, no finding of misrepresentation, and no sanction on the docket.

It is also worth resisting the maximalist read circulating online — that this ends OpenAI. It does not. Even a granted motion would not resolve fair use, which remains the dispositive legal question and on which OpenAI has real arguments. What sanctions would do is change the board: they would strip OpenAI of its ability to contest one of the factual predicates of infringement, and they would poison the well for every credibility judgment a jury makes afterward.

The broader signal matters more than this docket. Discovery in AI litigation has arrived at the model internals. Plaintiffs now know enough to ask for the training corpus index, the output logs, the dedup pipeline, the safety filters — and to notice when a company says those things do not exist. Every frontier lab with a copyright case pending should be reading this filing as a preview of its own next deposition.

The takeaway

The most consequential AI story of the week is not a model launch. It is a procedural motion, and its subject is not what OpenAI's model did but what OpenAI's lawyers said. If the publishers are right, the company spent two years telling a federal court it could not see inside its own system while running the searches internally. If OpenAI is right, this is a weakening plaintiff reaching for a privacy-invading Hail Mary.

Either way, a line has been crossed. The fair-use question — the one the entire industry has been waiting on — may now be decided in the shadow of a conduct fight, by a judge who has already once found OpenAI's designated witness unprepared. Companies that built their legal strategy on the premise that model internals are unknowable are discovering that courts are perfectly capable of asking who, exactly, was doing the knowing.

Sources: TechCrunch · AP via The Spokesman-Review · AP via Press Democrat