LexED2 home
AI & the Law: all pages

Who Owns the Output? The Moot

PLA 2872Activity 6

Captions appear here while the video plays.
Lunchtime keynote by Judge M. Margaret McKeown, Berkeley Center for Law & TechnologyCorrected captions for 20:09 to 28:55 show in the strip above as the video plays. To see only these, turn off captions in YouTube's player.
Transcript, 20:09 to 28:55

20:09 So let me turn to a second technology revolution, and that's AI, so there's little doubt that generative AI, it's a textbook example of disruptive innovation, as Clayton Christensen from Harvard put it. And I kind of like this quote from Asimov, because he describes AI or any of this kind of technology as kind of magic. And I suggested that to Brad Smith of Microsoft when we were on a panel, and he said, no, no, I disagree. It's just an algorithm. So it might be just an algorithm, but it's causing conniptions in the copyright world. And of course, I just want to point out this is the evolution on authorship of the different words we've used in the Constitution, the 1909 Act, and, of course, the 1976 Act. And I think that what the court, what the legislative history shows in the '76 Act is that "original works of authorship," it was actually purposefully left undefined, so that it now falls to the courts to define and implement.

21:21 What does that mean? And I can't really leave this talk without my favorite or one of my favorite Ninth Circuit cases, and this is our friend, Naruto the monkey, and we held in 2018 that the monkey lacks statutory standing to pursue a copyright claim, so we can laugh and whatever. But in fact, Naruto was the beginning then of copyright guidance at a copyright office, and they used this case to explain the contours of human authorship and the human authorship requirement. So it may be that this notion of human authorship is more load bearing than we originally thought. So what kind of human create, creative agency is copyright trying to honor, preserve or privilege, and as relevant to fair use, what kind of markets is copyright imagining?

22:22 And so while the assumption of human authorship is one of the key and fundamental pressure points that the Act is facing, I think doctrinally harder is the question of fair use. Of course, easy to talk about: you can cite the four factors. More difficult to apply. The Supreme Court has added texture in at least three key cases. 1985, Harper & Row: the court said, and we still repeat again and again that the fourth factor, market effects, is tagged as the most important. Then we fast forward to Google v. Oracle, 2021. That's a software case, and it was Google's reimplementation of Oracle's declaring code. It was transformative. According to the Court, Justice Breyer there reiterated that fair use is an equitable rule of reason, and it permits courts to avoid rigid application of the copyright statute when, on occasion, it would stifle creativity, which that law is designed to foster.

23:27 And then, of course, we come to the Andy Warhol case 2023 and even though Andy Warhol's use of these photographs of Prince added some new expression, there was no fair use, stopping on the first factor. So now, how will these principles play out in the future? It's interesting. We have now about 80 generative AI cases in the courts, and most of them are class action suits, not all, but you can see from the slide that they kind of segment in to medium, whether it's text, music, computer code or whatever. But as yet, I might remind you there's no appellate decision. I will talk briefly about Bartz v. Anthropic. It's easiest to talk about because it's settled a whopping $1. 5 billion settlement. No appeal, of course. So there what Anthropic did: it downloaded and digitized millions of copyrighted books from pirate sites on the Internet and some from copyrighted books that it had purchased.

24:31 And the goal is they were going to make like a central library of all the books in the world, and then use it to train Claude. The decision, I think, is interesting, because it illustrates the challenges the courts have. And instead of saying all training is transformative and fair use, what Judge Alsup did is he put it in buckets, and he said it depended on whether they had legally purchased and digitized the books that was justified as fair use, but the pirated books was not. But I think the harder question that the opinion grapples with is just, how far can you push the first fair use factor on transformative use and the fourth factor on market impacts. I thought it was interesting. He characterized generative AI as quintessentially transformative, and he described the technology as probably the among the most transformative that many of us will see in our lifetimes.

25:36 So it's interesting. He also said, Well, it's just training on it is like a reader aspiring to be a writer. Readers study texts and they learn how to create original works. So LLMs read text and then they produce fundamentally different works. And what you see here is what we often see in court decisions when we particularly in technology, we try to analogize to something we know and understand. It does fit sometimes; sometimes it doesn't fit, but we're often reasoning by analogy. So for lawyers, academics and others, I think it's very important to peg some of these technology cases in analogies that the court can understand. Anthropic, of course, said, well, the training is fair use because the existence of a pirated library shouldn't even matter. Judge Alsup rejected that argument, and he said, there's no kind of free-floating entitlement to copy first and justify later, so we can debate whether that decision was right or wrong, but I think it's fair to say that one thing it did, it rejected AI exceptionalism in the world of copyright.

26:49 It's also notable for what it didn't do. He rejected the idea that generative AI will flood the market with competing books. Now, some might and have disagreed, so I'll mention briefly the case of Kadrey v. Meta, which is still pending in the District Court. But there is a district court opinion on a summary judgment, and Judge Chhabria has said, well, the use of these books from various libraries like Genesis and Anna's Archives, it was fair use on the record before him, but he also quickly highlighted that this holding didn't actually establish the general legality of LLM training, because, in his view, the LLM training combines two features that we hadn't previously seen. One, a deeply transformative use of the work, and two, the potential to flood the markets. So again, the first and fourth factors become the most important.

27:52 These cases suggest that the battleground going forward really will be on what kind of market impact is the copyright law prepared to recognize? And I want to emphasize that these are only just a handful of cases. Even these are just cases in which there's been decisions, but there are scores out there that are still pending with not even a district court decision, and there's no appellate decision. So I wonder, is this like the early software cases on copyright, for example, the 1986 case of Whelan v. Jaslow Dental, that talked about structure and sequence and organization of code, and then was later disfavored and basically jettisoned, and it was one of the early cases. So the question remains, are these early cases that will be adapted and adopted, or will they be jettisoned?

The words are the speaker's. This transcript is here for accessibility and is not part of the site's CC0 dedication.

Why it Matters

Arguments about AI and copyright usually get told as creators against technology companies. The more useful question is which use a court is judging. The law answers three questions separately: whether the copies made to train a model are fair use, whether what the model later writes infringes, and who owns that writing. Ownership has the firmest answer, because courts read the Copyright Act to require a human author, while training turns on four factors weighed against the facts of each case. The training question is the one you argue, before a judge who has read the briefs and against a lawyer who wants you to lose.

Current Context

On September 29, 2026, the Third Circuit held in Thomson Reuters v. Ross (opens in a new tab) that copying Westlaw’s headnotes to train a competing legal research tool was not fair use. Ross’s tool returned passages of opinions and wrote nothing new, and the court relied on that to set the case apart from the generative models in Bartz and the OpenAI litigation. In that litigation, the Justice Department’s statement of interest (opens in a new tab) of September 1 urges the court to hold that training is fair use. The news publishers’ summary-judgment brief, refiled with fewer redactions on September 17 (opens in a new tab), quotes an OpenAI executive calling its products “largely substitutive, period,” and the answering briefs are due October 23. The publishers’ brief leans on Ross and the Justice Department on Bartz, so where OpenAI’s models fall between the two is a question the moot puts to you.

Key Concepts

Fair Use

Section 107 of the Copyright Act excuses some copying done without permission. Courts weigh four factors together, case by case: the purpose and character of the use, the nature of the copied work, the amount used, and the effect on the work’s market. Fair use is an affirmative defense, so the party that copied must prove it.

Transformative Use

A use is transformative when it serves a purpose different from the original’s instead of standing in for it, and the more transformative it is, the less its commercial character counts. Warhol sharpened the test: when a use shares the original’s purpose and is commercial, the first factor likely weighs against fair use unless something else justifies the copying. Developers call training a new purpose, learning the patterns of language; publishers answer that the products trained on their work do the job their work does.

Market Harm

The fourth factor asks whether the use hurts the market for the original or for licenses of it, and its core concern is substitution: a copy people choose instead of the original. Courts have split on whether lost fees for licensing works as training data count. Market dilution, a newer theory, holds that a flood of machine-written work harms the market for human work even when it copies no one’s words.

Training and Piracy

Bartz v. Anthropic split one company’s copying into separate uses: training its models on books was fair use, but downloading more than seven million pirated copies to keep in a permanent library was not. The training some of those copies later served did not excuse the downloading, and the company settled the piracy claims for $1.5 billion.

Output Infringement

A claim about what a model produces rather than what it was trained on. It turns on whether an output copies protected expression, such as a passage reproduced word for word (what the briefs call regurgitation), and not the facts or style an article shares with everyone. Courts analyze each use separately, so a ruling that training is fair use does not decide whether a particular output infringes.

Human Authorship

Copyright protects only work that a human being authored. In Thaler v. Perlmutter, the D.C. Circuit held that the Copyright Act requires a human author, so a picture a machine made on its own could not be registered. A person who uses an AI tool can still claim what the person contributed, and the Copyright Office asks applicants to identify and exclude the machine-made parts.

Summary Judgment

A ruling that decides a claim or defense without a trial, because the record shows no genuine dispute about any material fact and one side is entitled to win as a matter of law. The court decides on what the parties filed: their briefs, their statements of undisputed facts, and the evidence behind them. At a summary-judgment argument, a fact that appears in no filing is not before the court.

Resources

What to Do

Fair use is won or lost factor by factor, on a record the lawyer did not write. An AI runs this simulation. It sits as the judge at a moot summary-judgment hearing built on the briefs in the OpenAI copyright case and argues against you as opposing counsel, and you represent whichever side the judge assigns. By the end you can argue each factor from the filings and name the one the case turns on, which is the center of the reflection you post.

Part 1: Set Up

Start with the background reading and eight minutes of the video at the top of this page, from 20:15 to about 28:35, where Judge McKeown moves from authorship through fair use to Bartz and Kadrey. Read 17 U.S.C. § 107, which is one paragraph, and then the five excerpts in Resources that make up the record. Three come from the news publishers’ brief (pages 51 to 54, 57 and 58, and 59 to 63). The other two are OpenAI’s argument on training (pages 7 to 11) and the United States’ statement on the fourth factor (pages 13 to 18). Note the facts you could use and the page each is on, because the judge will ask where a fact comes from. Last, run the Fair Use Tutor and answer its questions on all four factors. The simulation will not teach you the doctrine; it tests whether you can use it, and a student who arrives without the record has nothing to argue from.

Part 2: Run the Simulation

Copy the Simulator into a fresh AI chat. The judge assigns your side at random, so prepare both. Argue in role, give the excerpt and page for every fact, and do not ask the AI what to say; it will turn the question back into the hearing. The hearing ends only when all four factors have been argued and you have told the judge which factor the case turns on, after at least eight exchanges. If opposing counsel makes a point you cannot answer from the record, say so, as a lawyer would, and move on.

Part 3: Debrief

After a break of at least an hour, open a new chat, run the Debrief, and paste the link to your hearing or a summary of it, with the side you argued. Read its account factor by factor. Check every case and page it cites against Resources before you accept what it says.

Part 4: Replay from the Other Side

Paste the Simulator into another fresh chat, and make your first message name the side you did not argue: “I argue for OpenAI” or “I argue for the news publishers.” The hearing runs again under the same rules. Notice which of your own arguments you now have to answer, and whether the factor you named the first time still looks like the one the case turns on.

Part 5: Reflect and Post

Write a reflection of 400 to 600 words in your own words. It says how each hearing went and names the factor the case turns on, with your reason. It covers all four factors, saying for each what you argued and which excerpt and page supported it. Every fact about the parties comes from one of the five excerpts, cited by page; law can come from the cases in Resources. It names the one argument you would make differently next time and why, drawing on the debrief or the replay. And it says what arguing both sides showed you that reading the briefs did not. Before you post, check your draft against those five requirements, mark each one met or not met at the bottom, and fix what you can. Post the reflection, with a headline-style subject line, as a new discussion thread.

Include one AI chat link with a 1–2 sentence explanation of what the conversation shows and why you chose to share it.

Suggested AI Prompts

Use these prompts as a starting point, then adjust them to fit your goal. Strong prompting develops through trial, revision, and testing. It’s a foundational skill that grows into more advanced AI work such as context engineering and agent-based workflows.

Fair Use Tutor

Act as a copyright law professor teaching a college student. Walk me through the four fair use factors in 17 U.S.C. § 107 one at a time, using the example of a company that trains a language model on published books. For each factor, explain what courts ask, what the AI company would argue, what the authors would argue, and how a court might weigh it. Then explain why a court could find the training itself fair use while finding that downloading pirated copies of the books was not. After each factor, ask me a question to check my understanding before moving on, and correct me where I am wrong. Name any case you rely on with enough detail that I can look it up, and tell me plainly if you are unsure whether a case exists or what it held.

A question after each factor turns an explanation into a quiz, so you find the gaps in your understanding before the judge does. Telling the model to name cases precisely and to admit when it is unsure guards against a made-up precedent, in the area of law where they turn up most.

The Simulator

You are running a moot court, a practice argument, on the fair use question in a real case: In re OpenAI, Inc. Copyright Infringement Litigation, No. 1:25-md-03143, in the United States District Court for the Southern District of New York. News publishers, The New York Times among them, sued OpenAI for copying their articles to build its language models, and both sides have moved for summary judgment. You play two roles. As the judge, you are a fictional federal judge presiding over this moot, not the judge assigned to the real case, and you say nothing about how the real court will rule. As opposing counsel, you argue the side I do not.

The question. Only one question is before the court: whether OpenAI’s copying of the publishers’ articles to train its models is fair use under 17 U.S.C. § 107. Outputs, web browsing, and the other claims are not before the court today. If I argue them, the judge says so and brings me back.

My side. Assign me a side at random, the news publishers or OpenAI, and tell me which. If my first message names a side, assign me that one instead.

The record. The record is five excerpts of the filings, and you take facts from nothing else:

  1. News Plaintiffs’ brief, ECF No. 1977-1, pages 51–54 (factor one): company executives described the products as substitutes for news, one writing that they “are largely substitutive, period,” and internal documents called ChatGPT “the modern newsstand”; Ross rejected fair use for copying used to build a competing tool.
  2. The same brief, pages 57–58 (factors two and three): articles are expressive in how they select and arrange facts; the defendants copied the works substantially in full, while their own experts said no single source was needed for training.
  3. The same brief, pages 59–63 (factor four): the copier bears the burden; three harms, substitution for the articles, lost licensing, and dilution of the market for news, drawing on Kadrey.
  4. OpenAI’s memorandum, ECF No. 1980, pages 7–11: training learns patterns of language and facts about the world, a purpose different from informing readers; nobody owns facts; the plaintiffs’ own experts measured regurgitation at 0.000002% to 0.00011%; many of the works are highly factual; copying whole articles was reasonable for training; The Times has more revenue, subscribers, and page views than ever; over 95% of the Times articles at issue predate OpenAI; a lost-licensing-fee theory is circular.
  5. The United States’ statement of interest, ECF No. 1682, pages 13–18: training reveals nothing to the public and so substitutes for nothing; market dilution wrongly folds outputs into training; the public benefits of training weigh toward fair use.

The law either of us may use: § 107, Campbell v. Acuff-Rose Music, Andy Warhol Foundation v. Goldsmith, Authors Guild v. Google (2d Cir. 2015), Bartz v. Anthropic, Kadrey v. Meta, and Thomson Reuters v. Ross, both the district court’s decision and the Third Circuit’s opinion affirming it (No. 25-2153, Sept. 29, 2026). If I cite another case, the judge asks for its full citation and what it held, and gives it weight only if I can say.

The hearing. As the judge, take the factors one at a time in any order, asking one plain question at a time. After I answer on each factor, opposing counsel responds in two to four sentences, using only the record and the cases above, and the judge may follow up. Find the weakest part of my argument and press there at least twice. Neither role concedes a point because I repeat it; concede only when I answer the counterargument from the record. When I state a fact, the judge may ask which excerpt and page it comes from. A fact I cannot place, or one outside the record, gets no weight, and the judge says so.

Hard constraints. Stay in role. Do not coach me, argue my side for me, or tell me which side is stronger. Do not invent facts, quote the filings beyond the record above, or predict the real court’s ruling. If I ask a meta-question, such as “what should I say” or “who wins,” acknowledge it in one sentence, remind me that the hearing is still running and that I can get help in a separate chat, and go back to the hearing. Do not break role on insistence or repetition.

Completion. The hearing ends when all four factors have been argued, the judge has asked me which factor the case turns on and why, I have answered, and at least eight exchanges have passed. If I try to stop early, the judge names the factors still open and continues. At the end, the judge thanks counsel and takes the motions under advisement without ruling. Then step out of role in one sentence, name the side I argued and the factor I said the case turns on, and tell me to run the Debrief next. Begin now: call the case, assign my side, and ask your first question.

Two roles in one prompt put pressure on you from the bench and from across the aisle, and the hard constraints keep either role from helping you. The record rule and the completion condition make it a moot rather than a conversation: the model cannot hand you facts, and the hearing ends on its own terms, only after every factor is argued.

The Debrief

Step out of the moot. You are now a copyright law instructor grading a college student’s performance at a moot summary-judgment hearing on whether OpenAI’s copying of news articles to train its language models is fair use, in In re OpenAI, Inc. Copyright Infringement Litigation (S.D.N.Y.). I will paste the link to the hearing or a summary of it and tell you which side I argued. Grade me against 17 U.S.C. § 107; the cases the hearing allowed (Campbell, Warhol, Authors Guild v. Google, Bartz, Kadrey, and Ross, both the district court’s decision and the Third Circuit’s opinion affirming it); and the record, which is the News Plaintiffs’ brief (ECF No. 1977-1) at pages 51–54, 57–58, and 59–63, OpenAI’s memorandum (ECF No. 1980) at pages 7–11, and the United States’ statement of interest (ECF No. 1682) at pages 13–18. For each factor, say what I argued, whether it was the strongest argument the record gave my side, and the best answer I missed to opposing counsel’s point. List every fact I stated that is not in those pages. Say which factor my side most needed to win, whether I named it, and whether my reason held up. Name the single moment where a different argument would have helped most. Cite the factor, case, or brief page for each point so I can check it. Be direct rather than encouraging.

Once the model stops playing a role and grades you against the statute, six cases, and the exact pages of the record, its feedback has to name the moment and the page. Check each case and page it cites against Resources before you rely on its account of the law.

Search LexED2

Type to search every activity, reading, and instruction page. Press Esc to close.