 All right, I think we're going to get started. Welcome to the Berkman Klein Center to our last luncheon speaker series, Tuesday Luncheon Speaker Series of the Economic Year 2017-2018. I'm Chris Babbitts from our Cyber Law Clinic. I'm honored to be able to introduce my friends and colleagues, Jess Fialde and Mason Courts, who are going to have a great talk today, art that imitates art, computational creativity, and creative contracting. Before we do that, just sort of the usual Tuesday housekeeping at the Berkman Klein Center Luncheon Speaker Series, we're being live-streamed and recorded for posterity. So hello, internet. And, oh, we're not recording today. Sorry, we're live. Just only recording, not live-streaming. Okay, good. We got it. But if you want to volunteer and ask a question, I think we'll have plenty of time for discussion. Just be mindful of that. And as I mentioned, it's our last one of these, I think, for the academic year. So we are not likely to have public events through the summer months. We'll be picking back up again in September. Keep an eye on the website for future announcements and, again, the series will start back up again in September. Again, I'm really happy to be able to introduce Jess and Mason, with my work in the Cyber Law Clinic. Jess Fialde is our Acting Assistant Director, a clinical instructor and a lecturer at HLS who works on a lot of our matters related to copyright and IP. Mason, a clinical fellow who, prior to joining us in the clinic, worked at the ACLU of Massachusetts and is a clerk in the District of Massachusetts, works on a pretty broad range of matters. That touch on IP as well as online speech, technology has a real firm data science background and their interests have come together around a set of issues that I think really bridge the gap between some of the research work that Berkman's doing on issues related to AI and sort of the nuts and bolts law practice work that we do in the clinic that relate to the intersection between AI and machine learning on the one hand and creativity, the creation of new copyrighted works on the other hand. So without further ado, I'm going to turn it over to the two of you. So thanks very much. Thank you. Hi, everyone. It's so great to see such a crowded room. So the way we have this structured, I guess we hear that like back when all the lunch talks used to be in this room, people would go around and introduce themselves but because it's so tight today, we want to make sure to leave a lot of time for a discussion. So we're not going to do that but if you do speak up later in the discussion if you could just preface your comment with your name and where you're coming to us from. That would be awesome. So the topic of today's talk starts with this idea of computational creativity which is a kind of catch all term for any sort of AI or algorithmic driven or assisted creative work. And there's a lot of this going on already. We're going to show you a few examples later on some of which are extremely cool. And this ranges from humans who are creating digital art with techniques that require high degrees of computation to get the effects they want to AIs that with a little bit of coaxing and coaching from a person can generate on their own essentially novel pieces of artwork in a given style, even some we're going to talk about that generate their own scene. So this raises an incredible array of really deep philosophical questions about what is the nature of creativity? At what point does the human cease to be the most important factor in generating the art? What does this say about our own human creativity if it can be replicated through machine learning, through AI? What is the nature? What's the difference between training and AI on a body of works and me going to the museum and just looking at a bunch of art and getting inspired and going home and painting something other than that the AI would do a much better job. So these are like these incredibly broad and deep questions that go to the very nature of human creativity and our relationship with machines. And today we are not really going to be talking about those questions. Instead we're going to be talking about some real concrete immediate legal issues, mainly when someone creates art using AI either as an assistant or as the main driver of the art. Who owns it and how can the rights in it be divided up? Yeah. And we think that this is the right approach because while these big questions are very compelling and I have to say they've been raised for like going on over 50 years, now the copyright office was presented with the first purportedly computer authored work to be registered before 1965 because they put that question out there to the scholarly community, like how should we deal with computer authored works first in 1965, which is incredible. So the law review articles have been bouncing around, people have been feeling these big questions, but what we're finding in our practice at the Cyberlaw Clinic is that people are encountering real world quite concrete questions about this today. And for that reason we think that the sort of concrete questions that they raise are just as interesting if not more interesting because we can delve really directly into the reality of how artists and engineers and computer scientists and neuroscientists and others are thinking about this work. So as a result of our first few forays into advising clients on issues related to AI generated artworks, Mason and I together started to develop what we called a schematic or an anatomy of AI generated artworks. And the reason that we felt that this was so important is that almost no one, right, not the artists, not the lawyers, basically no one, but the engineers and computer scientists that we were working with understood how these algorithms actually worked. So what we have on the screen here is a sort of largely visible reproduction of that schematic and Mason is gonna tell you a little bit more about the various parts. Sure, so in the upper left we have the inputs and these are the training data, right? This is what your machine learning algorithm and I should say this on the site here, we talk about this as AI and art. We're really focusing in this discussion on AI by way of machine learning. There are other potential ways to have AI's to generate art including evolutionary algorithms. In this case we're assuming some sort of input set that you're not starting from scratch. So you have the inputs, these are works that were generated probably by a human, maybe by another computer system at some point in the past. And it could be visual, it could be literary, it could be musical. I don't think we've seen sculptural yet but I wouldn't be surprised if that's coming down the pipeline. Yeah, to some degree in the next Rembrandt they did. That's true, there's some 3D scans of the paintings. Yeah. Yes, we can. Yeah. As Mason goes through them up. So it says the inputs provide training data for the trained algorithm. And the learning algorithm which says, analyzes inputs to produce the training algorithm. The learning algorithm, this is the machine learning algorithm. So this is your neural network or your symbolic logic network that is reading the inputs, whatever type they be and turning them into some sort of representation of relationships, what notes tend to follow, what notes, what kind of colors a particular artist uses, what sorts of scenes are pleasing to the eye. So the learning algorithm operates on the inputs and it produces what we're calling the trained algorithm. You might also call this the model, the data model. But this is essentially the statistical representation of everything the algorithm has learned by reading over the inputs. Again, could be a neural network, could be a Markov chain. We're really trying to stay as much as we can agnostic to the specific underlying technology. But it is a representation of some sort of numerous works that have been ingested. It probably doesn't look much like them. If you were to print out the code, you wouldn't see a bunch of Rembrandts in there. But it is a representation. And that's important when we talk about the status of it as a derivative work. It may be hard to see here, but at the highest level of abstraction, you can think of it as just a data table, right? So that's what the gray static behind trained algorithm is, is just columns of data. And then the trained algorithm can produce outputs. And those outputs are generally of the same type as the inputs. So if it was trained on music, it's gonna produce music. If it was trained on paintings, it's going to produce paintings. Optionally, you can have seed material. So you can supplement it with seed material, which will influence the character of the output. So for example, Google's deep dream, if anyone has seen this, they train an AI to recognize animal faces and then you put in an image, a photograph, or a painting as a seed input and it will transform that into the output based on what it has learned about detecting animal faces. Leads to some very strange results sometimes. It essentially looks at the input, identifies whatever might be an animal face and then puts an animal face over it. And you can run it through multiple times and you get more and more and more animal faces and deeper and deeper and deeper into what looks like a fairly hallucinatory experience. You haven't seen any of them. They're pretty incredible. They're not among the examples we have for you today. Yeah, that's true. But so the, and so one of the things that we think is really helpful about this model for bringing artists, developers, and lawyers together and getting on the same page is that under the law, we think each of these can and should be treated as a separate work for the purposes of copyright. I mean, the inputs, those are paintings, their music. Most people would generally consider those to be copyrighted works. The outputs, which also resemble the inputs, it's pretty intuitive to people that those are works but also the machine learning algorithm is code and code is to some extent copyrightable although it kind of depends the level of abstraction you're talking about. And also the trained algorithm, and this is where it gets really tricky, could be considered a derivative of the inputs and the trained algorithm in that it requires information and it's pulling information about the creative aspects of those inputs and storing it in some sort of statistical model. So Jess is gonna talk a little bit about what that means in terms of copyright. Yeah, but one thing we know about the trained algorithm is that in any case, there is at least some thin layer of copyright in compilations of data. So that's just worth a note. So there's potentially copyright in every single one of these things, including the seed material. And I just, we loop that up clearly with deep dream. So deep dream, the deep dream would be, the inputs would be the animal faces. The learning algorithm would be the deep dream software that Google developed. The trained algorithm is all set up on the internet free for you to use and the seed material would be whatever photograph you put in and tell deep dream, okay, now try to find animal faces in this. Okay. Questions, I guess. So what we'll do next is move into thinking about possible models for copyright, how copyright could get involved with this. But questions about the basic architecture of this. Yeah. Is there a particular metaphor or way of explaining that you found was most efficient or explaining that? Yeah. Yeah. So I think essentially the thing that we found was the biggest piece was that, was pulling apart the trained algorithm and really emphasizing that it's a separate work from either the inputs or the learning algorithm. And I think that the sort of key error that most people had when they were approaching it is that depending on which side they were aligned with, they either thought of the learning algorithm and the trained algorithm as co-extensive or the inputs and the trained algorithm as co-extensive. And so with those things lumped together, it's actually a lot harder to write a license for one of these pieces of artwork. Cause you just spend a lot of time arguing over what it is that you're talking about. Okay, so this is a simplified version of that same chart that I think just makes it a little bit clearer as we start to add elements. So as I was just saying, the advantage of each of sort of breaking apart these pieces is that we can think about the ownership and licensing of each of these pieces individually with a lot more clarity once we sort of pull them apart into an anatomy like this. And it gives, in our experience, the parties as they're trying to reach a resolution more levers to pull because it gives them more flexibility, right? It's not as though you're trying to negotiate a license over who owns this entire system. You can actually sort of tease apart, right? It's gonna be clear that whoever created or has subsequently acquired the copyright in the inputs owns that. Whoever created or has subsequently acquired the copyright in the learning algorithm owns that. And then you just sort of come down to talking about the trained algorithm and the outputs that are created by the individual product. So however, it's relatively rare that we get to work with people that we do have some now at the stage before the trained algorithm has been trained and the outputs are being produced. Typically, it's after the fact. And so we wanted to talk a little bit about the different scenarios for copyright infringement when this happens in the absence of a license. So in the green on the left, you see what we call infringement by training. So when you scan or otherwise input copyrighted inputs through the learning algorithm to create the trained algorithm, that is typically going to be a violation of the reproduction, right? You're typically gonna be making a copy essentially of the inputs and it might go a couple of ways. It could be a transitory copy that's just made for the purpose of scanning and then all you're left with is a sort of relatively clean dataset or depending on how the algorithm is set up, either a full copy of the inputs or partial copies of the inputs, copies of certain areas could be left. So in infringement by training, you might have as few as one instance of copying or you might have very many and they might repeat. So that's infringement by training. Anything else to say about that? No, okay. Infringement by output is a separate category of the sort of major category of possible infringements. And this one we think more about the derivative work right than we do think about the reproduction right. Now the reproduction right could be caught here. And although this is on the right, we're still largely thinking about how the inputs would be infringed by the outputs. There, you could have a reproduction, violation of the reproduction right if say you had such a poorly performing algorithm that it just straight up reproduced an input as an output, that would certainly be copyright infringement. But even if you had a relatively well functioning algorithm, if you had one that was designed to capture a whole like small chunks of the input works, you could still have a violation of the reproduction right by comparing an input to an output that had reproduced one of those chunks in full. More likely though, if the developers of the system have put effort into it, what you're gonna see is not so much a reproduction of the inputs as a work that's very closely connected to them and something that might be considered a derivative work. And when we think about derivative works, the shorthand is sort of uses for which the original owner would have expected you to get a license. Now that's a little bit tricky in the context of a new technology. I don't think even a couple of years ago, anyone who was creating art was thinking about whether they would be licensing their work for the purpose of training machine learning algorithms. But certainly if the inputs and outputs look closely related and we'll see a couple of examples as we move into discussion, there's a chance of an infringement of the derivative work right. And this is also a scenario where there could be very, very numerous instances of infringement depending on how often the trained algorithm is run to produce an output but it could easily run into the thousands or even the millions. And because in copyright statutory damages go by each instance of infringement, this could get expensive for the copyright infringer very fast. And the other thing that I think is really interesting about this is the trained algorithm, once trained, can easily be copied. So I could train an algorithm on some inputs if I put that up on GitHub or something like that. Other people could download it and start violating copyright. Now who's responsible for that? Are they violating copyright? If they've done, am I liable for contributory infringement because I enabled their infringement by training on copyrighted data? It gets really complex when we get to tools that produce art or AIs that produce art and themselves are easily reproducible. So we wanna spend the remaining sort of 35 minutes of this lunch presenting you with a series of examples of this kind of work and encourage you to share your reactions and your questions with us as we do. So we have four examples and this is the first one. Do I assume we wanna look at the video? I think we should play the video, it's really cool. So this is a still from the video. Do you wanna go over what it is? So this is a system that has been trained to recognize different patterns. So water, fire, flowers, I think a few others. And what's really interesting about this is that it has been trained on these images. It's created a trained algorithm that can recognize these patterns and replace them in images. But as seed material, instead of using a static image, it uses video. So in this case, the transformation is happening in real time and so this raises some, I think you have to set it up to mirror here. I'll work on it while I talk. So this raises some really interesting questions and I know I said we would not get into the deep philosophical stuff but I can't help myself just a little bit about the nature of inspiration. If I observe something and I draw a copy of it and if what I'm observing is itself copyrighted that might be a violation and if what I'm observing is just like a natural landscape or something it probably wouldn't, I'll talk while this goes on in the background. It goes for a while. The copy, the representation that's in my mind in transit between my vision and me writing something down isn't a copy and this is something that has been recognized in other instances in copyright law. For example, in the Cable Vision case where a very transitory copy that was made temporarily for the purpose of providing cable access, which DVR services, which lasted for 1.2 seconds was considered to be too transitory to count as being fixed in a medium. So what happens when we have AIs that can read things in in real time and immediately convert them? Does this create a challenge for copyright because essentially these systems can act so quickly that maybe any copy they're making is transitory or ephemeral? So this is an art piece called Gloomy Sunday. And I'll skip a little bit ahead. Yeah, so this is an AI that's watching live this footage over here. And it's for a few instances. The first one was trained only on seascapes. So you could see it was interpreting that like tablecloth as seascapes here. It's been trained only on images of fire. I just think this is like the most amazing thing that's ever happened to a phone charger. You're only on sky. I think Jess really summed this up very well in that, well, one, that this would not happen without AI tools. And it transforms a phone charger, which is I think Jess called it the most quotidian image you could imagine into this incredibly beautiful seascape. And I think this is a great example because it shows how valuable these tools are, but also raises some questions about if this were being run on something that was copyrighted and not just a video of someone moving their keys around. Well, it could be copyrighted, right? So the artist who made this, he's doing the live video on the left. So he's gonna have the copyright in that, but we would have concerns about where he got these flower images, right? Like it's possible that he pulled the images of the sea and the sky and the flowers from the public domain or CC0 sources. But if he didn't, then we have questions about. There are, so filters, like Photoshop filters are generally treated as tools. So wouldn't be treated generally any different than? What do you observe? Wouldn't be treated any different than what? Wouldn't be treated any differently than any other sort of artistic tool. You would look at basically whether there was a modicum of creativity, whether there was substantial similarity between the input and the output. I mean, before any of that, you would look at whether the input was even copyrighted in the first place. I think that the difference here is that those filters are hand-coded. They're not trained on existing works. This was trained by feeding at a bunch of existing works, some of which may have been copyrighted, some of which may have not. And so there's this potential for copyright violation by way of training, as well as copyright by way of producing the output. Also think the different outputs raise different scenarios, right? Like I think the seascapes, it's very, and the sky and even the fire, it's very hard in any of those three scenarios to imagine that you might recognize an image of fire in the output, whereas the flower one I think is a little bit different where it's possible that someone could say, like, hey, that's my Daisy, right? I took a photo of the Daisy with that exact same lighting and it's being reproduced over that area where the person's hand is now. So in the sort of sea, fire, sky scenarios, I think we have questions about infringement by training, more so than infringement by output. Yeah. As a person who's not trained in art, why would I go to mass art? Why would I go to music and art? I mean, this is amazing what you can do. I mean, I have a leftover drawer where I can pull out stuff and do it. The problem I have with this is that if you go, I go to the Whitney Museum in New York and I see a Basquiat or a Schnabel or a Jackson Pollock or something, what do you do with, this is obvious, seascape, sky, things that you recognize, but what do you do with the area of modern art? And also, what do you do? How do you compare something like a Picasso, something that comes from the Sotheby's and Christie's with the art world, with something that you could make in an afternoon with a really cool algorithm? What weight does the art world give to AI versus like actual talent? Yeah. Yeah. That's the next thing I'm gonna do. The artist behind this is the Sky Memo Act and I think is doing a computer science PhD, but is also clearly a very creative person and interested in making art. He owns this algorithm, right? He has not made it freely available. So none of us are in a position now where we can feed in our own inputs to this. He has a number of videos on Vimeo and I encourage you to Google them and check out the other ones. But right now I think the way that you would see this in a museum is that the museum would acquire this from him, right? Would acquire one of the finished videos or acquire like a piece of performance art where he would do it live. But yeah, right now to use this algorithm without, we don't have access to it. If someone got access to it, to use the algorithm without its permission but also we need instance of copyright infringement. So I don't, I'm not aware that he's trained them on anything but images, but it would be totally fascinating to see if he trained them on a series of Basquiat paintings or a series of Basquiat paintings or something like that. There are groups that have done this. Mathieu Pesches group out of Tubingen has made Depart. You can go online at Depart.io. It's a very similar mechanism of style transfer. So this is, you can kind of divide up a lot of this work into style transfer and generative camps where style transfer takes the style of a learned set of images here like the flowers. You could do this, but all of these are public and on Depart and it renders new content in those styles like a filter. So here he's providing new content as opposed to an algorithm out putting poorly generative model works that don't require some sort of input to be rendered in a trained style. So two notes on that. One, if you go to Depart.io.io, you can make your own. You can put in both the style that you want and the image that you want. So fun things to play around. Then too, Sarah, will you introduce yourself? It's for the recording. Sure, yes. Everyone can. I'm Sarah, I'm a PhD student in computational neuroscience in MIT and I do computational work as well. Thanks. Welcome to Van Gogh. Thank you. Next one? Yeah. So this is a song called Daddy's Car that was written using the Flow Machines software. So this software that was developed by Sony, it is generative music producing AI. This particular case, so Flow Machines is really interesting, it is designed not to just be purely generative but also to assist composers in making music. So it's really intended for human AI collaborative art. So let's listen to it a little bit. I think it's actually gonna be from my computer so it's gonna be kind of quiet, apologize. Listened and processed the music and then also input the lyrics textually. Yeah. And so this song was written, the melody was composed by the AI. The lyrics were written by the AI and then a human arranged it and produced the recording. So this raises some very different questions than the previous one in that the inputs and the outputs are very recognizable. And here, this kind of question of derivative work from the input to the output stage as opposed to copying from the input to the trained algorithm, it becomes an additional layer. So Jess did you wanna talk a little bit about those questions? I mean, I think, so there's a number of sort of similar examples of AI, are generating AI systems that are trained on a single artist's worker, in this case a single group's work. And they raised some compelling questions, particularly in the aftermath of the decision in the Blurred Lines case, which probably many of you are familiar with but that's the case about the Robin Thicke and for a William Song Blurred Lines that the Marvin Gaye's airs brought a copyright infringement suit alleging that it infringed one of Marvin Gaye's number one billboard hits. And the second circuit decided in the Supreme Court denied certiorari in a decision that essentially, while there was significant dispute among the musicologists as to whether, I think that there was basically evidence that there was no single one note to a second note identical in the song, the overall look and feel, the jury was justified in finding that there was copyright infringement nonetheless. Because essentially it copied the style, the instrumentation, some sort of the movements of the themes in the songs. And so you can imagine that in a world where Blurred Lines is a copyright infringement of that Marvin Gaye song, that this could easily be considered a copyright infringement of one or more Beatles songs. The other thing that's really interesting is here, this was, so Flow Machine is a music AI that's developed by a subsidiary of Sony. So I think that the work in this case was licensed. And I'm not aware that any of the living Beatles participated in this, but you could also imagine a scenario where the artist whose input is training the AI was actually directly involved and interested in collaborating with the engineers to figure out, to sort of train an AI deliberately to reproduce their works. And we've been thinking about that scenario in particular as a really interesting case, a sort of paradigmatic case in many ways, for joint ownership of both the trained algorithm and the outputs if both the designers of the learning algorithm and the inputs are deliberately collaborating to create a new work. So we think that that is a sort of exciting possibility. Any questions or thoughts on this one? Yeah, philosophical sort of discussion, which is how do you conceptualize a difference between inspiration and influence when it comes to musical creation? Right, so under the law of copyright, neither one of those is a violation, right? There has to be copying, which means that there is enough similarity between the original work and the copy to say that it actually took wholesale creative elements from the original rather than just being kind of one in the corpus of works that maybe influenced an artist. So that's not a great answer because it's not a technical hard line. But I think this, sorry. The interesting thing though, so there's this elaborate test for infringement and copyright that the first question you ask is access, right? Did they have access to the work that's allegedly infringed? And then the second question is this substantial similarity test that Mason raises, which has both an extrinsic and an intrinsic element. The extrinsic one says break it all down, look at every piece, see where there might be infringement and the intrinsic one says, okay, let's juror as a general sense when you listen to these two pieces of music side by side, do you like have a gut feeling that there's infringement? And the interesting thing is that when you have an AI, at least if it is an explainable AI and if its creators have documented what material they trained it on, there's no question about access. And there's also potentially no question about the extrinsic test. You would just be able to have the AI output its answers to that, right? And so it's really interesting to think about what in what ways copyright law was designed around this conclusion that the human brain, the human creator is a black box and to what degree we've actually romanticized that notion and to what degree it even forms part of the foundation for like what is creative, right? Like whether the fact that an AI could output so I used like 13% of this piece and I used 100% of the drums from this piece if it could tell us all of that, would we come to the conclusion that it was not a creative act, right? Imagine on the kind of the counter example human with a photographic memory an artist who could say, who could recite to you every piece of work they had ever seen and tell you when they painted, yes it is 0.1% this work that I saw 25 years ago, yeah, would we value that person's artwork less? And sort of the percentages of? Probably not, because that's what I'm asking about. We don't ask them. Even if they could we wouldn't ask them to but this is copyright is based around this assumption that at a certain level of the there's a certain level of detailed information we just can't get and that's why we have these fuzzy tests about gut feelings. Now we have tools coming along that make that fuzziness extremely, extremely clear and this we'll talk about a little bit about this at the end of the presentation kind of calls into some of the basic presumptions about why we have the copyright doctrine at all. Or they make original works? That's training, that's. Yeah and certainly we don't if that training were extremely evident in an output that they claimed was their own was original and creative and the underlying work was copyrighted that would be a violation. This just. Yo hey painting in the style of the kind of stuff. Yeah and I mean there are many things that are well recognized, the students of famous artists are often highlighted for having similar but subtly different styles. And I've tried some of that before. Different. Yes. So let's say a human artist decides then off of the AI to be inspired by something but it's a little too similar. How deep does the copyright protection go then? So you're talking about how, how like deeper thin is the copyright protection in the work by an AI? Yeah exactly. Yeah. And then if someone conference on that then are they they're cranking on something to agree with them? So our position would be at least as of now that there are no, there really aren't any AI systems that we're aware of that are sort of equivalent to human authors. And there's a number of reasons that that's far out. The copyright office essentially we would say like still doesn't have to answer that question asked in 1965, like is there, what do we do about computer authored works? The biggest thing is that like no computer as yet experiences that like spark of inspiration. Computers only create at the behest of humans, right? You always have to press the like go button, make an output. But so right now all of these like AI authored works which are really human authored works with substantial AI assistants are just human authored works and they get the same level of copyright protection that a human authored work without AI assistants would get. So that may be if they're very creative that's like thick copyright protection, right? If they are more like a bare data set that's thin copyright protection. But then I think your second question was sort of like if the AI work was infringing of copyright could the copy, the like piece inspired by it be infringing not only of its copyright but like the ones back possibly, yeah. And I think also with this idea of AI assisted human generation, I think that might in the future complicate matters of the question of finding substantial similarity because if an AI suggests a tune and then a human takes that and arranges that and maybe a band plays that, you have an AI that maybe is a kind of a contributory infringer here. There was no intent and certainly no willfulness maybe on the human person but what if by looking under the hood the human could have seen that this was trained on a single just trained all on the Beatles, is that person now a willful infringer such that someone could get statutory damages? All right. Yeah, go ahead. You can also take the Google advertising algorithm as like sort of slightly manipulating all artists who use internet search. Yeah, right, interesting. Recommendation engine for like what kind of content you're seeing. Totally. So in the interest of time, I wanna shift us forward because we have two more examples and they're both so cool. So. For manipulating them. Or I think it would be all of society. It was interesting to think about like one suit before. Yeah, I don't think that would be copyrighted. It wouldn't be copyrighted for a different bit. I don't know, like broad. Broad, it's a little bit. You'd have to have harm. Unwanted influence. Yeah, some kind of consumer protection. Yeah. Sorry, I'm finding myself much more fascinated by the kind of quiddity of reproduction rather than inspiration, which is not what I expected at all. So how does it become, well could it be the same thing? I mean, Google is crawling the web all the time, right? And there is some analytic process that's taking place there. Can't wanna imagine, I'm just interested in this reproduction problem of the training data that ultimately produces an algorithm and the question of whether there's infringement in that. That inheres in that. And really interested in how suddenly reproduction seems like an impoverished concept for what's going on here. Like how, couldn't we imagine an example like the first example you share, in which programmatically what's going on in terms of the training data is really very similar computationally to what any web crawler does. Where does this become reproduction instead of some other kind of analytic experience? One way around that is to concede reproduction by saying that it's fair use anyway. So this is something that has happened with a lot of computationally created compilations of information is that courts have said whether it's transformative enough that creates a new enough purpose is true of indexing search results to say that it's fair use. So that's one answer. Yeah, I mean, and I think that's right. And I think it sort of holds true with this as well. And the big, in our sort of schema, I think the two ways you would think about it is like the daddy's car on the one hand where all the inputs are one artist and really the goal is to create an output that is essentially like not that transformative, right? Like you want a song that sounds like a Beatles song. And the other one is something like the gloomy Sunday piece where the goal is to create something that looks like a seascape but not any particular seascape, just a seascape or, and this one's not a great example, but sort of massively multiple artist input where the goal is to teach it like everything that, for instance, a cityscape could look like. And then it, this one doesn't have inputs like that, which is what's cool about it. But then it just, it doesn't reproduce any particular cityscape. We think with the massively multiple artist inputs, it's much more likely to be a fair use, at least as far as the reproduction. Yeah. Yes. Yeah. We have things like, say the blues is that some things that occurred like three, it's almost like original seeds, but like original music or original, I don't know, original anything that is just sort of in the common sphere. And then somebody does something and copyrights it. And then it's like Basmati rights. Now, if you want it, it's copyrighted by the GMOs, but artworks can potentially, that can potentially get done with music and just co-opting like indigenous or African or sometimes with music. And then all of a sudden you've got a copyright on something, it isn't even yours. So I think there are a couple of copyright doctrines that I think try to address that. So one is that if you can prove that something existed, has existed for more than 70 years or so on, the exact time is constantly changing. You can say it is actually in the public domain. The other is this idea of sans-offaires. There are certain themes, and this can be in literature, in painting, in music, that are foundational elements of a genre and can't be copyrighted. I think that you raise a good point about cultural barriers because what is an apparent sans-offaire to a judge who has read a lot of Shakespeare may not be apparent as sans-offaire if that judge hasn't read. Yeah, or I would say what hasn't read a lot of indigenous literature. Yeah, so there might be definitely kind of the cross-cultural barriers of big issues. I really wanna get these two other examples in. So let's do Painting Fool really quickly and then go to the next one. Okay, so Painting Fool was trained on... Painting Fool is the name of an algorithm. The name of an algorithm that was trained on hundreds of thousands of different pieces of artwork to learn very broad styles. So not like the style of Matisse, but impressionistic, charcoal, sketching, so on. And then it was also trained to recognize what city skylines look like. And so now what it can do is it can do automatic scene generation in that not only does it generate the style to transfer onto the scene, it generates the scene itself. So here we have an AI that is really in the classification of generative in that it can essentially output something not wholly creative, because you can't say draw me a field, it has to be a cityscape. But it can do a cityscape, it can do it in multiple styles. And it's very hard now to trace that back to an individual human spark of creativity. So this is one of, I think, what we might call sort of a hard case if you wanted to produce copying or an easy case if you wanted to say there is no copying, that this is all fair use. Yeah, so I'm gonna say let's set this one up and the next one up and then we'll finish up this question. This is where I have to look up this video. The first step was to study the works of Rembrandt in order to create an extensive database. We gather the data from his collection of paintings from many different sources, including 3D scans and upscale images using a deep learning algorithm. Because a significant percentage of Rembrandt paintings were portraits, we analyzed the demography of the faces in these paintings looking at factors like gender, age, and head direction. The data led us to the conclusion that the subject should be a portrait of a Caucasian male with facial hair between 30 and 40 years old and dark clothing with a collar, wearing a hat, and facing to the right. From there, we started to extract features only with faces that were related to that specific profile. And we had to create a whole painting from just data and we used statistical analysis and Furry's algorithms to extract the features that make Rembrandt Rembrandt. We took parts of the face and we started to compare them. And then based on this, we're able to create a typical Rembrandt eye, or nose, or mouth, or ear. After generating the features, we were focusing on the face proportions. We used an algorithm that can detect over 60 points in a painting. We were able to align the faces and to estimate the distance between the eyes, the nose, and the mouth, and the ears. A painting is not a 2D picture. It's 3D. You can see the canvas, you can see the brushes, and that's what makes the painting come alive. A hive map is essential to make the painting a painting. We incorporated the height map into the painting and printed on a 3D printer that uses a special paint base UV ink. It printed many layers, one on top of the other, which resulted in the height and texture of the final painting. Sometimes a magical moment to see a painting for the first time. Even if it's computer generated, for me, it is something special. I would have believed if I would saw it in a museum that it would have been a real Rembrandt, just what I haven't seen before. We'll be interested to see Rembrandt looking at it. We will be happy that there are people trying to understand him and trying to create something out of that, so I think he will be happy. The next Rembrandt makes you think about where innovation can take us. What's next? The box concept of how creativity is done is romanticized. I think we're also seeing here another kind of romanticism for the idea that we could, if only we were able to measure at a high enough precision the world, that we would be able to predict the future or create new things that are the same. I think while we problematize the sort of black box romanticism, we might also want to be problematizing this concept. I totally agree, and I had never thought of that when I was watching this video before, but just until right now, the people in this video were like, we looked at the data, and it led us to paint a picture of ourselves. A bunch of white males between 30 and 40 decided to paint a white male between 30 and 40 on the basis of the data. I'd love to know more about this whole black box romanticizing thing, because as a singer or songwriter myself, I think that beyond the extensive inspiration that I got from 28 years of listening to all kinds of songs that I have no reference of, there is also a way of interpreting life through music that is described necessarily, but if you tell me to write a song that's a Beatles, I would do that, but if you tell me to write a happy song or a sad song, or write something about grief or something about love, that comes back to how I interpret life through notes and through music. Is there a way to get, because I know that some robots now have feelings and stuff, so is there a bringing together? I feel like they're starting to humanize them as much as possible. The painting fool actually learned to detect, that's the one, to detect emotion. There's a very good paper called Automatic Scene Generation with Intent, and then the follow-up paper, Automatic Scene Generation with More Intent, which is about, they hooked the painting fool up to a facial recognition algorithm that could detect whether people were happy or sad, and then it would paint it in a happy kind of cartoony features if the person was detected as being happy with bright colors, a more blue impressionist kind of painting if it detected that people were sad. Now that, of course, is an algorithm that was told by human trainers, this is what a happy person looks like, this is what a sad person looks like. We're still extremely far away, both from humans having feelings. What we're actually doing is we're projecting a lot of our feelings onto robots. Robots don't have emotions by any stretch, I would say. But I do think, you also raised a good point, which is I could train an AI on hundreds of thousands of pieces of art, more than I could ever see in my lifetime, but it would still only know art, it wouldn't have any other experiences, and we're very far away from what we call AGI, so artificial general intelligences, which could expand beyond a very specific, narrow domain that they're looking at, and learn from multiple different experiences and actually start to be spontaneous. Anybody we haven't heard from yet have questions? I wonder if there are any photographs, other photographers, photographs, and then put them in the museum as we're all in. I feel like there was discussion around whether that was space-sizing, and the things that we're seeing, actually, if you visit a fly now, that's the way you think about it. Is that kind of gesture of appropriation considered? I think it's an interesting point to raise. I think that as far as I know, I haven't seen a project like this that sort of intent, takes the posture of appropriation or intending to very closely approach its inputs. In fact, generally I think the intent in these systems and the definition of success on the part of their creators is to achieve some distance from the inputs. I think there are similar pieces of art, though, where, for example, the intent might be to, for instance, create discussion about the relationship between human and artificial creativity, and in that case, much like the satire and parody cases under copyright, you need a degree of recognizability in order for that to be a meaningful discussion. So I think that that could be... I think that is happening with some artists. Yeah, actually, that's a good example of that as the very first example we showed you of the water. Memo Octin says he calls the series Learning to See, because his point is like a political one about the AI only... like, when it looks at the... like, phone charger, whatever, it only knows about water, so it only sees water. And so he thinks similarly, like, as people, if our background is whatever, we only see ourselves in the next room. Yeah. I mean, I think that what you just showed us is actually very much... you know, it's not in the line of like, appropriation that it is in the line of... I mean, the fact that they call it creating the next rember is, in fact, discerning that we can express by people who are no longer alive and that... I think it, you know, obviously, this point Rembrandt's words are a must... I mean, I would think they're in the public domain. So I don't know about that. But, like, I guess, you know, and so one would say, it would be interesting, I think, to hear from you guys about how taking, you know, why would you say that this is, hey, the next Rembrandt? Where does the, you know, ownership fall for this if all the images are in the public domain? But ING commissioned these people to make this. And also, what, you know, what does this... you know, when you have, you know, even if we say, okay, it's all in the public domain, this is a... we've created new painting, how does that relate to, like, you know, as we think about all the the sort of struggles to authenticate original works and then juxtaposing with the ability by certain entities to create in this, you know, other than, I guess, by timing and dating of the paint and that sort of thing, distinguishable, stylistic work. Well, I don't, so I think so a couple of things about that. We use their next Rembrandt example in part because I think that that video shows you really well. I mean, there is, like, machine learning involved here, but it also shows you how intensely hands-on it was for the people, right? Like, there are a lot of people involved there making a lot of the key decisions, right? The AI did not decide to paint a white male in his 30s to 40s with a hat and a collar. People decided that. And then they worked very incrementally with the AI to, like, settle on an eye, settle on a mouth, decide the distance between those things. So that's part of why we show that example. So I think in that instance, the copyright is going to go to those people unless, of course, it's a work made for hire for ING, in which case ING gets the copyright. But then sorry, should we just... Oh, I just know we're, like, pretty close. I just wanted to get, like, a couple of people haven't had a chance to ask a question yet. I teach digital fabrication and my students are super interested in 3D scanning until they become reproductions. They're just not interested anymore. And I think a lot of what we're talking about here is kind of early in terms of copyright, because I think it's ultimately not that interesting for artists to reproduce things. It might be interesting to study, as we said earlier, to make reproductions to study a style or how you might technically do something. But I do find that this idea of seed material most of my students aren't grabbing things that are already available. They're way more interesting creating something original and then using the algorithm as some kind of tool. I think this fascination with the gadgets soon becomes very old. And I think this is big fear in art school too of this so-called democratization of creativity, where anyone can write music, anyone can make art. But I think at the end of the day, if there's no context, it's like data just remains data, only becomes information when it's contextualized. I think the same with art. Good art is something that has been contextualized and has actually some kind of meaning. And just the last thing I wanted to add is that I think there's a real gap in digital literacy and other people that are using these tools are very educated in programming and they can see behind the scenes what's going on, but for an artist it's a very high learning curve, so they're trying to use this as some kind of tool. And I find that most of my students are just trying to hack the system anyway. So it's sort of interesting to see what they're doing with this technology as opposed to what computer science, PhD people are doing or engineers. It's just a very different world. I think that's a great perspective and also sort of connects to how I was going to answer your second question, which is that when we see a really low-res image of that next Rembrandt I think to us, the whole video is designed with the music, whatever, to make you go like, ooh, oh my god, it's indistinguishable from all those other tiny fuzzy images of Rembrandt's that we saw. But I think that if you or your students or art historians looked at that, I think it's a very open question whether people who truly care about Rembrandt would find that painting to be compelling, right? Because, you know, just because you can doesn't mean you should. Right, well also just because the lighting is similar is the average of other lighting conditions in Rembrandt paintings doesn't mean that it has the same emotive affect on us and just because they reprint 3D-painted the canvas doesn't mean that they actually got it right about where the buildup should be and where the brush strokes should be. So I think it's a sort of open question as to whether and I think the question it's an open question but the answer is likely no if somebody who is a true scholar of Rembrandt would look at that and say it's like anywhere even close. One thing that then it becomes I mean there's no way they can because of the carbon dating but what if it's a modern artist instead of Rembrandt and what you have is a poor jury engine? Well I think that I mean right now so a lot of the things we've looked at I think all the examples we've looked at today I would classify largely under the idea of being their tech demos these are not being used they're not being put forward as we're making art they're being put forward as we're testing out techniques we're showing what is possible we're learning about the process of machine learning and so I agree with you 100% that there isn't a lot of interest as like kind of like art qua art here I do think they perform some of the pieces we've seen and I think the painting fool gets into this a bit and certainly this discussion about the next Rembrandt I think it provokes conversations about the relationship of art and you know as tools become easier to use what is there like a kind of like a smaller degree of creativity that is needed to create these works or is it just like a smaller amount of effort for the same amount of creativity and if so what do we feel about that? It's going to be reproducing as Steve Wright comes and you know tax the system so Should we take one last question maybe? It's Rachel's birthday Rachel's birthday Being able to take pictures of anything actually when photography first became widely available there was a question about whether photographs could ever be copyrighted because they said a human is not generating the art all the human is doing is choosing the subject And if a copyright could exist in a photograph who would own it like would it be the camera that question was really asked It seems like it seems kind of silly looking back of course like photography is a very creative art and people on the covering photography but yeah that was that was a question that had to be decided in the courts so I think yes we do have precedent for this kind of I don't want to call it a crisis because I think that overstates it for this kind of struggle with with applying copyrights in new technology Awesome Well thank you all so much for coming