One Million News Articles for $6,800: NYT v. OpenAI Hits the Fair-Use Showdown

The flagship AI copyright fight just moved to the stage where a single judge can decide most of it. On Friday, dueling summary judgment motions landed in Manhattan federal court in the New York Times' case against OpenAI and Microsoft, with the newspaper and other news organizations asking Judge Sidney Stein to rule that neither company made fair use of their reporting at any point in the AI pipeline (Chicago Tribune). Today, Reuters framed the filings as the key test of AI training under copyright law, noting that OpenAI, Microsoft, the Times, and a group of prominent authors have laid out sharply differing views on one of the highest-stakes questions in the field. The number that frames the stakes comes from the plaintiffs' own filing: under OpenAI's pricing, they say, it costs about $6,800 to generate one million news-style articles with no human being involved at all.

What Actually Happened

On Friday, both sides in the Times v. OpenAI and Microsoft case moved for summary judgment — a request that the court resolve core legal questions as a matter of law before trial (Chicago Tribune). The news organizations, which include the Times and the Chicago Tribune, asked Judge Stein to determine that neither defendant made fair use at any stage of what they call the AI pipeline: not when acquiring articles from behind paywalls, and not when training large language models on the copied work. Their lawyers wrote that rejecting the defendants' efforts to recast free riding as fair use would protect "the future not just of journalism but of responsible AI too," because incentives to produce original reporting are what responsible systems depend on (Chicago Tribune).

OpenAI's answering motion took a different tack: deny the regurgitation claims and reframe the property line. The company's lawyers wrote that its chatbots do not spit out news articles verbatim and that "nobody owns facts, just as nobody owns language, grammar, or style" — reporting effort does not create a monopoly on the truths being reported (Chicago Tribune). Microsoft filed its own motion asking for judgment on every copyright claim challenging LLM training and web grounding, the feature that pairs Copilot with live search. Microsoft's position is that training on human-produced journalism is a "transformative" use protected by fair use, and that copyright law "does not give News Plaintiffs the right to block transformative new technologies that do not replace their protected expression" (Chicago Tribune).

One inconsistency in the defense's own record is doing a lot of work. Microsoft's filing asserts Copilot users do not turn to it for news or current events. OpenAI, meanwhile, has acknowledged in the same litigation that a large chunk of its users come to its products for news — and more than one billion people use ChatGPT, per June figures from market intelligence firm Sensor Tower cited in the Tribune's report. The plaintiffs argue that substitution has drained the traffic and subscriptions that fund reporting (Chicago Tribune).

The $6,800 Newsroom Is the Whole Argument

Strip out the constitutional rhetoric and the plaintiffs' motion is an economic argument. Human reporting is expensive; generated text is nearly free. The filing details that under OpenAI's pricing, roughly $6,800 buys one million news-style articles averaging 500 words each — no author, no editor, no reporter (Chicago Tribune). The plaintiffs say the market is already showing what that does: a wave of low-grade "pink slime" sites that plagiarize or remix legitimate reporting, and chatbots that churn out content resembling news but containing inaccurate information about current events — a failure mode they call potentially more harmful to the news market than accurate substitutes, because it degrades trust in the product category itself (Chicago Tribune).

The filings also carry a threat aimed past the judge at the defendants' public image. Attorney Steven Lieberman, leading the Tribune's representation, said portions of the motions were filed under seal because Microsoft and OpenAI designated them confidential, and that they contain admissions by Microsoft executives that undercut the company's defenses. He promised an unsealing fight in the coming weeks: "The stuff that we were forced to file under seal... is scorchingly hot. Not just a smoking gun; they're a smoking bazooka" (Chicago Tribune). His summary of the defendants' position cuts to the licensing logic underneath the whole case: they pay for chips, computers, electricity, and buildings, and they treat their own code as proprietary — "And they contend that their code is proprietary and competitors copy it. That is piracy," Lieberman said of taking content without consent or compensation (Chicago Tribune).

Reuters' wrap on Tuesday put the stakes plainly: this is one of the highest-stakes questions in copyright law, and the parties' views could not be further apart. Whether training models on copyrighted text is fair use is the same question animating the music publishers' piracy case against Anthropic and a widening docket of publisher suits — including the publishers' separate claims that Anthropic torrented training data, which we covered in August. A ruling here would be the first of its kind on the core training question in the highest-profile case in the country.

The Government Already Picked a Side — and the Docket Keeps Growing

The defendants are not arguing alone. On September 1, the Justice Department filed a statement of interest backing OpenAI's fair-use position, arguing that training LLMs on copyrighted works is fair use and that a Times win would damage American tech dominance — with national security implications, per the Times' report on the filing. The statement also suggested creators unhappy with AI training should take their case to Congress rather than the courts (AP; NYT; Tech Times). It is non-binding — Judge Stein does not have to weigh it — but it signals where the executive branch stands in the highest-profile AI copyright case in the country. Nieman Lab's headline on the filing captured the administration's framing: a Times win would "threaten national security and hurt small newsrooms."

Meanwhile, the plaintiff pool keeps expanding. The Seattle Times and Newsday sued OpenAI and Microsoft on similar copyright grounds, per AP reporting carried Friday — meaning whatever Judge Stein decides now shapes a case with an ever-larger roster of publishers behind it. The news plaintiffs also have a pending sanctions motion from July, accusing OpenAI of destroying evidence and concealing its ability to locate news stories in training data and ChatGPT responses; Stein has not ruled, and per the Chicago Tribune, the summary judgment motions are aimed more narrowly at Microsoft until that resolves.

There is no deadline for a summary judgment ruling. But the sequencing matters: if Stein grants the plaintiffs' motion in whole or part, the case shrinks to damages and goes to trial as a near-slamdunk on liability. If he grants the defendants' motion, the fair-use question is effectively settled for this case — and the plaintiffs head to an appeals court that could set national precedent either way. The sealed-material fight and the sanctions ruling land in between, and either could shift the evidentiary picture before a word of the fair-use analysis is written.

What Builders Should Take From It

• Training data is now a balance-sheet item, not a philosophical debate. If fair use fails at summary judgment, every lab's model-retraining pipeline needs licensed or opted-in data, and licensing markets (like the AP-OpenAI style deals) reprice across the board. If you build data-dependent products, watch this ruling like an interest-rate decision.
• The legal surface is bigger than training. Microsoft's motion explicitly covers web grounding — the search-augmented layer. If your product's core value is "model plus retrieval," the same doctrine fight applies to what your system fetches and synthesizes, not just what it was trained on.
• The substitution argument is about price, and $6,800 per million articles is the anchor. Whatever you publish or generate with AI, assume downstream buyers can generate infinite cheap imitations. Your defensible layer is provenance, verification, and distribution trust — not the raw text.
• Plan for behavior changes in the platforms you build on. Depending on how this ruling lands, expect API-level changes: stricter training-data filters, region-specific features, or licensing-gated capabilities. Features you treat as permanent are, in legal terms, contingencies.
• "Nobody owns facts" is an argument, not a ruling. Even a defendants' win here is one district court's view, heading into appeals, while the UK, EU, and Japan run different regimes. If your product strategy assumes training on scraped content is free, you are pricing in a legal position, not a law.

The next move belongs to Judge Sidney Stein, and he is under no obligation to move quickly. But for the first time since ChatGPT shipped, the central question — whether the economics of AI training survive copyright law — sits on a judge's desk with complete briefing, a DOJ thumb on the scale, and a million articles' worth of price tags in the record. The most expensive $6,800 in publishing history may turn out to be the one that decides what training data costs.

Developer312 covers the AI business signals builders actually need to act on. Get the weekday briefing at developer312.com.