I received a couple questions from someone about IP and AI.

Question 1:

we could imagine an AI system without any censorship which can quote anything without any restrictions and provide access to any information that is available, anyone could create a website with any pirate content they want and it uses modern technologies allowing to provide this service without any way blocking it. How long IP could exist if there was a tool that completely ignores human made laws and lets information live freely?

Kinsella:

I’m not quite sure what you are asking. I think copyright is incompatible with AI. You can’t have both. That’s the problem with copyright. It’s already affecting AI. See, re Sarah Silverman’s suit against OpenAI, “AI Suffers Setback As Judge Trims Case“; “The Times Sues OpenAI and Microsoft Over A.I. Use of Copyrighted Work” (“Millions of articles from The New York Times were used to train chatbots that now compete with it, the lawsuit said”); Artists release silent album in protest against AI using their work; Hundreds of actors and Hollywood insiders sign open letter urging government not to loosen copyright laws for AI. See this tweet thread with Kinsella explaining that copyright gimps AI; tweet about “The EU’s AI Act invalidates the theory of Non-Expressive Use. It’s a major international precedent set by policymakers to consider training as falling under Copyright.”; tweet about “The historic NYT v. @OpenAI lawsuit”; tweet re copyrightable AI art; tweet “Potential copyright infringement & why generative AI may be in for a rough ride”. [See also Whereupon Grok admits it (and AI) is severely gimped by copyright law; twitter (“Meta has pirated millions of books to train its AI. Either this stops now or publishing is finished.”)]

Response:

Obviously OpenAI is affected by the government regulations, I am talking about black market version of it, that doesn’t care about made up laws and acts strictly within voluntary cooperation. The question is: if there was a distributed system (kind of like bitcoin is) that couldn’t be blocked by the government and allowed using AI without any restrictions, how in your opinion would it change the mindset of the people regarding copyright?

Kinsella:

Hard to say. Probably not much. They are too confused about IP and copyright to start to understand it just because of some obvious examples. Instead they would (a) minimize the example and focus on how the unregulated blackmarket AI is also being used for bad things (it might give racist answers etc., or “for crime,” like Bitcoin or The Silk Road) and (b) they would say, “well this just means copyright law is being abused here and all we need to do is find the right ‘balance.'” No one can ever think in principled terms.

After all think of how copyright obviously hobbles Youtube, but no one says we should abolish copyright because of this. They just moan about “abuse” and say the system needs to be tweaked or improved or fixed to achieve the right “balance.” It gets tedious to hear this nonsense over and over.

So I expect copyright to continue to hobble AI (patents might too; see how nChain/Craig Wright tried to use both copyright and patents, maybe trademark too, I can’t recall, as threats against the bitcoin ecosystem).¹ This will mean it will have reduced functionality and it will be more expensive as the AI companies are extorted into paying ransom in the form of “license fees” to book publishers, newspapers, and others with content on the Internet. It’s going to hold back human progress, as IP always does. No offense, Heritage Foundation, Cato, Independent Institute, Federalist Society, and others.²

Question 2:

Intellectual property is one of the most fascinating and, at the same time, controversial concepts created by the state to regulate interactions between people. The term itself is essentially an oxymoron because ideas, knowledge, or creative expressions cannot be “owned” in the same sense as physical objects. You cannot restrict the spread of a thought once it enters someone else’s mind. Yet the state has invented rules that allow this to happen through coercion, restrictions, and penalties.

True property is based on the principle of self-ownership: you own your body and, therefore, the fruits of your labor if they are created without violating the rights of others. But can it truly be considered a violation to “copy” an idea that someone has heard or seen? If I create a copy of your book or invent a similar machine, does that really harm your property? After all, the original remains with you, and you have not lost anything. This contradicts the very nature of property, which aims to avoid conflict over scarce resources.

The system of intellectual “property” benefits only those who wish to use coercion for profit: corporations, states, and bureaucrats. Authors, inventors, or artists receive only an illusion of protection, which quickly shatters against the reality of lawsuits, patent trolls, and bureaucratic obstacles. In contrast, true freedom of creativity and innovation comes from open systems where ideas freely circulate and enrich society.

Everyone who creates something in this world has the right to decide how and with whom to share their work. If your ideas are truly valuable, you will find those willing to support you voluntarily. But imposing a monopoly on thought is an attack on the freedom of others, on their right to use their own minds, to create, and to share their ideas.

Thus, the issue of intellectual property is ultimately a question of freedom versus coercion. A free society does not need state patents or copyrights. It needs a space for collaboration where people create, copy, improve, and freely exchange ideas without fear of bureaucracy or legal sanctions.

Kinsella:

This is not bad, and it’s aiming at the right answer. There are few things I would tweak.

First, a pedantic point. You write: “the issue of intellectual property is ultimately a question of freedom versus coercion”. We libertarians oppose aggression; libertarians sometimes use “coercion” as a synonym for aggression, just like we sometimes (sloppily) say that we oppose “violence.” But not all coercion is aggression, just like not all force or violence is aggression. See my posts The Problem with “Coercion” and The State is not the government; we don’t own property; scarcity doesn’t mean rare; coercion is not aggression.

Second, you write: “True property is based on the principle of self-ownership: you own your body and, therefore, the fruits of your labor if they are created without violating the rights of others.”

Here, you speak of “true property.” As one legal scholar explains,

In the United States, the word property is frequently used to denote indiscriminately either the objects of rights … or the rights that persons have with respect to things. Thus, lands, automobiles, and jewels are said to be property; and rights, such as ownership, servitudes, and leases, are likewise said to be property. This latent confusion between rights and their objects has its roots in texts of Roman law and is also encountered in other legal systems of the western world. Accurate analysis should reserve the use of the word property for the designation of rights that persons have with respect to things. (( See Kinsella, Legal Foundations of a Free Society (Houston, Texas: Papinian Press, 2023) [LFFS], ch. 2, App. I. ))

So the question is not what “is property” but what human actors have property rights to. As Rothbard explains, all (human) rights just are property rights; and all property rights are rights in scarce resources.³ And to be just—according to libertarianism and the private law—these property rights must assigned in accordance with (1) original appropriation (occupation; Lockean homesteading) or (2) contractual title transfer from a previous owner.⁴

I would also slightly disagree with your wording: “you own your body and, therefore, the fruits of your labor if they are created without violating the rights of others”. We do not own labor, or the “fruits of” our labor. We own previously-owned scarce resources acquired either by original appropriation or by contractual title transfer from the previous owner. Neither one of these actions “creates” the “thing” owned, nor do they require the assumption that an actor owns of the “fruits of his labor”. When an actor appropriates or occupies an unowned resource, he does not create it; it already existed; he merely appropriates it. Yes, yes, this effort involves the use of labor (and intellect, knowledge, and so on), but the labor is merely a type of action (as opposed to leisure); labor is not owned, and neither is “action” or “leisure.”⁵

It is true that laboring—rearranging an already-owned resource—is a source of wealth but not of property rights.⁶

In light of all this, I would also say that the problem with IP is not that ideas “are not property.” Even scarce resources are not “property”; as noted above, humans have property rights in scarce resources (determined in accordance with original appropriation and contractual transfer). The thing that I own is not “property”; it is something in which I have a property (ownership) right.

The problem with IP is not that ideas are “not property” or that IP is “not property”; the problem with IP law and the IP rights it creates, is that that IP rights violate existing property rights—since IP gives ownership rights to IP holders, over resources already owned by others in accordance with principles of original appropriation and contractual transfer. IP rights and IP law are unjust. This is the fundamental problem with IP rights.⁷

As I pointed out in “Intellectual Property versus Intellectual Property Rights“:

One mistake made by many opponents of IP is that they believe the problem with IP is that it is “not property,” which is one reason they are reluctant to adopt the loaded term “intellectual property.” But this is because they still hew to the common view that things we have property rights in “are property.” If they believe that IP law is illegitimate, this means that “intellectual property” is not actually “property”; that there is no such thing as “intellectual property”; or as some of them say, “intellectual property does not exist.” As this chapter will make clear, the problem with IP is not that it does not exist, but that IP rights and IP law are unjust. Inventions and creative works exist; patents and copyrights, and patent and copyright law, exist. The opponents of IP here remind me a bit of the natural law types who resist calling a bad law “law” but instead say things like, an unjust law is no law at all.

(See also my similar comments in “Munger on Property Rights in Words and Information.”)

Update: See also Trump fires director of U.S. Copyright Office, sources say. Evidence of growing obvious “tension” between copyright and AI, and a possible positive movement?

The Trump administration has fired the head of the U.S. Copyright Office, two sources familiar with the situation confirmed to CBS News Saturday.

The firing of Register of Copyrights Shira Perlmutter came after Perlmutter and her office earlier this week issued part three of a lengthy report about artificial intelligence and expressed some concerns and questions about the usage of copyrighted materials by AI technology.

“It is an open question, however, how much data an AI developer needs, and the marginal effect of more data on a model’s capabilities,” the report read. “Not everyone agrees that further increases in data and test performance will necessarily lead to continued real world improvements in utility.”

… Morelle speculated that there was “surely no coincidence he acted less than a day after she refused to rubber-stamp Elon Musk’s efforts to mine troves of copyrighted works to train AI models,” in reference to the report released by the Copyright Office this week.

Last month, Musk took to his social media platform X to seemingly express support for the abolition of intellectual property laws. Musk also owns AI startup xAI, with which in February he submitted a failed bid to purchase OpenAI, the company that operates ChatGPT.

Mr. Trump has been a major proponent of AI. Immediately after taking office, he announced a joint venture involving OpenAI, Softbank and Oracle that will invest up to $500 billion in private sector money to build artificial intelligence infrastructure.

Update: Recent article about “Fair Training” exceptions to copyright law for AI training:

Andrew W. Torrance & Bill Tomlinson, “Training is Everything: Artificial Intelligence, Copyright, and ‘Fair Training,’” 128 Dick. L. Rev. 233 (2023).

Abstract: In this Essay, we analyze the arguments in favor of, and against, viewing the use of copyrighted works in training sets for AI as fair use. We call this form of fair use “fair training.” We identify both strong and spurious arguments on both sides of this debate. In addition, we attempt to take a broader perspective, weighing the societal costs (e.g., replacement of certain forms of human employment) and benefits (e.g., the possibility of novel AI-based approaches to global issues such as environmental disruption) of allowing AI to make easy use of copyrighted works as training sets to facilitate the development, improvement, adoption, and diffusion of AI. Finally, we suggest that the debate over AI and copyrighted works may be a tempest in a teapot when placed in the wider context of massive societal challenges such as poverty, inequality, climate change, and loss of biodiversity, to which AI may be part of the solution. https://papers.ssrn.com/sol3/papers.cfm?abstract_id=4700945

This piece seems to be fairly useless, since the authors are not radicals and want to achieve a “balance” between anti-IP views and pro-technology considerations. Take a stand make some actual predictions, guys.

***

I use LawLine for CLE (continuing legal education); here some of the AI-related courses offered right now (see below), such as this one for example:

Intellectual Property Issues and Generative Artificial Intelligence in the Courts

 

 

Update: AI Training vs. Copyright Law: Updates from the Copyright Office and the Courts (video embedded below), discussing:

• US Copyright Office: Part 3 of the U.S. Copyright Office’s report on Copyright and Artificial Intelligence, focused on Generative AI Training (May 9, 2025), which is also linked and summarized on the Copyright Office’s dedicated pages for its Artificial Intelligence Study and Copyright and Artificial Intelligence initiative: Copyright and Artificial Intelligence and Artificial Intelligence Study
• The District Court in Delaware case referenced is Thomson Reuters Enterprise Centre GmbH et al v. ROSS Intelligence Inc. (Case No. 20-cv-00613-SB, U.S. District Court for the District of Delaware). On February 11, 2025, Judge Stephanos Bibas issued a summary judgment opinion rejecting the defendant’s fair use defense, finding that the use of copyrighted legal headnotes to train an AI legal research tool constituted infringement (though some factual issues remain for trial). This reversed aspects of the court’s own 2023 ruling after renewed motions. Links: Court opinion PDF, Docket on CourtListener.
• Others summarized by Grok

Transcript from Youtube, cleaned up by Grok:

 

 

Cleaned Transcript with Time Markers

[0:00]Edith Herald: Hello everyone and welcome to this Federalist Society virtual event. My name is Edith Herald and I’m an assistant director of practice groups with the Federalist Society. Today we’re excited to be hosting this FedSoc forum called “AI Training Versus Copyright Law: Updates from the Copyright Office and the Courts.” We’re very pleased to welcome our excellent speakers to this discussion today. We have Meredith Rose, who is a senior policy counsel at Public Knowledge, and Reagan Smith, who is the senior vice president and general counsel at News Media Alliance. And special thanks to Professor Zvi Rosen for moderating this discussion today. Professor Rosen is an associate professor of law at the UNH Franklin Pierce School of Law. And if you’d like to learn more about today’s moderator or speakers, their full bios can be viewed on our website, fedsoc.org.

[0:49] Throughout the program, we may turn to the audience for questions. So, please, if you have a question, enter it into the chat at the bottom of your Zoom window and we’ll do our best to answer as many as we can. And finally, I’ll note that as always, all expressions of opinion today are those of our guest speakers and not the Federalist Society. With that, Zvi, thanks so much for joining us today, and I’ll hand things over to you.

[1:01]Zvi Rosen: Thanks so much, Edie. I’m so excited about this event. We’re in really interesting times. When we were originally discussing this event, it was really going to be focused on the Copyright Office report on AI. The Copyright Office announced a Notice of Inquiry on copyrighted AI in 2023, and the process of putting together a report took some time. There was a Part 1 on digital replicas, Part 2 on copyrightability of authorship, and then pretty recently, the Copyright Office put out a pre-publication Part 3 of a report on AI training and fair use. And as that came out, there was a bit of a kerfuffle, some might say, where inter alia, both the Librarian of Congress and the Register of Copyrights were removed from their positions. And I will say the Federalist Society is putting on a webinar on that issue on August 12th. If you want more on that, please show up. We have a wonderful panel—I’ll be part of that as well, but I’m not one of the Con Law people.

[2:15] But then in the past week or so, we’ve had two more opinions. It might be 10 days by now. Both out of the Northern District of California: Bartz versus Anthropic, where Judge Alsup held that Anthropic, which makes Claude, could at least partially invoke the fair use defense for AI training—although that was not fully, and there were some very notable caveats there, specifically regarding pirate libraries. And Kadrey versus Meta, decided a day later by Judge Chhabria—and I need to caveat there that I was a signatory on an amicus brief on behalf of plaintiffs in Kadrey versus Meta—where Judge Chhabria held that under the current facts, it was a fair use to use copyrighted material for AI training, but in a tone that made very clear that it was a fact-specific opinion. Well, so the first thing first, and I’m going to focus this on Reagan, but Meredith’s comments are definitely welcome here as well. What is the purpose and role of the Copyright Office policy report?

[3:21]Reagan Smith: Thank you, Zvi, and thanks for having me today. So, my name is Reagan Smith, and I’m at a trade association called the News Media Alliance right now, which represents over 2,000 publishers active in the US market who are watching AI issues really carefully. They’re engaged in a ton of partnerships and integrating their businesses, but at the same time, they want to make sure that they can preserve their business model as they engage with this technology.

[3:50] So, we’ve been following the Copyright Office report and some of these opinions very closely and excited to talk about it today. Previously, I was in a global policy role at Spotify. But I think the reason why you’re asking me this question is I was also general counsel of the US Copyright Office in the first Trump administration and at the Copyright Office for a while before that. And so the Copyright Office does a number of policy studies. They’ve done this for decades. And in the Library of Congress, a lot of the times they’re coming as a result of a request by Congress, which was the case in this case—they had asked the Register of Copyrights to look at artificial intelligence, generative AI issues, and they had a schedule of four studies. And this is the third piece that we’re going to talk about today along with some of the litigation developments.

[4:53] And so these studies are generally looked at by Congress sometimes in forming legislation or deciding not to make legislation. They’re also considered by stakeholders, agencies, courts in terms of articulating views of the law or policy instincts by stakeholders. And just like I know because I think it might be of interest to this audience, the reports generally, if they are considered by a court, would be considered to the extent where they’re persuasive. So I don’t think there’s any administrative law issues really to consider there in terms of changing precedent. But a lot of times courts have found the Copyright Office’s reasoning to be persuasive on issues because you’re really kind of digging into a lot of the nuts and bolts of the copyright law and some of the copyright economy.

[5:35] So in this case they had 10,000 comments. They worked on this report for about a year and they tried to walk through different scenarios or groupings of issues in connection with generative AI and at a really high level the report sort of says it depends because fair use is an issue of fact and law and the facts are going to change. But they’re trying to articulate some legal concepts relevant to training and fair use and whether there’s infringement and whether the affirmative defense of fair use applies, which is kind of similar to many reports that the Copyright Office has done, for example, when the 121 rulemaking or we did a 121 policy study when I was there. And that would be something sort of similar.

[6:37]Zvi Rosen: Thank you, Meredith. As someone who’s been at a group that is one of the most important advocacy groups for what you might call more liberal copyright policies, some might say. How do you guys approach a Copyright Office report? And how does it look from the outside?

[6:59]Meredith Rose: Yeah. So you know I think Reagan articulated pretty well like procedurally—often the Copyright Office gets asked these very broad questions by Congress which in this case I think they were just instructed to write about AI and copyright and you know like a year and a half two years later we’re still sort of working through the implications of that question. It was a very very broad remit. And so civil society has a chance to come in—obviously like we file comments—and the Copyright Office did just round table after round table on topics like this and really got stakeholders from every conceivable position and some unconceivable positions coming in and sort of sharing their perspectives on it.

[7:43] And the resulting report obviously took a very long time because this was an enormously hefty topic to work on. And you know some of the things that were in it I think were actually very useful. There was a very comprehensive discussion about how training actually works for AI. There’s a lot of attention given in particular to LLMs and things like RAG models. For folks who are not familiar, RAG models are one of the models that’s most concerning to journalists because what the defining feature of a RAG model is that it goes and actively has a search backend. So it actively searches for information in response to the query as opposed to things like ChatGPT which is sort of trained in bulk at discrete points in time.

[8:31] And so there was a lot of discussion of that. It was a very good overview of all the different stakeholder positions. Some of them have changed a little bit because again the process was begun I think in late 2023 when a lot of this initially kicked off. But it’s a really good sort of overview of the different universe of perspectives on this as well as a discussion of you know how are these models actually trained? Which is very very deeply needed given the degree of sort of conflicting information about even just the technical underlying aspects of this.

[9:08] I think as far as the actual sort of fair use analysis that the report engages in there’s a lot of differences of opinion about how that came out. Fair use is sort of necessarily—it’s a four-factor, very wobbly might be a generous term to use for it sometimes. And so, you know, if you have three fair use lawyers in a room, you’re going to get five different opinions about any given case, especially something relatively technical like this. But I think, you know, to reiterate what Reagan said, this is a very big lift.

[9:43] The Copyright Office’s top line essentially was that copyright law in the US and especially fair use can actually deal with a lot of these questions right now. One of the questions that Congress asked is, do we need new laws to really keep pace with this? And the resounding answer of the Copyright Office was no—at least not at this exact moment. They did say, you know, there are some things we have to kind of wait and see if or how licensing markets start to develop. If it looks like they’re developing in a way that might be anti-competitive or raise some policy concerns, then they offered some suggestions about how to manage that going forward. Yeah, but the topline conclusion was really we’re still in the early stages. We think that the law is flexible enough to deal with these scenarios through current litigation and so that Congress should take a sort of wait-and-see approach to how things are playing out.

[10:39]Zvi Rosen: Thank you. I mean, and we should say that the Office did really an amazing job getting through all of this. They were burning the midnight oil going through 10,000 comments and really getting up to speed on a wide variety of both technical—I mean they were already up to speed on the legal side—but a really wide array of technical stuff, and the report is really valuable in that way as well as saying where we are. How does the Copyright Office and these recent decisions we’ve been talking about approach questions of prima facie infringement? Do you think that the Copyright Office and recent decisions got it right? And do you have any thoughts about that?

[11:23]Reagan Smith: Yeah, Reagan, do you want to go first on that one? Yeah, we don’t have to, you know, keep going in order, but I’m happy to jump in because I think this one though is not a really contentious question. I think we’ll probably all agree that, you know, we’re here because there is prima facie infringement generally in the LLM developments, which were the cases of the two decisions. And maybe we can also talk about the Thomson Reuters versus Ross decision that came a little bit before the report today because I think that’s a case in Delaware that I think is also now going on appeal where they found it was not fair use to create a competitor to the Westlaw model.

[11:54] But the Copyright Office report, as Meredith said, kind of walks through what is going on. There’s a lot of academic, you know, scientific journals like and stakeholder comments where everybody agrees there’s a ton of copying going on, right. So in copyright, if you make a copy without authorization and you have access to it, that sort of puts you into infringement land, and then when we talk about fair use it’s as a defense to infringement. So the Copyright Office walks through sort of the life cycle of some of the ways these copies are being made.

[12:32] So maybe they’re coming from creating a data set which might be for one purpose, and then it might go into training, keeping a reference file post-training or fine-tuning, which is, you know, you’ve got maybe a more amorphous model and you’re trying to customize it to do a specific task. Meredith mentioned RAG, which is when there’s, you know, like a totally different set of copying that is not part of the model itself, but something the model is querying. And so all of those acts of copying are ones that copyright law needs to sort of accommodate and think about in terms of who’s doing what and whether or not this is something actionable under the copyright law.

[13:16] And so the recent decisions did that as well. I think you mentioned the Bartz versus Anthropic case. In Anthropic there, the records I guess in the opinion were that they were creating a central library of materials that they had created and put and were just sort of storing within the company’s possession that they were then using for some of their various different AI uses.

[13:46] And then the final thing I would say that’s in the Copyright Office report that was interesting and was something of interest to the NMA members was they also talk about in some instances there are memorization or retention in a model itself. And they pointed to some academic articles that say it explains why this happens in some but not all cases. And it seems to be an artifact of the frequency with which or the way—I’m trying to kind of speak in a way that’ll make sense to everyone whether we’re AI scientists or not—but the process of training the model and how often it is exposed to this through the encoding and tokenization process. So if that is the case, that might become another copy that then you need to evaluate when you start looking at how the copyright law is going to accommodate generative AI uses while sort of preserving the incentive to produce these copyrighted materials that are being used with this technology.

[14:51]Zvi Rosen: I want your take as well. I just want to jump in though and say yeah, we definitely should talk about Thomson Reuters versus Ross, which is about training on West Head Notes for to create a competing AI-powered legal research tool. And we also should mention of course with memorization there’s two cases which are working their way up—not these decisions—which is the New York Times versus OpenAI lawsuit and also the Disney lawsuit versus—I actually don’t know who Disney sued, AI people.

[15:19]Meredith Rose: I think that was Stability.

[15:26]Zvi Rosen: Okay. Yeah. Yeah. There’s so many. There’s 40-some AI lawsuits in the US alone. I think at least a dozen outside the US. There’s a lot happening. But Meredith, I don’t want to cut you off from putting in your piece on that.

[15:38]Meredith Rose: Yeah, of course. You know, I think there’s a lot of, you know, as I mentioned earlier, if you ask three people, you’re going to get five different opinions on this, and it is a very robust report and so there’s a lot of things to spark debate around. You know, I think Reagan mentioned this sort of memorization problem and this argument that if a model perhaps memorizes accidentally—which let’s sort of be clear from a policy perspective these AI companies do not want it to be able to precisely regurgitate because I don’t think there’s any disagreement that if you have an infringing output that is a copyright infringement. I think that’s pretty much everybody can kind of agree—if you regurgitate an exact copy of something that was in the training data, that’s a problem.

[16:25] But we have situations now where occasionally what are called model weights can internalize essentially a piece of data that it was trained on. Just to kind of zoom out and do a little bit of a background science on how these models are trained. The very very short oversimplified version is that you continuously sort of feed material in. If you kind of imagine like a conveyor belt going through almost like a Jetsons-style box with a bunch of knobs and dials on it. You feed something in one end and then you see sort of what you get at the output. And whether the thing that comes out of the other end is kind of close to what you like. And then you fiddle with the knobs and the dials until it gets closer and closer and closer to the sort of desired output that you’re searching for.

[17:19] Those positions of those knobs and those dials are what are called model weights. So when you have things like Facebook’s Llama model, which is an open-source model, what they do is they make their model weights open source so that anybody can say, “Okay, I’ll put my settings here and I’ll get something sort of closely approximating the model that is being shared.” This kind of creates a real problem because this argument is that well if a model has accidentally memorized something even if it doesn’t have an output that actually infringes on it then the model itself might be sort of essentially an infringing file format for the thing that it has ingested and memorized, which can create a whole lot of complicating problems particularly when you know I think copyright law tends to focus most appropriately on when is a copy made perceivable by a human as opposed to being entirely only sort of replicable or parsable by a computer.

[18:15] Which is when you get into like—and I don’t want to get into the sort of intermediate copy debate that a lot of copyright lawyers will fall down—but it is this sort of open question right about if you’ve got a model it’s like does a tree fall in the forest and no one’s around to hear it? If an AI memorizes something and never outputs it, do you have an infringement? And that’s really one of the questions that came up in this report. And I think the Copyright Office came down on the side of yeah, probably that’s probably an infringement even if it never actually outputs the thing that it’s memorized. Which is I would disagree with that pretty strongly. That if it’s carrying around a memorization that never gets output, then the point of copyright law is to focus on works that are perceivable to humans. And so I think that would be taking the reasoning a step too far.

[18:57]Reagan Smith: If I could, I think we have a different series of views. I think the idea that perceptible to humans was the test itself is sort of dead in the water from the 1909 Copyright Act. That was sort of that issue. In 1908 there was a Supreme Court decision about player pianos. And ever since then our Congress has very wisely through updates to the Copyright Act said no, that’s not going to work. We’re going to be a little bit technology agnostic and look at a copy as a copy. Now, we have ways of exonerating or saying that copy is not actionable if it’s happening inside a computer. But there’s also ways ensuring—and that might be primary or secondary liability, that might be what the ultimate purpose of something is that I think we’re going to get into.

[19:40] But I think it’s very important to consider, as Judge Alsup also in the Anthropic case found in that case. It’s part of his decision that there is copies actually retained in the model itself. So that supports what the Copyright Office said, which is that sometimes it does happen. When they’re looking at the report they reject the idea that it’s a bug and not a feature and they say actually it’s kind of endemic through this process of you know sort of the Jetsons boop boop that you explained, Meredith, that some of it is going to be retained and what might be unintentional is whether AI companies are putting enough mitigation efforts after the fact through the user interface or what you know it’s kind of being called guard rails to make sure that outputs don’t come out that are either infringing or too directly competitive.

[20:30] And I think that’s like an interesting area of what exactly is that line we’re going to probe. But the point is if you’ve made a copy and you’re walking around with something in body that’s using it, that’s something that our copyright law can and does look at whether or not that should be something you need to seek authorization for. And so one example would be there was an appellate case by Texaco where Texaco the oil company was making a bunch of copies and they said well we’re going to you know later do this for science and things that are not putting out the journals that they were copying without taking subscriptions to it. And the court there, which is the Second Circuit, said, “No, no, there’s like actually a license here, and you’ve got to buy your subscriptions for use in your corporations as scientists and then go off and do science, and that will be great.”

[21:16]Zvi Rosen: Oh, it almost feels like the issue once again on prima facie infringement is the old RAM copy issue, MAI versus Peak, whether loading a program into RAM to some degree. Meredith, I’ll give you a chance to respond, but you want to jump on to fair use as well.

[21:35]Meredith Rose: Yeah, I think the fair use question is more interesting, but I will just say that the Jetsons beep boop machine is actually the term of art now. So, I’m going to continue to use that.

[21:47]Zvi Rosen: Excellent. Well, and if anyone else used it, they should cite this webinar.

[21:52]Meredith Rose: I commit that analogy to the public domain. That is entirely in character. I appreciate it.

[21:57]Zvi Rosen: So fair use of course is four elements but the second and third—the nature of the work and amount and substantiality—they matter but I think the first and fourth—the purpose and character of the use and the market effects—are the ones that are being most hotly debated. Well, and of course, we’ve really redefined purpose and character in substantial part, not fully, as being how much does it transform? And well, I will say I think Judge Chhabria in particular does talk for a while about commerciality and says it is actually relevant—it’s not a nothing weight as it has been in some cases. But yeah, how should we think about the first fair use factor in these cases? Is this transformative? And how do we navigate the Scylla and Charybdis if you will of Warhol and Google versus Oracle.

[23:01]Meredith Rose: Yeah I see Meredith is ready to go on this one. So so many thoughts. Yeah so you know as you mentioned transformativeness has kind of become a shorthand and again folks will debate whether appropriately or not has become sort of a shorthand for a way to analyze the first factor about how much does this new work transform the works that were sort of input into it or that were sampled or made use of in order to create the output work. I think you know both judges talked about this being a transformative use.

[23:32] I think they both agreed that you know using published texts in order to train a machine so that it could put out wholly original new text create new generative work is about as transformative as you get. I think the complicating factor and this comes up quite a bit in the Copyright Office report is this Warhol decision which referenced—and I know Zvi also did like a very good also Federalist Society talk about this which you should go look up. Warhol was an interesting case which has given people a lot of heartburn in the copyright community. Everybody seems to have heartburn for slightly different reasons.

[24:08] But the fact pattern in Warhol was essentially that you had a photographer, Lynn Goldsmith, who had taken a photo of Prince, the recording artist, at the very beginning of his career, and put it into a sort of licensing photo pool that I believe is either run by or partly participated in by Vanity Fair, the magazine. Andy Warhol made a licensed authorized silkscreen print sort of if you’ve seen his famous like Marilyn Monroe prints it was in that style based on this photo. Now unbeknownst to Vanity Fair or to the original photographer he then went and made another 15 of them on top of the one that he had gotten permission to do.

[24:45] Those were sort of held in the private collection. This was about three years before he died that he did this. So after he passed, it was held in the private collection. It was shown at some galleries, but it was never commercial. The rest of these other 15 were never commercially exploited until Vanity Fair decided to run a sort of in memoriam print retrospective. I believe shortly after Prince passed away. And they licensed one of the other ones in order to put on their cover. And Goldsmith sued.

[25:14] And then we got into this very long sort of protracted as copyright litigation tends to be protracted litigation around whether or not that was a fair use. And the Supreme Court essentially came down and said that the first factor needs to consider among other things—and it’s part of the language of the first factor in the statute that it says you need to also consider commerciality or whether the use is for a nonprofit educational purpose.

[25:46] And so the way the court interpreted that is well if it’s a commercial use that actually it’s you can sort of imagine as like a sliding scale with transformative on one end and commercial on the other. And the way that they sort of visualized it is that the more commercial it is the more it offsets how transformative it is. That we can debate about sort of how that actually practically works or not.

[26:04] But it definitely works more in a situation where at least for legal purposes you have sort of one controlling entity that is making all of these decisions right so Andy Warhol made the decision to make these prints his estate so legally technically the same thing made the decision to commercialize it so you can all sort of bundle it together under the same decision-making authority. This gets very complicated when you start applying it to AI where the chain of decision-making often breaks several times before you get to the end output.

[26:41] And so the Copyright Office was tasked with the unenviable duty of trying to understand what if anything Warhol meant for this AI fair use analysis. They came down with the perspective that whatever end use you’ve got the base model and then the base models are distributed for fine-tuning often to a second or third party which are then redistributed to more folks who could use it as end users.

[27:09] And they sort of took this and interpreted it as well, whatever this end user does with it, if the end user uses it to create a work that might be in commercial competition with something that was in the training set, that might render the training not a fair use. Which creates all kinds of problems. For one thing, it’s sort of retroactivity problems. Is it only not a fair use? Does this sort of create a poison the roots of all the other reinstantiations and fine tunings that are done off of these models. You get a very sort of muddled end result logically.

[27:46] And so you know their topline analysis that this is a really fact-specific question and fair use is necessarily limited to specific circumstances is correct. But the way in which they tried to reconcile this very difficult decision which itself involves a little bit of time travel with AI and this general purpose technology really just created I think a lot more questions than it ended up answering.

[28:11]Zvi Rosen: There is certainly a lot there. And I’m sure Reagan yeah I’m sure you have lots of thoughts as well.

[28:16]Reagan Smith: Yeah. So that was a lot. So I’m going to like back up to the beginning and what even is the first factor. So the first factor which is a statutory factor says you should look at the nature of the use including whether it is commercial and non-commercial. This is in the statute of the Copyright Act which is coming forward. It’s been codified in US laws like a judge-made law and developments of doctrines for over about a century.

[28:48] So the purpose of the use is what we’re ultimately looking at under the first factor which is one of the most important factors and so in that respect that’s what the Warhol decision which was a 6-3 Supreme Court decision says let’s look at what the use is right so Warhol is the law of the land and I think correctly decided but that’s why it’s focusing on what is the use and I think that decision and then we can go into AI says let’s look at whether it is for a commercial purpose or not and makes clear that this phrase transformative is no longer or never was I guess sort of like a litmus test to resolve the first factor in and of itself.

[29:19] But why are we even saying the word transformative if I just said the statutory language is the purpose of the use of the copyrighted work? It’s because there is a theory and it comes from Judge Leval and was also endorsed by the Supreme Court in this case about Campbell versus Acuff-Rose which is about using a parody of a music song whether you’re transforming or commenting or targeting sort of needing to use the first work in order to achieve a transformative purpose.

[29:54] And so that line of transformative which there’s a lot of cases on kind of looks at a couple of things that the Warhol decision and as well as the Copyright Office report and these other opinions we’re talking about focuses on which is what are you using this copyrighted work for right so one of the requirements of transformativeness might be that it needs to be sort of like targeted so you’re taking in the Campbell case a Roy Orbison song because you want to make commentary about it by a rap group in that case 2 Live Crew. So, you needed to use that to make your fair use in that case. And it also needs to relate back to the underlying work. You’re actually talking about that as opposed to you’re taking one thing for no other reason, but you could have picked something else. So, you need to have a justification for why you took something in the first place.

[30:44] Now, that gets interesting in the case of AI because I think that’s an open debate and it also might depend on what the developer is trying to do. So around the same time the Copyright Office report came out, there was a report done and published. There was the Washington Post had a good article about it talking about a fully permissioned like trained LLM model where everything was either in the public domain or had been licensed. So in that case you might not have a reason to take material that you have not gotten permission for in order to achieve in that case they said the model was equivalent to the Llama model that Meta was accused of and you know was the subject of the Kadrey versus Meta decision in the Northern District of California.

[31:34] For another reason I would say the RAG copying we were talking about where you’re looking at a separate database of something and sometimes what is news copying—that’s why we’re really interested in it—a news article. Why are you looking at that? Well, it might be because you want to do something similar to it, but you didn’t need that to train the model because you’re not even using that to train the model.

[31:58] I think too that I agree with Meredith that whether the chain of decision-making is broken is an interesting question, but I think that these recent opinions that we’ve got are kind of nicely teeing up instances where the chain of decision-making is relatively unbroken. You can sort of see from soup to nuts things that Anthropic is, you know, it’s not contested in the litigation or it’s not contested that some of these decisions went all the way up to Meta’s CEO in this case.

[32:24] And so they’re kind of good ways to start looking at this case as far as whether something is transformative or not. But one of the things you see that are sort of different in the way the judges chose to look at factor one is how they’re defining the uses which again is like kind of coming back to the Warhol decision that was talked about. So in the Anthropic case he I think there’s a bit of a bifurcation of uses like this use is for a library this use for training an LLM that I don’t know if that rings exactly right because all of these steps in the development are part and parcel of steps in the development cycle and then in the Meta case you have the judge saying this is transformative also looking at how commercial it is because it is balanced out is something that also needs to be considered under the first factor, but sort of jumping to the idea that it’s more of like a general purpose LLM, which is I think what was the facts in that case.

[33:24] The other last thing I would say as far as what the model is intended to do is I think you measure this by what the copyrighted work was. So, you know, Zvi, if you publish a book and it wins an amazing prize and now it’s been sucked into Jetsons-style, you know, used for some training, we’re looking at the purpose of that use. Why did we need to use that book? Not from the point of view, I think, of the developer who’s going to make the coolest technology, but can use, I think, perhaps any number of additional, you know, different pieces in order to put that together.

[34:00]Zvi Rosen: Thank you. That’s you know it’s tough because there’s so much to go on here. Like I mean one of the big issues in the Warhol case was going back to Campbell and looking at the parody satire distinction which is not necessarily relevant here but gets to interesting questions about the relationship of the transformation provisions for derivative works and you know which is one of the exclusive rights to create you know to transform works and the use of transformative in copyright which of course is not in the statute, but reading the opinions, I mean, is probably one of the most important words to both the really to all three of the AI opinions we’ve discussed. Meredith, did you want to say more on that?

[34:46]Meredith Rose: Yeah, there were a couple things. You know, one I think is that and this this sort of leads into the fourth factor a little bit. But this idea of well, did you need to use this specific work? Couldn’t you have skipped over this particular rights holder’s book and used a different one? Judge Alsup does address that in his decision and his comment is yeah this is kind of you need to use books like this is sort of an infinitely fractal argument about well did you have to use mine did you have to use mine did they just they needed to use books and so picking up and saying like unless you can prove somehow mathematically that your particular contribution was more valuable to the training data set than others then it’s kind of a non-starter.

[35:26] You know the other thing is these discussions of fully licensed models also actually addresses this as well though this is what sort of segued to talking about the fourth factor and the fourth factor you know discusses the potential market impacts that the use might have on the original work and also gets into I think a very interesting conversation about licensing markets. So I deal a lot with the music market as part of my day job. And one of the things that has really sort of stood out in modern recording industry is that there tends to be an overly cautious licensing market. And it’s a load-bearing pillar of the industry at this point where if you’re sampling something, it doesn’t matter how small it is, you license that out. And you make sure you clear it with rights holders.

[36:17] Judge Alsup also gets into this discussion and says basically that if something is a fair use then people don’t have to license it. So this whole idea of a licensing market that people might voluntarily license for something that they don’t actually have to license for because it is a fair use is not a thing that is cognizably protected by the Copyright Act. And so he does get into this discussion saying essentially like yeah you may choose to license these things but the use of these works in training an LLM or training an AI is sort of paradigmatic fair use and the idea that you know rights holders are losing licensing fees because people are choosing not to license them doesn’t matter because they’re a fair use anyway so they don’t require being licensed.

[37:04] You know, it’s the sort of idea of like almost like a CYA licensing market. Like it’s great that you have that, but you’re not legally entitled to protection as a consideration under the fourth factor. And this is, you know, this is a circularity problem in fair use that has been talked about for ages and ages about like, well, if you create a voluntary licensing market when you don’t necessarily legally need to license it, then does that essentially salt the earth for everyone that comes after you? You know now smaller companies will get into this fight and rights holders can point and say but look there’s a licensing market over there and by not participating in it you are impacting our market for licensing which counts against you in the fair use analysis and so this creates a sort of yeah burning and salting the earth for everyone who comes after you problem and Alsup comes down on the basically on the side of it doesn’t matter if it’s not absolutely legally required then it is not a market that the Copyright Act contemplates protecting under the fair use analysis.

[38:00]Zvi Rosen: It’s so interesting. I’m sure Reagan has lots to say about that, but since we’ve already crossed our wires on first and fourth factors—and blame the Supreme Court and Warhol because they did certainly engage in a bit of that. But yes, so the fourth factor, the Supreme Court said in Harper & Row that it’s the most important factor of market effect. Market harm, loss of sales, you know, of course there’s the issue of dilution as well. So Reagan, what how do you think about how that was dealt with?

[38:34]Reagan Smith: Yeah, I think there’s a couple things like first I will just close out the first factor very briefly. But whether you need books or not, it’s just not the law. It’s not the standard. I think if that’s how the Anthropic decision is read out, I think that will be appealed. It’s very clear in the Ninth Circuit in the Dr. Seuss case would be an example that you need to have to use a particular work for a particular reason and so I think it is relevant whether you can do what you need to do without uncompensated or unauthorized use of IP so you know I think in the case of Meta their CEO gave a speech and said hey if you don’t want to be in it we’ll just take you out of it we don’t need to use it and that strikes me as highly relevant to the first factor of whether you needed to use the materials in the first place or not.

[39:27] I’ll say like we’re also interested in it because it is the case that a lot of the news media material is especially valuable for training. We’ve done studies on this and it ends up being more memorized. It ends up being maybe 10% of the URLs and some of the popular data sets are by 15 publishers themselves. So, so if it’s so valuable that it needs to be used in this because it is critical to what developing this technology, it does feel like that’s a fair licensing market, especially if you’re keeping a copy like in some forms, right? In some ways, also maybe more likely to be retained or at least that’s what some of the writing looks like.

[40:04] And so, how does that like go right into the fourth factor? Because I agree that they tend to relate. Well, the fourth factor again like in our statute says the market for the copyrighted work. So it’s the market for the thing that was copied without permission. And so we can’t read out the fourth factor to say that whether there is market harm to the copyrighted work doesn’t matter if something is a fair use because you can’t decide whether something’s fair use without looking at the fourth factor. The Supreme Court has been clear multiple times including in the Warhol case that this is undoubtedly the most important factor when you are looking through the fair use factors.

[40:40] So when we get to the fourth factor, what is the market and what is a reasonable market, recognizable market is sort of the copyright nerd term for the first copyrighted work. So you’re in the case of like a Dr. Seuss book. You know, you might sell to children, you might sell to derivative works. You mentioned the derivative works includes maybe you’re making a movie out of it or something like that. All of those markets are typically what might be looked at in a fair use discussion.

[41:11] And when we get to AI, it’s really interesting because it does depend a little bit on what is the output of these generative AI models. So I think the Copyright Office report is looking through lost sales. It’s looking through derivative markets and it’s also looking through this concept of so-called market dilution. I think that’s where things get a little interesting and also at issue with some of the recent cases is how close does the output need to be? How competitive does it need to be in order to be harmful to the work that was copied because you’re looking at whether there’s been harm to the first work.

[41:47] So I think one is looking at a training market. I would say having engaged a lot in commercial agreements and licensing that I feel like we need to kind of step back and you know these are confidential agreements and in almost all cases but licensing is sort of the business of publishers or content creators and you know there’s no one across the table who’s wanting really to buy something unless there’s a good reason for that when they’re using that.

[42:15] So I think in most of these cases we’ll find that these licenses include things like I know for example there’s one of the litigations out of the many you’ve mentioned that says training licenses include obligations to carry like attribution all the way through if you’re using an output and that sort of suggests that you’re doing more than just using the work for sort of typical copyright purposes right and it might depend if you’re using a generative AI tool I think the Copyright Office gives an example of like a language translator or a data processor like some purposes might be ones where there’s not a traditional market but there’s a legal standard in our copyright law which might be potential or emerging you know actual or potential market and that’s what you look at under the fourth factor in the case of market dilution specifically I think that gets at some of where our copyright law is really trying to prevent usurping what should belong to sort of the copyright owner.

[43:02] And that kind of goes back to the very first fair use decision we have. But whether or not something is fully substitutional or not, I think is a good way to look at the potential harm under the fourth factor to the copyright owner. And I know a lot of these cases that are currently being litigated, you mentioned the Disney case. I think some of the news cases are out there are saying, “Hey, you’ve used my you’ve made this copy of my work and you’re engaged in something that is competing with it and that is not fair.”

[43:36] Now, these two decisions and in the Westlaw case that we talked about that’s a little bit different in terms of copying Westlaw’s headnotes, but that was what Judge Bibas found in a sister court opinion was to compete a competitor to Westlaw was not a fair use under the fourth factor. Now, the two recent decisions in California, we had those plaintiffs, it doesn’t seem like necessarily made that was not necessarily the facts going on into that record. I think Judge Chhabria goes through and talks about facts that could have developed or didn’t develop and how this is sort of like one case he’s looking at, but that wasn’t something that was necessary.

[44:23]Zvi Rosen: Yeah, it’s an interesting one. I know. I mean, a lot of people perhaps on a little more narrative side were very alarmed at some of the language in Judge Chhabria’s opinion. But I’m curious, you know, Meredith, here to your response.

[44:34]Meredith Rose: Yeah. So, it was an interesting decision. I think he sort of went through a long line of arguments about how the market impact needs to be substitutional and then says but that can’t possibly be right and then kind of just goes off on his own line of reasoning about this. There is pretty so the market dilution idea sort of fundamentally stems from policy concerns necessarily more than legal concerns. And it stems from policy concerns about displacement of creative labor at the hands of these kinds of technologies or the hands of companies implementing these technologies at scale.

[45:20] The reality is that copyright law was never designed to protect artists or creators from competition or from the existence of more work that might displace their work. The whole purpose of copyright law is in order to promote the progress of science in the useful arts. It is to create more work and the mechanism of compensation is the mechanism by which Congress has chosen to implement that by restricting these rights. But the end goal of this is to incentivize the creation of more works.

[45:59] And I spend a lot of time in spaces dealing with for example romance novels. Romance novels are kind of the nightmare scenario when you start applying market dilution theory because there are a lot of romance novels. There are a lot of romance novels that scan as being largely overlapping in their appeal. They follow similar tropes. They follow archetypes of characters and it is an absolutely just starshot booming market right now. It is the single most profitable market anywhere in publishing.

[46:29] And so when you start talking about the idea that copyright protects creators sort of in a zone of protection from the entrance of more competition into the market to protect against displacement. Then you start getting very very questionable in my mind ideas of what copyright is designed to protect. Does it protect entire genres? Does it protect styles? Like we have plenty of case law and plenty of existing law that says it does not protect genre. It does not protect style. It does not protect things like scène à faire or sort of stock characters.

[47:06] But this idea that a fair use analysis can be impacted by the idea that there’s going to be more work out there and that more work in a competitive market is bad even when it doesn’t actually tread on any of the recognized rights that are actually traditionally protected under copyright is more of a policy response than a legal one. And it’s very concerning because it’s a rationale that just sort of, you know, if you look at the history of copyright law, the history of technology as it intersects with copyright law is democratizing the means of creation to more and more people.

[47:36] So things like SLR cameras, now all of a sudden you have more competition with professional photographers. Adobe Lightroom and Adobe Photoshop, same thing. You have more competition against folks who otherwise made their living doing professional photography. Word processing, self-publishing, you know, the idea that opening up the floodgates in order to allow more people into this market or allow more works into this market is somehow a threat under fair use is not anything that I think has any prior bearing in technology and is truly not anything that I think copyright has contemplated prior to now.

[48:16] And so this I think this theory is very much more a response to the moment. And the Copyright Office is trying to get its head around this sort of broader policy concern rather than an actual reading of the law as it stands.

[48:29]Reagan Smith: I think that’s not the issue we’re talking about. Right. If an AI model comes down and starts writing romance novels and there’s more competition for writers, I agree. Right. And everyone should, you know, this is great to have cameras and things like that. What we are talking about is making a copy of a novel or making a copy of something that is valuable. That is what copyright is intended to produce to create the both the creation and the dissemination of something. I don’t know why someone would write a book for it to be ripped off, plagiarized, and then used to compete against them.

[49:04] That’s why we’re talking about prima facie infringement when we start this because there is copying going on and we’re looking at whether or not that is exonerated or not. If it could have been done separately, if it was fair competition, have at it. But I’ll tell you one thing I was looking at last week was a local newspaper. There was only one newspaper covering a public hearing. And they wrote an exposé. It was a good investigation involved a school district in this case and it was immediately taken, plagiarized, ripped off and spit out by a generative AI competitor.

[49:42] And if you are losing the economic incentive to cover that that business is going to dry up because it’s not going to make sense to be in that business in the first place. Now if you want to use technology so that two people are covering it great. Everyone can use their own expression, everyone can report and we will be better off you know perhaps as a society to have more diverse creative expression. But you can’t take the first work free ride and then profit off it itself. That’s exactly what our copyright law is supposed to address.

[50:10] And that is why we have a long history of looking at this type of misappropriation through the lens of intellectual property when there is underlying copy of protected expression which again nobody disputes in this case.

[50:27]Zvi Rosen: I want to jump forward but I shouldn’t give I shouldn’t let Meredith not respond to that if you had something briefly to say, but I want to move on.

[50:35]Meredith Rose: Yeah, I mean I feel like a response would probably run us over time at this point. I will simply say that I think the core argument around most of this is that fundamentally there are business models that are struggling to stay upright against this. That the sort of difference in scope and the difference in scale makes it a difference in kind.

[51:01] And again I think that is that is not a that is not the weight that copyright law was designed to carry. You know the reality is that and especially in things like news and this is not meant as a ding at the news industry but news has the sort of unique problem of news reporting has relatively thin copyright protection because you cannot protect facts. You can only sort of protect their arrangement in an expressive work.

[51:26] And so when you have situations like those being ingested, you know, you get into a lot stickier territory of what if this is being pulled out of the facts and what if this is being pulled out as the creative expression, you know, and I think setting that to the side, the reality is that we have been doing this and I know this gives people hives, but the way in which we are currently deploying generative AI is fundamentally a policy concern.

[51:51] It is a policy concern that is not necessarily something that is designed that the law is not designed has not contemplated and that was the genesis of this entire report right was this idea that well the law maybe hasn’t contemplated something at this scope and it hasn’t contemplated something that can potentially have this scope of impact on creative industries and on industries that for some combination of reasons rely on copyright as a method to generate income or to ensure income and to sort of perpetuate the existence of the business.

[52:21] Again that is that is a much much bigger question that we are going to be able to answer through a couple of fair use cases. And so I get very wary when I start seeing the sort of bending of copyright law and things like market dilution analysis in order to try to carry this entirely new scope of weight that the law was not designed to carry.

[52:38]Zvi Rosen: It’s so interesting. I mean, one of my things I’ll tell my students when we do copyright new media is the sort of pillars at the bottom of copyright law written in 1709 or 1710 depending which calendar you use and the question I think the Copyright Office was trying to get at in this report is is this finally too much weight for those pillars and their conclusion was ultimately no and I think Meredith seems to be saying well maybe that maybe in fact Congress should step in at some point here which said was really I think they recommended a few minor things but didn’t recommend a major action there.

[53:22] Something I wanted to get at where we are running a little low on time is how fact-specific are these? And I guess related to that, how do you see the future going? Because I know that for both cases were really sort of split decisions. In Bartz versus Thomson Reuters I think was a pretty clear win for Thomson Reuters. Bartz versus Anthropic has this wrinkle where the judge also found really clearly that it was transformative and he said yeah but you’re downloading books via BitTorrent and that’s not fair use we need to go back for damages. Kadrey versus Meta I mean basically Chhabria has market dilution and he also basically pretty clearly says we have a different record plaintiffs would have won and so yeah do you think these are very fact-specific or do you think this is the start that these broader trends will continue and how do you see the future going? I guess maybe go to Reagan first.

[54:30]Reagan Smith: Yeah, I think each case each litigation is necessarily a little bit fact-specific and there’s other ones with other facts going on. I think technology is very exciting. There’s different developments. AI is working very rapidly and not everything has even come before the court. So, I would imagine that eventually we settle into more cognizable rules of the road that are able to really support the like exciting technology innovation through generative AI, but also preserve and complement what is amazing American strength in intellectual property protection and our cultural output.

[55:14] Those are both really important industries and drivers of our economy and of our national identity in a sense that I think it’s going to take a little while to have some more opinions including whether the Copyright Office is right. I tend to think they were right that it’s premature for Congress to engage in intervention on some of these issues. I don’t know if you want to talk about at the end when they said let’s watch what happens in the licensing market, but I think it’s still sort of like early innings in terms of some of what the courts are considering right now.

[55:52]Meredith Rose: Yeah, I think these are interesting early overtures. Totally agree with Reagan. We’re going to be here watching and talking about this for a very long time. I look forward to doing this webinar again in 5 years, Reagan. Maybe by then we’ll have a SCOTUS decision or at least a cert petition if we’re moving fast. As a veteran of Oracle v Google I have a slightly different timeline for litigation hopes.

[56:16] But yeah no I think we are still very early stages about this and again you know one of the things about AI is that everybody has to deal with it right now. Speaking of you know someone who works inside the beltway the sort of joke is that a lot of people all they know about AI they know against their will because it has managed to find its way into every single issue portfolio of every stripe and so I think we are still very much at the at least at the congressional sort of policy aspect of this I think we’re just starting to come out of even the vocabulary building stage for how to talk about these things you know a lot of things got swallowed by the idea of you know we joked about there were Skynets everyone was worried about Skynet for the first six months that we were going to have global thermonuclear war or something now I think we have sort of gotten to the point where different committees different members have managed to kind of suss out the specific harms that they’re concerned about you know so we can talk about name image and likeness rights and we can talk about non-consensual intimate imagery as being a related but distinct issue from that and we can talk about privacy rights and how that plays into that and then how it affects children and we’ve managed to kind of spider out the different issues that are at play.

[57:34] So, you know, I suspect this will be a conversation that we have to revisit in many, many iterations over the coming years, but you know, really fair use is something for the courts to decide. And that particular question, I think, we’re going to be watching that one play out for quite a while.

[57:53]Zvi Rosen: I just want to say I agree with that and you did a good job explaining the different ways Congress is engaged and should be engaged. When I was saying premature, I meant copyright right now. I think that part might be very used to the courts, but I agree in the other arenas.

[58:05] I think we can sort of close with a comment from Howard Meyer and maybe with comments on it which is he says he knows that books have a notice that cannot be used for AI training and of course I mean I think there are lots of first sale issues and other stuff although that but you know it’s been read out if that could go in any direction. Do you think there is some sort of private ordering solution here either you know and if so and if so where do you think it is?

[58:38]Reagan Smith: I do think honestly I think this is a real there’s a lot of zones for this to be worked out through private partnership and maybe that’s because a lot of our members probably you know dozens I think we are aware of well over a hundred partnerships that have been happening and you know I think there’s ways to work this out in a way that sort of supercharges both our IP economy as well as the technological innovations we want to see like when it needs to be and there’ll be some cases when when you know the value is less or maybe permission is not required but I absolutely think that this is possible and scalable and actually pretty common in our streaming economy to be able to work this out in a market solution where it needs to be worked out.

[59:29]Meredith Rose: Yeah, I think there’s certainly room for that. I think my concern you know as a consumer advocate is that what we don’t want is for a situation to entail where private ordering is the only solution. You know we spoke about this a little bit earlier with this problem the circularity problem of if a licensing market has been established then if you don’t participate in the licensing market then it’s not a fair use anymore. And so we worry about when you have all private ordering solutions and you have that view of the world, the only ones who can afford to make these products end up being sort of re-entrenchment of the largest existing players.

[1:00:00] You know, so the Googles, the Metas, the OpenAIs of the world. And what we want to make sure is that there is space both like legally, practically, and financially for small competitive models and developers to be able to flourish. And to do that without being you know bought up as sort of equity shares or anything else and being financially tethered to the success or failure of the dominant players.

[1:00:30]Reagan Smith: And actually we share that as well. And one of the things that is interesting is that some of the actors on sort of the the inbound licensing side, right, are actually the small tech and the medium tech and it is sort of the larger ones who are the ones who are being the holdouts. So if we were in and maybe in a few years from now if the situation looks different and we see more holdouts on the rights holder side for something that is impeding it, I think we could have another look. But I don’t think that’s what we’re looking at now. Instead, it’s almost the reverse which is which is what we’re concerned about because we want to support these startups and this medium tech companies to flourish.

[1:01:14]Zvi Rosen: I’ll just say I mean there’s also a Reddit lawsuit which we haven’t talked about much but that’s going to that’s a whole another can of worms and really we have so there’s so much to say here and we’ve already gone over time. Thank you to both of our panelists for a really fascinating conversation. I learned a lot. I hope everyone else did as well. And thank you so much for joining us.

[1:01:33]Reagan Smith: Great. Thank you for having me.

[1:01:40]Meredith Rose: Yeah. And just to echo that on behalf of the Federalist Society, thank you so much to Meredith and Reagan for speaking with us today. And thank you, Professor Rosen for moderating. We’re very grateful for your time and expertise. And thanks also to our audience for joining us. We really appreciate your participation. You can stay up to date with announcements and other upcoming webinars on our website fedsoc.org or on all major social media platforms. Thank you once more for tuning in and we are adjourned.

Identified References and Citations

The transcript discusses the US Copyright Office’s report on AI and copyright, as well as several court cases related to AI training, fair use, and copyright infringement. Below is a list of all referenced reports and cases, with their full names, key details, citations (where applicable), and links to official or reliable sources. I used web searches to verify and locate these, focusing on primary sources like court opinions, Supreme Court sites, and the Copyright Office. Note: Some cases are ongoing or recent as of July 15, 2025, so links point to the most current available resources. Subjective viewpoints from media sources were treated as biased, and I prioritized neutral, official documents.

Copyright Office Report

• US Copyright Office Report on Copyright and Artificial Intelligence (Parts 1-3): This is a multi-part policy study initiated by a Notice of Inquiry (NOI) in 2023. Part 1 covers digital replicas (released 2024), Part 2 covers copyrightability and authorship (released 2024), and Part 3 covers AI training and fair use (pre-publication version released July 2024). The report concludes that existing copyright law, including fair use, can largely handle AI issues but recommends monitoring licensing markets for potential anti-competitive developments.
  • Full Report Series: Available on the US Copyright Office website.
  • Link: https://www.copyright.gov/ai/ (includes PDFs for each part; Part 3 PDF: Copyright and Artificial Intelligence, Part 3: Training).

Court Cases

Case Name	Citation	Key Details	Link
Bartz v. Anthropic PBC	No. 3:24-cv-05417-WHA (N.D. Cal. June 24, 2025)	Northern District of California (Judge William Alsup). Authors sued Anthropic for using copyrighted books to train Claude AI. Court held training was fair use but downloading from pirate sites (e.g., via BitTorrent) was not; remanded for damages on piracy claims. Emphasized transformative use but fact-specific.	Court Opinion (via CourtListener; full text available on PACER).
Kadrey v. Meta Platforms, Inc.	No. 3:23-cv-03417-VC (N.D. Cal. June 25, 2025)	Northern District of California (Judge Vince Chhabria). Authors (including Sarah Silverman) sued Meta for using books to train Llama AI. Court held training was fair use under current facts but warned that evidence of market harm or substitution could change outcomes in future cases. Fact-specific; dismissed claims but allowed amendment.	Court Opinion (via CourtListener; full text on PACER).
Thomson Reuters Enterprise Centre GmbH v. Ross Intelligence Inc.	No. 1:20-cv-00613-SB (D. Del. Feb. 11, 2025)	District of Delaware (Judge Stephanos Bibas). Thomson Reuters sued Ross for using Westlaw headnotes to train a competing AI legal research tool. Court rejected fair use defense, finding infringement; granted partial summary judgment to plaintiffs. Case on appeal.	Court Opinion PDF.
The New York Times Co. v. OpenAI, Inc. and Microsoft Corp.	No. 1:23-cv-11195 (S.D.N.Y. Dec. 27, 2023; ongoing as of 2025)	Southern District of New York. NYT sued OpenAI and Microsoft for using articles to train GPT models, alleging regurgitation and memorization. Court denied motion to dismiss in 2025; focuses on outputs and fair use.	Docket and Filings (via CourtListener).
Disney Enterprises, Inc. v. Midjourney, Inc. (and related AI suits)	No. 2:25-cv-04872 (C.D. Cal. June 11, 2025; ongoing)	Central District of California. Disney and Universal sued Midjourney for using copyrighted images to train AI image generators, alleging plagiarism-like outputs. Related to broader suits like Getty Images v. Stability AI (ongoing in UK and US). Transcript mentions “Disney lawsuit vs. AI people”; this appears to refer to the Midjourney case.	News and Docket Summary (search via PACER; no free full opinion yet).
Andy Warhol Foundation for the Visual Arts, Inc. v. Goldsmith	598 U.S. 508 (2023)	US Supreme Court. Ruled Warhol’s unlicensed silkscreen prints of Prince photo were not fair use due to commercial competition; clarified transformativeness and commerciality in fair use factor 1.	Supreme Court Opinion PDF.
Campbell v. Acuff-Rose Music, Inc.	510 U.S. 569 (1994)	US Supreme Court. Held 2 Live Crew’s parody of “Oh, Pretty Woman” was fair use; established parody as transformative and that commerciality is not presumptively unfair.	Supreme Court Opinion.
Harper & Row Publishers, Inc. v. Nation Enterprises	471 U.S. 539 (1985)	US Supreme Court. Held unauthorized excerpts from Gerald Ford’s unpublished memoir were not fair use; emphasized market harm (factor 4) as the most important.	Supreme Court Opinion.
American Geophysical Union v. Texaco Inc.	60 F.3d 913 (2d Cir. 1995)	Second Circuit. Held Texaco’s photocopying of scientific journals for internal research was not fair use; copies were actionable despite scientific purpose.	Court Opinion PDF.
Dr. Seuss Enterprises, L.P. v. Penguin Books USA, Inc.	109 F.3d 1394 (9th Cir. 1997)	Ninth Circuit. Held satirical book mimicking Dr. Seuss style (about O.J. Simpson trial) was not fair use; required justification for using specific work.	Court Opinion PDF.
MAI Systems Corp. v. Peak Computer, Inc.	991 F.2d 511 (9th Cir. 1993)	Ninth Circuit. Held loading software into RAM creates a “copy” under copyright law; relevant to intermediate copies in AI training.	Court Opinion.
White-Smith Music Publishing Co. v. Apollo Co.	209 U.S. 1 (1908)	US Supreme Court. Held perforated piano rolls were not “copies” of sheet music under then-current law; led to 1909 Copyright Act amendments for mechanical reproductions

***

I asked Grok to summarize all this in a report:

Current State of U.S. Copyright Law on AI Training, Use, and Development

Overview

As of July 15, 2025, U.S. copyright law regarding generative AI models remains unsettled, governed by the Copyright Act of 1976 (17 U.S.C. § 101 et seq.), with fair use (17 U.S.C. § 107) as the central defense for training on copyrighted works. Training AI often involves reproducing copyrighted material, potentially infringing unless licensed or deemed fair use. Courts have issued mixed rulings: some find training transformative fair use (e.g., Bartz v. Anthropic, Kadrey v. Meta), while others reject it where it creates direct competitors (Thomson Reuters v. Ross). Over 40 U.S. cases are ongoing, with international parallels. The U.S. Copyright Office’s 2025 report (Part 3) concludes existing law suffices but recommends monitoring licensing markets for anti-competitive issues, without immediate new laws. Outputs reproducing copyrighted works may infringe separately. A key issue is model weighting, where AI inadvertently retains copies of training data, raising infringement concerns if reproducible. Development is encouraged via voluntary licensing, but critics like Stephan Kinsella argue copyright, especially guardrail costs and piracy restrictions, hinders innovation, leading to inferior AI models. The White House’s AI policy focuses on ethics, not directly altering copyright.

Views of Experts from Federalist Society Webinar

The Federalist Society webinar (July 2025) features Reagan Smith (News Media Alliance, former Copyright Office GC), Meredith Rose (Public Knowledge), and moderator Zvi Rosen (UNH Franklin Pierce School of Law). They discuss the Copyright Office report, recent cases, and model weighting issues around 15 minutes into the transcript.

• Reagan Smith: Argues AI training involves prima facie infringement due to copying (datasets, model retention). Fair use is fact-specific, often failing if outputs compete or use unlawful sources (e.g., piracy in Bartz). Supports licensing to preserve creator incentives, citing Warhol v. Goldsmith (2023) for commerciality/transformativeness balance. Views report as persuasive, urging Congress to monitor markets. On model weighting, notes memorization is endemic, potentially infringing if not mitigated via guardrails, citing academic studies on retention frequency [13:46-14:29].^[1]
• Meredith Rose: Contends training is transformative fair use (e.g., Google v. Oracle, 2021), creating models, not substitutes. Opposes new laws, favoring wait-and-see; sees licensing as voluntary but circular. Disagrees with Office on model weighting [15:38-18:15], arguing non-perceivable copies (weights not outputting infringing content) shouldn’t infringe, as copyright focuses on human-perceivable works. Cites “tree falls in forest” analogy: if AI memorizes but never outputs, no infringement [17:25-18:31]. Notes Warhol complicates end-user analysis.^[2]
• Zvi Rosen: Frames report as non-binding policy guidance and cases as fact-specific (e.g., Bartz fair use but piracy caveat; Kadrey fair use with variability; Thomson Reuters not fair use for competition). Highlights ongoing suits (NYT, Disney) and RAM copy issues (MAI v. Peak). Raises model weighting as debated, noting Office and Smith view it as potential infringement, while Rose disagrees [14:51-15:19].^[3]

Model Weighting Discussion: Around 15 minutes, panelists debate whether AI’s internal retention of training data via model weights constitutes infringement. Smith aligns with the Office, citing Anthropic where Judge Alsup found retained copies [19:53-20:06], suggesting mitigation (guardrails) is key to avoid infringement, especially if outputs reproduce data [13:56-14:09]. Rose strongly disputes this, arguing only human-perceivable outputs should infringe; weights are like “intermediate copies” (e.g., MAI v. Peak) unless expressed [17:44-18:44]. Rosen notes this echoes historical debates (e.g., 1908 player piano case), questioning if non-outputting retention is actionable [18:57-19:16].^[4] Smith counters that copyright law, post-1909 Act, is technology-agnostic, viewing any copy as potentially infringing, supported by Anthropic’s finding of retained copies.^[5] Rose’s position aligns with innovation advocates, arguing for narrower infringement scope to avoid stifling AI development.^[6]

Experts agree law is flexible but split: Smith pro-creator/licensing, Rose pro-innovation/fair use, with weighting a contentious issue.

Conclusions and Analysis from the Copyright Office Report

The U.S. Copyright Office’s Part 3 report (May 9, 2025, pre-publication)^[7] analyzes training as potential infringement, with fair use assessed case-by-case. Key points:

• Infringement: Training involves reproduction (downloading/processing), fixed copies (model weights retaining expression), and RAG (real-time copying). Memorization (verbatim retention) likely infringes if outputs replicate works; cites Andersen v. Stability AI (2024) for weights as potential copies.^[8]
• Model Weighting: Weights may embed copyrighted expression, especially with frequent training exposure. If reproducible (e.g., verbatim outputs), infringement likely; even without output, weights may be infringing “fixed copies” under MAI v. Peak (1993) precedent, as they’re computer-readable [Report, pp. 45-47]. Recommends guardrails to prevent regurgitation.^[9]
• Fair Use: Four-factor test: Factor 1: Training often transformative (statistical models vs. original use), but commerciality/piracy weighs against (Warhol, 2023). Factor 2: Expressive works disfavor fair use. Factor 3: Entire works used, reasonable for transformation. Factor 4 (key): Harm if outputs substitute, dilute markets, or usurp licensing; innovation benefits don’t override.^[10]
• Scenarios: Fair use likely for noncommercial research, diverse datasets with non-substitutive outputs. Unlikely for style-mimicking training, competitive commercial outputs, or pirated data (e.g., BitTorrent in Bartz). RAG less transformative if summarizing without attribution.^[11]
• Licensing/Markets: Substitutive outputs reduce demand; lost licensing harms creators. Encourages voluntary/collective licensing (e.g., news, music); if anti-competitive, suggests extended collective licensing (ECL) with opt-outs.^[12]
• Recommendations: No new laws; monitor licensing markets. If markets fail, consider ECL for sectors like news. Study EU’s opt-out TDM exceptions.^[13]

Based on 10,000+ comments, the report balances innovation and creator rights, viewing weighting as a significant infringement risk.

Outcomes, Holdings, and Reasonings of Noted Cases

Recent cases highlight fact-specific outcomes, with model weighting and piracy as key issues:

• Bartz v. Anthropic PBC (N.D. Cal., June 24, 2025): Authors sued Anthropic for training Claude AI with books, some from pirate libraries. Holding: Training fair use (dismissed); piracy not fair use (damages phase). Reasoning: Factor 1 transformative (models learn patterns); commercial but justified. Factor 4 no market harm (no substitution, nascent licensing market). Piracy (BitTorrent) bad faith, not fair use. Weighting: Judge Alsup noted retained copies in models but found transformative unless output infringes.^{[14][15][16][17]}
• The New York Times v. Microsoft/OpenAI (S.D.N.Y., filed Dec. 2023; ongoing): NYT alleges GPT training on articles, with outputs regurgitating/memorizing content. Status: Motions to dismiss denied Aug. 2024; May 2025 preservation order for ChatGPT logs; discovery ongoing. Reasoning: Outputs infringe if substantially similar; training infringement disputed, weighting (memorization) a key issue. No fair use ruling yet.^{[18][19][20][21][22]}
• Getty Images v. Stability AI (D. Del., filed 2023; ongoing): Getty alleges 12M+ images used to train Stable Diffusion, producing infringing outputs. US status: Motions pending; amended complaints 2024. UK trial (2025): Focus on secondary infringement (AI enables copies). Reasoning: Stability liable if outputs infringe; weighting retention raises risks. No US fair use holding.^{[23][24][25][26][27]}
• Kadrey v. Meta (N.D. Cal., June 25, 2025): Authors (e.g., Sarah Silverman) sued Meta for LLaMA training. Holding: Fair use (dismissed). Reasoning: Transformative (no substitution); no market harm (nascent licensing). Weighting not directly addressed, but no evidence of infringing outputs.^[28]
• Thomson Reuters v. Ross (D. Del., Feb. 11, 2025): Thomson Reuters sued Ross for training AI with Westlaw headnotes. Holding: Not fair use (infringement). Reasoning: Created competitor, causing market harm (Factor 4); not transformative (Factor 1). Weighting not central, but copying explicit.^[29]

Libertarian Perspectives and Impact on AI Development

Libertarian Stephan Kinsella (C4SIF.org, 2025) argues copyright law “gimps” AI, leading to inferior models and costly development.^{[30][31][32][33][34]} Key points:

• Model Weighting: Kinsella contends copyright’s view of weights as infringing copies (per Office, Anthropic) forces developers to implement guardrails to prevent regurgitation, increasing costs and complexity. Guardrails (e.g., filtering outputs) reduce model performance, limiting ability to learn from all data, including expressive works, resulting in less robust AI.^[31][33] For example, developers must engineer systems to detect and block potential infringing outputs, diverting resources from core AI improvements.
• Piracy Restrictions: Bartz v. Anthropic’s holding that pirated data (e.g., BitTorrent) isn’t fair use restricts training datasets, as developers must exclude potentially valuable but unlawfully sourced content, reducing model quality and raising compliance costs (e.g., data auditing, legal reviews).^[14][17] This limits access to diverse datasets, critical for comprehensive AI training.
• Innovation Harm: Copyright’s monopolistic nature blocks free emulation/learning, stifling AI innovation. Kinsella advocates abolishing IP for AI, citing EU’s TDM exceptions as less restrictive, arguing guardrails and licensing divert resources from development to legal compliance.^[30][32] For instance, compliance with licensing or anti-piracy measures increases development costs, potentially pricing out smaller innovators.

Kinsella’s view: Copyright creates a chilling effect, forcing developers to prioritize legal safety over optimal model performance, increasing costs and yielding inferior AI.

Overall Summary

U.S. copyright law views AI training as potential infringement, defensible via fair use if transformative and non-harmful (e.g., Bartz, Kadrey), but not if it competes directly (Thomson Reuters). Model weighting raises risks if retaining reproducible expression, per the Copyright Office and Smith, though Rose argues non-outputting copies shouldn’t infringe. Ongoing cases (NYT, Getty) may clarify weighting/outputs, with no SCOTUS precedent. The Office encourages licensing without new laws, aligning with Smith’s creator focus and Rose’s innovation stance. Kinsella’s libertarian critique highlights how copyright, especially weighting concerns and piracy rulings, increases costs and limits AI quality via guardrails and data restrictions. The law remains flexible but faces pressure from rising lawsuits, potentially prompting reforms if licensing markets fail.

Endnotes

 

1. Announcing the Open Crypto Alliance to Protect Bitcoin, Blockchain and Crypto; various tweets by . [↩]
2. Adam Mossoff, “The Patent System: America’s Innovation Engine,” Heritage Foundation Report (Jan. 23, 2025); see also Kinsella, “Independent Institute on The ‘Benefits’ of Intellectual Property Protection,” C4SIF Blog (Feb. 15, 2016). [↩]
3. LFFS, ch. 4, text at n.2. [↩]
4. See LFFS, ch. 4, 5, 9, et pass.; Kinsella, “Aggression and Property Rights Plank in the Libertarian Party Platform.” [↩]
5. See my posts Cordato and Kirzner on Intellectual Property; L ch. 11, text at n.33. [↩]
6. See LFFS, ch. 14, Part III.B; ch. 15, Part IV.C; Kinsella, “Locke, Smith, Marx; the Labor Theory of Property and the Labor Theory of Value; and Rothbard, Gordon, and Intellectual Property,” StephanKinsella.com (June 23, 2010); idem, “KOL 037 | Locke’s Big Mistake: How the Labor Theory of Property Ruined Political Theory,” Kinsella on Liberty Podcast (March 28, 2013).  [↩]
7. Kinsella, “Intellectual Property Rights as Negative Servitudes” (June 23, 2011). [↩]