Who owns what AI learned from?

The full story

The episode's narration, word for word. Numbers link to the sources below; timestamps play that moment.

The question 0:00

Last time, we ended on a simple fact: these models were trained on other people's work. In 2025, Anthropic, the company that makes Claude, agreed to pay one and a half billion US dollars over about half a million books ^1,2. Other companies are still in court ^3,4,5. So who owns what AI learned from? This is about law that is still being written, and none of it is legal advice.

Where the copying happens 0:29

Season six showed how models learn from huge amounts of text and pictures. To do that, companies collect copies, store them and feed them through training ^6,7. What's left is a model: a large file of numbers ⁷. So the fights are about three things: the copies made to collect and train, whether the finished model is itself a copy, and whether what comes out copies someone's work ^6,7,8.

What copyright protects 0:56

First, the basics. Copyright doesn't protect ideas or information, only the way they're expressed ⁹. What you can't do is copy someone else's actual words without permission, unless the law makes an exception ⁹. And the exceptions are where the two countries in this story split ^10,9.

America's open test 1:17

In the United States, the exception is called fair use, and it's an open test with four factors ¹⁰. Why are you using it? What kind of work is it? How much did you take? And does your use hurt the market for the original ¹⁰? Courts favour uses that are transformative: they add something new, with a different purpose, and don't substitute for the original ¹⁰.

Australia's closed list 1:43

Australia has no fair use ^9,11. It has fair dealing, which only covers a fixed list of purposes, like research, criticism, news and parody ⁹. Training a commercial AI model isn't on the list ⁹. In October 2025, the Attorney-General ruled out adding an exception that would let AI developers use Australian creators' work for free and without permission ¹².

A library of all the books 2:09

The case that settled began with three authors suing Anthropic ⁶. In June 2025, Judge William Alsup set out the evidence ⁶. In 2021 and 2022, Anthropic downloaded more than seven million copies of books from pirate websites, including at least five million from one called Library Genesis ⁶. The judge wrote that Anthropic preferred to steal them rather than buy them, to avoid what its chief executive, Dario Amodei, called a legal, practice and business slog ⁶.

Then it bought and scanned 2:44

Later, Anthropic changed course ⁶. It spent many millions of dollars buying millions of print books, often second-hand ⁶. Contractors stripped off the bindings, scanned every page and threw the paper away ⁶. The aim, in Anthropic's words quoted by the court, was a central library of all the books in the world, kept forever ⁶.

One ruling, three answers 3:07

The judge split the case three ways ⁶. Training on the books was fair use, he ruled, calling it quintessentially transformative ⁶. Like any reader aspiring to be a writer, he wrote, the models learned from the books to create something different, not to replace them ⁶. Scanning books it had bought was fair use too ⁶. But the pirated library was not, and buying a book later wouldn't undo the theft ⁶. That part was headed for trial ⁶.

One and a half billion dollars 3:39

Instead, in September 2025, Anthropic agreed to settle for one and a half billion dollars ². In July 2026, a different judge gave final approval ¹. It covers four hundred and eighty-two thousand works, about three thousand dollars each before costs and fees ¹. US law sets damages of seven hundred and fifty dollars to a hundred and fifty thousand per work, so a trial could have cost far less, or far more ^1,13. By April 2026, ninety-one per cent of the books had been claimed ¹.

What the money covers 4:17

What did that money buy? Only the past ¹. The authors gave up claims about downloading and copying their books up to August 2025 ¹. They kept any claims about what Claude outputs, and about anything Anthropic does after that ¹. Anthropic must destroy the files from Library Genesis and the other big pirate library ¹. It told the court those two collections were never in the training data of any model it released, though Judge Alsup had found parts of them were used to train models in development ^6,1.

Two days later, Meta wins 4:51

Two days later, in the same court, Meta, which had also downloaded books from pirate libraries, won the training part of a similar case brought by thirteen authors ¹⁴. But Judge Vince Chhabria's ruling read more like a warning ¹⁴. In many circumstances, he wrote, it will be illegal to train AI on copyrighted work without permission ¹⁴. Meta won only because these authors made the wrong arguments, and the ruling covered just those thirteen authors ¹⁴.

A reader, or a flood? 5:22

On the market, Alsup said a flood of competing books would be like teaching schoolchildren to write well, and that copyright doesn't protect authors against competition ⁶. Chhabria replied that this is not remotely like a product one person could use to churn out countless competing works ¹⁴. He called that harm market dilution, and said these authors simply hadn't proved it ¹⁴.

The first appeals court 5:47

In September 2026, a US appeals court ruled that a start-up called Ross wasn't making fair use when it copied more than two thousand legal summaries from Westlaw to train its search tool ^15,16. It wasn't fair, the court said, because Ross was building a direct rival to Westlaw ¹⁵. But this AI only fetched existing passages, and the court stressed it wasn't generative AI ^15,16.

The New York Times case 6:14

The New York Times case is still waiting ^15,3. The New York Times sued OpenAI and Microsoft in December 2023, saying they copied its journalism to train their models ^3,17. In 2025, the judge let most of the case go ahead ^3,18. In September 2026, in the combined case, the US Justice Department filed a statement backing the argument that training models like these is transformative ¹⁵. As of October 2026, no court has ruled on fair use in that case ^15,3.

Is a copy inside the model? 6:48

In Britain, Getty Images sued Stability AI over pictures ⁷. Getty had no evidence the training happened in the UK, so it dropped that claim ^7,19. In November 2025, the High Court found the model's weights don't store Getty's images, so the model isn't an infringing copy ⁷. In December 2025, Getty was given permission to appeal ¹⁹. One week later, a court in Munich reached the opposite kind of answer ⁸. ChatGPT had memorised song lyrics so well it could recite them, so the court said the lyrics were copied into the model ⁸.

Pictures and music 7:26

Artists sued Stability AI and Midjourney in January 2023, and as of October 2026 that case is still unresolved ⁵. In June 2024, the three biggest record companies sued the song generators Suno and Udio ^20,4. Late in 2025, Universal settled with Udio, and Warner settled with both, in deals for licensed music models ²⁰. Other claims against Suno are still in court ^20,4.

Paying, and opting out 7:59

Some companies pay ²¹. In 2023, OpenAI licensed part of the Associated Press archive ²¹. In 2025, the US Copyright Office said licensing deals were growing fast in some fields, but were patchy ²². Others offer an opt-out ^23,24,25. A website can add a line to a small file called robots.txt, to ask a company's crawler not to use its pages for training ^23,24,25. Google's and Anthropic's own wording talks about future training, not about what they've already collected ^24,25. And Australia's SBS says these signals are routinely ignored or bypassed ²⁶.

Australia, 2026 8:41

In 2026, the question came back to Canberra ^27,28. Submissions to a parliamentary committee on AI closed in September, and it must report by the end of November ²⁸. Anthropic accepted the exception had been ruled out, and proposed letting AI companies train on Australian work unless creators opt out, which it called a narrow form of conditional approval ²⁷. OpenAI asked for a balanced framework that lets models learn from publicly available information ²⁹. SBS replied that the onus should be on AI companies to avoid wrongdoing ²⁶.

The strongest case, both ways 9:17

For training without permission, as Judge Alsup and the US Justice Department argue: the models learn from books in order to make something new, and the UK court found no copies inside ^6,15,7. Against, as the Authors Guild and Judge Chhabria argue: companies profit from millions of works they didn't pay for, and the outputs can compete with them ^30,14. In 2025, the US Copyright Office said some training will be fair use and some won't ²².

And who owns what it makes? 9:49

The other direction is clearer. In 2025, a US appeals court ruled that a computer program couldn't be the author of a picture it made, because the law requires a human author ³¹. In March 2026, the Supreme Court declined to hear an appeal ³². The US Copyright Office says a human must decide the expressive parts, and that prompts alone aren't enough ³³. In Australia, a 2010 court case on computer-made phone directories suggests the same: no human author, no copyright ³⁴.

Where it stands 10:24

So, as of October 2026, who owns what AI learned from? The authors still own their books ^1,9. So far, whether a company may train on them without asking has depended on the country, on how the copies were obtained, and on what the model can reproduce ^6,14,7,8. The courts haven't settled it, and Australia's parliament is still deciding ^28,15. How much energy and water does AI really use?

Sources

Every factual claim in the episode is tied to one of these. Spotted an error? Tell us.

1. Bartz v. Anthropic PBC, No. 3:24-cv-05417-AMO (N.D. Cal.), Order Granting Final… — cdn.arstechnica.net
2. JURIST, "Anthropic agrees to $1.5B settlement in AI piracy lawsuit" (7 September 2025) — jurist.org
3. Axios, "NYT case against OpenAI and Microsoft can advance" (1 April 2025) — axios.com
4. CourtListener, UMG Recordings, Inc. v. Suno, Inc., 1:24-cv-11611 (D. Mass., filed 24… — courtlistener.com
5. CourtListener, Andersen v. Stability AI Ltd., 3:23-cv-00201 (N.D. Cal., filed 13… — courtlistener.com
6. Bartz v. Anthropic PBC, No. C 24-05417 WHA (N.D. Cal.), Order on Fair Use, Judge… — storage.courtlistener.com
7. Getty Images (US) Inc and others v Stability AI Ltd [2025] EWHC 2863 (Ch), Joanna… — judiciary.uk
8. MediaLaws, "Gema v. OpenAI – Decision of the Munich Regional Court" (Munich Regional… — medialaws.eu
9. Attorney-General's Department (Australia), "Copyright basics" (read 16 October 2026) — ag.gov.au
10. U.S. Copyright Office, "U.S. Copyright Office Fair Use Index" (read 16 October 2026) — copyright.gov
11. Australian Law Reform Commission, Copyright and the Digital Economy (ALRC Report… — alrc.gov.au
12. Attorney-General the Hon Michelle Rowland MP, "Albanese Government to ensure Australia… — ministers.ag.gov.au
13. U.S. Copyright Office, Copyright Law of the United States, Chapter 5, 17 U.S.C.… — copyright.gov
14. Kadrey v. Meta Platforms, Inc., No. 23-cv-03417-VC (N.D. Cal.), Order on cross-motions… — storage.courtlistener.com
15. Thomson Reuters Enterprise Centre GmbH v. ROSS Intelligence Inc., No. 25-2153 (3d… — lawnext.com
16. B. Ambrogi, "'No More Than an Ordinary Copyright Case': 3rd Circuit Unseals Its… — lawnext.com
17. CourtListener, The New York Times Company v. Microsoft Corporation, 1:23-cv-11195… — courtlistener.com
18. B. Allyn, "Judge allows 'New York Times' copyright case against OpenAI to go forward",… — nhpr.org
19. H. Jewson and E. Roberts, "Getty Images v Stability AI – Getty granted permission to… — burges-salmon.com
20. Music Business Worldwide, "Warner Music Group strikes 'landmark' deal with Suno;… — musicbusinessworldwide.com
21. Associated Press, "AP, Open AI agree to share select news content and technology in… — ap.org
22. U.S. Copyright Office, Copyright and Artificial Intelligence, Part 3: Generative AI… — copyright.gov
23. OpenAI, "Overview of OpenAI Crawlers" (platform docs, read 17 October 2026) — platform.openai.com
24. Google Search Central, "Google's common crawlers": Google-Extended (read 17 October 2026) — developers.google.com
25. Anthropic, "Does Anthropic crawl data from the web, and how can site owners block the… — support.claude.com
26. SBS, submission to the Joint Select Committee on Artificial Intelligence (2026) — sbs.com.au
27. J. Butler, "Anthropic pushes for opt-out model for Australian content as ABC warns of… — theguardian.com
28. Parliament of Australia, Joint Select Committee on Artificial Intelligence: terms of… — aph.gov.au
29. Capital Brief, "ChatGPT, Claude makers OpenAI and Anthropic formalise push for… — capitalbrief.com
30. The Authors Guild, statement on the proposed Anthropic settlement (September 2025) — authorsguild.org
31. Thaler v. Perlmutter, No. 23-5233 (D.C. Cir., decided 18 March 2025) — media.cadc.uscourts.gov
32. SCOTUSblog, Thaler v. Perlmutter, No. 25-449: petition for certiorari denied 2 March 2026 — scotusblog.com
33. U.S. Copyright Office, NewsNet Issue 1060, "Copyright Office Releases Part 2 of… — copyright.gov
34. Telstra Corporation Limited v Phone Directories Company Pty Ltd [2010] FCAFC 149 (Full… — judgments.fedcourt.gov.au

Researched and scripted with AI assistance, fact-checked claim by claim, with synthetic narration and diagrams drawn in code. How we make episodes.