The Evidence Weighing on NYT v. OpenAI: Does 'Regurgitation' Defeat AI Fair Use?

As the summer of 2026 approaches, the most consequential legal battle in the history of artificial intelligence has entered a critical and highly technical discovery phase. Inside the Southern District of New York, a sprawling multidistrict litigation (MDL) has consolidated sixteen separate copyright infringement lawsuits against OpenAI and Microsoft. The plaintiffs, led by The New York Times and the Authors Guild, are seeking to prove that the foundation models powering ChatGPT were built on mass copyright infringement. The outcome of this litigation will not merely determine financial damages; it threatens to dictate the fundamental architecture of how machine learning models are trained and deployed globally.[3]

At the heart of the dispute is the U.S. legal doctrine of "fair use," a framework designed to balance the rights of creators with the public's interest in innovation and free expression. Historically, fair use has protected technologies that utilize copyrighted works in a "transformative" manner, such as search engines indexing the web or databases digitizing books for text-mining. However, generative AI presents an entirely novel challenge. The courts must decide whether ingesting billions of words to train a neural network is akin to a human reading a library to learn grammar, or if it constitutes the creation of an unlicensed, competing database.

Early judicial signals in 2025 suggested a potential pathway for AI developers. In cases like Bartz v. Anthropic, courts leaned toward the interpretation that the sheer act of training a model on lawfully obtained data is inherently transformative. This perspective argues that the models are extracting statistical relationships and uncopyrightable facts, rather than storing the expressive text itself. Consequently, the legal battleground has shifted dramatically. The focus is no longer solely on the "input" phase—how the data was gathered—but has pivoted sharply to the "output" phase, scrutinizing exactly what these models produce when prompted by everyday users.[3]

This shift has elevated a technical phenomenon known as "regurgitation" or "memorization" to the center of the legal stage. The New York Times and other plaintiffs allege that large language models do not merely learn abstract concepts; they memorize specific, highly weighted training data and can spit out near-verbatim copies of protected reporting and literature. If a user can prompt ChatGPT to bypass a paywall and generate an exact replica of a Pulitzer-winning investigation, the plaintiffs argue, the AI ceases to be a transformative tool and becomes an illegal market substitute.[2]

To prove this claim, the plaintiffs demanded hard evidence of how ChatGPT operates in the wild, leading to a massive discovery dispute. In January 2026, U.S. District Judge Sidney Stein affirmed a magistrate judge's order compelling OpenAI to produce a staggering 20 million ChatGPT user logs. The plaintiffs had initially requested 120 million logs from the tens of billions preserved by the company, but the court settled on the 20 million sample—representing roughly 0.5 percent of the total—as a statistically significant dataset to test the regurgitation theory.

OpenAI fought fiercely against the subpoena, arguing that handing over millions of user conversations would constitute a massive invasion of privacy. The company contended that the vast majority of these logs had nothing to do with the plaintiffs' copyrighted works and that producing them would turn uninvolved users into collateral damage. However, the court rejected this privacy gambit. Judge Stein ruled that the logs could be safely de-identified and protected under strict confidentiality orders, prioritizing the plaintiffs' need to gather evidence on the model's real-world outputs.[3]

The 20 million logs are now the most scrutinized dataset in the legal tech world. Plaintiffs' experts are combing through the de-identified conversations to determine whether regurgitation is a routine occurrence or a rare anomaly. OpenAI has long maintained that exact memorization is a technical bug, not a feature, and that it typically only occurs when users deploy highly specific, "adversarial" prompts designed to force the model to break its own guardrails. The plaintiffs, however, hope the logs will reveal that everyday users are regularly receiving outputs that compete directly with original works.[2]

This evidentiary hunt is entirely focused on the fourth factor of the fair use test: the effect of the use upon the potential market for or value of the copyrighted work. In copyright law, a use is rarely considered fair if it serves as a direct market substitute that deprives the original creator of revenue. If the discovery process proves that ChatGPT is frequently used as a replacement for paid news subscriptions or book purchases, OpenAI's fair use defense will face a nearly insurmountable hurdle.

While the output logs represent the primary offensive front for the plaintiffs, they are also probing the historical inputs used to train OpenAI's earlier models. A significant point of contention involves the "Books1" and "Books2" datasets, which were allegedly compiled from shadow libraries containing pirated literature, such as Library Genesis. OpenAI deleted these datasets prior to the commencement of the current litigation, citing "non-use," but the plaintiffs have aggressively sought discovery regarding the internal decision-making process that led to their destruction.[1][3]

In a notable procedural victory for OpenAI, the court recently shielded the company's internal communications regarding the dataset deletion. In February 2026, Judge Stein reversed a prior magistrate ruling, declaring that OpenAI's discussions with its legal counsel about the "Books1" and "Books2" datasets are protected by attorney-client privilege. The judge ruled that merely stating the datasets were deleted for "non-use" did not waive the privilege, preventing the plaintiffs from accessing potentially damaging internal emails about the company's early data-sourcing strategies.[1]

As the evidence mounts on both sides, the broader technology industry is watching with intense anxiety. A ruling that generative AI outputs inherently violate copyright law would send shockwaves through the sector. Developers might be forced to implement draconian output filters, fundamentally degrading the utility of the models, or face the prospect of destroying their existing foundation models entirely—a remedy The New York Times explicitly requested in its initial complaint.[2]

In response to this existential legal threat, a two-tiered system of AI development is rapidly emerging. Rather than waiting for a judicial verdict, several major publishers, including Reddit, News Corp, and The Associated Press, have signed lucrative, multi-million-dollar licensing agreements with AI developers. These deals grant tech companies explicit permission to train on the publishers' archives, effectively bypassing the fair use debate entirely. However, this licensing economy has sparked its own fierce debate about the future of the internet.[3]

Legal scholars and open-source advocates warn that dismantling the fair use defense could have disastrous unintended consequences for competition. If the courts rule that every piece of training data must be explicitly licensed, the barrier to entry for AI development will become impossibly high. Only trillion-dollar tech conglomerates will possess the capital required to secure comprehensive data licenses, effectively outlawing independent academic research and open-source foundation models. This dynamic could hand a permanent, insurmountable monopoly to the very companies currently being sued.

Conversely, copyright holders argue that the survival of independent journalism and human authorship cannot be sacrificed on the altar of technological progress. They contend that if AI companies are allowed to build multi-billion-dollar empires by freely extracting the value of human labor, the economic incentives to create original work will collapse. For these creators, the fair use doctrine was never intended to protect autonomous machines that ingest the entirety of human culture only to sell it back to the public at a premium.[2]

As the multidistrict litigation grinds through the summer of 2026, the burden of proof rests heavily on the data extracted from the 20 million ChatGPT logs. The impending summary judgment motions will require the court to weigh the empirical evidence of regurgitation against the transformative potential of artificial intelligence. The resulting precedent will not only define the boundaries of digital copyright for the next century but will fundamentally shape the economic relationship between the humans who create knowledge and the machines designed to synthesize it.[3]