A fresh escalation in the long-running fight over whether ChatGPT can be trained and used without infringing copyright is now turning on a question that sits adjacent to the core legal theory: what evidence was disclosed, when, and in what form.
According to reporting tied to a new filing, news publishers—including the New York Times—are asking a court to impose sanctions on OpenAI. The allegation is not simply that the model may have produced outputs that resemble copyrighted journalism. Instead, the publishers claim OpenAI withheld tools and datasets that could help identify whether copyrighted reporting appears in ChatGPT’s responses, and that those materials would be relevant to how infringement is evaluated in the case.
The motion for sanctions signals that the dispute is moving beyond the familiar arguments about training data, output similarity, and fair use. It is increasingly about litigation process itself: transparency, documentation, and the obligations parties have to preserve and produce evidence. In other words, the fight is becoming as much about discovery and disclosure as it is about copyright doctrine.
What the publishers say was hidden
At the center of the new motion is the claim that OpenAI concealed specific categories of evidence—described as tools and datasets—that could be used to determine whether copyrighted journalism is present in ChatGPT outputs. The publishers argue that these materials would not be “nice to have” but directly relevant to the factual questions the court must resolve.
In copyright cases involving generative systems, the evidentiary challenge is unusually complex. Unlike a traditional infringement dispute where a plaintiff can point to a specific copied passage and show access and substantial similarity, generative AI introduces layers of uncertainty: the model’s internal representations, the probabilistic nature of text generation, and the fact that outputs can be paraphrased, summarized, or stitched together in ways that make direct copying harder to detect.
That complexity is precisely why publishers are emphasizing tools and datasets that can support systematic evaluation. If a party has methods for identifying overlaps between outputs and copyrighted works—or methods for mapping which sources influenced particular generations—those methods can become pivotal. The publishers’ argument, as described in the reporting, is that OpenAI had such capabilities and did not fully provide them during the discovery process.
The unique pressure point here is that the withheld materials are framed as being able to help “identify copyrighted journalism in ChatGPT outputs.” That phrasing matters. It suggests the dispute is not only about whether OpenAI’s training data included copyrighted material, but also about whether the model’s behavior can be tested in a way that meaningfully distinguishes lawful transformation from unlawful reproduction.
Why sanctions are a big deal
Sanctions motions are not routine. They typically come into play when one side believes the other has failed to comply with discovery obligations, violated court orders, or engaged in conduct that undermines the integrity of the process. In high-stakes technology litigation, sanctions are often the mechanism plaintiffs use to argue that the court should correct an imbalance created by non-disclosure.
If the court agrees that evidence was improperly withheld, the consequences can range from orders compelling production to more severe remedies, depending on the jurisdiction and the severity of the alleged misconduct. Even when sanctions do not result in the most dramatic outcomes, the mere filing can change the tone of the case. It can lead to tighter scrutiny of what each side claims it can prove, and it can force the court to confront whether the record is complete enough to decide the substantive copyright issues.
There is also a strategic dimension. Publishers are not just asking for a ruling on infringement; they are asking the court to treat the discovery dispute as part of the merits. That can influence how judges view credibility and compliance, and it can affect what evidence becomes admissible or persuasive later.
The broader context: generative AI and the “black box” problem
The publishers’ allegations fit into a larger pattern that has emerged across generative AI litigation: courts and plaintiffs are struggling with the “black box” nature of modern models. OpenAI and other developers often argue that they cannot disclose certain internal details due to trade secrets, security concerns, or the sheer difficulty of translating proprietary systems into courtroom-ready evidence.
But plaintiffs counter that without meaningful access to evaluation methods, they cannot test claims about how the model behaves. In practice, this creates a tension between two competing needs: protecting proprietary technology and ensuring fairness in litigation.
The motion for sanctions, as described, is essentially a claim that OpenAI crossed a line from legitimate confidentiality into improper concealment. The publishers are arguing that the withheld tools and datasets were not merely internal secrets, but evidence that could help answer the question the court is being asked to decide.
This is where the case becomes especially consequential for the industry. If courts accept that certain evaluation tools are required for fair adjudication, it could reshape how AI companies prepare for litigation. It could also influence how they design their internal compliance and documentation processes—because the ability to produce evidence quickly and transparently may become as important as the underlying technical performance.
How infringement evaluation works in AI cases
To understand why tools and datasets matter so much, it helps to look at how infringement is evaluated when the alleged copying is not a direct quote.
In many copyright disputes, plaintiffs rely on comparisons: identifying passages that match or closely track protected expression. With generative AI, however, the output may not reproduce the original text verbatim. It might summarize, paraphrase, or blend multiple sources. That makes “substantial similarity” harder to demonstrate using simple side-by-side comparisons.
As a result, plaintiffs often seek computational methods that can quantify similarity, detect overlap, or trace whether outputs are derived from specific works. These methods can include:
1) Similarity detection approaches that compare generated text to known copyrighted works.
2) Attribution or retrieval-style analyses that attempt to determine whether a model’s output resembles content from particular sources.
3) Evaluation datasets designed to test whether the model reproduces protected material under different prompts and conditions.
If OpenAI possessed tools or datasets that could support these kinds of evaluations—and if those were withheld—publishers argue that the court is being deprived of the best available evidence. That is a serious claim because it goes to the reliability of the factual record.
The publishers’ motion, as characterized in the report, suggests they believe the withheld materials would help the court understand whether copyrighted journalism appears in outputs. That implies the tools are not generic; they are tailored to the kind of analysis needed for this case.
A shift from “training” to “use” and “output”
One reason this dispute has drawn attention is that it touches both training and output. Many public debates focus on training data: whether copyrighted works were used to train models without permission. But the practical harm publishers worry about often shows up in output: when users ask for summaries, rewrites, or “news-like” text, and the model produces something that competes with or replicates protected reporting.
The sanctions motion, by focusing on tools and datasets that could identify copyrighted journalism in outputs, underscores that the publishers are pressing on the “use” side of the equation. They want to show not only that training may have involved copyrighted material, but that the model’s behavior can reproduce protected expression in ways that matter legally.
This is also why the discovery dispute is so central. If the publishers cannot test outputs effectively, they may struggle to prove infringement. And if OpenAI can test outputs internally but does not share the relevant methods, the publishers argue that the playing field is tilted.
The “documentation gap” problem
Another angle—often overlooked in public coverage—is the role of documentation. In complex AI systems, there can be multiple versions of models, multiple training runs, and multiple changes to safety filters and generation settings. Even if two parties agree on the general concept of “the model,” they may be talking about different artifacts.
In litigation, that can create a documentation gap: what exactly was used, when, and under what conditions? If a company can produce internal records that clarify these points, it can strengthen its defense. If it cannot—or if it fails to produce them—plaintiffs may argue that the record is incomplete.
The publishers’ claim that OpenAI hid tools and datasets can be read as part of this broader documentation issue. Tools and datasets are not just technical assets; they are evidence of how the company evaluated its own system. If those evaluation artifacts exist, they can help establish what the company knew, what it measured, and what it concluded.
That is why sanctions are being sought. Sanctions are often the remedy when a party’s failure to disclose prevents the other side from obtaining a fair opportunity to test the evidence.
What happens next
While the reporting describes the motion and the allegations, the next steps will depend on how the court responds. Typically, the court will consider whether the motion meets the threshold for sanctions and whether the alleged withholding violated specific discovery obligations or court orders.
OpenAI’s response will likely focus on several themes common in these disputes: the scope of discovery requests, the distinction between proprietary information and discoverable evidence, and whether the withheld materials were actually required for the case. The company may also argue that it produced sufficient information through other channels, or that the requested tools and datasets are not necessary to evaluate infringement.
Publishers, in turn, will likely emphasize that the withheld materials are directly relevant to infringement evaluation and that the failure to disclose undermined their ability to test the model’s outputs. They may also argue that the timing of disclosure matters: even if some information exists, late production can still prejudice the opposing party.
Courts also tend to weigh intent and prejudice. If a judge concludes that the nondisclosure was inadvertent or that the publishers suffered no meaningful harm, sanctions may be limited. If the judge concludes that the nondisclosure was willful or materially prejudicial, the court could order stronger remedies.
Why this case matters beyond the parties
Even for readers who are not following every twist of AI copyright litigation, this sanctions motion is significant because it highlights a recurring fault line in the legal treatment of generative AI.
First, it raises the stakes for evidence transparency. If courts demand that AI companies provide evaluation tools and datasets that can test