Bartz v. Anthropic: The Most Pivotal AI Lawsuit You’ve Never Heard Of
By Wyatt Smith ‘26
Introduction
The question of what is to be done with regard to training Large-Language AI models (LLMs) on copyrighted data is a question that has permeated the minds of content creators around the world, and none more so than authors. LLMs function by absorbing vast quantities of data and internalizing the value patterns in that data so that they may predict the response to a given prompt. When an LLM has been properly trained, its prediction models are so sophisticated that its outputs can be nearly indistinguishable from an intelligent response. To reach this level, the model must consume, or be “trained on,” unconscionable amounts of information. GPT 5, for example, was trained on roughly 479 terabytes of raw data. Naturally, much of that data consists of copyrighted works, such as books. Bartz v. Anthropic centers on one such LLM, and it is arguably the most pivotal piece of AI litigation to date.
The Facts
Claude is a general-purpose LLM trained and owned by “unicorn” AI company Anthropic PBC. Claude was trained on a “central library” of books that Anthropic built for it during its development. Some of these books, of which there are millions, were purchased by Anthropic as physical copies and then digitized into their central library for training purposes. Some of the books, however, were not paid for at all; instead, Anthropic illegally downloaded approximately seven million books, five million of which came from “Library Genesis,” with the remaining two million coming from a website called “the Pirate Library Mirror.” This piracy is not disputed by Anthropic. In response, New York Times bestselling author Andrea Bartz filed a class action lawsuit in the Federal District of Northern California.
The Issue and the Rule
With both sides agreeing on the facts, the issue presented in this case is whether or not the use of illegally-obtained books to train an LLM constitutes fair use under United States copyright law. 17 U.S.C. §107 lays out a four-factor test for determining whether the use of copyrighted material by a party other than the rights-holder is legitimate, considering:
The purpose and character of the use - Whether the use is commercial or non-commercial, and most importantly, whether the use is transformative or derivative. That is, does the use transform the original in some way, such as a parody, or is the deviation from the original insufficient, such as a song cover?
Nature of copyrighted work - A factual work, such as a newspaper article, is more amenable to fair use than a creative work like a book.
Amount and substantiality used - How much of the copyrighted work is being appropriated?
Effect on the potential market - Also known as “market harm,” the court asks whether the unauthorized use of the copyrighted material could undermine the market for the original work.
Fair use has underpinned a significant plurality of AI litigation. Time and time again, AI developers have been sued by content creators who allege that the use of their work in training AI models constitutes infringement. Time and time again, AI companies have defended their actions with two words, often successfully: “fair use.” This case, however, comes with a major caveat: the illegal procurement of the books. This distinguishing factor has placed Anthropic PBC at a major disadvantage in its argument that fair use protects its training of Claude.
The Situation as of Now
The plaintiffs began their suit with two claims: that the use of legally purchased and copied books to train Claude did not constitute fair use under extant copyright law, and that neither did the use of illegally obtained books. In June 2025, Judge William Alsup, who is currently presiding over the case, issued a summary judgment. He found that AI companies are indeed allowed to train models on copyrighted material, so long as this material is legally obtained. This ruling is important because it sets a precedent for a number of other ongoing cases, such as Author’s Guild v. OpenAI Inc. On the flip side, Alsup ruled that the use of illegally obtained works does not constitute fair use and is a blatant violation of copyright law in the U.S.
Following this development, the plaintiffs proceeded as a class including all U.S. authors and copyright holders whose books had been found in Anthropic’s pirated libraries. Judge Alsup certified the class in July 2025, allowing the lawsuit to move forward. The class of authors holds the rights to 482,000 distinct works. With the defendant Anthropic now at a clear disadvantage, the parties began moving towards a settlement, reaching an agreement by September 5th.
This settlement would entail a minimum payout of $1.5 billion from Anthropic, including $3000 for every unique work that had been pirated. This money would be distributed to the class via an opt-in system, as opposed to the usual opt-out; that is, any author in the class wishing to receive the money would have to expressly ask to be included in the suit. Authors not accepting the $3000 payout would be free to file their own lawsuits against Anthropic. Judge Alsup insisted on this style to prevent any one author from “getting the shaft” (his words). The settlement would also include an injunction against any more piracy by Anthropic.
When the parties presented this settlement to Alsup on September 8th, he reacted with substantial questions and no small portion of skepticism. For one, Alsup questioned why the plaintiffs had not provided a comprehensive list of affected works and class members, which he demanded they procure before any settlement could move forward. Alsup’s primary concern was with the claims notice process, which he excoriated as being “forced down the throat” of authors who may not understand it. All in, Judge Alsup provided a list of 34 grievances for the parties to fix, and admonished the settlement as not even close to finished.
The parties scrambled to amend the settlement in the wake of these criticisms. By September 15th, they submitted a comprehensive filing addressing the Judge’s concerns and delivered the final list of affected works and class members. When the parties presented the revised settlement on the 25th, Judge Alsup was satisfied enough to grant preliminary approval. It appears as though the case will be settled for roughly $1.5 billion, the same amount suggested in the first settlement agreement, and by far the largest payout in an AI-related lawsuit to date.
What This Means for the Industry
This case, if it is settled (which in all likelihood it will be), will have massive implications for all of its stakeholders, and for the AI and writing industries as a whole. This impact will go far beyond the monetary.
First, every AI company that deals with copyright law in the U.S. has taken note of the June summary judgment. Particularly, they have noted that the use of legally purchased copyrighted works to train AI models does, in fact, constitute fair use. This is excellent news for AI companies, which have been under both public and legal scrutiny for their unconstrained use of such works. While this ruling is by no means binding, it may serve to alleviate some of that legal scrutiny, even if it won’t help public opinion. At the very least, it will be a case that attorneys representing AI developers can point to as precedent.
For American authors, this case is a mixed victory. The core claim that training LLMs on copyrighted works is infringement has not survived in this case, and seems unlikely to survive in any other cases either (such as Author’s Guild v. OpenAI Inc.). AI developers, it seems, will continue to train their models on copyrighted works. However, the authors did succeed in narrowing the scope of what can be considered fair use in the context of AI training. From now on, courts will regard the use of illegally procured material in training as infringement, which will protect content creators to a certain extent. Not to mention, authors in the United States are about to get $1.5 billion richer. This is not only the highest number paid out in an AI lawsuit, but also the highest ever financial recovery in a copyright case. That is no small feat.
For Anthropic, on the other hand, this is going to hurt a bit, but it won’t be the end. While Anthropic is a private company (meaning access to their financial data is limited), they are backed by several powerful tech firms, including Amazon and Google, who have both invested ten-figure sums into the startup. Anthropic was valued at $183 billion as of early September as well. While their cash flows are likely nowhere close to $1.5 billion, they will be able to raise enough capital to absorb this blow.
What’s Next?
This case is just the beginning of a long string of cases to come that will clarify, narrow, and set boundaries on what is and is not acceptable for AI companies to do. The artificial intelligence industry, despite its massive growth and relative proliferation, is still an infant and has yet to mature in any meaningful way. As the industry matures, the courts will hopefully continue to make landmark rulings such as this (preliminary) one, protecting both the industry and the public from the dangers posed by lawless corporatism. It is an exciting time for intellectual property law in the United States, for better or for worse.
Wyatt Smith is a senior majoring in operations and supply chain management.
Sources
17 U.S.C. § 107 (2024)
Authors Guild v. OpenAI Inc., 1:23-cv-08292, (S.D.N.Y.)
Anthropic PBC. (2025, September 2). Anthropic raises $13B Series F at $183B post-money valuation. Anthropic.com. https://www.anthropic.com/news/anthropic-raises-series-f-at-usd183b-post-money-valuation
Benched.ai Editorial Team. (n.d.). GPT-5 Dataset. Benched.ai. https://benched.ai/guides/gpt-5-dataset
Bulger, D. W., Finkelstein, M. L., Weis, N. (2025, September 17). Judge Pauses Anthropic’s $1.5 Billion Copyright Settlement. Arent Fox Schiff Law. https://www.afslaw.com/perspectives/ai-law-blog/judge-pauses-anthropics-15-billion-copyrightsettlement#:~:text=However%2C%20on%20September%208%2C%20Judge,be%20incorporated% 20into%20training%20datasets
D’Angelo, F., De Santis, E. (2025, June 23). Bartz v. Anthropic PBC. Loeb & Loeb LLP. https://www.loeb.com/en/insights/publications/2025/07/bartz-v-anthropic-pbc
Hansen, D. (2025, September 28). Bartz v Anthropic Settlement Gets Preliminary Approval – Key Takeaways. Authors Alliance. https://www.authorsalliance.org/2025/09/28/bartz-v-anthropic-settlement-gets-preliminary-approval-key-takeaways/#:~:text=narrowed%20in%20scope%20%E2%80%94%20the,5
Levy, A. (2025, September 8). Anthropic Judge Blasts $1.5 Billion AI Copyright Settlement. Bloomberg Law. https://news.bloomberglaw.com/ip-law/anthropic-judge-blasts-copyright-pact-as-nowhere-close-to-done
Lieff, Cabraser, Hermann & Bernstein Attorneys at Law. (2025, September 22). Class Plaintiffs File Proposed Plan for Notice and Distribution Phase of Historic $1.5 Billion Settlement in Anthropic AI Litigation. lieffcabraser.com.
Moon, M. (2025, September 8). Judge rejects Anthropic's record-breaking $1.5 billion settlement for AI copyright lawsuit. Yahoo!tech. https://tech.yahoo.com/ai/article/judge-rejects-anthropics-record-breaking-033512463.html?guccounter=1&guce_referrer=aHR0cHM6Ly9kdWNrZHVja2dvLmNvbS8&guce_referrer_sig=AQAAADRCqxN3X_vmizsa3xUWndbea5QfoXoN19iAZlzVX6u9IspIPtyR2lDIkq94daDa6PM4RCCUJo1-DW3YirZPx_-fK6-CklM02PRwIolC_R7NyWOM6Hs7eZQV5VJavPE11awu-CtrPIOZEYbkLusC6ePz3elAw0sefep7VfgYbB3v
Ortutay, B. (2025, September 25). Judge approves $1.5 billion copyright settlement between AI company Anthropic and authors. Associated Press. https://apnews.com/article/anthropic-authors-copyright-judge-artificial-intelligence-9643064e847a5e88ef6ee8b620b3a44c#:~:text=U,not%20apply%20to%20future%20works
Smith, W. (2025). Tracking Challenges to The Legality of Training Generative-AI Models on Copyrighted Data. The Undergraduate Law Review at Charlotte, 1(1), 57-72. https://static1.squarespace.com/static/671472ba24b34f5c76805572/t/6843b2e1780237079c7f9e88/1749267170752/SPRING2025PUBLICATION+%281%29.pdf
