On July 9, 2026, The New York Times, the New York Daily News, and more than a dozen other publishers — including the Center for Investigative Reporting, The Intercept, Ziff Davis, and the Santa Rosa Press Democrat — filed a 52-page sanctions motion in the Southern District of New York accusing OpenAI of misleading the court for over two years about its ability to search ChatGPT's training data and output logs for their copyrighted journalism (TechCrunch, MediaNama). The filing sits inside the consolidated copyright litigation built around the Times' December 2023 suit against OpenAI and Microsoft — a case that survived a motion to dismiss in an April 4, 2025 SDNY opinion allowing direct and contributory infringement claims to proceed (SDNY opinion).
The core allegation isn't that OpenAI trained on copyrighted news — that dispute predates this motion by two years. It's that OpenAI told the court, repeatedly, that searching its systems for the publishers' content was infeasible, burdensome, and would invade user privacy, while internally it had already run exactly those searches. According to the motion, a second deposition of OpenAI privacy engineer Vinnie Monaco in April 2026 revealed the company had assembled a database of roughly 78 million de-identified ChatGPT conversations to gauge its own infringement exposure, and had built a "Bloom filter" inside a tool called "Project Giraffe" to detect and log regurgitated content — deployed shortly after the Times sued (TechCrunch, TheWrap). Publishers also allege OpenAI deleted or compressed billions of ChatGPT conversations into an unsearchable state, undermining a court preservation order.
The Steelman: Discovery Misconduct at This Scale Deserves the Court's Hardest Tools
Publishers aren't reaching for an unusual remedy. Rule 37 of the Federal Rules of Civil Procedure and a court's inherent authority both permit exactly what this motion asks for: barring a party from relying on evidence it obtained through gamesmanship, instructing a jury to presume that withheld evidence was unfavorable, and shifting fees onto the party that misled the court. If OpenAI told a federal judge its systems couldn't do something it had already done — and had built internal tooling to do at scale — that isn't a gray area. Litigants who misrepresent the existence of searchable evidence undermine discovery for every case, not just this one, and the information asymmetry between a foundation-model company and outside plaintiffs makes candor obligations especially important: publishers have no independent way to verify what OpenAI's infrastructure can or cannot do. The remedies sought — barring OpenAI from selectively relying on the 20-million-conversation sample it fought to shrink, plus adverse-inference instructions and fee-shifting — are calibrated to the alleged conduct, not punitive overreach.
Why This Doesn't Need a New AI Statute
None of that argues for the sweeping AI-specific transparency mandates that have circulated in Congress. The Generative AI Copyright Disclosure Act (H.R. 7913, 118th Congress), introduced by Rep. Adam Schiff in April 2024, would have required every developer of a training dataset to file a public notice with the Register of Copyrights describing the copyrighted works used, backed by civil penalties starting at $5,000 per violation (H.R. 7913 text). It never made it out of committee, and this case is a useful illustration of why that gap hasn't been the disaster transparency advocates feared. The sanctions motion publishers filed on July 9 is doing, through ordinary civil procedure, much of what a disclosure statute was built to do — forcing an AI company to reveal what it actually knows about its own training data and outputs — without Congress standing up a registration bureaucracy inside the Copyright Office that would apply the same fixed compliance burden to a frontier lab and a three-person startup alike. A blanket disclosure mandate bites regardless of a company's conduct; targeted sanctions bite only when a company actually misleads a court. That is the more proportionate instrument, and it already exists.
The broader discovery fight underscores the point. Publishers originally sought 120 million ChatGPT logs; OpenAI countered with 20 million out of the "tens of billions" of logs it had preserved. Magistrate Judge Ona Wang's order compelling that production, affirmed by Judge Sidney Stein on January 5, 2026, built in real privacy safeguards — a sample cut down by orders of magnitude, mandatory de-identification, and a protective order restricting how the data can be used (discovery-order coverage). That is a workable middle ground between plaintiffs' evidentiary needs and user privacy, reached without any new statute. What broke down, on the publishers' telling, wasn't the discovery framework — it was OpenAI's candor within it.
What to Watch
OpenAI has denied the allegations, with a spokesperson calling them "blatantly false" and accusing the Times of trying "to invade the privacy of people who have nothing to do with this case" (TheWrap). The court hasn't ruled on the sanctions motion, and OpenAI will get a full opportunity to respond in briefing before Judge Stein decides anything. If the allegations hold up, the appropriate response is exactly what's already in front of the court: case-specific sanctions scaled to the misconduct, not a legislative overhaul. If they don't, the existing process will have tested the claims and rejected them — which is also the system working as intended.
Either outcome argues against grafting a new AI-transparency statute onto a discovery process that is, in this instance, already doing its job.