Fair Use in the Copyright Crosshairs: What’s Next for Both

0:00Alrighty, so I believe we are live. Welcome to today's Legally Speaking Podcast Live, proudly sponsored by Clio. And today we're taking on one of the most decisive legal battles of our time. What does fair use become in a world where AI learns from copyrighted works at industrial scale without permission, payment or oversight? And to help us break it all down, I'm joined by a leading voice shaping this global debate, Andrew Berger, who's a return guest to the show, who's one of the US's most respected copyright litigators. So a very warm welcome back, Andrew. Good morning. Good to see you again and looking forward to today's discussion.

0:44So let's get into it. Let's start with copyright's purpose in the age of sort of AI, because you've said AI is ingesting copyrighted works at an industrial scale without compensation. So what's the biggest misconception about copyright's purpose right now for folks to be aware of?

1:03Andrew Berger:Well, Rob, there are 65 suits by copyright holders against AI developers and more every day. And the biggest misconception is that copyright is some giant stop sign that will bar the models from ingesting their content and compel developers to compensate for unauthorized use. Hey, creator is going to be compensated, as they were with a$1.5 billion settlement in the Anthropic case, but not for every unlicensed use of their works. Looking down the road, developers are going to likely continue to convince courts that training with lawfully acquired content is transformative. But here is the exception.

1:46Andrew Berger:When the AI output captures the creative essence of a copyrighted work, forget it. Copyright is going to say no. Nor will judges excuse the model's training with pirated content. All the mega developers in their race to build these models vacuumed up pirated content from what are called shadow libraries that exist on the Internet. And these are the same developers who say, hey, we've at all time acted in good faith. But by claiming innocence, Rob, they open the door for plaintiffs to show developers knew the legal risks associated with stolen goods. a jury if this case these cases ever get that far are going to punish a thief wow fascinating and yeah just the level of cases already you've highlighted there it's uh yeah why i was so keen to have this conversation today and look judges are split on whether copyright protects public knowledge creative incentives or intellectual capital so why does that divide matter so much for ai specifically well it really does because you know it makes It's, you know, predictions of a fair use outcome have never been for the faint of heart.

3:02Andrew Berger:But now, oh, my God, they become a real educated hunch. So let me just give you briefly the three fair use decisions and what they signal for the AI ecosystem. And I'll make this as brief as possible. We've got a public knowledge protector in Judge Alsop in the case of Barts versus Anthropic. He views copyright's mission not to protect authors, but it's expanding public knowledge. AI is then aiding the output by synthesizing information and democratizing authorship. Anybody can now use AI to write anything. Now, so with this positive view, Judge Alsop finds AIs copying and trading on lawfully acquired books to be transformative.

3:51Andrew Berger:But the judge draws a line around pirated content, condemns their downloading, etc. It labels the exploitation of pirated content, and listen to this, inherently, irredeemably infringing the three I's. And OK, so that's Judge Alsop. Now, Judge Jabria, who decided Kadri versus Meta, has a different perspective. He sees copyright's purpose to protect and incentivize those who create the output. So AI is not just a benign assistant at your desktop, but a digital monster churning out an increasing number of non-infringing low-cost books that flood the marketplace. And he also introduces market dilution.

4:38Andrew Berger:Even if AI-created books don't copy a single sentence from your books, they dilute the market so badly, authors can't make a living. And then finally, the intellectual capital protector is Judge Bebas in Reuters versus Ross. He protects West's headnotes when defendant uses them without payment in a legal research tool that It competes with Westlaw. The judge held that Ross has harmed West's potential licensing market for training data, and that's enough to defeat fair use. Okay, so those are the three cases in a nutshell. But what does this all mean for the AI ecosystem? The answer really is one big question mark.

5:18Andrew Berger:If ELSA views prevail, developers will have no incentive to license training rights, so long as the training produces a transformative outcome. Now, think about this. If Chabria's view prevails and the potential for mass scale indirect competition is enough to swing the for use fact for the fourth fair use factor against fair use, no model training could ever be safe. That view risks innovation as infringement. Now, think about this. Remember your flip phone? You probably had one of those years ago. Absolutely. I know. It lost its value not because the iPhone infringed it, but because the iPhone was just better.

6:00Andrew Berger:So if Chabria's logic holds any AI that dilutes the market for that old stuff you have around, that AI could possibly be liable. In that case, copyright, instead of incubating progress, will derail it. So if Bebas's view, the last judge prevails, the source of the content doesn't matter whether it's contented or scraped or purchased. You need a license. And that means without that ticket, your ride is illegal. And that means that only the million dollar incumbents can afford the licensing cost. Okay, so we've got these three rather scary developments, and who can finally determine who's right?

6:49Andrew Berger:Well, the Supreme Court. They may eventually unify the law, but that may take years. So until then, creators, developers, and investors, if you're listening, you're going to have to accept uncertainty and risk. the law rob you get may depend on the judge you draw thank fascinating and thank you for for sharing that it is clear there's some divide there and you know really helpful to give some some context i want to kind of jump to fair use under under pressure now and sticking with the courts if i may because the courts are diverging on a sort of four-factor analysis here so what is the most dangerous interpretation you're seeing right now, which is going to enlighten us?

7:34Well, the most dangerous one is from Judge Chavria, because what he's saying is, hey,

7:42Andrew Berger:the ends justify the means. That's the approach he took in Kadri. It's dangerous because what that does is it retroactively greenlights piracy. In Kadri, the court looked at the final output from the model, because the final output was transformative, the judge excused the initial acts of downloading and trading with pirated stuff. You know, pirated stuff is stolen stuff. I mean, this is, you know, that approach is wrong because Chabria is violating the Supreme Court's guidance in Warhol. Warhol says, let's look at each use independently. In other words, you don't get a free pass just because the final output is transformative.

8:24Andrew Berger:Now consider this example. If a startup makes a significant breakthrough with stolen content, that breakthrough doesn't excuse the theft. Progress via theft is still theft. Now contrast, Chabria's ends justify the means approach with that of Judge Alsop. Judge Alsop examined each use as he should. When looking at the downloading of the pirated data set, separately, he said, hey, wait a second. Even if the download and training on wafflingly acquired books is fair, downloading and maintaining a permanent library of stolen books is a separate act of infringement. So Chabria's approach Chabria's approach creates a permissionless system.

9:17Andrew Berger:All copying from any source is fine if the output is transformative. But if other courts follow Chabria, developers will have no legal incentive to acquire data lawfully. Why pay for a license law if fair use treats piracy and lawful acquisition the same? Fascinating stuff. And again, thank you for being so thorough of this and kind of getting my mind spinning with the sort of where we are today with all of these thoughts. One thing you've also highlighted is market dilution, which you kind of touched on, but as a potential winning argument. So how is that different from classic market harm? Sure.

9:57Andrew Berger:Well, market dilution is about saturation. Classic market harm is about substitution, the two S's. So market harm says, hey, well, consumers buy the defendant's product instead of the original. It's a one-to-one inquiry. You pirated my song, so I lost it download. It's easy to measure. But market dilution about saturation asks, hey, is your training process weakening the economic value of a whole class of goods? That's a new theory endorsed in CADRE only. that argument isn't that AI generates a similar copy of my client's romance novel. It's that AI generates 10 ,000 new romance novels in the style of my clients, flooding the zone and making my client's product invisible.

10:53Andrew Berger:You know, in other words, it's just there in spades and, you know, your product is now on the bottom, bottom shelf. You know, think of the grocery store. So think of an algae bloom choking a lake. You know, as the summer goes on, the algae bloom spreads and spreads. Now, there's some support for this market dilution in section 107.4 of the copyright statute. It speaks broadly. It examines, quote, the effect of the use upon the potential market for or the value of the copyrighted work. So the statute examines any effect on value from any use. So it doesn't limit the use to direct copying for the harm to be real.

11:43Andrew Berger:You know, therefore, there's some support for this. Market dilution, nevertheless, is a profound shift. And you can bet that developers are going to ferociously push back against it. Yes, you absolutely can bet on that for sure. Again, you touched on it, but I want to talk now about the anthrotic 1.5 billion settlement and potentially what comes next. Because you said it's not a rounding error, it's a warning shot. So what makes it so important? Yeah, well, you know, it's the largest copyright settlement ever by far. And first, it exposes the enormous risk of training on pirated content. And this risk is shared not just by developers, but by their investors.

12:36Andrew Berger:You probably heard the technologist mantra, hurry up and break things. I mean, that's Silicon Valley's motto. But there's an AI variant. Never ask for permission and only seek forgiveness if you must. But the anthropic settlement shows that forgiveness costs more than a billion dollars. And most IP insurance policies don't cover claims of that magnitude. They're large, but not that large. So investors, beware. Any uninsured portion of a settlement may come from your equity.

13:21Andrew Berger:So investors, if developers stay away from pirated content, if they don't stay away from pirated content, they're going to be gambling with your capital. and if developers don't want to follow Anthropic and pay for forgiveness, they're going to have to license and their excuses to do so now are diminishing. They've long argued, hey, it's just too hard to license. There's just too many works and too many transaction costs. But look at the deals made recently in the last two months in the music industry. Universal with Udio, Spotify with Sony and Universal and Warner and Suno with Warner. These deals all in the past two months showed that large-scale rights aggregation is possible when both sides see value.

14:17Andrew Berger:In other words, there's possibility of more than just fighting here. Judge Chabria agreed in cadre. He found it, quote, hard to believe a less easy market wouldn't, you know, wouldn't take off. He added that if these works are valuable enough to train a billion-dollar model, developers are going to find a way to do it. So when you view Anthropic settlement alongside these music deals, you see a pincer movement. The cost of piracy is catastrophic, 1.5 billion, and the pathway to licensing is being paved. So the combination is going to force developers, especially the smaller ones to come to the table.

14:58Andrew Berger:And that's a positive for all. So no rounding error, Rob. It's the first inning of a new market reality. Yeah, I guess that leads nicely to what I was going to ask next in terms of do you expect a wave of similar claims against, I don't know, Apple, Meta, Google, Microsoft and OpenAI? Yeah, I mean, there's no question about it. But, you know, what we have in this open AI case in New York, Judge Stein recognized that downloading as infringement is enough, at least to pass through a motion to dismiss. And that decision really changes the risk calculation. If Judge Stein ultimately finds in that mega case, no fair use, that downloads of pirated or scraped material is infringing, but so are downloads of legally acquired works if they're later used for training.

15:59Andrew Berger:That's because, hey, when you buy a book, you don't buy a training license, you know, etc. You know, you can't just take that book and, you know, give it to your model and say, train on it. You just have the right to read the book. So all this potential liability turns a copyright dispute into a math problem where statutory damages may be off the charts. because we're talking about millions of books. So the possibility of this kind of judgment raises the stakes for developers and everybody else. It forces them to rethink their reserves. It increases the settlement value, and it gives creators the right to go.

16:42Andrew Berger:And this is really interesting, deeper in discovery. They can hunt for documents that might show, hey, there's no fair use before the developers even raise the fair use issue. And for instance, in the OpenAI case, it's been ordered to produce 20 million chat logs which are the digital record of an interaction between a user and chat gpt 20 million chat logs they have to be animized that is you know for privacy reasons so all eyes are going to be on judge stein when he decides the fair use issue likely later on in 2026. The plot thickens, as they say, Andrew. And I guess you want to flip back now to the courts, because the courts versus Congress, in terms of who really decides the future, because you've been clear that the major legislative reform is unlikely.

17:38So what's broken politically, in your view?

17:42Andrew Berger:Yeah, I mean, what's really interesting is that, you know, years ago, when we had the 1976 Copyright Act. There was an alignment of all the players in the copyright industry, you know, the content providers and the publishers and aggregators, et cetera. And that produced the 76 Act. And again, in 1990, when we had what's called the DMCA, the content industries and the online platforms got together. But today that coalition's collapsed. AI divides every market sector. Authors and rights holders demand consent and compensation. You know, in some cases, they've gotten it. But tech platforms say, no, no, no, we want unlicensed training.

18:27Andrew Berger:We don't want to pay you anything. And publishers and labels are kind of like in the middle. They're split between opportunity on the one hand and threat on the other. So there's no united lobbying front, no shared vocabulary, and no politically safe grounded. What about Congress? Well, we know it's too polarized. As a result, it holds hearings, but you know, so far hasn't thought much. The most we're going to see are some minor procedural fixes, maybe some data set disclosure obligations, more transparency. But that's it. So that vacuum leaves the courts to wrestle. And oh boy, what a problem that is.

19:06Andrew Berger:Because we're dealing with fair use, substantial similarity, and other doctrines that are meant really designed to look at human copying, not industrial scale machine ingestion. You know, we've got a whole new paradigm. And so good judges, as we just discussed, are reaching different conclusions. and even second-guessing themselves. Judge Bebas, a very good judge who's handled the Reuters-Ross case, just certified the case to the Third Circuit when he said, quote, I may be wrong. Judges don't usually do that publicly, but that captures the uncertainty of the moment. So question marks continue, and we'll see what happens.

19:55We will see what happens. And I know you've mentioned quite a few cases and references throughout the discussion, which has been super helpful in framing things. Is there a specific case right now or one that you've mentioned that you believe will shape the next decade? Yeah.

20:11Andrew Berger:Yeah, there is. And that's this mega case. Some people refer to it as the monster or the Frankenstein case in New York, the OpenAI case. And why is that? Because it consolidates a dozen separate actions and, unlike any other, tests the whole entire liability universe from input to output. I mean, a whole ball of wax. So on the input side, it says, hey, is downloading of even illegally acquired work an infringement, as I mentioned before, unprotected by fair use? I mean, just downloading might be enough. If that's the case, oh, my God. On the output side, it's going to decide not only whether ChatGPT regurgitates, but also summarizes.

20:59Andrew Berger:You know, there's a difference between outputting, you know, kind of a memorization and a summary, but the summary issue now is also before the court. A finding either of regurgitation or summarization might reshape model architecture. I mean, these models are going to have to say, wait a second, I'm going to change a whole bunch of things. Further, the Times case is the first to subject the developers. And this is really interesting, too, what's called the Hobson's choice, which are two mutually exclusive options, each of which could have catastrophic consequences. Let me explain what I mean by that.

21:37Andrew Berger:Developers face this choice because their defense to downloading, as I mentioned before, is always innocence. Hey, we didn't do anything. Everything we did is fair use, and so we acted in good faith. But that problem with that is defense puts their state of mind at issue, permitting plaintiffs to ask, hey, why did you think that? Who told you that? So when the developers say, no, no, we can't tell you that we're evoking the attorney client privilege to block the inquiry, courts are likely to follow magistrate Judge Wang's approach in this mega open AI case, where she said either gentleman waive the privilege and produce your communications with your lawyers.

22:16Andrew Berger:or waive your good faith defense. Oh boy, that's a dangerous find to be in because waiving the privilege, guess what happens? It exposes the company's internal legal thinking and nobody wants to do that. But waiving the good faith defense weakens the developer's attempt to defeat willfulness. And that's also a big deal because willfulness could have catastrophic damage. Because this case presents class-wide statutory damage exposure, where each of the work carries a statutory penalty of up to$150 ,000 in damages. Any rulings adverse to developers are going to place monumental pressure on them to settle.

23:00Andrew Berger:So the combination of full-spectrum possible liability on input outside, architectural changes, a dangerous bind to be in with real financial stakes makes this case. the case to watch. And by next July, August, there's going to be a fair use motion made. And oh my God, all eyes are going to be on this one. Yeah. And just as you said there, there's some big, big decisions that can have to be made that can have some, like you say, shattering effects. I want to go back a bit now in terms of lessons from past technology shifts that we've seen and why AI is different? Because I believe you draw on photography, sampling and file sharing.

23:46What lessons apply and which fail entirely in your opinion?

23:51Andrew Berger:I mean, these are really interesting issues because we've always had technological change. And guess what? We've always had copyright and they've always coexisted. They've always found a way to live with each other to make technology possible and benefit all of us. So let's go photography first. You know, back in the, you know, 19th century, many people feel photography would kill art. Oh, my God. You know, we got this photographic, you know, and no longer we're going to have impressionist art or anything else. But the Supreme Court in a case called Burrow-Giles, we disagree. They granted copyright protection to photography because they found that photographers made creative choices in lighting and staging and the moment to shoot.

24:36Andrew Berger:In other words, the court held that copyright protects one's mental conception, even if a machine does the work. You got the idea? So mental conception machine. Now, the copyright office and the courts are still in the pre-Borogio stage. They view AI outputs as mechanically created and lacking human authorship. And so they see AI as a slot machine, not as a camera, you know, a slot machine. God knows what's going to come out. But I believe the courts are going to eventually recognize AI's creativity when they view prompts as a creative direction, not an algorithmic command as a creative direction.

25:21Andrew Berger:Now, when is that going to happen? Well, you know, if you output one of the models with a simple prompt, hey, that's not going to do anything. But think about a session where the user defines again and again the prompts, you know, make the light softer, you know, let's make it Art Deco, you know, add some new angles, etc. All that iteration is legally identical, I suggest to you, to what the photographer's role in Burrow-Giles was. In other words, you know, let's make it, let's use the machine in a way that we wish to have that machine operate. So once a court is convinced that prompt engineering, which is what it's called, provides sufficient control to make the output predictable, copyright protection is going to follow.

26:09Andrew Berger:And for a lot of people, hopefully that's sooner than later. yeah absolutely because you know livelihoods are at stake with this stuff as well it's big stuff and again that kind of leads to where i wanted to go next in terms of sort of cultural consequences if creators stop creating because compensations collapses and your thoughts are around that well you know the problem with that is if you know creators need an incentive to produce and, you know, copyright creation is risky and lengthy. And unless they're compensated, you know, in some kind of way, the creative class is going to say, hey, I don't want to do this anymore.

26:51Andrew Berger:You know, we just, no, that's not going to happen right away. But that's going to happen over time. And, you know, right now, you know, the developers are saying, we're not going to pay you anything except in these cases I talked about before. or the music situations, because they need billions of pieces of content. And they just don't want to spend the time or the effort to reach out to you and me and say, hey, we'll pay you for your blog posts, we'll pay you for your videos, we'll pay you for whatever. And as a result, you've got a lot of creators really facing this sort of existential problem.

27:30Andrew Berger:Hey, why should we stay in the game? the progress clause on our constitution says hey we're going to give creators exclusive rights to incentivize them to produce but if ai is not um it's not respecting those exclusive rights um you know the progress clause gets stood on its head uh and copyright holders say hey i'm gonna do something else yeah fascinating and we've covered a few different directions We've just gone back looking, as you mentioned, their photography. What I want to end on now is looking forwards and in terms of looking ahead, because if you were advising a judge tomorrow, where would you draw the legal line between innovation and exploitation?