Everything About AI
NextFin News - The New York Times is accusing OpenAI of an uncomfortable contradiction: its own staff understood how damaging generative AI could be to publishers, even as the company built its business on training models with publishers' work. The allegation, raised in court filings, escalates a two-year copyright war that now threatens OpenAI's path to a public listing and could force the destruction of models trained on copyrighted news.
The dispute has moved far beyond a single newspaper's grievance. The Times, The New York Daily News and 15 other media organizations told a federal court in July that OpenAI withheld evidence showing how its AI systems are trained and used — evidence the publishers say sits in the company's own training data sets and ChatGPT output logs. "But instead of just producing that evidence at the start of the case and focusing on the merits of its fair use defense, OpenAI chose obstruction," the publishers said in their motion.
The stakes are existential on both sides of the caption. In a September filing seeking summary judgment, the Authors Guild and co-plaintiff authors argue that "OpenAI's GPT models pose an existential threat to those who write and publish books," a threat that "threatens the incentives to create, the livelihoods of authors, and the future of book publishing." OpenAI and its partner Microsoft counter that training AI on publicly available text is fair use, that the publishers manipulated the products to manufacture evidence, and that the latest filings are a privacy invasion dressed up as discovery. The outcome will decide not just who pays whom, but whether the AI industry's foundational practice — ingest everything, ask questions later — can survive.
The Discovery Battle Is Where the Case Is Really Being Fought
The central question in the litigation is no longer only whether copying occurred. It is whether OpenAI knew what it was doing and then concealed it. That distinction matters because knowledge turns ordinary infringement into willful infringement, and willfulness is where statutory damages multiply.
In the consolidated copyright dispute now before U.S. District Court Judge Sidney Stein — In Re: OpenAI Inc. Copyright Infringement Litigation, S.D.N.Y., No. 25-md-03143 — U.S. Magistrate Judge Ona Wang ordered OpenAI to turn over internal communications about its deletion of two datasets, known as Books1 and Books2, that allegedly contained pirated books from Library Genesis. OpenAI had told the court the datasets were removed "due to their non-use." Judge Wang was not persuaded that the reasons for that deletion should stay secret. "OpenAI has waived privilege by making a moving target of its privilege assertions," she wrote.
The plaintiffs have gone further, asking the judge to invoke a "crime-fraud" exception to attorney-client privilege so they can see communications between OpenAI and its lawyers. Their theory: if counsel told OpenAI to delete the dataset, that could amount to intentional destruction of evidence. If proven, the company could face enhanced penalties of up to $150,000 per work, on top of court sanctions.
"OpenAI has consistently maintained that the reasons for the removal are privileged because they were legal decisions made in consultation with counsel. At no point has OpenAI disclosed or relied on the privileged reasons or retracted its position, and it has made clear that there are no non-privileged reasons."
The transmission mechanism here is straightforward, and it is why the internal-message allegation cuts deeper than the fair-use debate. Fair use is a defense about the character of the copying. Knowledge and concealment are facts about the character of the company. A court can forgive copying it deems transformative; it is far less forgiving of a defendant that understood the harm, hid the evidence, and then asked for the law's protection.
Knowledge, Not Copying, Is the New Legal Fault Line
This is not the first time internal communications have become the most damaging evidence in an AI copyright case. During discovery in a separate copyright dispute, Meta's researchers described Library Genesis as "a data set we know to be pirated." According to the filings, the issue was escalated to "MZ," who approved the pirated library's use anyway. That pattern — researcher flags the problem, executive approves the shortcut — is exactly what publishers are hunting for inside OpenAI's Slack channels and emails.
The strategy has already produced a price tag that every AI lab now underwrites in its risk models. In June, a federal judge in the Northern District of California ruled that Anthropic did not violate authors' copyrights when it trained its Claude model on books the company actually purchased and scanned. But the same judge found that works Anthropic ingested from a pirated dataset known as "The Pile" did not qualify as fair use, setting the stage for a separate trial. In July, Judge Araeli Martinez-Olguín in San Francisco granted final approval to Anthropic's $1.5 billion settlement with authors — the largest copyright class-action settlement in history — with individual authors receiving roughly $3,000 per work after fees.
The legal fault line has therefore shifted. The first wave of AI copyright litigation asked: did the model copy? The second wave asks: did the company know the source was tainted, and did it hide that knowledge? That shift is dangerous for OpenAI because its fair-use defense rests on an implicit claim of good faith — that the training was an automated, indiscriminate, transformative reading of the public web. Evidence that staff recognized publishers as a severe threat while simultaneously training on their work without permission punctures that narrative. It suggests the copying was not inadvertent; it was a calculated business decision.
There is also a commercial mechanism at work beneath the legal one. If internal messages show OpenAI understood the damage to publishers, then the company cannot credibly claim there is no market harm — and market harm is one of the four fair-use factors courts weigh most heavily. The internal acknowledgment becomes the plaintiff's best evidence on the very factor that could decide the case.
The Government and the Market Have Already Picked a Side
While the discovery fight plays out in Manhattan, the broader policy battle has drawn in the federal government. In early September, the U.S. Department of Justice filed a statement of interest supporting OpenAI, arguing that training AI models on copyrighted text does not generally constitute copyright infringement. It was the first time the U.S. government has taken a position on copyright litigation over the use of copyrighted material to train AI.
The Justice Department's reasoning cuts against the publishers' theory of the case: if training-stage copying is lawful, then what OpenAI's staff knew or feared about publishers matters far less. The department also warned that a New York Times victory would "threaten national security" and hurt small newsrooms, arguing that large publishers would disproportionately benefit if AI companies were forced to pay for copyrighted content. That framing matters because it gives the court a policy reason to rule narrowly — protecting the AI industry's training practices even while leaving output-level harms for another day.
The market, meanwhile, is treating the legal overhang as a pre-IPO problem to be managed rather than an existential one. Anthropic and OpenAI are moving toward what could be two of the largest initial public offerings in history. Investors are betting the companies can settle, license, or litigate their way to a clean enough balance sheet for a listing. OpenAI's chief information security officer, Dane Stuckey, said in a September 3 post that the newspaper was demanding 20 million private ChatGPT conversations — a demand the company says it will fight on privacy grounds.
OpenAI has also pushed back on the factual premise of the publishers' motion. In an August 14 filing, the company told the court that news organizations are seeking sanctions because they have "no evidence whatsoever" that ChatGPT systematically regurgitates their articles. An internal OpenAI project searched only a "rough-and-ready" sample of 78 million conversations, the company said, and its existing tools cannot execute the comprehensive searches the publishers demand. Drew Pusateri, an OpenAI spokesman, called the latest allegations "blatantly false" and said the company would "continue defending our users' privacy and the long-established principles of fair use."
Cyclical Friction or Structural Regime Change?
The key judgment for investors and publishers alike is whether this is a cyclical dispute that will settle into a licensing routine, or a structural break that rewrites the rules of the industry. The evidence points to structural change, and it will not revert on its own.
Three facts support that call. First, a licensing market already exists: German publishing giant Axel Springer reached a deal with OpenAI weeks before the Times sued, and other publishers have followed. The existence of an actual market for news content undercuts the fair-use argument that no market for the use existed to be harmed. Second, the government has now intervened — not to protect publishers, but to protect the AI industry's training practices. That intervention signals the dispute has moved from contract negotiations to national policy. Third, the wave of litigation keeps widening: more than 30 local newspaper publishers banded together in June to sue, and The Seattle Times and Newsday filed a joint lawsuit in early September. This is not one plaintiff with a grievance; it is an industry establishing a litigation posture.
The cyclical reading has merit, too. Copyright law has absorbed disruptive technologies before — from photocopiers to search engines — and fair use has usually bent to accommodate the new use. The Justice Department's filing is a strong signal that courts may do the same here, limiting any ruling to output-level conduct while leaving training largely unencumbered. If that happens, the current turmoil is a costly but temporary negotiation over price, not a regime change.
But even that outcome changes the structure of the industry. Whether through court order or settlement, AI companies are already signing licensing deals and building attribution systems. The "ingest everything, ask questions later" model is being replaced, deal by deal, with a paid-access regime. The question is no longer whether AI will pay for content; it is how much, and how soon.
What Would Prove This Wrong
The strongest counter-thesis is simple: the sanctions motion is a litigation tactic born of weak evidence, and the fair-use tradition will protect training-stage copying. OpenAI says the publishers have "no evidence whatsoever" of systematic regurgitation, and the Department of Justice has aligned with that position. A narrow ruling that training is fair use would leave the industry's core practice intact and reduce the internal-message controversy to a sideshow.
The specific signal that would falsify the structural-change call is a court order denying the publishers' sanctions motion combined with a summary-judgment ruling adopting the Justice Department's fair-use position for training. If Judge Wang rejects the crime-fraud exception and the court holds that training on copyrighted text is transformative as a matter of law, the regime-change thesis fails. Conversely, if the court grants the sanctions motion and orders production of internal communications showing OpenAI leadership understood the threat to publishers and concealed it, the structural call is confirmed — and the damages math changes dramatically.
What Comes Next
In the short term, expect volatility tied to discovery rulings and IPO timing. Every order from Judge Wang on privilege and every filing on the sanctions motion will move the perceived risk premium on AI valuations. In the medium term, licensing costs will rise regardless of the legal outcome — the Axel Springer model is spreading because it is cheaper than litigation. In the long term, the industry is moving toward a paid-licensing regime for training data, whether by court order, settlement, or statute.
Three scenarios frame the path ahead. The base case is continued discovery, partial production of internal messages, and an expanding web of licensing deals. The upside case for publishers is a willful-infringement finding, billions in damages, and a court order to destroy models trained on their content — the remedy the Times has already requested. The downside case for publishers is a broad fair-use ruling that leaves training largely free and reduces their leverage to bargaining over price.
The real verdict in this fight will not be whether AI can read the news. It is whether the news industry can make AI pay for the privilege — and the answer is becoming clear, one internal message at a time.
Explore more exclusive insights at nextfin.ai.
