Leadership3 distinct publishers3 min readPublished
Three summary judgment briefs in the consolidated news publisher case describe the same conduct using denominators that never meet, and the one the court adopts will set what licensed text is worth to every buyer.
The Board Room · Leadership desk
Compiled by The Board RoomSomething wrong?How this is made
The two records do not conflict on the facts so much as on the unit of harm each one measures. OpenAI counts at the point of output, sampling what users actually received in discovery-produced chat logs [2]. The publishers count at the point of ingestion, alleging that entire archives were copied at five separate stages of an artificial intelligence pipeline and that the resulting products answer the questions readers used to bring to news sites [8]. Both descriptions can hold at once, which is part of why the briefs barely engage each other. It also matters that the log sample was never designed as a measurement of the corpus: those 20 million anonymised conversations exist because a magistrate ordered OpenAI to produce them in November 2025, with a stay request denied on November 13 [15]. Judge Stein's choice of denominator therefore does more work than any single figure in either filing.
Arithmetic drawn from OpenAI's own brief shows how far apart the two frames put the exposure. OpenAI reports that the plaintiffs' experts used more than 250 million prompts against non-public GPT models and configurations available only through the API, and still triggered partial regurgitation in fewer than 2.6 percent of asserted works, extracting under 4.4 percent of each [13]. Applied to the asserted total, 2.6 percent is a ceiling of roughly 281,000 works [3], which is the number OpenAI is content to concede, and about 21,700 times the 0.00012 percent rate its brief leads with [4]. OpenAI sets its 4.4 percent figure against the 16 percent that plaintiffs' experts extracted in the Google Books litigation [14].
The board-deck version of Friday's filings is that OpenAI won the memorisation argument, since the plaintiffs' own experts put the rate at 0.00011 percent for The Times and the Daily News titles and 0.000002 percent for The Intercept [7], so licensed text becomes optional. The Times theory, though, never depended on regurgitation: its complaint frames the injury as free-riding on its investment to build substitutive products without permission or payment [21], and the Justice Department's brief accepts that frame when it tells the court that the creative possibilities and public benefits of training "far outweigh any competitive harm" [18]. The Justice Department's brief weighs that harm against the benefits of training; it does not deny the harm exists.
OpenAI's own study points the same way on the output side: prompts about 710 randomly selected asserted works produced more than 68,000 outputs with no regurgitation at all [11]. That answers whether ChatGPT functions as a distribution channel for articles, but not whether the works OpenAI concedes were ingested needed a licence, and OpenAI's own position is that no evidence of copying exists for roughly 4.5 million of them, about 42 percent of the asserted total [5][1]. A defendant arguing over which 58 percent of a corpus went into training [2] is arguing about scale.
No filing supplies a price. For anyone negotiating content licences this quarter, the operative uncertainty is which denominator survives summary judgment: an output test makes licensing a filtering cost, while an ingestion test makes it a per-work price, with one consolidated case setting the arithmetic every other rights-holder brings to the table. Whether fair use covers training is a question for this decade; what a licence is worth before that answer arrives is a question for this quarter, and on the public record that value turns less on what the model reproduces than on which stage of the pipeline the court decides to count.
Ranked by verification strength, evidence, and original report placement.
According to the news plaintiffs, the defendants copied entire archives at five separate stages of an artificial intelligence pipeline and built commercial products that answer the questions readers used to bring to news sites.
In court papers filed Tuesday, the Justice Department supported OpenAI's argument that training its models on writings found on the internet is protected by fair use, writing that "the creative possibilities and public benefits" of such training "far outweigh any competitive harm."
The Justice Department told the court that siding with The New York Times and other publishers would thwart "creative and scientific progress while hindering American prosperity and economic mobility."
Graham James, a spokesperson for The New York Times, said the administration is siding with a handful of trillion-dollar AI companies at the expense of American creators, and that letting companies take content without permission or compensation would undermine the sustainability of human-created content.
In its original lawsuit and an amended complaint, The Times focused on the unfair competition of companies that seek to free-ride on its investment in journalism by using it to build substitutive products without permission or payment.
The first two judges to consider whether AI training is fair use issued diverging rulings last year.
Distinct publishers with included, body-backed reporting in this cluster.
1 article · September 4, 2026
1 article · September 5, 2026
1 article · September 2, 2026
Follow any of these and your For You feed starts watching them — no settings page required.
invest
DOJ's 20-page brief backs OpenAI, argues AI training on copyrighted work is fair use5 distinct publishers
invest
The Justice Department moves the AI copyright fight to the licensing table1 distinct publisher
invest
WikiHow's suit against OpenAI asks for a remedy that outlives the damages award1 distinct publisher
product
Trump administration files in support of OpenAI's fair-use defence in the Times case3 distinct publishers
Evidence-backed comparisons of source perspectives and observed adoption signals. Read the methodology
Which Builder, Operator, and Investor concerns the observed source mix emphasized—not a truth score.
Evidence, demonstrated adoption, hype gap, incentives, and confidence are assessed independently, each on its own current evidence. How these are measured.
Party filings, single relay
Almost every number in this story reaches us through OpenAI's own summary judgment memorandum as read by PPC Land, including the figures OpenAI attributes to the opposing experts, and no expert report has been examined outside the parties. The government's brief is on firmer ground: the Guardian and ABC's wire copy carry the same passages separately. The publishers' side is thinnest, paraphrased in one account, with the traffic and revenue comparisons that would test the substitution theory blacked out.
Docket milestones only
What can actually be counted is institutional. One government brief has adopted OpenAI's fair-use framing, three motions have put the question to Judge Stein, and 20 million conversations sit in the record because of a production order OpenAI could not stay. No licence price, signed deal or measured traffic change appears anywhere in this reporting, so the market consequence of the ruling remains an argument rather than an observation.
Numbers outrun the question
The framing that these briefs use denominators which never meet survives contact with the documents, and both sides oblige. OpenAI's 0.00012 percent describes ordinary conversation; the plaintiffs' sub-2.6 percent describes what 250 million prompts against non-public models can force out. Neither settles whether copying at training time was lawful, which is what Judge Stein has been asked. On the political half, only the Guardian says out loud that the government's brief advises rather than binds, so elsewhere it reads heavier than it is.
Every figure has a side
Each quantitative claim originates with a litigant moving for judgment in its own favour, and the qualitative ones come from a plaintiff's spokesperson and from an administration that has made AI leadership declared policy. The Justice Department argued public benefit in court the same week the commerce secretary was urging G20 officials to embrace fair use, as the Guardian reports. PPC Land's reconstruction is careful, but it is a reading of adversarial briefs, and OpenAI's assertion that the plaintiffs have grown since ChatGPT launched rests on figures the redactions remove.
Firm on the brief, thin on the record
Two independent publishers agree on the government filing and the Times' reply, which makes that half of the story solid. The document-level record rests on one trade publication's reading of partly redacted briefs, and its account of the Browse dispute stops before the plaintiffs answer. Dates, the docket number and the composition of the plaintiff group are specific enough for anyone to check against the file.