Newly unredacted filings in an active copyright case surface internal language that is less about law than about economics โ and the structural problem those documents describe predates any ruling.
Active litigation โ read carefully. The New York Times v. OpenAI and Microsoft is ongoing. Every document discussed below is an excerpt selected and presented by one side in a lawsuit. Nothing has been proven. No court has ruled. No finding of infringement or wrongdoing exists. Internal documents quoted in litigation filings are chosen by the filing party and may lack their surrounding context. This article offers no view on fair use, infringement, the merits of the case, or its likely outcome. It is not legal advice and not investment advice.
What Happened
Newly unredacted filings in The New York Times v. OpenAI and Microsoft โ active litigation filed roughly three years before this writing โ have surfaced internal documents that plaintiff counsel is presenting as evidence. As reported by TechCrunch on September 17, 2026, those filings quote two separate documents attributed to Brent Hecht, identified as a Microsoft Director of Applied Science. The first, a January 2023 internal memo, used the phrases “an astonishing theft of unprecedented proportions” and “the largest theft of labor in human history.” The second, a January 2024 internal presentation, stated: “It is highly unusual that an end-product threatens the economic foundations of its essential suppliersโฆ” These are distinct documents written a year apart; they are not interchangeable.
Microsoft responded on the record. A spokesperson described both documents as “one employee’s individual perspective” that “do not represent the company’s views,” and characterised Nadella’s deposition testimony โ in which he stated that “anything that is paywalled should be licensed by anyone who wants to use itโฆfor grounding or training” โ as “observations.” OpenAI did not immediately respond to a request for comment. The filings also quote Nick Turley, OpenAI’s head of ChatGPT, describing publishers as facing an “existential threat” from products that are “largely substitutive,” and Greg Brockman, OpenAI’s president, describing models as “excellent at news.” Steven Lieberman, counsel for the New York Daily News โ a lawyer arguing his own side’s case โ told TechCrunch that “the evidence revealed here for the first time shows that OpenAI and Microsoft knew that what they were doing was wrong.” That is his characterisation as plaintiff counsel; it is not adopted here.
The filings present three dataset counts, all figures offered by the party bringing the case and not accepted by any court: over 91,692 copies of works from The Times, the Daily News, and the Center for Investigative Reporting in mid-training datasets; over 2 million documents from nytimes.com in a Common Crawl-derived dataset; and at least 160,903 unique works from news publishers in a dataset referred to as Project Mango. These figures describe what the plaintiffs allege sits in training corpora. They say nothing about what any model produces on demand โ what is in a corpus and what a system reproduces are separate questions.
Reminder: Everything above reflects documents selected and presented by one side. Nothing is proven. No court has ruled. No finding of infringement or wrongdoing exists.
The key insight: The damaging document in a copyright case is rarely the one about copyright. It is the one where someone inside describes the economics plainly โ and what these filings surface, read carefully, is people on several sides describing an economic relationship in economic terms. Whether that description meets any legal test is a question for courts. Whether it describes a real market structure is a question worth examining on its own.
The Structural Read
Start with what the documents actually discuss, as distinct from what they prove. The January 2023 Hecht memo and the January 2024 Hecht presentation โ two separate documents, written a year apart, not to be merged โ both reach for economic language: labor, theft, suppliers, economic foundations. Whether that language was accurate, hyperbolic, or strategically chosen inside a corporation is not a question this analysis can or should answer. What it signals structurally is that the people inside these organisations were thinking in supply-chain terms. Journalists are input suppliers. AI products are downstream outputs. The question of whether the supplier gets paid is not primarily a moral question; it is a market-structure question about whether a price mechanism exists.
The Nadella deposition line is the most structurally interesting thing in these filings โ and the most easily misread. His stated view that “anything that is paywalled should be licensed by anyone who wants to use itโฆfor grounding or training” is a stated principle about how a market ought to work. Microsoft explicitly characterised it as “observations.” It is not an admission. It is not a concession. It is not evidence that any specific thing was or was not done. Read carefully, it is a description of a market that does not fully exist yet: one where paywall boundaries function as pricing signals and licensing infrastructure routes payment to content producers. That market is not built. What fills the gap when that infrastructure is absent is litigation โ one counterparty, one court, one ruling at a time.
Turley’s “existential threat” and “largely substitutive” language, and Brockman’s “excellent at news” description, are the documents doing the substitution analysis that economists and lawyers will spend years formalising. This article makes no claim that substitution has occurred, that publishers have been harmed, or that any legal test for substitution is met. The observation is narrower: that people close to these products were reaching for substitution as the right frame. Whether a product substitutes for the thing it was trained on is an economic question prior to being a legal one, and these documents suggest it was already being posed in those terms internally.
Business Engineer โ Harness Theory Lens
Litigation as the Default Pricing Mechanism for an Unpriced Input
When an input has no established price, buyers acquire it where the price is lowest or the counterparty is weakest โ a general property of unpriced inputs, not a characterisation of any party’s conduct here. Disputes then migrate toward whoever can afford to bring them. Courts end up doing the work a licensing clearing house would do in a mature market: setting a price, establishing a precedent, defining the relationship between input supplier and downstream product. They do this slowly, one counterparty at a time. A Nadella deposition articulating the correct principle is structurally interesting precisely because the infrastructure to act on that principle does not yet exist at scale. The principle and the plumbing are not the same thing.
Three Implications
IMPLICATION 1 โ THE ECONOMIC DOCUMENT MATTERS MORE THAN THE LEGAL ONE
The Hecht documents โ two separate items, a January 2023 memo and a January 2024 presentation โ are quoted in these filings not because they say anything definitive about copyright law, but because they describe the economic relationship in plain terms. Microsoft’s disclaimer (“one employee’s individual perspective”) is a standard and legitimate response; it does not change the structural observation that organisations debating these inputs internally were doing so in supply-chain language, not legal language. That framing will shape how courts eventually price the question, whatever their ruling.
IMPLICATION 2 โ THE LICENSING INFRASTRUCTURE GAP IS THE REAL MARKET PROBLEM
Nadella’s deposition describes a principle โ paywalled material should be licensed for grounding or training โ that, if operationalised, would require a functioning licensing market: rights registries, rate structures, clearing mechanisms, enforcement. None of that infrastructure exists at the scale AI training requires. In its absence, pricing happens through litigation by default, and the counterparties who can sustain multi-year federal cases define the early precedents. Smaller publishers do not participate on equal terms. No claim is made here that any party intended this or exploited it โ it is a structural property of immature input markets.
IMPLICATION 3 โ CORPUS COUNTS AND MODEL OUTPUTS ARE DIFFERENT QUESTIONS
The three dataset figures โ 91,692+ copies in mid-training datasets, 2 million+ documents from nytimes.com in a Common Crawl-derived dataset, 160,903+ unique works in Project Mango โ are counts of what the plaintiffs allege sat in training corpora. They are not aggregated here and are not converted into any measure of harm, revenue, or value. More importantly, what is present in a training corpus and what a deployed model reproduces on demand are distinct technical and legal questions. Conflating them is the most common misreading of AI copyright cases, and these figures do not resolve that distinction in either direction.
