Apple tells a US court it trained a research model on a books dataset, and calls it fair use

Apple has told a United States court that it used a subset of the RedPajama dataset containing books to train its OpenELM language models, and that the conduct at issue is fair use rather than copyright infringement.

The admission appears in Apple's answer to the consolidated class action in the US District Court for the Northern District of California, filed on 30 March 2026, according to the litigation tracker Chat GPT Is Eating the World, which published the filing and reported the admission on 31 March 2026.

Apple also drew a line the plaintiffs will have to fight over. In the same filing, according to that report, Apple characterises OpenELM as a non-commercial research model that is not part of Apple Intelligence, the consumer feature set to which the lawsuit has been publicly attached since it was brought.

The lead complaint was filed on 5 September 2025 by the novelists Grady Hendrix and Jennifer Roberson, Publishers Weekly reported at the time. Keller Rohrback, one of the firms now running the case, says on its case page that Hendrix was consolidated with Martinez-Conde v Apple and Alexander v Apple on 14 November 2025, and that the court appointed William K. Dreher of Keller Rohrback and Rohit Nath of Susman Godfrey as interim co-lead counsel on 30 January 2026. A consolidated complaint followed on 13 February 2026, adding John Hornor Jacobs and Eboni McKinnon as named plaintiffs, so the case is no longer the two-author action it began as.

The central allegation concerns Books3, a compilation of roughly 196,640 copyrighted titles said to derive from the shadow library Bibliotik, according to ClassAction.org's summary of the filing. Publishers Weekly and eWeek both quoted the complaint as saying that Apple used Books3 to train OpenELM and "also likely" trained its Foundation Language Models on the same dataset.

That hedge matters. The Foundation Language Models are the ones inside Apple Intelligence. On the pleadings as reported, the plaintiffs infer rather than assert that connection, and Apple's answer denies it. Whether discovery closes the gap between a research model and a shipped product is the practical question in the case, and it is the question that determines how large the exposure is.

The authors are working from a template that has already produced the largest copyright recovery on record. In June 2025 Judge William Alsup held that training on books was fair use but that downloading and retaining pirated copies from shadow libraries was not, according to Norton Rose Fulbright's March 2026 survey of AI copyright litigation. That split is what made Anthropic's position untenable.

The resulting settlement is now final. The Authors Guild reported that Judge Araceli Martinez-Olguin granted final approval to the US$1.5 billion Bartz v Anthropic settlement on 20 July 2026, at roughly US$3,000 a work, a figure the court described as four times the US$750 statutory minimum for ordinary infringement, with attorneys' fees of about US$101.56 million, or 6.8 per cent of the fund.

Apple's case will not move at that speed. Keller Rohrback says the court has not set a trial date. PacerMonitor lists Judge Yvonne Gonzalez Rogers as the assigned judge. Chat GPT Is Eating the World reported on 22 April 2026 that the parties' proposed schedule runs into 2028.

The wider board has shifted too. Norton Rose Fulbright records that the Supreme Court denied certiorari in Thaler v Perlmutter on 2 March 2026, leaving the human authorship requirement intact; that Thomson Reuters v Ross Intelligence is on appeal to the Third Circuit after the district court rejected a fair use defence in February 2025; that in Kadrey v Meta the court granted partial dismissal on fair use grounds for training while claims over torrent seeding remain live; and that the consolidated OpenAI copyright proceeding is in discovery after a motion to dismiss was denied on 27 October 2025.

Apple's own AI direction has meanwhile moved outside its walls. AppleInsider reported on 22 April 2026 that Google Cloud's chief executive Thomas Kurian said his company is working with Apple to develop the next generation of Apple Foundation Models based on Gemini technology. That does not retire liability for what was trained earlier.

What remains unresolved is narrow and consequential. The court must decide whether training a research model that never ships in a product is treated differently from training one that does, and whether the source of the copies weighs more heavily than the use made of them. Apple has not yet had to defend the Foundation Language Models on the merits, because the plaintiffs have not yet shown they were trained on Books3. Until discovery settles that, the size of Apple's exposure is unknown, and on the current schedule it will stay unknown for at least another year.