News publishers suing OpenAI over the development and operation of ChatGPT have opened a new front in the closely watched copyright litigation, accusing the artificial-intelligence company of concealing evidence that could show whether its systems reproduced protected journalism.

The New York Times, the Daily News and other media organizations filed a motion on July 9 asking a federal judge in Manhattan to impose sanctions for what they characterize as sustained discovery misconduct. According to reports on the filing, the publishers contend that OpenAI made inaccurate representations about its ability to search training datasets and ChatGPT conversation records, while withholding information about internal tools that were already being used to investigate model outputs.

The request does not determine whether OpenAI infringed the publishers’ copyrights. Instead, it asks the court to address how evidence was preserved, identified and produced during the lawsuit. That procedural distinction is important because sanctions can alter the shape of a case before a court reaches the underlying questions of copying, fair use, market harm or damages.

The publishers are seeking several forms of relief. They want the court to prevent OpenAI from relying on a disputed sample of 20 million ChatGPT conversations, arguing that the production is unreliable. They are also asking for findings or evidentiary restrictions that would limit OpenAI’s ability to argue that the conversation records do not demonstrate substantial reproduction or grounding in the publishers’ content. Their motion additionally seeks reimbursement for legal expenses associated with pursuing material they say should have been disclosed earlier.

Such remedies would be significant, but they are not automatic. The judge must assess the specific discovery history, OpenAI’s explanations, the relevance of the disputed systems and records, and whether any failures were negligent, disproportionate or intentional. The court could deny the motion, order narrower corrective discovery, award fees, exclude evidence or recommend that certain disputed propositions be treated as established.

The latest allegations center in part on testimony by John Vincent “Vinnie” Monaco, an OpenAI data-privacy engineer who appeared as a corporate witness on issues involving model regurgitation and an internal initiative known as Project Giraffe. The publishers say an April deposition revealed that OpenAI had capabilities and datasets that contradicted earlier claims about the technical difficulty of locating copyrighted works or searching large collections of model interactions.

The court had already expressed concern about Monaco’s preparation during an earlier stage of the discovery process. In an April 7 opinion, U.S. Magistrate Judge Ona T. Wang found that Monaco had not been adequately prepared for a corporate deposition and that a pattern of objections and incomplete answers had impeded the examination. The court authorized additional deposition time and deferred a decision on further sanctions or discovery remedies until that process was completed.

That earlier order did not establish that OpenAI had destroyed or deliberately hidden evidence. It did, however, create a documented procedural backdrop for the publishers’ latest request. The July motion argues that the subsequent testimony provided information that should have been available earlier and that the deficiencies extended beyond an underprepared witness.

According to the publishers’ account, OpenAI had conducted internal searches of training materials for copyrighted journalism despite previously emphasizing the difficulty of searching its training corpus. They also allege that the company assembled a database containing about 78 million de-identified ChatGPT conversations before The Times filed its lawsuit.

The purpose and construction of that database could become central to the sanctions dispute. If it was used to measure or investigate the reproduction of protected works, the publishers are likely to argue that it represents evidence directly relevant to their claims. OpenAI may respond that internal safety, quality or privacy evaluations are technically and legally different from the broad discovery searches requested by opposing litigants.

The publishers also point to Project Giraffe and an associated Bloom filter, which they describe as part of a system for detecting and recording possible regurgitation in model outputs. A Bloom filter is a probabilistic data structure that can be used to test whether an item may belong to a larger dataset. In an AI-content context, such a mechanism could help identify matches or potential overlap without performing a conventional search across every underlying record.

The existence of such a tool does not by itself prove copyright infringement. AI developers routinely test models for memorization, unsafe output, data leakage and benchmark contamination. The legal importance will depend on what the tool was designed to detect, which datasets it examined, what results were retained and whether those materials were responsive to the publishers’ discovery requests.

The publishers argue that OpenAI’s internal monitoring capabilities undercut its position that locating relevant evidence was prohibitively difficult. Their motion portrays the dispute as one in which the company already possessed organized information about possible reproduction but required the plaintiffs to pursue prolonged and expensive discovery before learning that the systems existed.

OpenAI has denied the allegations. A company spokesperson said the publishers were making false claims while continuing efforts to access private conversations belonging to users who have no connection to the litigation. OpenAI said it would continue defending user privacy and its position that training generative models on publicly available material is protected by established principles of fair use.

Privacy has been a recurring part of OpenAI’s response to demands for ChatGPT records. The company has argued that large-scale production can expose sensitive personal or commercial information even when records are de-identified. It has also maintained that most conversations in a broad sample are unlikely to have any connection to the publishers’ copyrighted works.

The publishers counter that properly protected and anonymized records are necessary to measure how frequently users received outputs based on their journalism. Conversation logs could potentially show whether reproduction occurred only under unusual, deliberately engineered prompts or arose through ordinary use. That distinction may affect arguments about the practical market impact of ChatGPT and whether the system functions as a substitute for original publications.

The production of 20 million conversations followed extensive negotiations over the scope of discovery. The publishers initially sought a substantially larger sample before the request was narrowed. OpenAI ultimately produced the 20 million-record set under court supervision, but the publishers now contend that extensive redactions, substitutions or other limitations made the data unsuitable for reliable analysis.

They further allege that OpenAI deleted billions of outputs after the litigation began and replaced millions of conversations within the requested sample. Those claims are among the most serious in the motion because parties generally have a duty to preserve potentially relevant evidence once litigation is reasonably anticipated. Whether the records were subject to a preservation obligation, deleted under ordinary retention policies or recoverable from other systems will be closely examined.

OpenAI has previously disputed allegations that it improperly destroyed material and has framed its deletion practices as part of standard privacy commitments. The company has said that deleted consumer chats are ordinarily scheduled for removal unless a legal or security requirement applies. Litigation holds can complicate those commitments by forcing technology providers to retain material that users expected to be removed.

The clash therefore places two policy concerns in direct tension. Publishers argue that AI companies cannot invoke privacy or engineering complexity to prevent access to evidence about commercial systems built from copyrighted material. OpenAI argues that copyright plaintiffs should not receive sweeping access to private conversations merely because a small portion might contain relevant examples.

The underlying lawsuit began in December 2023, when The New York Times accused OpenAI and Microsoft of using millions of its articles to develop AI products that compete with the newspaper as sources of information. Other publishers later brought or consolidated related claims, widening the dispute beyond a single newspaper and turning the litigation into a major test of how copyright law applies to generative AI.

The plaintiffs’ theory goes beyond the ingestion of articles during training. They allege that ChatGPT can generate outputs that reproduce protected expression, summarize paywalled reporting or provide substitutes for the original works. They also contend that AI-generated answers can reduce referral traffic and weaken the economic incentives supporting professional journalism.

OpenAI maintains that model training is transformative and protected by fair use. It argues that models learn statistical relationships from large collections of data rather than functioning as databases that retrieve and distribute complete articles. The company has also said that plaintiffs have relied on artificial prompts designed to force reproduction instead of demonstrating typical product behavior.

The distinction between training and output remains one of the most important unresolved issues in AI copyright law. A court could treat the analysis of training inputs differently from specific instances in which a model produces text closely resembling a protected work. Evidence about internal regurgitation testing, user interactions and mitigation systems could therefore matter even if the broader training process is ultimately found to be lawful.

For the technology sector, the sanctions motion highlights the expanding operational cost of copyright litigation. AI developers are not only defending their legal theories about fair use; they must also document how datasets were assembled, how models were evaluated, what logs were retained and how internal experiments relate to claims made in court.

Companies with proprietary training pipelines may face growing demands to create auditable records without revealing trade secrets or exposing user data. Internal teams responsible for model safety, privacy, evaluation and legal compliance may need more integrated retention policies so that tools created for one purpose can be located and reviewed when litigation begins.

The controversy could also affect negotiations between AI companies and publishers. A number of media organizations have opted for licensing partnerships that allow their material to be incorporated into AI products under commercial agreements. Others have pursued litigation, arguing that voluntary deals cannot replace compensation for earlier or unauthorized use.

If the publishers obtain strong sanctions, their bargaining position could improve by increasing the legal and evidentiary risks faced by model developers. If OpenAI defeats the motion and demonstrates that the disputed information was properly handled, the decision could reinforce limits on how far copyright plaintiffs may go in demanding customer data and internal AI telemetry.

The immediate question is narrower than the broader future of generative AI: whether OpenAI met its obligations in this particular litigation. The judge will have to separate allegations about the merits of the copyright claims from claims about the integrity of the discovery process. Evidence that a model may have reproduced journalism does not automatically establish discovery misconduct, just as procedural misconduct would not, by itself, resolve every element of infringement.

Still, discovery disputes can have substantial commercial consequences. Adverse findings may increase settlement pressure, complicate OpenAI’s fair-use defense and encourage similar motions in other copyright cases. A rejection of the publishers’ claims could strengthen OpenAI’s argument that broad requests for user conversations create privacy risks without producing proportionate evidence.

The sanctions motion is now pending, and OpenAI will have an opportunity to formally respond. Until the court evaluates the competing accounts, the publishers’ assertions regarding hidden datasets, deleted outputs and substituted logs remain allegations rather than judicial findings. The ruling will nevertheless be closely watched across the AI, media and enterprise-software industries as companies reassess how model behavior, copyrighted content and user data must be documented when commercial AI systems enter the courtroom.