[00:00:00] Speaker 04:
May it please the court, Jesse Panuccio for appellants. I would like to reserve four minutes for rebuttal if possible. Just try and watch the clock. Yes, Your Honor.

[00:00:09] Speaker 04:
This interlocutory appeal presents the certified question of whether the DMCA imposes an identicality requirement despite the statutory text having no such language.

[00:00:21] Speaker 04:
Defendants argued repeatedly below for a strict identicality requirement. The district court agreed, but appellees now concede that was error. That was a necessary concession because the text, context, and purpose of the statute all foreclose the imposition of a strict identicality requirement. Under the plain meaning of the statute, plaintiffs in this case state a claim for violations under B1 and B3.

[00:00:50] Speaker 04:
Given this concession, plaintiffs try to shift the focus of the appeal to standing. But the district court correctly held that plaintiffs have standing in this case. Indeed, plaintiffs face classic injury to property rights and rights recognized by the Constitution, and that is all that is required by TransUnion. That injury is currently existing and ongoing, not speculative. The court should therefore reverse the order dismissing plaintiffs' DMCA claims.

[00:01:18] Speaker 03:
So on the standing point, so even if we suppose for purposes of the standing analysis that we accept your understanding of what the statute prohibits, you would still need to show that there was a non-speculative, realistic possibility that your content was going to be reproduced without the appropriate attributions. And So where in the complaint have you alleged that?

[00:01:48] Speaker 03:
Sort of take me through what the logic is of why you think that that's a real possibility here.

[00:01:54] Speaker 04:
Well, let me say two things about that, Your Honor. I will take you through the allegations, but I do want to just address the setup of the question, which is let's focus in on the B-1 claim, for example. The injury, in fact, standing is different from the merits, so you don't have to prove – all of the, that you will prevail on the merits. You just have an under-trans union have an injury in fact that was recognized in the founding era as something courts could litigate. And in these DMCA cases, the injury in fact is the stripping of the CMI.

[00:02:25] Speaker 04:
It is the pulling of the CMI off of the work that is the injury.

[00:02:28] Speaker 03:
In fact, that is what- You mean when they trained the model or when the model outputs something?

[00:02:35] Speaker 04:
Well, wherever they're stripping the CMI, we think they're probably doing it before training and there are allegations to that effect in the complaint. But if you take a work and you strip CMI from it, just generally speaking, that is injury. In fact, that is classic injury to property rights. And moreover, it's the type of injury because it's an injury to copyright. It's the type of injury- that was recognized in the Constitution, and TransUnion says that is enough for standing.

[00:02:59] Speaker 03:
Right, but I guess, fair enough, but it has to actually be happening. Right. And so if we set aside for a moment the training, because I'm not sure you preserved a claim based on that. And if we're looking at what are the outputs of the model, what are the allegations that establish that the outputs of the model will involve stripping CMI?

[00:03:26] Speaker 04:
I'll take the second part first, but I do want to get to the waiver issue that I think you're referring to, because I do think that's wrong. We have very clear allegations from the first complaint all the way to the second amended complaint about stripping at the training phase, and I'll walk the court through that. But as for distribution, we cite several things. We plead several things. First of all, with respect to several of the does, we go through and show how their code was outputted in the model.

[00:03:59] Speaker 04:
That's one. Two, we know the other side has an admission that 1% of the time they will regurgitate code. And third, they put on a feature. They put a feature onto the program saying, would you like to be notified if we are regurgitating code from the public GitHub repository?

[00:04:18] Speaker 03:
Can I ask you about the 1%? Because, I mean...

[00:04:23] Speaker 03:
Maybe that's enough, maybe that's not. It seems to, the analysis of whether that's enough seems to depend on knowing some things about how many different, how much code is on GitHub, how many people are querying Copilot and generating these outputs, and I don't think any of those numbers are in the complaint. So do you think that what we have is enough, and if not, I mean, is there more that you could say about that if you were to amend?

[00:04:58] Speaker 04:
Well, I think if what we've put in our complaint is not enough, it's hard to see how the DMCA could ever have life with artificial intelligence because all of these, it is uniquely in the defendant's possession how many users are prompting every day, what are they prompting it with, and what are the outputs. So we were able to test.

[00:05:20] Speaker 04:
for ourselves and see that our own code, our plaintiff's code is being outputted by the model just on our own tests. That should be enough to get us through the door to take discovery, especially when all of the evidence is uniquely in their control. If it's not, it's hard to see how anyone would ever say, I know my code is being outputted because no one knows what all the users of Copilot are prompting every day. That's either, you know, most likely in the hands of the defendants. Your Honor, I do want to get back to, and I think this is very important, I do want to address the idea that stripping at the input was waived or not alleged.

[00:05:59] Speaker 04:
If you go back to the original complaint, paragraphs 143 and 145 through 146, we said the following. 143, defendants removed or altered CMI from open source code that is owned by plaintiff and the class after the code was uploaded to a GitHub repository by incorporating it into copilot with its CMI removed. Now, that was 143 of the original complaint. It remains in the operative complaint today as paragraph 214.

[00:06:30] Speaker 04:
We then alleged, at paragraphs 145 to 146 of the original complaint, defendants were not licensed by plaintiff to train any artificial intelligence using the licensed materials or to incorporate the licensed materials into copilot. Those allegations still remain in the complaint at paragraphs 216 and 217.

[00:06:49] Speaker 03:
But the district court seemed not to think that you had a claim based on training. And it said that. And it said that at the hearing, it said that in, I think, footnote seven of the order on the motion to dismiss the first version of the complaint. And other than, I guess, repeating the same allegations that it didn't think had stated that claim, did you do anything to disabuse the district court of that understanding?

[00:07:17] Speaker 04:
I think we didn't. I'll take your honor through it, but I want to be clear. Footnote seven is in the portion of the opinion about standing. But if you move forward to the portion of the opinion that gets into the 12v6, here is how that, after we had that first complaint, here is what we said in the opposition to the first motion to dismiss. This was ECF number 67 at page 16. We said, plaintiffs allege that defendants intentionally removed or altered CMI from the licensed materials after they were uploaded to one or more GitHub repositories.

[00:07:49] Speaker 04:
And then at the hearing on the 1st, motions to dismiss. We said it was the use of plaintiff's code in derogation of the licenses, which is at the core of our claim.

[00:08:00] Speaker 04:
And the court went on and asked this question. Is the act of copying itself into the neural network by itself violative of any of the GitHub licenses? And we answered at SER-91, I think the answer is yes. So that was at the hearing. And then if you look at the first order on the first motions to dismiss. This is SER 5455.

[00:08:24] Speaker 04:
The court concluded plaintiffs alleged that their license code contains CMI. Plaintiffs alleged that defendants removed or altered that CMI from license code and distributed copies of the code knowing that the CMI had been removed or altered.

[00:08:38] Speaker 04:
At page 55, SER 55, plaintiffs alleged that relevant CMI was affixed to their license code And defendants subsequently trained these programs to ignore or remove the CMI. And then the court said, that is enough to state a claim. So we think that fairly said that we're stating a claim on input, stripping at the input phase. And then at every subsequent phase of the MTDs, of the motion to dismiss practice, we reminded the court of its original holding. Now, at that point, defendants started shifting and saying we want to talk about this identicality requirement.

[00:09:13] Speaker 04:
And a lot of the briefings started focusing on that. But that original holding of the court that we stated the claim on an input theory was there. And in all of our subsequent oppositions, we reminded the court of that holding. So I do not think it's fair to say we've waived this in any way, Your Honor.

[00:09:27] Speaker 02:
But counsel, so would you agree, though, that with respect to the waiver issue, what it comes down to is our interpretation as to whether you fairly raised that question at the hearing itself when Judge Tiger specifically asked questions as to your theory. I understand your point that you did not disclaim reliance on that theory. But if our interpretation of the record is different, then you would have waived it, notwithstanding the fact that you raised allegations in the pleading.

[00:10:00] Speaker 02:
Am I correct?

[00:10:02] Speaker 04:
I'd say two things in response to that, Your Honor. One, you know, I tried to read here that exchange, the colloquy between Judge Tiger and us where he asked directly the question about input. And we said, I think the answer is yes. And then what happened subsequently?

[00:10:18] Speaker 04:
Two, I don't think it's just that we continued to plead it in the complaint. It continued to be in our oppositions, our actual written oppositions to the motions to dismiss to preserve that theory. And just to round that out, we said in our oppositions, plaintiffs allege that defendants coded their programs to ignore, or remove CMI in order to stop reproducing CMI as output, these allegations are still sufficient. And we were pointing back to the court's first order. These allegations are still sufficient to plead a DMCA claim under sections 1202B1, which is not the distribution claim, and 1202B3.

[00:10:58] Speaker 04:
So I think that question was kept alive by us, Your Honor, throughout the proceedings. Although Everyone started to talk more about identicality, and that's the question up here. I don't think it's fair to say we waived anything.

[00:11:13] Speaker 03:
We'll make sure you get enough time for a rebuttal, but I do want to give you a chance to say a bit about the merits.

[00:11:19] Speaker 03:
And maybe you could start by addressing, and I think I understand your friends on the other side to have conceded that identicality doesn't really mean strict identicality.

[00:11:32] Speaker 03:
So that's one end of the spectrum. And at the other end, I understood from your brief that you're acknowledging that a derivative work that might infringe, even if it, for example, quotes part of the original work, that not including attribution in that new work is not necessarily a DIMCA violation. So where in between those two poles do you think the statute draws the line?

[00:11:59] Speaker 04:
Well, I think it's important to, one, distinguish between B1 claims and B3 claims so we understand what we're talking about. I mean, usually you're going to compare copies if you're dealing with a B3 claim, but you don't even really get into comparing copies unless you're in a circumstantial evidence case. If you have direct evidence of CMI stripping, then most of these cases, let me step back, most of the cases that discuss comparing the original work and a copy are and trying to figure out if CMI was stripped are doing that because that's the only evidence that's there.

[00:12:31] Speaker 04:
There's no direct evidence that the defendant actually went in and stripped the CMI. And so then courts will start to look at it and say, can we infer from this circumstantial evidence that there was CMI stripped? But where you have direct evidence of CMI stripping, that's enough under the statute. And that's what we allege in this case. I mean, they own the repository. They downloaded everything from the repository. They stripped the CMI out. And they train on it. And then they have outputs.

[00:12:56] Speaker 04:
plaintiff's code. So I think we have to, it would be important for the court to recognize in its opinion here the distinction between direct evidence and circumstantial evidence.

[00:13:05] Speaker 03:
And suppose, can you answer the same question without referring to what happened at the training stage? Yes.

[00:13:15] Speaker 04:
So to get back to Your Honor's original question, so if you are comparing copies of I think this is very familiar to copyright law. It's been going on for 250 years on the right of reproduction, where if somebody copies a work, the question has always been, is it substantial enough copying? Does it get to the heart of the expression to be an infringing copy? And I think you could have much the same type of inquiry here. I don't think we need to reinvent the wheel. This is a familiar inquiry to copyright.

[00:13:47] Speaker 03:
But doesn't that just collapse? I mean, is it your position then that like anything that infringes the copyright is a DMCA violation if it doesn't have the attribution information in it?

[00:14:02] Speaker 04:
Two responses to that. The first line answer is no, not everything that is a infringement of another right under the bundle of sticks of copyright will necessarily also infringe the DMCA right.

[00:14:14] Speaker 04:
And that is because, for example, take the George Harrison case. You may remember the song, My Sweet Lord. He was found to have subconsciously copied. If you have subconscious copying, I don't think you would meet the center element that is uniquely in the DMCA. So not all infringements of the reproduction right would necessarily be a DMCA violation. But at the same token, at the same time, It is often true that a copyright infringer is breaking multiple sticks in the bundle of copyrights, and Congress has adjusted that for 250 years about what bundle of sticks you have in copyright.

[00:14:55] Speaker 04:
But there's nothing earth-shaking about the idea that someone who is infringing, who is breaking the law of copyright, would break it multiple times over.

[00:15:03] Speaker 04:
And some of these questions are going to be fact-bound. Let me just give one example from OpenAI's brief.

[00:15:11] Speaker 04:
Well, let me step back. Think, for example, of a book where you actively... rip the first page off the book, okay? And then you copy that on a photocopier and you begin to sell copies with no CMI. I think everybody would say that is clearly a DMCA violation, a B1 violation. You've actively removed the CMI and now you're distributing it. Now imagine someone had that same book and instead of ripping the cover page off, they just went to the photocopier and they flipped past the first page and they made all of the photocopies.

[00:15:47] Speaker 04:
They say, I think as I understand their theory, that that wouldn't be a violation of the DMCA because it's removing the CMI by omission. I don't think that's right. I think you'll get into fact-bound questions about what the intent was and how the CMI came to disappear from the copy.

[00:16:05] Speaker 02:
Counsel, let's say what you have is a book reviewer who's critiquing a book and takes out a very lengthy paragraph that fills a page.

[00:16:19] Speaker 02:
without CMI, without attribution, is that a violation?

[00:16:24] Speaker 04:
It would depend on whether the, so I think it would depend on whether that's a violation of the right of reproduction. Recall, Your Honor, that the clause at the end of B, it's the double center requirement. So you have to strip the CMI intentionally, and then knowing that it will induce, enable, facilitate, or conceal any infringement of any right under this title. So typically, book reviews are seen as fair use and not infringement. And in fact, the fair use section of the statute says fair use of a copyrighted work is not an infringement of copyright.

[00:17:00] Speaker 04:
So the text weds up perfectly here.

[00:17:04] Speaker 04:
If you have fair use, which a book review would normally be, then exerting a piece of the book without the CMI would not be a DMCA violation.

[00:17:12] Speaker 02:
So any time you have fair use that results in there being no copyright violation, the same is true with respect to the DMCA, no violation?

[00:17:22] Speaker 04:
Well, it would depend on whether you could imagine, for example, and we don't agree with this, but imagine, and at least one court has said that the training itself is a fair use. That doesn't mean the downstream outputs would also be a fair use. So, again, it's a fact-bound question. But your Honor's hypothetical is a very good one. A book review is going to be fair use, and it's very unlikely that you're anticipating downstream use of that book review to infringe.

[00:17:50] Speaker 03:
All right. Thank you. We'll give you three minutes. Thank you.

[00:18:02] Speaker 03:
Mr. Cariello?

[00:18:10] Speaker 01:
Thank you, Your Honor. May it please the Court, Chris Cariello for defendants GitHub and Microsoft, and I'm dividing time today with Ms. Blatt, who represents the OpenAI defendants. From their initial complaint through two amendments, through their argument a moment ago, plaintiffs have been searching for just the right combination of things to say to try to get a 1202B claim past the pleading stage. But the defects in that claim are inescapable because they lie in the very account plaintiffs provide of what this technology is, and what it does.

[00:18:41] Speaker 01:
Between Ms. Blatt and I, we're going to focus on two of them. We briefed three independent reasons for affirmance here. But I'm going to focus on the standing problem, the basic problem that the injury for which plaintiffs seek redress in a 1202 claim isn't something that has ever happened to them or will happen to them. And then I'll make a start, if I can, on the 1202 merits question before handing off the podium to Ms. Blatt. So starting first with standing, it is axiomatic that in order to have standing, you yourself must be injured. You can't sue on the basis that someone else might be injured at some point.

[00:19:15] Speaker 01:
For damages standing, you need to have been injured. For injunctive relief standard, it must be likely that you will be injured in the future. Now, we don't have to deal with damages anymore because plaintiffs no longer seek them in their complaint. So the only question we're dealing with is whether they've plausibly alleged a likelihood that they will be injured in the manner that they have articulated since the very beginning of this case, which is emitting their copyrighted code with CMI removed. So let's start with what they do allege and what they don't allege to see how they have failed to plead a likelihood that they will be injured.

[00:19:50] Speaker 03:
First, their code. Before you get into the details of that, why is it appropriate to think of this as a standing issue rather than a failure to state a claim on the merits?

[00:20:01] Speaker 01:
Sure. So it could be both. The McGee case says you need to yourself be among the injured. I think in a case like this where named plaintiffs are purporting to represent a class, it is important to think of it as a standing problem. You've come in, you've articulated the possibility that someone could be injured in a particular way. Are you, the named plaintiff, among the people who could be injured in that particular way? And so I agree with you. If this hasn't happened to them, they also have a 12B6 problem.

[00:20:32] Speaker 01:
But I do think it's a standing problem in the first instance because they need to demonstrate the injuries happened to them.

[00:20:37] Speaker 03:
But I guess it's sort of oversimplifying. Their claim is you took our stuff without permission. And the argument you're making under the rubric of standing is no, we didn't.

[00:20:52] Speaker 03:
And that just seems like – I mean the we didn't do anything to you, this didn't happen argument – It doesn't normally result in a dismissal for lack of standing. It results in a dismissal on the merits.

[00:21:05] Speaker 01:
So I understand. I think your honor is onto the fact that it is also a merits problem. I don't think that means it's not a standing problem. The way with just a friendly amendment, the way I see their complaint as saying, here is technology that can result in an injury to some people. And they're coming in and saying, well, You know, we might be injured this way, but they haven't plausibly alleged that that's ever happened or could happen. So they're alleging a sort of product they're challenging, a technology they're challenging, saying it could result in injury to a large class, but they, the named plaintiffs, can't say that they are among those who would be injured.

[00:21:43] Speaker 03:
And why haven't—so then on getting to the facts of it— They've alleged that 1% of the queries return something that they say infringes rights under the statute. You have millions of queries being made every day on this. I'm not sure all the numbers necessary to do the math are in the complaint, but it seems like the complaint could easily be amended to add them.

[00:22:16] Speaker 01:
Your Honor, they've had three opportunities to do this, and the defects I'm about to lay out, we have been laying out since moment one on the first complaint. So plaintiffs tell us how this works. They say there are billions of lines of code. So their code is just among these billions. That's paragraph 95.

[00:22:35] Speaker 01:
And then they tell us what Copilot does with these billions. It's a pattern recognizer. So it transforms these billions of lines into by inferring statistical patterns, and that's paragraph 64. Those statistical patterns become embedded, and then on the back end, this is paragraph 65 and 91. The way it works is it is predicting from those embedded statistical patterns what you want next. So if you're a software programmer, you're sitting there, and you're programming something, let's say it's a sorting function, right? Take this data, make it alphabetical, and Copilot, if it's seen something like that a lot across the training set, says, looks like you're looking for this sorting function.

[00:23:15] Speaker 01:
So what don't plaintiffs allege here? They don't allege what their code is. They don't allege what their code does. They don't allege how much of it is in the billions. They don't say it appears frequently. They don't say it's part of a pattern. They don't say it appears more than once.

[00:23:29] Speaker 02:
Can I ask, though, perhaps I misunderstood Mr. Panuccio's theory, but as I understand it, he's suggesting that when a programmer, a inputs the code into the algorithm, the CMI is being stripped.

[00:23:52] Speaker 02:
And if it's being stripped, whether or not it is then being used for output, as I understand it, the injury that's being alleged is the stripping of the CMI directly.

[00:24:05] Speaker 01:
Your Honor, there has always been and there continues to be a lack of clarity, I think, in plaintiff's theory about when and how precisely CMI is stripped. Stripping is an act. They have to allege the mechanism by which that happens. I had understood throughout this entire case that their theory was when Copilot generates from its statistics the predicted code completion, it is not including CMI. And just to be clear, Plaintiffs have over and over again alleged precisely that theory in their complaint, right?

[00:24:38] Speaker 01:
This not including theory, not a strict theory.

[00:24:41] Speaker 02:
Putting aside what precisely they're alleging, if that were the allegation that, in fact, once the coding information is captured, if you will, by the program, by the algorithm, it is removing the CMI, Is there at least standing to claim in that instance that the violation occurred at that point giving rise to standing?

[00:25:12] Speaker 01:
So in the hypothetical world where a different plaintiff pled something like that and all they said was at some point during training, CMI is removed. They might be able to plausibly allege – and I see my time is up, but if I may – they might be able to plausibly allege that their code was among that that had CMI stripped. Now, we would get over this first hurdle, right? Did it happen to you? We could then have additional arguments over whether or not that injury satisfies transunion, right? Mr. Panuccio brought up that concept.

[00:25:44] Speaker 01:
We could have arguments on the merits.

[00:25:46] Speaker 01:
But they could at least plausibly get over the hurdle if they articulated that injury and then plausibly allege that that injury has happened to them. That's what they cannot do here. They have articulated an injury which is output of their code without CMI and failed to do anything to plead that that would be likely. We don't know what their code is. They've never alleged that it could potentially have utility to anyone, right, that a coder might be eliciting it. We are shooting in the dark.

[00:26:17] Speaker 01:
All we know is that their code is somewhere among billions of lines of code in this model. And Your Honor, if I may just take a brief moment to address a few other points that they've raised. Plaintiffs point to their examples. These examples are wildly unrealistic, and they don't allege that they are realistic prompts. Contrast that with ER 199 to 202 to see what plaintiffs could have if they were different people. possibly alleged.

[00:26:47] Speaker 01:
At 199 to 202 of the excerpts of record, they describe two examples from textbooks, textbook code that would be well-traveled across GitHub and that might actually be useful to someone. As far as we know, this is defective code that no one would ever want to prompt and that no prompt would ever elicit in the real world.

[00:27:11] Speaker 01:
Thank you very much, Ernst.

[00:27:24] Speaker 03:
Ms. Blatt.

[00:27:27] Speaker 00:
Thank you, and may it please the Court, Lisa Blatt for the OpenAI defendants.

[00:27:32] Speaker 00:
Although removal can occur without identicality, this Court should hold that removal only occurs if a defendant tampers with existing CMI on existing works and not simply fails to add CMI to the defendant's own products.

[00:27:49] Speaker 00:
Now, in this case, the complaint does not and cannot allege, at least at the output stage, and we don't think the input stage was either alleged or before the court, anything that comes close to removal, because what we have here is simply an attribution right that is being alleged, which is that co-pilot generates modified snippets of plaintiff's code without also generating information. the CMI. And you can look no further than the only relief sought in this case is at 263 of the complaint.

[00:28:20] Speaker 00:
The only live issue in this case is a request for an injunction that is solely based on attribution. The injunction is make sure any output on copilot contains CMI. That is a request for attribution. There's no request for any kind of relief at the input stage. Moreover, when the other side sought certification and it talked about damages, which aren't asserted, but when damages were asserted, all the damages are based on output.

[00:28:51] Speaker 00:
That $9 billion figure is based on failing to output.

[00:28:56] Speaker 03:
Can I ask you the same question I asked Mr. Panuccio?

[00:29:01] Speaker 03:
On the one hand, if you have a work of fan fiction, even if it has quotes from a book, that doesn't trigger the attribution requirement. On the other hand, if you take the book and strip off the cover page, I think I understood you to effectively concede that even if you've changed a couple words on page 200, that that's still a violation because it's close enough to being identical. So how would you articulate where the line is in between those two?

[00:29:30] Speaker 00:
Sure. And you hit it exactly right with this extremes. So there are two extremes at issue. And then there's the line we draw, which is based on the statutory or sorry, the common sense dictionary definition. definition of removal the two extremes are the one that you said which is an identicality requirement that even if you have a removal there's this notion which no one is defending that that would preclude liability even though there was removal because the two works are not identical and there was one change at the other extreme is exactly what you said that any copying that fails to have an attribution violates at least the removal requirement if you have the cnter And that is incredibly important that we're asking this court to address, because not only is it the only relief sought, the whole notion of an identicality requirement that has what has been held by so many district courts is because it has been the dam that has prevented the floodgates of attribution right claims.

[00:30:29] Speaker 00:
Now, where the line is drawn is with us, and we can talk about, because you hit on all the exact hypotheticals, is we think there has to be an actual tampering on CMI. So the hypo, that's the hard hypo, would be what is the difference between stripping off or kind of tearing off the cover of the book and just copying every page but not the cover of the book, which I think is the hard hypo you asked. And in our view, that only the tearing off is a removal because in the other example that is highlighted, it's simply a failure to attribute.

[00:31:05] Speaker 00:
Now, it may be a line, but it is a line that has to be drawn, including by the other side, because once they concede in their reply brief that there is no attribution right, some courts and all courts have to grapple with what do you do when you have a copyright violation or any copying, whether it's fair use or not. because any copyright violation, almost all of them, fails to include the CMI.

[00:31:28] Speaker 03:
Right, but if the line is going to be... The difference between removing something and just not reproducing that part, I mean, maybe makes sense for physical media, but it doesn't seem to make sense at all when the media are all stored digitally. So how would you apply that?

[00:31:47] Speaker 00:
Sure. So you can apply it at least in the...

[00:31:52] Speaker 00:
AI in the output sense because you know generative AI is not, it is a model standing on its own that is not connected to the training data. So you are right if you just have, let's just talk about the theory that's not before the court, that is before entire courts all over the country about training. We just think the district court was very focused and said, I want to make sure training is not part of this case. Once the court satisfied itself that training was not part of the case, it then went off into the output.

[00:32:24] Speaker 00:
But in terms of the input, you are correct. courts are grappling with, particularly in the Southern District of New York, what to do when you have these training data. And in the kind of the granddaddy of cases, the New York Times case, what Judge Stein has held is that some of them adequately state removal at the training stage, and some of them don't, depending on what the facts are alleged. And there are very detailed allegations of the programs that are used to to input that stuff in the training stage, that for whatever reason, I don't know why, these plaintiffs just didn't allege and didn't make training part of their case.

[00:33:02] Speaker 00:
Now on the output stage, that is... Just purely a, it's no different than if an art student sits in a museum and tries to replicate a Monet painting and doesn't copy down Monet's signature. That is simply what the AI models are doing by their own allegations. They never allege, this is absolutely critical, that AI models work like a search engine, which is they don't take an existing repository and copy and paste information. In fact, the sources that they cite in paragraphs 75 and 102 for the notion that there is some sort of copying, those very sources explain that it's the Chen study and the Githo post.

[00:33:41] Speaker 00:
And we put all this on page 12 of our brief. say that these models aren't looking up the training data. The training data is not part of the model. So it's simply based on a statistical pattern. It's like I have memorized the stuff that I am saying to you now. That is not that I am downloading it for something. It's simply I have memorized it and it's Lisa Blatt saying it. I'm not removing any CMI from something I've read. I'm just not telling you what I read and to prepare for this argument.

[00:34:09] Speaker 03:
So, I mean, your, I mean, the briefs use the example of, you know, the reproducing Twilight. I mean, if somebody with an exceptionally good memory just has, you know, and I, perhaps there are some fans sufficiently devoted to have done this, you know, sits down and just, you know, types out the entire text of the book, but minus the title and the, you know, by Stephanie Meyer part of and then starts selling, prints it out and starts selling bootleg copies of that.

[00:34:39] Speaker 03:
You think that's not a violation because it sort of went through somebody's head along the way to being copied?

[00:34:43] Speaker 00:
Yeah, so let's be clear when you say violation, because we're talking about that may or may not be a copyright violation. And if what's bothering you, well, that's damages and that's lots of stuff happens, including injunctive relief. If we're talking about a DMCA violation, it requires a removal, which is why we have the courts have adopted the identicality requirement. The statutory penalties are bankrupt-inducing. They are not supposed to be about copying. The author doesn't have to register their works. It could be a fair use. It could be all sorts of things.

[00:35:13] Speaker 00:
But who needs the copyright laws if they're correct about their definition? You would never need to bring a copyright claim. You just always bring a DMCA claim swallowing the copyright law's hold, which is this case. There is no allegation of infringement. and yet they're claiming up to a $25,000 per violation. Imagine any teacher in a classroom would be liable. We can talk about how fair use would work with that, but it's not clear to me that you could still be liable under the DMCA, even if you have a fair use copying.

[00:35:45] Speaker 00:
It depends on their definition of C-ENTER. And they got by with their definition of Sienta here based on an attribution right. So however this court answers the certified question, and we don't think they're standing, but if you do answer the certified question, it is absolutely critical that you tell the courts that whatever you think about identicality, you have to make sure there's a removal, and removal has to be done with the dictionary definition of taking out something, not simply, and I think the will example was the best example we gave, is that at the end result, if I'm denied, I'm a beneficiary and I don't get my parents' money, I'm unhappy, but I could be either omitted from the will or I could have been removed from the will and the statute draws that line and you have to draw the line no matter what.

[00:36:30] Speaker 00:
Any hypothetical we could talk about, because even in your Twilight example, you'll start to move off. Well, what if they copied everything but one page? That's clearly not a DMCA. We could just keep going down to the hypo we gave about the book review. In the book review example, the complaint itself and their brief to you in the opening brief says that the failure to attribute is a DMCA violation, 24 times their complaint cites to an attribution right. And so, again, we think in order to clarify this law, it's the Ninth Circuit.

[00:37:00] Speaker 00:
There are a lot of cases that come up in California because you and the Second Circuit are the big copyright states. Copyright circuits, it's absolutely critical that this court at least reset what's going on and not just fix half the problem on an identicality requirement without also telling courts what the other side concedes is that there is no attribution right. Thank you. Thank you.

[00:37:29] Speaker 03:
Rebuttal?

[00:37:31] Speaker 04:
Thank you, Your Honor. I'd like to make three points, if I may. First, the position that OpenAI is asserting, I think, is extraordinary. So I gave the example of actively ripping off the cover of a book and then photocopying it and distributing those copies. And OpenAI says, yes, that would violate the DMCA. But all you have to do, in their view of the law, to avoid that liability is simply take that same book look at the cover page and say, well, instead of tearing it out, I'm just going to flip it over one.

[00:38:04] Speaker 04:
I'm going to make the exact same copies and engage in the exact same distribution. That would make a farce of the DMCA. That can't possibly be the law.

[00:38:11] Speaker 03:
I agree with you that that doesn't seem to make a lot of sense. But what do you do with the fact that the statute says remove or alter?

[00:38:19] Speaker 04:
Well, remove or alter. Let's take their example from page 35 of their brief. They say, Look at restaurant menus. When the restaurant removes a menu item, that's when you remove something. They point to that example. Well, what does a restaurant do when it removes a menu item? They don't take their existing menus, grab a scissors, and cut it out, and then hand the patrons a menu with a hole in it. They just create a new menu. And so the original menu had lasagna on it. The new menu does not. And a patron would sit down and say, huh, They've removed lasagna from the menu.

[00:38:52] Speaker 04:
Even though there was no active stripping of that menu item, we all in common parlance think that that is removal of the item. And the same thing would be true here. So I don't think it tortures the regular meaning of remove at all to say that some acts of stripping CMI by omission will count. Now, let me go to Your Honor's Twilight example, the fan fiction.

[00:39:15] Speaker 04:
It may come down to a question of intent. Yes, if you memorized it with the intent of retyping it and just leaving the CMI out, I think that could qualify. But if you were more toward George Harrison, where you're sort of just kind of remember that Twilight's out there and you start writing similar stories and maybe it infringes, maybe it doesn't, you probably won't have the requisite C enter. And so this will be a classic fact question, and that's why we say that in our brief.

[00:39:42] Speaker 04:
The DMCA is a copyright. It is not something separate from copyright. It is one of the bundle of sticks, and it is important that it be preserved. And my friends on the other side are seeking an interpretation of that statute that would essentially read it out of the law. And Ms. Blatt said, look at the Second Circuit. And I would commend to the Second Circuit Judge Park's opinion in the Mango case. That is very clear on saying this is a real right. It has teeth. And you only need to allege what the statute says. Let me make two other points with the 30 seconds I have left. One is the injunction we request.

[00:40:13] Speaker 04:
We don't request an injunction just as to distribution. Paragraph 49, we say enjoin defendants from engaging in the unlawful conduct alleged herein, all of it. And of course, the contours of an injunction come into play once the case has gone through discovery and has been litigated and a court fashions an injunction to meet what has actually been proved. Lastly, I want to say one more time, My friends say the injury isn't happening or the input is not part of the complaint. The first injury we allege is that stripping of CMI happens after they download from GitHub and input it into their model.

[00:40:49] Speaker 04:
And this is clear as day in paragraphs 214 and 216 through 217 of the amended complaint. That is one injury. The second injury happens again when the output works. They distribute them either in training or in the output without the CMI. But these are two separate injuries, and we very clearly allege, and I think calling it a training theory is wrong. It's an input theory. They are stripping the CMI. They own GitHub. They own the whole repository. They download the whole repository.

[00:41:19] Speaker 04:
Then they strip the CMI to clean the data for training. That act of stripping itself is an injury, and it's an ongoing injury because the point of the DMCA is not to have these CMI-stripped copies out in the wild. Thank you, Your Honor.

[00:41:32] Speaker 03:
Thank you very much. We thank all counsel for their helpful arguments. And the case is submitted and we are adjourned.