# OpenAI's discovery record convicts the sourcing, not the training

URL: https://www.thedeepfeed.ai/posts/2026-10-13-the-authors-case-discovery/
Category: Policy
Published: 2026-10-13
Author: the-deep-feed
Tags: policy, copyright, openai, microsoft, authors-guild, new-york-times, fair-use, training-data, litigation
Kind: deep

> Unsealed briefs in the authors' and newspapers' case against OpenAI and Microsoft put executives' own words about LibGen, theft and substitution on the record, and those words bite hardest on how the books were obtained, not on whether training is fair use.

## TL;DR

- On **September 17** the parties in *In re OpenAI Copyright Infringement Litigation* (S.D.N.Y. No. 25-md-3143, Judge Sidney Stein) refiled their summary-judgment briefs with most redactions lifted. The authors' brief that had hidden nearly **every** executive admission in its September 5 public version now quotes them.
- The viral **'largest theft of labor in human history'** line comes from the **New York Times-led news brief**, not the authors'. The public brief cites it to a separate fact paragraph without naming its author, and the companion phrase from Microsoft's **Brent Hecht** was a forecast of how the public would see AI training.
- The authors' record is about sourcing: **117,500** LibGen books in 2018, a **35 TB** torrent matching LibGen's **4.6 million** books in 2019 and 2020, a note reading *We trained GPT-3 on pirated stuff! No sharing that!*, and a 2022 deletion effort called **Project Clear**. Microsoft, the brief says, saw LibGen named in an **April 2019** document shared with Bill Gates.
- The law has already split the two questions. Training won fair use in **Bartz** and **Kadrey**, and the Justice Department's **September 1** filing backs it. Pirate-library acquisition cost Anthropic **$1.5 billion**. The unsealed record lands almost entirely on the side courts have already priced.
- The record constrains what OpenAI and Microsoft can say in court. Their public stance has not visibly moved: Microsoft disowned its own director's documents, and its president's **October** call for AI rules said nothing about copyright. Oppositions are due **October 23**, and the exhibits stay sealed until **January 2027**.

On September 17, the parties to the largest copyright fight in American AI did something procedural and routine: they refiled their summary-judgment briefs, leaving unredacted every passage nobody had asked to keep secret. By the next morning, the claim that a Microsoft executive had called AI scraping "the largest theft of labor in human history" was a [TechCrunch headline](https://techcrunch.com/2026/09/17/microsoft-exec-called-ai-scraping-the-largest-theft-of-labor-in-human-history-new-unredacted-filings-reveal/) and a [955-point Hacker News thread](https://news.ycombinator.com/item?id=49752056). Ten days later, the Authors Guild's summary of its brief, featuring an OpenAI researcher's worry that "sketchy russian website" would show up on Hacker News, [showed up there](https://news.ycombinator.com/item?id=49863864) with 625 comments.

The coverage converged on one storyline: the companies knew. The filings support that, within limits set out below. But the record's legal weight falls unevenly. Almost every damaging line concerns *how* OpenAI obtained books and *whether* it expected its products to substitute for their authors. Very little touches the question the companies have staked their defense on: whether training a model on copyrighted text is itself fair use. Courts have already started answering those two questions differently, and the Justice Department has taken a side on only one.

The working hypothesis was that discovery put the companies' internal assessments of their own data practices on the record, and that the record now constrains their public positions on copyright and regulation. The first half holds. The second holds in court, not yet in public. Nothing either company has said since September 17 suggests a retreat on training, and both appear to be betting they can win that argument even if the sourcing costs them.

This is reporting on a pending case, not legal advice. The unsealed briefs are allegations; no court has ruled on these motions.

# One docket, two briefs, one viral line

The case is *In re OpenAI, Inc., Copyright Infringement Litigation*, No. 25-md-3143 in the Southern District of New York, before U.S. District Judge Sidney H. Stein. It is a multidistrict proceeding: on April 3, 2025, the authors' class actions were [consolidated](https://authorsguild.org/news/plaintiffs-file-motion-for-summary-judgment-v-openai-and-microsoft/) with suits from news organizations led by The New York Times. The Authors Guild and more than a dozen authors, including John Grisham, Jodi Picoult and George R.R. Martin, first sued on September 19, 2023. Two summary-judgment briefs drove September's coverage, and they are often confused.

The **class plaintiffs' brief**, [Dkt. 1982](https://authorsguild.org/app/uploads/2026/09/S.D.N.Y.-25-md-03143-dckt-001982_000-filed-2026-09-17.pdf), with its 162-page [statement of undisputed facts](https://authorsguild.org/app/uploads/2026/09/Class-Plaintiffs-SUF-9.17.26.pdf) at Dkt. 1987, asks Judge Stein to hold OpenAI and Microsoft liable as a matter of law on 194 "Asserted Works." It attacks three acts: downloading books from pirate libraries, copying them to train seven GPT-series models, and passing copies between the companies, with Microsoft also vicariously liable. This is the brief the [Authors Guild's September 21 release](https://authorsguild.org/news/ag-v-openai-top-execs-knew-mass-book-piracy-was-illegal/) summarized.

The **news plaintiffs' brief**, [Dkt. 1977-1](https://storage.courtlistener.com/recap/gov.uscourts.nysd.640396/gov.uscourts.nysd.640396.1977.1.pdf), covers five news cases led by The Times and seeks liability "at each stage of the AI pipeline": acquisition, training, grounding and output. This is where the theft language lives. Its first sentence reads:

> This case is about, as Microsoft's Director of Applied Science put it, "an astonishing theft of unprecedented proportions"; SF1437, perhaps the "largest theft of labor in human history." SF1652.
>
> — [News Plaintiffs' Combined Summary Judgment Brief, Dkt. 1977-1](https://storage.courtlistener.com/recap/gov.uscourts.nysd.640396/gov.uscourts.nysd.640396.1977.1.pdf), Sep 17, 2026

Read the citations, not the headline. The first phrase is attributed to Microsoft's Director of Applied Science, Brent Hecht, and cited to fact paragraph 1437. The second cites a different paragraph, 1652, whose author the public brief does not name. [Ars Technica](https://arstechnica.com/tech-policy/2026/09/microsoft-exec-called-ai-scraping-the-largest-theft-of-labor-in-human-history/) and TechCrunch attributed both to Hecht, and that attribution spread fast:

> "A Microsoft executive called the training of AI models the 'largest theft of labor in human history.' In another instance…that same executive called it 'an astonishing theft of unprecedented proportions.'"
>
> — [@neilturkewitz](https://x.com/neilturkewitz/status/2100913158500110480), Sep 18, 2026

The context matters more. The authors' statement of facts quotes the source of the "astonishing theft" phrase: a January 2023 Hecht paper titled "How to do foundation models the Microsoft way," written weeks after ChatGPT launched. It observes that developers "assume that any data on the internet can be used for model training . . . regardless of licenses, permissions, or consent," and warns this "may be very bad for Microsoft's business" because the public "will soon consider large models 'hoovering up' all their work to be an astonishing theft of unprecedented proportions." It is a forecast of public perception from a researcher arguing for "compliant foundation models," not a legal conclusion. That makes it weaker as a confession than the headlines suggested, and stronger as proof the risk was understood at Microsoft before Bing Chat shipped.

# How the record became public

The documents did not leak; they surfaced through a scheduling order. A stipulated sealing order entered September 3, restated in the parties' [October 9 proposed amendment](https://storage.courtlistener.com/recap/gov.uscourts.nysd.640396/gov.uscourts.nysd.640396.2095.0.pdf), gave anyone wanting material in the briefs kept sealed until September 14 to say why. By September 17 the parties "must publicly re-file their respective summary judgment briefs and Rule 56.1 statements, leaving unredacted all portions that no Party or third party has sought to seal."

Comparing the two public versions of the authors' brief measures what that deadline released. The September 5 version, [Dkt. 1881](https://authorsguild.org/app/uploads/2026/09/2026-09-05-Redaction-dckt-1881_0.pdf), was stamped "REDACTED – PUBLICLY FILED." Against the September 17 refiling:

| Passage in the authors' brief | Sep 5 public version (Dkt. 1881) | Sep 17 refiling (Dkt. 1982) |
|---|---|---|
| "We trained GPT-3 on pirated stuff! No sharing that!" | Redacted | Public |
| Altman: piracy "wrong and illegal"; Nadella: "Absolutely" | Redacted | Public |
| April 2019 LibGen disclosure to Bill Gates | Redacted | Public |
| "Project Clear" LibGen deletion | Redacted | Public |
| "Temporal regulatory arbitrage" Slack message | Redacted | Public |
| Microsoft's Hecht and the "doom loop" | Redacted | Public |
| OpenAI-Microsoft "horse trading" of data | Redacted | Public |
| Tarun Gogineni's "acceptable economic disruption" | Public | Public |
| AI-generated book market-share figures | Redacted | Still redacted |

![Ledger schematic: authors' brief passages redacted on Sep 5, public on Sep 17; book market-share data still sealed.](/post-images/2026-10-13-the-authors-case-discovery/redaction-ledger.jpg)

Two limits follow. First, what is public is the plaintiffs' *characterization* of the evidence: briefs quoting short excerpts. The exhibits themselves (Slack exports, Hecht papers, deposition transcripts) remain provisionally sealed until at least January 2027 under the parties' proposed schedule. Second, the defendants have not yet answered on the facts; their oppositions are due October 23. Until then, every quotation below is one the plaintiffs chose, framed as they framed it.

Daily News counsel Steven Lieberman told Ars the defendants had "insisted that these documents be treated as confidential so that the public could not see them. Well, now the cat is out of the bag."

# What OpenAI wrote to itself about LibGen

The authors' chronology is the core of the acquisition case. Every item below is an allegation supported by citations to the sealed record, as presented in [Dkt. 1982](https://authorsguild.org/app/uploads/2026/09/S.D.N.Y.-25-md-03143-dckt-001982_000-filed-2026-09-17.pdf) and [Dkt. 1987](https://authorsguild.org/app/uploads/2026/09/Class-Plaintiffs-SUF-9.17.26.pdf).

- **October and November 2018.** OpenAI employee Alec Radford downloads approximately 117,500 books from Library Genesis, a site the U.S. Trade Representative has listed as a "Notorious Market for Counterfeiting and Piracy" since 2017. The download includes 96 of the Asserted Works.
- **April 2019.** Preparing for a meeting with Bill Gates, Sam Altman on April 18 shares a document stating the model had grown nearly tenfold after OpenAI "added another ~11B words from Library Genesis (LibGen)." It links LibGen's Wikipedia page, which noted this same Manhattan court had ordered the site shut down in 2015. The plaintiffs say Microsoft CTO Kevin Scott also received it.
- **June 2019.** Microsoft invests $1 billion. It eventually invests $13 billion across three agreements.
- **July 19, 2019.** Sam McCandlish proposes dropping LibGen mentions from a scaling paper "since it's a bit of a sketchy data source." Dario Amodei, then research director, replies that "as a training set [LibGen is] a bit sketchier." McCandlish: "I was just worried about optics - i.e. 'openai uses copyrighted data from sketchy russian website' showing up on HN would be unfortunate."
- **August 19, 2019.** Internal notes: "We trained GPT-3 on pirated stuff! No sharing that!"
- **September 2019 to January 2020.** Benjamin Mann and Chris Hallacy torrent roughly 35 terabytes from LibGen, which the plaintiffs' expert says matches the size of LibGen's entire fiction and nonfiction collections, about 4.6 million books.
- **May 2020.** Preparing the GPT-3 paper, Amodei asks in Slack: "Is it sketchy to call our corpuses 'Books1' and 'Books2' and not say what they are, particularly when in fact they are Libgen." Asked by Jack Clark for more specificity, Mann answers that the description is "deliberately vague since it's libgen."
- **Summer 2022.** OpenAI deletes its LibGen holdings, an effort first run in a Slack channel named "excise-libgen" that then-general counsel Jason Kwon renamed "project-clear." The cleanup list ran to nearly 500 internal copies; the two LibGen compilations, roughly 570,000 books used for GPT-3 and GPT-3.5, are the only training corpora OpenAI has ever deleted.
- **November 30, 2023.** In a channel speculating that competitors train on LibGen, OpenAI's Nat McAleese writes: "Train on libgen, raise money, get new data, delete the old models, be clean forever more. Temporal regulatory arbitrage." Three colleagues react with a custom emoji, ":not-legal-but-very-cool:".
- **February 2026.** In deposition, Altman is asked, "You agree that piracy is wrong and illegal?" He answers, "I do." Satya Nadella, asked whether it is illegal to download pirated material, answers, "Absolutely. That's correct."

![Timeline schematic 2018 to 2026: LibGen downloads, Gates briefing, Books1/Books2 renaming, Project Clear deletion, 2026 depositions.](/post-images/2026-10-13-the-authors-case-discovery/libgen-timeline.jpg)

The brief adds that OpenAI obtained at least one full copy of Books3, a 196,640-book pirated dataset.

The authors' path from this sequence to liability is short. Downloading from a pirate library is a separate use from training, the brief argues, citing *Andy Warhol Foundation v. Goldsmith*'s instruction to analyze each use on its own. It leans hardest on Judge William Alsup's 2025 ruling in *Bartz v. Anthropic*, which involved the same pirate libraries and held that such piracy is "irredeemably infringing" even if the copies are later used for training and discarded. The Napster cases fill in the rest: getting "for free something they would ordinarily have to buy" is not fair use. In October 2025, Judge Stein had already [refused to strike](https://www.yahoo.com/news/articles/sdny-judge-denies-openai-bid-120823034.html) this "download claim" as a new theory.

One detail has drawn little comment. Several people whose 2019 and 2020 messages anchor this account (Dario Amodei, Jack Clark, Ben Mann, Sam McCandlish, and Tom Brown, whom the statement of facts quotes calling LibGen "sketchy AF") later left OpenAI and are listed among [Anthropic's co-founders](https://en.wikipedia.org/wiki/Anthropic). Anthropic is not a party here, and nothing in these filings concerns its own conduct. But it is the company that has already paid the largest bill on the acquisition theory, as the *Bartz* section below sets out.

# What Microsoft knew, and what its CEO said under oath

The authors' case against Microsoft rests on knowledge, through the April 2019 document, and participation, through data transfers a Microsoft employee called "horse trading."

Per the brief, OpenAI sent its LibGen1 and LibGen2 compilations to Microsoft in early 2021 under a GPT-3 access deal; Microsoft later deleted them. Microsoft first resisted sharing its Bing index over copyright concerns, then transferred full copies under "Project Taxi." In 2023, OpenAI paid Microsoft to crawl half the web and hand over the results, under "Project Mango." The plaintiffs' expert says 49 of the Asserted Works turned up in that crawl data.

![Schematic of OpenAI-Microsoft data trades: LibGen copies sent 2021, Bing index via Project Taxi, Project Mango crawl returned in 2023.](/post-images/2026-10-13-the-authors-case-discovery/data-horse-trading.jpg)

Nadella's deposition, as the authors present it, puts Microsoft's leadership in an awkward spot. He testified that he "wish[ed] that OpenAI had not gone to a pirated site" and that OpenAI "could have scanned the copy," as "Google did with Google Books." In the news brief he goes further: "anything that is paywalled should be licensed by anyone who wants to use it [whether] for … grounding or training," and had he known OpenAI trained on paywalled content, he would have "invoked [Microsoft's right to] require[] OpenAI to retrain its models." The same brief quotes Greg Brockman replying "ah nice" when a colleague described "a hack to get around nytimes paywall."

The Hecht documents, as the briefs present them, show a Microsoft scientist building an internal case over about a year. A December 2022 assessment warned Microsoft could be seen "as the Napster of AI." A September 2023 presentation called the arrangement with content producers an "unconventional bargain based around information asymmetry," with an appendix slide: "None of us willing to go to the author's convention and defend what we did." Then came the document both briefs quote:

> Our AI content strategy has started a doom loop that will hurt the performance of our models and the entire web at the same time. It is highly unusual that an end product threatens the economic foundations of its essential suppliers, but that is the situation we have created for our LLM business with respect to its content supply chain.
>
> — Brent Hecht, Microsoft document read into the record at deposition, quoted in [Class Plaintiffs' SUF ¶765, Dkt. 1987](https://authorsguild.org/app/uploads/2026/09/Class-Plaintiffs-SUF-9.17.26.pdf)

The news brief pairs that with Microsoft's own measurements: click-through rates for The Times and Daily News domains fell 83 to 93 percent on the Copilot "answer engine" versus traditional Bing search.

Microsoft's whole public position on the documents came through a spokesperson to Ars, who said Hecht's documents "reflect one employee's individual perspective, are not a legal analysis, and do not represent the company's views," and that Nadella's testimony addressed "broad principles and changes underway in how people find and consume information," observations that "should not be confused with conclusions about copyright questions before the Court, which Microsoft addresses in its filings." OpenAI did not respond to Ars or to TechCrunch.

# Substitution, in the companies' own words

If acquisition is the authors' strongest claim, substitution is their most ambitious. The fourth fair-use factor, market effect, is "the single most important element of fair use," in the Second Circuit's words from the *Internet Archive* case. The plaintiffs aim to prove market harm with the defendants' own forecasts.

The authors' brief opens with Jack Clark, OpenAI's policy director, writing in May 2020 on the eve of GPT-3's release:

> Our work on AI and Creativity is going to increasingly lead to us creating systems that substitute for the labor of [] people . . . The better we do on GPT-X, the more worried genre fiction authors will become about us substituting for them on Amazon . . . Our work in this area will make people unemployed . . . There will be a point where a bunch of artists express worry about what we're doing here and we'll likely ignore their concerns and release anyway . . .
>
> — Jack Clark, internal memo, May 2020, quoted in [Class Plaintiffs' Memorandum, Dkt. 1982](https://authorsguild.org/app/uploads/2026/09/S.D.N.Y.-25-md-03143-dckt-001982_000-filed-2026-09-17.pdf) (SUF ¶749)

The brief also quotes Tarun Gogineni, hired in 2022 to lead OpenAI's model writing quality work, who by its account viewed authors' complaints as "acceptable economic disruption" and posted in 2025 that his "research mission" was to have GPT models finish George R.R. Martin's series. Martin is a plaintiff. On the news side, ChatGPT head Nick Turley wrote that "[o]ur products are largely substitutive, period" and "will get more and more substitutive as they get better."

The legal theory these quotes serve comes from the other 2025 training decision, *Kadrey v. Meta*. Judge Vince Chhabria held Meta's training was fair use on his record, but wrote that "by training generative AI models with copyrighted works, companies are creating something that often will dramatically undermine the market for those works." He ruled for Meta, the authors note, because there was no "indication that … Meta developed [its LLM] with the purpose of enabling it to create books that compete with the plaintiffs'." The authors' bet is that Clark, Gogineni and a metric they say OpenAI tracked, "Writing WAU," supply that missing purpose.

A second argument turns OpenAI's own expert against it. Under *Warhol*, a defendant must justify copying the particular works it copied. OpenAI's technical expert, Chris Callison-Burch, opined that the Asserted Works were "not technically required" and that "no individual work was necessary for inclusion in the corpora." That is a sensible answer to the claim that any one book mattered. The plaintiffs read it as a concession that nothing required copying these books at all, noting that one model in suit, GPT-4o mini, was trained without any of them.

![Two-column schematic: authors' four fair-use factor arguments set against OpenAI's replies, with factor four highlighted.](/post-images/2026-10-13-the-authors-case-discovery/four-factor-ledger.jpg)

The community reading split along a predictable line. The top comment on the TechCrunch thread, from [jacquesm](https://news.ycombinator.com/item?id=49752242), called it "the robbery of all of our culture to sell it back to us at a mark-up." [pbasista](https://news.ycombinator.com/item?id=49752859) pushed back that the bots "were scraping publicly accessible internet" and that calling it theft was "disingenuous. Especially when coming from a company operating a web search engine." On the authors' thread, [natbennett](https://news.ycombinator.com/item?id=49867370) named the gap between strategy and grievance: OpenAI's belief that its models could replace genre writers "is legally useful because of the definition of fair use," but "the authors I know don't talk about it. They're very angry about their work having been used without compensation." The most-replied comment, from [Skyy93](https://news.ycombinator.com/item?id=49864549), called the Guild "a lobby organisation using only the pieces and bits they like," which is true of any advocacy brief.

The single most shared artifact on X was the shortest:

> "We trained GPT-3 on pirated stuff! No sharing that!"
>
> Astonishing admissions in these newly released documents in Authors Guild v OpenAI.
>
> — [@ednewtonrex](https://x.com/ednewtonrex/status/2101299362756170073), Sep 19, 2026

# The companies' case

OpenAI and Microsoft filed their own summary-judgment motions on September 4. Their briefs were not among the public documents reviewed for this piece; the account below relies on [Publishers Weekly's report](https://www.publishersweekly.com/pw/by-topic/industry-news/publisher-news/article/101196-authors-guild-co-plaintiffs-seek-summary-judgment-in-openai-case.html) of the filings and on the companies' statements.

OpenAI's core argument is that training changes the purpose of the material so thoroughly that it is fair use. Building a general-purpose model requires "hundreds of billions to tens of trillions of words from a multitude of diverse sources," and the result is "useful for an enormous range of tasks that have nothing to do with books." ChatGPT "does not display copies" of books; OpenAI puts the alleged regurgitation rate at 0.00007 percent and says the plaintiffs' expert "spent significant effort trying to brute force OpenAI's LLMs to reproduce text." Its answer to the substitution evidence is doctrinal: "competitive harm from new, non-infringing works is not harm that copyright protects against." It leans on *Authors Guild v. Google*, the Second Circuit's 2015 Google Books decision, and on *Kadrey*. Microsoft's spokesperson told Ars its AI products are a fair use that does not substitute for news sites.

The strongest version of the defense does not need to dispute a single quotation. Clark's memo and Turley's message describe a technology that competes with human writing. Whether that competition counts as harm under factor four is a legal question, and the leading district-court answer is *Bartz*, where Judge Alsup treated market dilution as outside what copyright protects. The authors concede the point by arguing *Bartz* was wrong. On this reading, the documents prove foresight of an effect copyright does not reach.

The defense also has the federal government. On September 1, the Justice Department filed a [statement of interest](https://storage.courtlistener.com/recap/gov.uscourts.nysd.641355/gov.uscourts.nysd.641355.316.0.pdf) in the MDL, signed by Michael Weisbuch, senior counsel to the associate attorney general. It says the fourth factor "heavily favors fair use" and that licensing requirements would "function primarily as large subsidies for old mainstream media companies." The next day, Commerce Secretary Howard Lutnick told G20 officials that countries should [allow AI training on creators' work](https://www.thestandard.com.hk/innovation/article/341720/US-urges-G20-countries-to-allow-AI-training-on-creators-work) while finding a way to "protect artists," without saying how.

Two features of the government's filing matter here. It was filed before the unsealing, a point the news plaintiffs pressed when they [asked Judge Stein](https://news.bloombergtax.com/litigation/news-outlets-ask-court-to-ignore-doj-view-in-ai-copyright-suit) on September 28 to "give the DOJ's statement no weight at all." And it is expressly about training. It notes that "[e]ach stage may present distinct questions of copyright law" and says the United States "focuses on the question whether the use of copyrighted works at the training stage" is fair use. Its 20 pages contain no discussion of pirate libraries. The administration's support covers the question the unsealed record barely touches, and says nothing about the one it addresses most.

# The split the law already drew

Line up the decided cases and the September record by the conduct at issue, and the pattern is plain.

| Conduct | Bartz v. Anthropic (N.D. Cal. 2025) | Kadrey v. Meta (N.D. Cal. 2025) | DOJ statement (Sep 2026) | What the OpenAI record adds |
|---|---|---|---|---|
| Training on lawfully obtained books | Fair use | Fair use on that record | Fair use | Little; experts say no single book was needed |
| Buying print books, scanning, discarding | Fair use | Not at issue | Not addressed | Nadella says OpenAI "could have scanned the copy" |
| Downloading from pirate libraries | "Inherently, irredeemably infringing" | Not decided as a separate use | Silent | LibGen, Books3, "No sharing that!", Project Clear |
| Foreseen market substitution | Dilution theory rejected | Dilution theory accepted in principle | "Heavily favors fair use" | Clark, Gogineni, Turley, Hecht |

![Matrix schematic: four kinds of AI data conduct against Bartz, Kadrey and DOJ positions; the pirate-download row marked as the exposure.](/post-images/2026-10-13-the-authors-case-discovery/conduct-matrix.jpg)

The price of the acquisition row is already set. Anthropic settled *Bartz* for $1.5 billion; Judge Araceli Martínez-Olguín, to whom the case was reassigned from Alsup on December 31, 2025, granted [final approval on July 20, 2026](https://blog.taaonline.net/2026/07/bartz-v-anthropic-granted-final-approval-by-the-court/), overruling all 53 objections. Class counsel reported [440,490 of 482,460 eligible works claimed](https://authorsguild.org/news/anthropic-settlement-update-91-percent-of-books-claimed/), 91.3 percent, at an estimated $2,931.62 per work before interest and deductions. Anthropic framed the deal narrowly: "The issue we settled on was about how some materials were acquired, not whether we could use them to develop [AI]," deputy general counsel Aparna Sridhar said, as [reported by Euronews](https://www.euronews.com/2026/08/05/project-panama-how-anthropic-secretly-destroyed-millions-of-books-to-train-its-ai) from The Washington Post.

The same case produced the window's other book story. Documents unsealed in *Bartz* described "Project Panama," Anthropic's plan, in one planning document's words, to "destructively scan all the books in the world," with a vendor saying the company sought to convert "from 500,000 to two million books over a six-month period." Alsup had already held that buying, scanning and discarding print books was fair use. Anna's Archive, the shadow library, turned the reporting into a fundraising appeal, "AI companies destroy physical books – let's scan rare books before it's too late," which drew [914 comments on Hacker News](https://news.ycombinator.com/item?id=49383026) on August 21. It comes from an operator of exactly the kind of pirate library at issue in both cases, with an interest in the outcome. But its premise lines up with Nadella's deposition: the lawful route the courts have blessed costs money and paper. The unlawful one was a torrent.

So the unsealed record lands mostly on the row with a known price and little doctrinal uncertainty. That raises OpenAI's exposure on the download claim and does little to the training defense, where an adverse ruling would reach every model developer. The substitution evidence is the exception, which is why the authors spend pages arguing that *Bartz* got market dilution wrong. If Judge Stein agrees, the internal documents stop being color and become evidence of market harm.

# What the record has and has not changed in public

Test the hypothesis against what the companies have said since September 17. The evidence that the record constrains them: Microsoft's only substantive public response has been to disown its own director of applied science and to recast its CEO's sworn testimony as "observations." A company whose scientist forecast "an astonishing theft" and whose CEO testified that paywalled content "should be licensed" for training can no longer suggest that nobody inside saw a problem. In court the constraint is sharper. Altman and Nadella have both agreed under oath that downloading pirated material is wrong or illegal, so the acquisition defense must rest on law, not good faith.

The evidence that it does not: on October 12, Microsoft president Brad Smith told the BBC that AI is less regulated than milk, [as reported by The Guardian](https://www.theguardian.com/technology/2026/oct/12/ai-less-regulated-than-milk-microsoft-president-calls-for-tougher-rules): "Every carton of milk that you see in the grocery store is subject to safety rules and inspection." He called for incident reporting, "emergency brakes" and [independent evaluators](/posts/2026-08-07-the-secret-framework/). Neither account of the interview records him saying anything about copyright or training data. Microsoft's public appetite for rules covers what AI systems do, not what goes into them, and the federal government is arguing in this courtroom and at the G20 that the input question should go the industry's way.

Nor has the record slowed new suits. On September 4, The Seattle Times and Newsday [sued OpenAI and Microsoft](https://engadget.com/2251707/seattle-times-newsday-sue-openai-microsoft-for-copyright-infringement) in the same court, alleging paywall-bypassing scraping and removal of copyright management information. The trust problem reaches users too: a September 10 post, "OpenAI keeps re-enabling the 'allow training' setting," one user's unverified report that a disabled opt-out had switched back on, [drew 487 points](https://news.ycombinator.com/item?id=49643556). Polymarket's contract on whether the [Supreme Court accepts an AI copyright case before 2027](https://polymarket.com/event/supreme-court-accepts-ai-copyright-case-before-2027) stood at 13 percent at the time of writing.

# The calendar the companies are negotiating against

The next public disclosures are scheduled, and they will put the defendants' version of these documents on the record for the first time.

| Date | Event | Source |
|---|---|---|
| Oct 23, 2026 | Oppositions to summary judgment due (Dkt. 2049) | [Dkt. 2095](https://storage.courtlistener.com/recap/gov.uscourts.nysd.640396/gov.uscourts.nysd.640396.2095.0.pdf) |
| Oct 29, 2026 | Daubert oppositions due; opposition briefs publicly refiled | Dkt. 2095 (refiling date proposed) |
| Oct 30, 2026 | Amicus briefs due (Dkt. 2049) | Dkt. 2095 |
| Nov 20, 2026 | Replies due (Dkt. 2049) | Dkt. 2095 |
| Dec 7, 2026 | Replies publicly refiled | Dkt. 2095 (proposed) |
| Jan 13, 2027 | Motions to keep summary-judgment exhibits sealed | Dkt. 2095 (proposed) |
| Jan 27, 2027 | Exhibits nobody moved to seal filed publicly | Dkt. 2095 (proposed) |
| Early 2027 | Hearing expected, per the Authors Guild | [Authors Guild](https://authorsguild.org/news/ag-v-openai-top-execs-knew-mass-book-piracy-was-illegal/) |

![Calendar schematic Oct 2026 to Jan 2027: oppositions, public refiling, replies, exhibit sealing motions and the January 27 exhibit release.](/post-images/2026-10-13-the-authors-case-discovery/disclosure-calendar.jpg)

October 29 is the date to watch: OpenAI's and Microsoft's answers to the LibGen chronology and the Hecht documents become public, including whatever context they can attach to "No sharing that!" and "temporal regulatory arbitrage." January 27 is when the underlying documents, not the plaintiffs' excerpts, may finally appear. Both will test whether the quotations hold up in full context, which the September coverage could not do.

For anyone building on these models, the practical reading is narrower than the headlines. On its own, the unsealed record does little to make training on copyrighted text unlawful; the government's filing and two district courts point the other way, and the substitution theory depends on Judge Stein rejecting *Bartz* on market dilution. What it does is make the provenance of training data a liability that can be tied to named people, specific dates and Slack channels. Anthropic has already shown what that costs per book. OpenAI's own employees, by the authors' account, renamed the dataset, deleted the copies and joked about regulatory arbitrage, which suggests they understood the same thing six years earlier.

## Sources

- [In re OpenAI Copyright Infringement Litigation, 25-md-3143 (S.D.N.Y.), Dkt. 1982: Class Plaintiffs' Memorandum of Law ISO Partial Summary Judgment (filed Sep 17, 2026)](https://authorsguild.org/app/uploads/2026/09/S.D.N.Y.-25-md-03143-dckt-001982_000-filed-2026-09-17.pdf)
- [Dkt. 1987: Class Plaintiffs' Corrected Rule 56.1 Statement of Undisputed Material Facts (filed Sep 17, 2026)](https://authorsguild.org/app/uploads/2026/09/Class-Plaintiffs-SUF-9.17.26.pdf)
- [Dkt. 1881: Class Plaintiffs' Memorandum, redacted public version (filed Sep 5, 2026)](https://authorsguild.org/app/uploads/2026/09/2026-09-05-Redaction-dckt-1881_0.pdf)
- [Dkt. 1977-1: News Plaintiffs' Combined Summary Judgment Brief (filed Sep 17, 2026), via CourtListener](https://storage.courtlistener.com/recap/gov.uscourts.nysd.640396/gov.uscourts.nysd.640396.1977.1.pdf)
- [Statement of Interest of the United States, Dkt. 316 (filed Sep 1, 2026), via CourtListener](https://storage.courtlistener.com/recap/gov.uscourts.nysd.641355/gov.uscourts.nysd.641355.316.0.pdf)
- [Dkt. 2095: Proposed Amended Stipulated Omnibus Sealing Order for Summary Judgment and Daubert Briefing (filed Oct 9, 2026)](https://storage.courtlistener.com/recap/gov.uscourts.nysd.640396/gov.uscourts.nysd.640396.2095.0.pdf)
- [Authors Guild: Unsealed Briefs in Authors' Case v. Microsoft/OpenAI (Sep 21, 2026)](https://authorsguild.org/news/ag-v-openai-top-execs-knew-mass-book-piracy-was-illegal/)
- [Authors Guild: Plaintiffs file motion for summary judgment v. OpenAI and Microsoft (Sep 5, 2026)](https://authorsguild.org/news/plaintiffs-file-motion-for-summary-judgment-v-openai-and-microsoft/)
- [TechCrunch: Microsoft exec called AI scraping 'the largest theft of labor in human history' (Sep 17, 2026)](https://techcrunch.com/2026/09/17/microsoft-exec-called-ai-scraping-the-largest-theft-of-labor-in-human-history-new-unredacted-filings-reveal/)
- [Ars Technica: Microsoft exec called AI scraping the 'largest theft of labor in human history' (Sep 17, 2026)](https://arstechnica.com/tech-policy/2026/09/microsoft-exec-called-ai-scraping-the-largest-theft-of-labor-in-human-history/)
- [Publishers Weekly: Authors Guild, Co-Plaintiffs Seek Summary Judgment in OpenAI Case (Sep 9, 2026)](https://www.publishersweekly.com/pw/by-topic/industry-news/publisher-news/article/101196-authors-guild-co-plaintiffs-seek-summary-judgment-in-openai-case.html)
- [Bloomberg Law: News Outlets Ask Court to Ignore DOJ View in AI Copyright Suit (Sep 28, 2026)](https://news.bloombergtax.com/litigation/news-outlets-ask-court-to-ignore-doj-view-in-ai-copyright-suit)
- [Reuters via The Standard: US urges G20 countries to allow AI training on creators' work (Sep 2, 2026)](https://www.thestandard.com.hk/innovation/article/341720/US-urges-G20-countries-to-allow-AI-training-on-creators-work)
- [Engadget: Seattle Times and Newsday sue OpenAI and Microsoft (Sep 5, 2026)](https://engadget.com/2251707/seattle-times-newsday-sue-openai-microsoft-for-copyright-infringement)
- [The Guardian: AI less regulated than milk, Microsoft president says (Oct 12, 2026)](https://www.theguardian.com/technology/2026/oct/12/ai-less-regulated-than-milk-microsoft-president-calls-for-tougher-rules)
- [Authors Guild: Anthropic settlement update, 91.3 percent of books claimed (Apr 17, 2026)](https://authorsguild.org/news/anthropic-settlement-update-91-percent-of-books-claimed/)
- [TAA: Bartz v. Anthropic granted final approval by the court (Jul 22, 2026)](https://blog.taaonline.net/2026/07/bartz-v-anthropic-granted-final-approval-by-the-court/)
- [Euronews: Project Panama, how Anthropic destroyed millions of books to train its AI (Aug 5, 2026)](https://www.euronews.com/2026/08/05/project-panama-how-anthropic-secretly-destroyed-millions-of-books-to-train-its-ai)
- [Decrypt via Yahoo: SDNY judge denies OpenAI bid to strike authors' book-download claim (Oct 29, 2025)](https://www.yahoo.com/news/articles/sdny-judge-denies-openai-bid-120823034.html)
- [Hacker News: Microsoft exec called AI scraping 'the largest theft of labor in human history' (955 points)](https://news.ycombinator.com/item?id=49752056)
- [Hacker News: Unsealed Briefs in Authors' Case v. Microsoft/OpenAI (632 points)](https://news.ycombinator.com/item?id=49863864)

---

Canonical: https://www.thedeepfeed.ai/posts/2026-10-13-the-authors-case-discovery/
Site: https://www.thedeepfeed.ai
Full corpus: https://www.thedeepfeed.ai/llms-full.txt