AI companies argue that training on copyrighted books is transformative. But does learning from a book justify reproducing it without permission?
I received an email from a law firm informing me that one of my books had been identified in at least one of three shadow library indexes.
A shadow library is an online collection that provides access to books, research papers, and copyrighted materials, without authorization from the copyright holders. Library Genesis, or LibGen, Z-Library, and Books3 are examples of datasets associated with unauthorized books.
The law firm was pursuing copyright infringement claims against AI companies that may have used pirated books to train their AI models.
My reaction wasn't about how much money I might receive.I wanted to know what the evidence has established. Was my book illegally downloaded? Was the text copied? Was that text then used to train an AI model?
In a previous episode, I talked about how I investigated a claim involving music being stolen for AI training.
In that investigation, I discovered that the databases cited contained metadata about songs—not the actual recordings. I wondered whether something similar might be true for books.
Was my book actually copied?
Were Books Actually Copied?
The shadow libraries LibGen and Z-Library contain digital copies of books. Books3 was a dataset containing approximately 196,000 books. In litigation involving Anthropic, a federal judge described the company's acquisition of millions of books from sources including LibGen and Pirate Library Mirror.
Court records in litigation against OpenAI establish that an employee downloaded pirated books from LibGen.
Actual copyrighted books were copied.
That answers one question. There's another.
Doesn't Copyright Protect Against Reproduction?.
Copyright gives authors certain exclusive rights. If I write and copyright a book, someone else generally cannot reproduce it without my permission unless a legal exception applies.
How can a company reproduce millions of copyrighted books and use them to develop a commercial product?
The AI companies involved argue that certain uses qualify as fair use.
What does fair use mean?
Under Section 107 of the U.S. Copyright Act, certain uses of copyrighted works may be permitted without the copyright owner's authorization. The statute identifies purposes such as criticism, commentary, news reporting, teaching, scholarship, and research.
I previously thought that fair use was primarily restricted to use by journalists quoting portions of copyrighted material. Turns out it's not limited to journalism. And it isn't limited to copying only small portions of a work. In some circumstances, courts have found that reproducing an entire copyrighted work qualifies as fair use.
That surprised me.
What If I Published Someone Else's Book on My Website?
Suppose I purchased a copyrighted textbook, scanned every page, and posted the complete book on my website. I could argue that my website is educational. I'm teaching people. I'm making knowledge accessible.
Would that make the reproduction fair use?
Generally, no.
Making an entire book freely available would provide readers with a substitute for purchasing the original. Calling the activity educational wouldn't automatically make it lawful.
So why might the situation be different if an AI company copied the same book to train a model? The company's argument is that it isn't distributing the original book. It's using the text for a different purpose: developing an AI system. It may turn out that difference matters under copyright law.
It raises another question:
Does creating something different justify the copying required to create it?
What Does It Mean for AI to Learn From a Book?
AI companies describe training as a learning process.
A human reads books, acquires knowledge, develops an understanding, and uses that understanding to produce new work.
An AI model, the argument goes, does something similar.
What actually happens?
In language-model training, a developer obtains digital text—a book, for example—breaks it into smaller units, and uses those units as training examples. The model makes predictions and adjusts them based on the differences between predictions and the training material. Repeated across enormous quantities of text, this process changes how the model generates responses.
The resulting model doesn't contain a complete, separately stored copy of every book used for training. It retain patterns, information, and sometimes memorized passages. This is what AI companies mean when they say a model learns from copyrighted material.
Let's compare AI training with something people do.
We read books. We learn from them. We apply what we've learned. No one expects us to obtain permission from every author whose work influenced our thinking.
There's a difference.
A person can read and learn from a book without making a copy of the book.
AI developers reproduce and process digital text to train their models.
When we describe that entire process simply as learning, we lose sight of the copying involved. That doesn't make the copying unlawful. It's a separate question that deserves examination.
In 2015, the U.S. Court of Appeals for the Second Circuit ruled that Google's scanning of millions of copyrighted books to create a searchable index qualified as fair use.
Google copied entire books.
the court concluded that the search function served a different purpose from the books themselves, and that its limited display of text did not provide readers with a meaningful substitute for the originals.
AI companies are relying on similar reasoning. They argue that they aren't copying books to sell the books - they're using the contents to train models capable of performing tasks.
In June 2025, a federal judge accepted a version of that argument in a case involving Anthropic. The judge found that Anthropic's use of books for AI training was transformative and therefore qualified as fair use.
The court drew a distinction.
Anthropic had also acquired millions of pirated books and retained them in a central library. The court rejected the argument that maintaining that library was justified simply because some of those books were used for training. It distinguished books Anthropic had purchased and converted into digital files from books it had downloaded from piracy sources.
A court can find that using books for AI training is fair use, while rejecting fair use as a justification for acquiring and retaining pirated copies.
As we'll see later, Anthropic subsequently agreed to pay $1.5 billion to settle the copyright claims involving those pirated books.
That settlement raises some interesting questions of its own.
Is AI Learning Similar to Plagiarism?
As I explored this issue, I began to wonder whether AI training has something in common with plagiarism.
Suppose someone reads my book, learns from it, and develops their own understanding.
That's learning.
Suppose that person takes the distinctive arguments in my book, rewrites them in different words, and presents them as their own original thinking. That could be plagiarism, even if they haven't copied my exact sentences.
Plagiarism and copyright infringement aren't identical. Plagiarism concerns attribution and intellectual honesty. Copyright infringement concerns the unauthorized use of protected expression.
An AI system may generate an explanation influenced by a book without identifying that book or its author. We cannot determine whether a particular answer came from one book, several books, or other sources
Even if an AI system gave an author credit, attribution wouldn't automatically authorize copying the book for training.
We have two separate questions.
Was the copyrighted work reproduced lawfully?
And:
Should the resulting AI system acknowledge the sources that contributed to its answers?
Perhaps We're Asking the Wrong Question
The public debate is often framed as a choice between two positions. One side argues that AI companies stole copyrighted material. The other argues that AI models learn from books just as humans do, making training transformative and permissible.
I think we need to distinguish different activities.
First, obtaining a copy of a copyrighted book.
Second, reproducing and storing that book.
Third, processing the text during AI training.
Fourth, developing a model whose parameters reflect that training.
Fifth, generating responses that may or may not reproduce protected expression.
A model can be transformative without every step used to create it being lawful. That's why I think the phrase AI learns from books is misleading. It may describe the intended function of training.
It doesn't tell us whether the copying required to accomplish that function was permissible.
What Should We Be Asking?
I started this investigation because someone claimed that my copyrighted book had appeared in a shadow library index. I don't know whether the actual contents of my book were copied or used to train an AI model. The investigation revealed something more important.
There is evidence that AI companies acquired complete copyrighted books from piracy sources.
There are court decisions distinguishing the acquisition and retention - storage- of books from their use in training.
There are ongoing legal disputes over whether the reproduction involved in AI training qualifies as fair use.
The question isn't whether AI should be allowed to learn from copyrighted works. Learning isn't an exclusive right protected by copyright. Reproduction is.
The question is: Under what circumstances should a company be permitted to reproduce someone else's copyrighted work, without permission, to develop a commercial product?
We need to understand exactly what was copied, why it was copied, and how it was used.
What Happened to OpenAI's Books1 and Books2 Datasets?
According to federal court records, in 2018 an OpenAI employee downloaded pirated books from Library Genesis. Plaintiffs in the copyright litigation assert that OpenAI created datasets from those books, later known as Books1 and Books2. The court record establishes that Books1 and Books2 were used to train GPT-3 and GPT-3.5.
In 2022, approximately a year before the relevant litigation began, OpenAI deleted those datasets. OpenAI initially described the deletion as being due to “non-use.” That explanation became the subject of a discovery dispute during litigation. The plaintiffs wanted to know why the datasets were deleted. OpenAI asserted attorney-client privilege over communications concerning the deletion.
In November 2025, a magistrate judge ruled that OpenAI had waived (attorney-client) privilege and ordered disclosure.
In February 2026, a federal district judge reversed that ruling, finding that OpenAI had not waived the privilege.
According to the court record, OpenAI recovered versions of Books1 and Books2 after their deletion and produced them to the plaintiffs. The plaintiffs have argued that those recovered versions are incomplete.
If I were an attorney involved in this case, one of my first questions would be:
Can I examine the recovered datasets?
What I want to know is whether my book was among the books downloaded. If it was, what happened to that text? Was the entire dataset used for training? Were selected portions copied into other datasets and then used for training? Were additional subsets created for training or model validation?
We know the datasets were used for training. That doesn't tell me whether my particular book was included or exactly how its contents might have been used. I find those questions more interesting than speculating about why the datasets were deleted.
If my book is there, that establishes it was copied.
Then we can ask whether that copying was lawful.
What Does a $1.5 Billion Settlement Mean?
In July 2026, a federal judge approved a $1.5 billion settlement between Anthropic and authors and publishers whose books had been downloaded from piracy sources.
Anthropic agreed to destroy the pirated files and copies derived from them.
The legal dispute was resolved, the authors and publishers will receive compensation, and the case ended without a trial on the remaining issues.
Does settling a lawsuit mean we've discovered what happened?
Not necessarily. We know Anthropic downloaded pirated books. We know it used copyrighted books to train AI models. What happened to the individual books? Which books were used to train which models? What information did those models retain?
Some of those questions may have been answered during discovery. If so, a settlement doesn't make those answers public.
How Can a Company Afford to Pay $1.5 Billion?
Anthropic agreed to pay one and a half billion dollars. How can a company afford that? My assumption was that Anthropic must be making enormous profits. That's not necessarily true.
In February 2026, Anthropic announced that it had raised another $30 billion in investment, at a company valuation of $380 billion. It reported an annualized revenue run rate of $14 billion.
That doesn't mean the company earned $14 billion in profit. It doesn't mean it had collected $14 billion over the previous year. A revenue run rate is a projection based on the company's current rate of sales.
Anthropic has enormous expenses. Developing AI models requires computers, data centers, electricity, engineers, and substantial ongoing investment. The company can be generating billions in revenue while still losing money. Investors are willing to provide billions because they believe the company may become more valuable.
Why?
Because AI isn't simply a chatbot that people pay $20 a month to use.
Companies are paying for AI to write software, analyze documents, automate work, and perform tasks that previously required employees or outside services. Anthropic sells access to its AI models through subscriptions, developer tools, and business services.
A single company might spend hundreds of thousands or millions of dollars a year using those tools. The potential market is enormous. That changes how we might look at a $1.5 billion settlement.
Could Copyright Infringement Become a Business Expense?
Suppose a company believes it can create a product worth tens or hundreds of billions of dollars by using copyrighted material. What happens if the company believes that the eventual cost of resolving copyright claims will be smaller than the value of the business it creates?
Could copyright infringement become a calculated business risk?
I'm not suggesting that's what Anthropic intended. Businesses routinely evaluate financial risks. They consider the cost of developing a product, the potential revenue, and the liabilities they might face. A $1.5 billion settlement looks different depending on the size of the company paying it. For a company valued at $380 billion, it's less than one-half of one percent of that valuation.
Anthropic raised $30 billion in a single funding round.
The settlement represents five percent of that amount. And the payments are spread over time.
Are the financial consequences are large enough to change a company's behavior?
Large AI companies may be able to afford settlements, licensing agreements, and legal battles that smaller competitors cannot. Could stronger copyright enforcement unintentionally make it easier for the largest companies to dominate the AI industry?
That's a question worth asking.
What Does an Author Actually Receive?
The Anthropic settlement covers roughly 500,000 books - approximately $3,000 per book, before deductions and the details of how payments are allocated.
For some authors, that may be more money than their book ever earned. For others, it may be far less than the value they believe was taken.
The settlement compensates authors for past conduct. It doesn't establish how much a book is worth as training material. It doesn't establish what an author should receive if that book helps develop a commercial AI product.
An author might spend several years researching and writing a book.
The book may sell a few thousand copies.
Then an AI company obtains the text, combines it with millions of other works, and uses it to develop a model that generates revenue for years.
What is that author's contribution worth?
I'm not suggesting we can simply divide an AI company's revenue by the number of books used for training. Some books may contribute more useful information than others. Much of a model's value comes from engineering, computing, other training data, and the way the product is delivered. If the company benefits commercially from reproducing copyrighted works, should the people who created those works receive a share of that value?
How would we calculate it?
That's a different question from whether the original copying qualified as fair use.
What Happens to the Market for Books?
I wrote a book to help people understand how to avoid unnecessary remediation expenses. Suppose someone has a problem and wants to know whether a proposal makes sense. They could purchase my book. Or they could ask an AI system.
If the AI provides the information they need, perhaps they never purchase the book.
Did I lose a sale?
Maybe.
Maybe they never would have purchased the book anyway. Perhaps they discover my work through AI and decide to buy the book or hire me. We can't know which outcome occurred simply because someone asked an AI system a question. Imagine that millions of people increasingly turn to AI instead of purchasing informational books.
What happens to the market for those books?
What happens to the people who write them?
There are different kinds of books. A novel is not the same as a technical reference book. Someone may read a novel for the experience of reading it. Someone purchasing a technical book may primarily want the information it contains.
If AI can provide that information directly, does the reader still need the book? For some kinds of books, perhaps not. That could change the market for nonfiction, reference books, textbooks, and other works people purchase primarily to obtain knowledge.
There is another possibility. AI companies may become customers for those works. Instead of selling books only to readers, authors might license their work for AI training. Perhaps a specialized technical book becomes more valuable because its contents are useful to AI developers. Perhaps authors are paid to produce new material specifically for training AI systems.
Does Copyright Protect an Author's Intellectual Property?
I don't own the facts in my books. I don't own the principles of remediation. Copyright protects my original expression of those ideas. Someone can read my book, learn from it, and explain what they've learned. That's permissible.
Reproducing the book is a different activity. This is why the Anthropic ruling is interesting. The judge found that Anthropic's use of copyrighted books for AI training qualified as fair use. The court rejected the company's fair-use justification for maintaining a library of pirated books. An author may have enforceable rights against unauthorized copying and piracy, while having less control over how a lawfully acquired book is used to train AI.
What does that mean for the future of copyright?
If an AI company purchases one copy of a book, should it be permitted to use that book to develop a product serving millions of customers? If the company must reproduce the book to train the model, does fair use justify that reproduction? If the resulting AI system competes with the original book, should that change the answer?
Those are questions the courts are still working through.
What Happens to the Creation of New Knowledge?
Here's what may be the most important question of all. Where will future AI systems get their information? AI models learn from material people have created. Books. Research papers. Articles. Music. Software. Countless works.
Producing original material takes time, effort, and often money. Someone has to conduct the research, make the discovery, test an idea, develop an explanation, create something new. AI can help with those activities. But generating a convincing explanation isn't the same as establishing that the explanation is true.
Suppose AI reduces the financial incentive to write books. Perhaps fewer purchase technical books because they can get answers directly from AI. Perhaps authors decide that spending years researching and writing a book is no longer worth the effort.
What happens then?
Will AI companies pay authors and researchers to produce new material?
Will universities, governments, businesses, and independent experts continue generating enough reliable information?
Will AI-generated content increasingly become part of the material used to train future AI systems?
How will we distinguish genuinely new knowledge from information that has simply been repeated or reformulated?
I'm not predicting that people will stop writing books. People create for many reasons besides money. AI may make writing and research easier, creating opportunities that didn't exist before. The economic incentives matter. If we want future AI systems to have access to reliable, original knowledge, someone has to keep producing it.
And if AI changes the market for that knowledge, we should understand how it changes the incentive to create it.
Are AI Models Learning From the Best Books—or the Most Popular Ones?
There's a question that occurred to me while investigating how AI companies obtained books and music: How did they decide which works to use? Were they looking for the most accurate books? The most original ideas? The best music? Were they using collections that disproportionately contained the most popular titles?
A book can become a bestseller because it's exceptionally good. A book can also become a bestseller because of marketing, publicity, a famous author, or being released at the right time. An excellent book can sell very few copies because nobody knows it exists. The same is true of music.
Now imagine selecting training material because it's popular. What happens to the resulting AI model? Does it learn the best available information? Or does it learn the information that happened to become the most widely distributed?
Here's what the Anthropic court records tell us. Anthropic acquired millions of books and then experimented with different subsets for training its models. The company evaluated different combinations to improve model performance. So there was selection and testing.
What were they optimizing for? Were they evaluating the accuracy of the information in each book? The originality of the ideas? Or were they measuring how well the resulting model performed on particular tests?
Suppose a hundred popular books repeat the same explanation of a scientific problem. And one obscure book presents a different explanation supported by better evidence. If the obscure book isn't in the training data, the model may never encounter that explanation through those books. Even if it is included, the more common explanation may still dominate what the model generates.
The result could be an AI system that's good at reproducing conventional explanations, without being good at identifying when those explanations are wrong.
This raises an interesting possibility.
What if AI developers are optimizing models to produce answers people consider useful, rather than answers that are accurate?
Consider what happens when people increasingly rely on AI to discover books, music, and new ideas. If the AI has learned disproportionately from popular material, will it tend to recommend and reproduce what is already popular? Could that make it harder for new authors, independent musicians, and unconventional thinkers to be discovered? Could creators begin producing work that resembles what AI systems favor, because that's what the market increasingly rewards?
We could end up with a feedback loop: Popular works influence AI. AI influences what people discover and consume. And what people consume influences what creators produce next.
The quality of an AI model depends partly on the information it learns from.
The fact that information is widely available doesn't mean it's correct.
If we want AI to help us discover better answers, we need to ask whether we're training it on the best available evidence—or information that's easiest to find.
What Did the Settlement Solve?
Let's return to the $1.5 billion settlement. Anthropic agreed to compensate authors and publishers for the covered claims. Did it establish a sustainable market for licensing books to AI companies? No.
Did it determine how much each book contributes to the value of an AI model? No.
Did it establish whether AI will increase or decrease future book sales? No.
Did it answer every question about what happened to the books? Not publicly.
The settlement resolved a legal dispute. It didn't resolve all the economic questions that the dispute exposed.
That's what I find so interesting about this investigation. I started by asking whether my book had been copied. Then I wanted to know whether that copying was lawful.
Now I'm wondering whether the entire business model for creating and distributing knowledge is changing.
And whether copyright law, as it exists today, addresses the problems we're going to face.
links:
June 2025 Bartz v. Anthropic ruling, the February 2026 OpenAI discovery ruling, Authors Guild v. Google, and 17 U.S.C. § 107.
Authors Guild et al v. OpenAI Inc. et al, 1:2023cv08292 - Document 782 (S.D.N.Y. 2025) https://law.justia.com/cases/federal/district-courts/new-york/nysdce/1:2023cv08292/606655/782/