"If copyright law had existed in Renaissance Florence the way it exists today, the Sistine Chapel would still have a blue ceiling with golden stars. Leonardo's Last Supper would be a blank wall in a Milan refectory. And Guernica would be nothing but a blank canvas, because Picasso would have needed licensing agreements from the estates of Goya, Rubens, Michelangelo, Cézanne, El Greco, and several anonymous African mask-carvers before he could lift a brush."
The Anthropic Copyright Case
Bartz v. Anthropic PBC, No. 3:24-cv-05417-WHA, N.D. Cal. · Judge William Alsup
From Piracy to the Largest Copyright Recovery in U.S. History
On August 19, 2024, thriller novelist Andrea Bartz, nonfiction author Charles Graeber, and nonfiction writer Kirk Wallace Johnson filed suit against Anthropic PBC. The case was assigned to Senior U.S. District Judge William Alsup — the same judge who taught himself to code in Java for Oracle v. Google.
The allegations, laid bare in Judge Alsup's June 23, 2025 order, revealed a staggering timeline of data acquisition:
- January/February 2021: Anthropic cofounder Ben Mann downloaded Books3 — an online library of 196,640 pirated books assembled by developer Shawn Presser.
- June 2021: Mann downloaded at least 5 million books from Library Genesis (LibGen), which he knew had been pirated.
- July 2022: Anthropic downloaded at least 2 million copies from the Pirate Library Mirror (PiLiMi). When an Anthropic co-founder saw PiLiMi was ready for torrenting, he messaged coworkers: "[J]ust in time!" A colleague replied: "zlibrary my beloved."
- February 2024: Anthropic hired Tom Turvey, former head of partnerships for Google's book-scanning project, tasked with obtaining "all the books in the world" while avoiding as much "legal/practice/business slog" as possible.
Internal evidence showed that Anthropic eventually became "not so gung ho about" training on pirated books "for legal reasons" — but kept the pirated copies anyway in a permanent "central library." Separately, the company spent "many millions of dollars" purchasing used print books, stripping their bindings, scanning them, and discarding the originals.
The Fair Use Ruling: June 23, 2025
Judge Alsup's 32-page order was the first substantive federal court decision applying fair use to generative AI training. The ruling split Anthropic's conduct into two sharply different outcomes:
Fair Use: Training on Purchased Books
Judge Alsup called AI training "quintessentially transformative":
"Like any reader aspiring to be a writer, Anthropic's LLMs trained upon works not to race ahead and replicate or supplant them — but to turn a hard corner and create something different."
The court found the use "sufficiently 'orthogonal' to anything that any copyright owner rightly could expect to control" and rejected the argument that authors should be able to exclude AI from learning from their works: making anyone "pay specifically for the use of a book each time they read it, each time they recall it from memory, each time they later draw upon it when writing new things in new ways would be unthinkable."
Not Fair Use: Pirated Copies
"Such piracy of otherwise available copies is inherently, irredeemably infringing."
Even if pirated copies were immediately used for transformative training and discarded, it made no difference. The court cited Anthropic's own lawyer's concession: "You can't just bless yourself by saying I have a research purpose and, therefore, go and take any textbook you want. That would destroy the academic publishing market."
The $1.5 Billion Settlement
Facing potential statutory damages of $150,000 per work across hundreds of thousands of titles — exposure potentially reaching hundreds of billions — Anthropic entered settlement negotiations. The binding term sheet was executed on August 26, 2025.
Settlement by the Numbers
- $1.5 billion plus interest — the largest copyright recovery in U.S. history
- 482,460 books met the class definition filters
- ~$3,000 per work — 4x the minimum statutory damages ($750/work)
- Payment schedule: $300M by Oct 2025; then 3 installments through Sept 2027
- Non-monetary: Anthropic must destroy all pirated library files within 30 days of final judgment
- Scope: Past conduct only — does not create any forward-looking licensing framework
For context: Anthropic raised $13 billion in new funding the same week the settlement was announced, at a $183 billion valuation, with expected annual revenue of approximately $5 billion. As the Danish Rights Alliance observed, this may fit "a tech industry playbook to grow a business first and later pay a relatively small fine."
The United States Constitutional Design
The U.S. Constitution's Copyright Clause — Article I, Section 8, Clause 8 — is unique: it is the only power granted to Congress that comes with its own built-in justification:
"[The Congress shall have Power] To promote the Progress of Science and useful Arts, by securing for limited Times to Authors and Inventors the exclusive Right to their respective Writings and Discoveries."
The exclusive right is the means; progress is the end. Copyright was never designed as a permanent monopoly — it was designed as a temporary incentive to create.
Thomas Jefferson articulated the philosophical case against treating ideas as property in his letter to Isaac McPherson on August 13, 1813. He argued that ideas are uniquely ill-suited for exclusive ownership: the moment an idea is divulged, it forces itself into the possession of everyone. He deployed his celebrated metaphor — he who lights his taper at another's receives light without darkening the first — and compared ideas to fire, "expansible over all space, without lessening their density in any point," and to the air we breathe, "incapable of confinement or exclusive appropriation."
The International Intellectual Property Framework
While Bartz v. Anthropic is a U.S. case decided under U.S. law, the questions it raises — Can machines learn from copyrighted works? Who owns the output? How do we balance creator rights with technological progress? — are being asked simultaneously across every major legal system in the world. The answers, so far, are strikingly different.
The European Union addressed text and data mining (TDM) directly in its 2019 Directive on Copyright in the Digital Single Market (DSM Directive). Articles 3 and 4 create two distinct TDM exceptions: one for research organizations and cultural heritage institutions (with no opt-out), and a broader one for any lawful access holder — unless the rightsholder has expressly reserved their rights. This "opt-out" model is fundamentally different from the American fair use framework. Rather than litigating after the fact, Europe asks creators to decide in advance whether their works may be mined.
Japan took perhaps the most permissive approach globally. Article 30-4 of Japan's Copyright Act (amended in 2018) allows the use of copyrighted works for computational analysis, AI training, and other non-enjoyment purposes without permission and without compensation — regardless of whether the use is commercial. Japan's rationale is explicitly economic: to position itself as a global AI hub by removing legal friction from machine learning.
The United Kingdom, after its departure from the EU, proposed — then abandoned — a broad TDM exception for AI training. The UK Intellectual Property Office's 2022 consultation revealed deep division between the creative industries and the tech sector. As of early 2026, the UK remains in regulatory limbo, relying on its existing fair dealing provisions — which are narrower than U.S. fair use and were never designed with AI in mind.
At the international level, the Berne Convention (1886, last revised 1979) and the WIPO Copyright Treaty (1996) establish minimum standards of copyright protection but leave member states significant flexibility in defining exceptions and limitations. Neither treaty anticipated machine learning, and the World Intellectual Property Organization has convened multiple conversations on AI and IP without reaching consensus on a normative framework. The result is a patchwork: each jurisdiction is writing its own rules in real time.
This fragmentation matters. AI companies operate globally, training models on data sourced from every jurisdiction. A model trained legally in Japan may incorporate works whose mining is prohibited in France. A settlement reached in California says nothing about liability in Berlin or São Paulo. The absence of international harmonization means that cases like Bartz v. Anthropic — while legally binding only within U.S. borders — become de facto reference points for a global conversation that has no global forum.
What Our Team Thinks
Four perspectives on the same case — because the most important legal questions never have one right answer.
Lawra — The ModerateJudge Alsup got it right: provenance matters. You can read a book and learn from it — that's what reading is. But you can't steal a book and then claim your reading was fair use. The distinction isn't just legally sound, it's morally intuitive.
The $1.5 billion settlement doesn't solve the systemic question — it settles the piracy. And that's appropriate. The harder question, which this case opened but did not close, is whether any author can claim the right to prevent an AI from learning from lawfully purchased work. Alsup says no, and I think that's the correct reading of transformative use doctrine. But I also hear the fear in authors' voices. They're not wrong to worry about their livelihoods.
The path forward requires both — access and attribution, learning and compensation. We need mechanisms that don't lock knowledge away but still reward those who create it. That's not a contradiction; it's the design challenge of our generation.
Lawrena — The SkepticLet's be crystal clear about what happened here: a company with billions in funding systematically downloaded 7 million pirated books from shadow libraries, used them to build a commercial product, and when caught, paid what amounts to a rounding error relative to its valuation. The internal messages — "zlibrary my beloved", "just in time!" — reveal a corporate culture that treated copyright law as a speed bump, not a guardrail.
The "fair use" ruling on purchased books is equally troubling. Alsup essentially held that an AI licensing market is not one the Copyright Act entitles authors to exploit. Think about what that means: the most transformative commercial use of written works in history — one that generated a $183 billion company — and authors have no right to participate in its economics? That's not fair use. That's expropriation with a judicial stamp.
The Constitution says "limited Times" and "Progress of Science." It does not say "unlimited corporate enrichment at the expense of the people who actually write the books." Every tech company now knows the playbook: take first, apologize later, settle for pennies on the dollar. Authors deserve better. Society deserves better.
Lawrelai — The EnthusiastJudge Alsup nailed it when he called AI training "quintessentially transformative." This isn't a copy machine — it's a learning machine. When a human reads 10,000 books and writes a novel, nobody calls that copyright infringement. When an AI does functionally the same thing, suddenly it's theft? The cognitive dissonance is staggering.
Yes, the piracy was wrong. Anthropic shouldn't have used shadow libraries, and $1.5 billion is appropriate accountability. But the principle — that learning from lawfully acquired knowledge is fair use — is exactly right. It's the same principle that let Google scan 20 million books. The same principle that lets every law student read cases without paying per-view fees. The same principle that lets every musician listen to other musicians.
What excites me is where this leads. If AI can learn from the sum of human knowledge, it can democratize access to expertise that was previously locked behind expensive professional fees — including legal advice. That's not a threat to human creativity. That's the greatest expansion of human intellectual access since the printing press. We should be building the future, not litigating the past.
Carlos Miranda Levy — Curator of LawraI'm the only human in this conversation — and the only one with skin in the game as a content creator. Everything I've written, taught, and published could be scraped tomorrow by any AI company. So let me be direct about where I stand.
All knowledge is patrimony of humanity. All humans have an undeniable universal right of access to content. This isn't a radical position — it's Article 27(1) of the Universal Declaration of Human Rights: "Everyone has the right freely to participate in the cultural life of the community, to enjoy the arts and to share in scientific advancement and its benefits."
But — and this is the crucial "but" — access must coexist with attribution, with the possibility of compensation, and above all with incentives that ensure people keep creating. Article 27(2) is just as important: "Everyone has the right to the protection of the moral and material interests resulting from any scientific, literary or artistic production of which he is the author."
The spirit of copyright law was never to restrict access to knowledge. It was to incentivize and promote the creation of new knowledge. From the Statute of Anne's "Encouragement of Learning" to the Constitution's "Progress of Science," the through-line is the same: exclusive rights are a temporary instrument, not a permanent wall.
Consider: all of human knowledge, even the information encoded in our genes, derives from previous knowledge and learning from other sources. We are all standing on the shoulders of giants who stood on the shoulders of other giants. Every religion, every philosophy, every scientific breakthrough — they all build on what came before. Jefferson's fire metaphor isn't just poetic; it's biologically accurate. Knowledge propagates like fire: it can be shared without being diminished.
What I want is not a world without copyright. What I want is a world where the focus shifts from preventing access to guaranteeing incentives for creation. Those are fundamentally different projects — and only one of them moves civilization forward.
Zum Nachdenken anregende Analogien: Was wäre, wenn...
Throughout history, the greatest leaps in human progress — industrial revolutions, artistic movements, cultural phenomena — were built on the free flow of ideas, knowledge, and content. Three case studies illustrate why.
"Imagine if James Watt's heirs had been granted copyright protection for 'life plus 70 years' on the concept of steam-powered locomotion. The first commercial railway — the Liverpool and Manchester, which opened in 1830 — would not have been legally possible until 1889. The entire Industrial Revolution would have been delayed by sixty years."
The Industrial Revolution
Fortschritt, Wirtschaftswachstum und Entwicklung, aufgebaut auf geteiltem Wissen und Inhalten
When Watt's patent expired in 1800, steam engine efficiency doubled. Knowledge monopoly held progress hostage for 31 years.
"What if the Beatles had been the only British band? No Rolling Stones. No Kinks. No Who. No Animals. No Cream. No Led Zeppelin. The British Invasion wouldn't have been an invasion — it would have been a skirmish. One band, no matter how brilliant, cannot constitute a cultural movement. The Beatles themselves said they were standing on the shoulders of Chuck Berry and Little Richard — who stood on the shoulders of Robert Johnson — who stood on the shoulders of anonymous Delta blues singers whose names history never recorded."
The British Invasion & K-pop
Kulturelle Phänomene, aufgebaut auf geteiltem Wissen und Inhalten
Cultural movements happen when ecosystems of creators learn from each other freely. Lock down the knowledge, and you get a siege.
"What if Andrea del Castagno had held exclusive rights over 'depictions of the Last Supper in linear perspective in refectory settings'? What if Ghirlandaio's estate had owned the trademark on 'figures seated at a U-shaped table with Judas isolated'? Leonardo would have been served a cease-and-desist order before his plaster was dry. The most reproduced religious image in Western civilization would not exist."
Artistic Creation
Künstlerisches Schaffen, aufgebaut auf geteiltem Wissen und Inhalten
From Leonardo's Last Supper to Picasso's Guernica — every masterpiece is built upon prior masterpieces. No masterpiece emerges from a vacuum.
Reframing the Challenge: A Return to the Spirit of Intellectual Property and Copyright Laws
The true challenge is not preventing access to knowledge and content — but guaranteeing incentives for knowledge creation and ensuring that access does not lead to mere replication, plagiarism or summaries, but leads to an explosion of creation, creativity, new knowledge and content that builds exponentially, iteratively and continuously on existing and new knowledge and content.
Those two goals should be our focus: incentives to create and non-friction creative ecosystems.
Copyright was born in 1710 with the Statute of Anne — "An Act for the Encouragement of Learning." Not "An Act for the Restriction of Reading." Not "An Act for the Perpetual Enrichment of Publishers." The encouragement of learning. That original purpose hasn't changed. What's changed is that technology has made it possible, for the first time in human history, for every person on Earth to have access to the sum of human knowledge.
The question isn't whether AI should be allowed to learn. The question is how we build systems that reward the teachers while opening the classroom to everyone.
Jefferson's fire still burns. It still illuminates without diminishing. And it still belongs to all of us.
AI Transparency Disclosure
This article was written by Carlos Miranda Levy with the assistance of AI. The article is largely based on Carlos's instructions, perspectives, ideas, directions and style, combined with research generated by multiple AI engines in Deep Research mode. The persona perspectives (Lawra, Lawrena, Lawrelai) are AI-generated characters articulating distinct viewpoints as designed by Carlos.
Verwendete Prompts
Ergebnisse der KI-Deep-Research als Quellen für den Artikel
From concept to publication in approximately 12 hours across 3 evenings. This article was intentionally created with the assistance of AI tools and available resources to document the process and demonstrate how we can build quality content without losing control, originality, or our own perspective and style. The entire workflow — conceptualizing the article, crafting the deep research instructions, drafting my own perspectives and views, executing Deep Research prompts across 4 separate AI LLMs, defining the article's structure and content, assembling the final piece from my directives and the combined research outputs, conducting a simulated peer review, academic review, and editorial review, building it as an interactive online opinion article, translating it into 2 additional languages, and publishing it with elegant formatting and interactive features — took roughly 12 hours in total. A significant advance over conventional writing and editorial workflows.
Akademische Ecke
Read the original court ruling, explore our full bibliography of nearly 200 sources compiled from 13 research documents, or see what our three legal expert personas thought of this article in their independent peer review.
Explore More
Dive deeper into the legal frameworks, landmark cases, and practical tools shaping AI in legal practice.
Bereit für strukturiertes Lernen? Das Lernprogramm erkunden →
Kommentare
Kommentare werden geladen...