🔍 Read the full analysis: Could Claude’s Training Data Become An IPO Issue For Anthropic? on ThorstenMeyerAI.com
Prime made for students and young adults
- Fast, free delivery for dorm and study essentials
- Prime Video and Amazon Music included
- Member-only deals
TL;DR
Anthropic filed to go public as disputes over its use of copyrighted books continue, putting the company’s training-data history under investor scrutiny. A 2025 ruling found that training on books could be fair use but rejected keeping pirated copies; later lawsuits remain unresolved, and Anthropic’s S-1 does not say Claude is sentient or can blackmail people.
Anthropic filed to go public on October 27 as authors’ copyright claims over books copied for its AI development remain active, bringing the company’s training-data practices into focus for prospective investors. The record does not show a new court ruling that Anthropic illegally trained Claude: a 2025 judge found model training on books fair use, while ruling that keeping pirated copies was not.
The distinction matters because public debate has sometimes compressed two separate issues into the phrase “illegal training.” In Bartz v. Anthropic, U.S. District Judge William Alsup’s June 23, 2025 order said training Claude on books was fair use. He separately found that downloading and retaining pirated books to build a permanent library was not fair use. The case settled before a jury decided damages.
The settlement was approved in July 2026 for $1.5 billion, covering roughly 500,000 registered books downloaded from the LibGen and PiLiMi collections. That works out to about $3,000 per book before fees. The settlement requires destruction of the downloaded collections, subject to legal-hold obligations, and releases specified past claims; it does not grant a general license for future training or release claims about Claude’s outputs.
Anthropic disputes wrongdoing. In the settlement, the company represented that neither the LibGen nor PiLiMi datasets, nor portions of them, were in the training corpus of any commercially released model. That statement sits alongside the court record’s account of pirated copies being downloaded and portions of the collections being used to train some models. The distinction between copies used for training and the contents of released models is central to understanding the dispute.
From download to IPO
Five years, one ruling, no trial
The piracy happened in 2021 and 2022. The only ruling came in 2025. The new lawsuits are years away from a verdict.
- Jun 2021 · Jul 2022DownloadsAt least 5 million copies from LibGen (2021), at least 2 million from PiLiMi (2022); 196,640 from Books3 earlier.
- Aug 2024Bartz v. Anthropic filedThree authors sue Anthropic, on behalf of a class of book owners.
- 23 Jun 2025Judge Alsup's orderTraining on books: fair use. Scanning bought books: fair use. Pirated library copies: not fair use, to trial.
- Sep 2025$1.5 billion settlementNo trial. Anthropic denies wrongdoing; release covers conduct up to 25 Aug 2025 only.
- 13 + 15 May 2026Opt-out suitsCruz (28 plaintiffs) and Kwon (incl. Chang-rae Lee) join Cambronne and two publisher suits.
- Jul 2026Final approvalSettlement approved. Anthropic: no opt-out deal "above the Bartz claim value".
- 24 Sep 2026Motion to dismissAnthropic: claims too late, no infringing output shown, many works already released.
- 26–27 Oct 2026NowS-1 and the headlineIPO filing cites Bartz as a risk; The Stanford Daily names Stanford authors.
- 17 Dec 2026ScheduledHearingCourt hears the motion to dismiss.
- May 2027 · Jan 2028ScheduledDiscovery, then trial settingFact discovery closes 13 May 2027; trial-setting conference 11 Jan 2028.
Sources: Bartz v. Anthropic, order of 23 June 2025; settlement FAQ; joint case statement, 27 Aug 2026; Anthropic letter, 27 Jul 2026; case management order, 14 Sep 2026; Anthropic S-1, 26 Oct 2026; The Stanford Daily, 27 Oct 2026. Scheduled dates can move.
Powered by Thorsten Meyer AI
Copyright Risk in the IPO Filing
For investors, the issue is not simply whether a model learned from books. The 2025 ruling treated training as fair use on the record before it, but found the acquisition and retention of pirated copies unlawful. The resulting settlement shows that data sourcing can carry substantial legal costs even where the court’s treatment of model training itself is favorable.
Anthropic’s S-1 filing also makes the dispute relevant to how the company describes business risks to public-market investors. The source account says the filing discusses copyright exposure and concerns about model behavior. It does not state that Claude is sentient or that the company has confirmed it can blackmail people. It describes behavior “resembling blackmail” in a staged safety test, and notes that debate about AI sentience could prompt regulation and that others have raised existential concerns.
That precision matters: claims in a filing are not court findings, and risk disclosures do not establish that a feared event has happened. Investors and readers need to distinguish the settled past claims from the new lawsuits still being litigated, as well as Anthropic’s statements from judicial decisions.
Popular right now
Apple AirPods Pro 3 Wireless Earbuds with Active Noise Cancellation
KitchenAid Artisan Series 5-Quart Tilt-Head Stand Mixer, Dried Rose (KSM150PS)
Samsung 42-Inch S90H OLED 4K Smart TV Bundle
As an affiliate, we earn on qualifying purchases.
The original case, Bartz v. Anthropic, was filed by three authors in August 2024. Alsup’s 2025 order distinguished books Anthropic bought and scanned from pirated copies it downloaded. He described the company’s scanning of purchased books as fair use, but found that downloading more than seven million copies from Books3, LibGen and PiLiMi to create a permanent library was not. He wrote that buying a copy later would not erase liability for the theft.
The case settled rather than proceeding to a jury trial on damages. The agreement covers claims for past conduct through August 25, 2025, according to the source material. In May 2026, authors filed separate opt-out lawsuits; that was not a month in which a judge ruled that Anthropic had illegally trained Claude. A hearing on Anthropic’s motion to dismiss those claims is set for December 17, 2026.
A Stanford Daily report named authors whose books it said were among the pirated copies, including Tobias Wolff, Abraham Verghese, Chang-rae Lee and Daniel Mason. Those identifications should be treated as the newspaper’s reporting, not as a finding in the court order. The source account says Chang-rae Lee is listed as a plaintiff in one new suit, while it could not verify Wolff as a plaintiff in the five related cases reviewed through August 2026.
““spectacularly so””
— Judge William Alsup, in the June 23, 2025 order
Open Questions in the Copyright Disputes
The new lawsuits have not been decided. Anthropic’s motion to dismiss is pending, and the December hearing could shape which claims proceed. It remains unclear whether the plaintiffs will establish that specific books were used in ways that support their claims, and how a court would assess any requested damages.
The exact relationship between pirated copies, training datasets and commercially released models is also disputed in public descriptions. Anthropic’s settlement representation addresses the training corpus of commercially released models; it does not by itself answer every question about earlier research models, dataset preparation or other alleged uses. The settlement’s scope is limited to specified past claims, rather than serving as a blanket authorization for training on copyrighted material.
The source account also says it did not verify the Stanford Daily’s book list against the settlement database. The list and any claim that a particular author is a plaintiff should therefore be described with attribution. The IPO filing’s full implications for the company’s legal exposure depend on the filing’s disclosures and the outcome of pending cases—not on viral summaries of its language.
December Hearing and IPO Review
The next scheduled court milestone is the December 17, 2026 hearing on Anthropic’s motion to dismiss the new authors’ claims. Until then, those claims remain allegations, not findings of liability. The settlement’s remaining payments are scheduled to continue under its terms, including a final $450 million installment by September 27, 2027, according to the source material.
As the IPO process proceeds, investors can review Anthropic’s S-1 for its account of legal, regulatory and model-safety risks. Any later amendment, court decision or settlement in the opt-out cases could change the picture. For now, the confirmed record is narrower: the 2025 court order distinguished fair-use training from unlawful retention of pirated copies, the related class action settled, and separate author claims have yet to be resolved.
Key Questions
Did a judge rule in 2026 that Anthropic illegally trained Claude?
No. The relevant ruling was issued in June 2025 in Bartz v. Anthropic. Judge Alsup found the training at issue to be fair use but ruled that downloading and retaining pirated books was not. The newer lawsuits have not been decided.
What does the $1.5 billion settlement cover?
The settlement covers specified past claims involving roughly 500,000 registered books downloaded from LibGen or PiLiMi. It requires destruction of those downloaded books, subject to legal holds, and does not grant a general license for future training.
Does Anthropic admit that pirated books were in Claude’s commercial training data?
Anthropic disputes wrongdoing and represented in the settlement that the LibGen and PiLiMi datasets, or portions of them, were not in the training corpus of any commercially released model. The court record separately describes pirated copies being downloaded and portions of the collections being used to train some models. Those statements address different aspects of the dispute.
Does Anthropic’s IPO filing say Claude is sentient or can blackmail people?
Not according to the source account. The filing reportedly discusses behavior “resembling blackmail” in a staged safety test, the possibility of regulation connected to debate about AI sentience, and existential concerns raised by others. Those are risk disclosures, not declarations that Claude is sentient or has blackmailed a person.
Source: ThorstenMeyerAI.com
