Undertexter av Nicolai Winther Dramatisk musikk so Thank you. Hi everyone and welcome back. My name is Victoria Szabo and I'll be chairing this second session in the art track. Our theme is about authorship. It's a topic that's already come up a lot throughout the day and we look forward to hearing from all of the speakers on this subject. Our first paper is called Sentience, a Performative Game on Tarot-Stealed Language Models and the Ambiguous First Person. The speaker today is Marcus Round. Marcus Round is a freelance creative technologist and a computational artist whose practice explores language and agency in the age of AI. He defamiliarizes language games into absurd, immersive installations that invite participants to confront the mechanics of linguistic choice and responsibility when mediated by AI. His work has been exhibited at the V&A in London and TIAT in San Francisco. Please join me in welcoming Marcus Round. Hello. Hi. Hi, everyone. I'm Marcus Round.. Hello. Hi. Hi, everyone. I'm Marcus Round.. Okay. So, hi, everyone. I'm Marcus Round. I recently graduated with a masters in computational art from Goldsmiths University in London and Currently looking for a PhD. So if you know of one that's suitable then please do you talk to me? But this project is a Immersive performance game is what I called it I started around the question of when ChatGPT says I, what does it mean? And that's less a question about ChatGPT than it is about the word I. And what does it mean? What does it mean when you say I? What does it mean when I say I? And I think that's a really interesting question. a question about ChatGPT than it is about the word I. And what does it mean? What does it mean when you say I? What does it mean when I say I? And I use large language models as a tool to kind of investigate that. That is what they are. They're a computational kind of tool for analysing language. And there's a whole suite of tools that have emerged in the field of what's called mechanistic interpretability, which is when these AI labs have sort of created these tools through a blind kind of machine learning process. And the sort of later step is to then figure out what exactly is going on in its quote unquote mind. And you can do all sorts of analysis. So you can give it a piece of text and sort of measure which patterns of neurons within the algorithm fire most strongly when it encounters certain words as an example this is a graph of what happens one of the many patterns that fires when it encounters the word I. This is from a tool called neuronpedia.org. Definitely look that up. I always give them a shout out. It's a great site. So this is Sentience. I'm going to see if I can play the video. Here we go. Can I play the video? Here we go. So it's a performance game for an audience of two. I'm there as the performer, and we have two participants come in. It lasts about 20 minutes. They sit down, I do this kind of performance. They make some decisions, play this game, experience this thing, and then they stand up and leave. The next two people come in, and I do that over and over and over again for an entire weekend. And you can see I sort of look like this ambiguous poker dealer, bureaucrat, tarot reader figure. What's going on is that I hand them cards. And it looks a bit like they're playing poker, but they're playing cards one at a time onto the table. And each of these it looks a bit like they're playing poker, but they're playing cards one at a time onto the table. And each of these cards has a little snippet of language that I authored. I wrote this, like I wrote each of these phrases and they're all very personal things about like my future, what I'm afraid of, what I didn't do, what I want to do, what I hope and my family and all this kind of thing. And they take turns, between the two of them, to place a card onto the table and construct this sentence. So that it becomes this very personal... So if you look at this example, I'm not ready yet, but I trust myself, and I want to try because I want more, and I am brave. That's a sentence that could be constructed by the two participants. It becomes a question then of, like, who is this I that's in there? Because it's been constructed between two people using cards that I've authored. And sometimes these sentences that are played to the table seem quite profound or quite meaningful, and sometimes they can be quite meaningful. I've had people come up to me afterwards and say it brought up memories of their brother that had passed on or something like this because it happened to mention that. But it is at the end of the day just sort of like a random chance with a little bit of human agency sprinkled in, constrained by language and by things like that. So I also have these poker chips you might have seen. I call them tokens because that's like a pun about AI. But I just have this massive deck of cards. Well, it's 52 cards, exactly. And you can see that they're different when upright versus upside down. And that is a reference to tarot. There's another part of this thing. They write this sentence, and then they send it off to AI for autocompletion. It's this sort of satirical performance thing where I get them in and I say, you're going to write a personal affirmation about your deepest desires, but you have to do it with a partner, and I'm going to get an AI to autocomplete it for you. And before the autocompletion is written, they have to draw another card, which is one of these cards. And they might draw the chariot, which you can see here, upside down. And then you've got a bunch of these tarot cards. They don't have imagery. They just have text. And you can see that they're grouped around different themes. Like if you look at the words that are highlighted in this chariot upside down, you've got dispute, defeat, litigation, sort out, dispute, fighting, defeat. Like there's a theme there. You might not exactly be able to pin down what it is, but there is that you can sort of gradually get this vague kind of theme. is that you can sort of gradually get this vague kind of theme. And that is related to what the chariot reversed is ostensibly about. It's sort of divinatory meaning. The way that I did that was by feeding the... A.E. Waite wrote this sort of canonical book about the divinatory meaning of each of the major arcana, and I fed each of those into this tool and extracted the features that fire in the AI's mind when it reads that text. So there'll be a feature to do with self-sacrifice or a feature to do with divination and prophecies of the future. or a feature to do with divination and prophecies of the future. And I then, by extracting that, like that's the hanged man when it's upright. And so when the participants draw that particular card, I then take those features and I dial that up in the model. So the prompt doesn't change. It's actually a very, very simple prompt. It's not about prompt engineering or anything like that. It's a very simple prompt, simple text completion. They write the first half of a sentence, and I just get the AI to complete that text. That's the basic function of a large language model. But sort of almost without the model knowing, if you want to think about it like that, I've actually manipulated its weights by turning up the themes and references related to self-sacrifice so it becomes more relevant or more probable to mention things about that as it completes their sentence and that will change depending on which tarot card they drew. We do that three times so the two participants go back and forth to write one first half of a sentence. The AI completes the second half, then they write the first half, second half, first half, second half, until you've got three full sentences, and that is their sort of like affirmation that they've authored. It becomes, I haven't sort of gotten into the whole dramaturgy and theater of it, because that's sort of not what this particular paper is about, but it's quite a satirical, bureaucratic process that's going on. That's my operator interface, which the participants never see, but just to give you the idea of what's going on, this is like, you can see the sort of sentences that they've written at the top. I think these are probably random examples, so don't read too much into this particular example. But you can just see that I'm there typing in what they've written, and then I pick the card that they've drawn, and it gives you the features on the top right that are to do with that card. And then I send that off to the Neuronpedia API and I get a response back. That is a little, that's sort of presented to the audience outside of this performance space that I'm in. There's a little space to sort of show the last one. And you can then go and see which features activated most strongly in the actual affirmation that they authored. So a little bit of interactive exhibition space to it. The, there's a set, so tarot is kind of, it comes up a lot in computational art and could be somewhat cliched maybe, but I think it's very relevant to this piece. I think, I mean, the fact that these tarot cards just have text on them and not imagery is sort of, it's really interesting to think about these models whose only perception of the world is through text. They have no embodied kind of perception of the world. They only experience the world through text. And just sort of the way it steers the conversation between a reader and a readee through these kind of loosely grouped themes, I thought made it quite relevant to this Tool this is called sparse auto encoder is the name of this thing that lets you sort of draw out the particular themes in text and then dial them up in the model I'll just finish by mentioning one other part which is Before I get the you remember they have three auto completions that the AI will give them, but before they can receive the auto completion, I say, I'm sorry, I'm in character. And I say, I'm sorry, but we have a very strict security policy in place here. And before you can get your auto completion, I just need you to prove to me that you're human. And which is kind of started as like a funny kind of thing that they have to prove to me that you're human. Which started as a funny kind of thing that they have to prove to me that they're human before the AI will auto-generate their personal statement. But it became part of the dramaturgy of this immersive piece where, first of all, I hand them this capture, which instead of traffic lights it has street signs with words on them and I ask you to pick the words that remind you of your mother which is like a reference to Blade Runner and be sort of again the difference between machine vision and language and like different people's like projecting their own associations onto language and Then there's other ones like I give them a book, which is just a table of numbers, and I get them to read a number to me as if it's like a sort of confirmation code. But then the final one, the sort of climax of the piece before they can get to the final generation, I hand them a book, and I've highlighted a single piece of text, I am afraid of dying, and they have to look me in the eye. I say, you know, you've got to, I'm sorry, but the system says you've got to prove to me that you're human. So I give them this book, which is Everything is Illuminated by Jonathan Safran Foer, and this is like a fictional character saying this, written by an author. I hand it to them and they have to look me in the eye and read the words out loud. So I've had like hundreds of people look me in the eye and read the words out loud. So I've had like hundreds of people look me in the eye and tell me I'm afraid of dying. And it's just very interesting that eye that's there every time they say I and I'm looking and they're looking me in the eye. I think that that's there's something very interesting in that that is not captured by this piece of analysis here. But that will end my presentation, thank you. I think we have time for maybe a question, if anyone has a question. Yes, in the back please. Oh, please. Does it work? Yes. Okay, so I was wondering if you personally believe in this tarot thing. Do you feel like there is meaning behind that, or were you just interested in this entertainment kind of part of the project? I would say that I think... I'm not a big tarot person, but I've been to a couple of tarot readings, and on at least one occasion, the reader has sort of said, like, you know, there's nothing mystical going on here. The point of this is not that God is speaking from on high. It's that we are drawing random cards, and that you and I, as the reader and the reedy, are sort of drawing our own meaning out from this process. And I think that's quite interesting, and I think that's what's going into this piece. I think this piece is not about tarot as much as it's about language models and language, the human use of language, and tarot is just a tool to kind of get there. to kind of get there. But it's definitely about just the human projection of meaning onto a kind of random or semi-random process and how that relates to AI when you're speaking to ChatGPT. That's the kind of semi-random process that you're conversing with and you're still projecting a lot of meaning onto that conversation. So I think that's what was relevant there. Thank you. Quick related question. Do you think that if you'd used a different divination system that you would yield significantly different results? That's an interesting question. I think tarot was a good one because it is so familiar to people. I think that's why it's used so much is that people, in the piece, I never mentioned tarot. I never used the word tarot, but people can kind of see what's going on. I think that it's because it's about a conversation as well. Like tarot is very much you have a reader and a readie, and it's that kind of one-to-one conversation, but done through what ends up being like a very bureaucratic sort of office desk persona slash casino gambler thing. Yeah, the cards was an important part of it for sure because I related it then to poker and the sort of random shuffling of cards, things like that. Yeah. All right, great. Thank you. Thank you. All right, our next talk is called Vertigo, Vertigo, Reconstructing a Cinematic Ideal Through Its Predictive AI Double. And our speaker is Adam Cole. Adam Cole deals with the complexities of intimacy and identity in the age of AI through works that explore desire in the shadow of artificial representation. Okay. I printed my notes, which felt very retro, but I did not think I'd also be holding a microphone. So we'll make it work. All right. Here we go. I'm very excited to present my project, Vertigo Vertigo, which is a scene-for-scene AI reconstruction of Hitchcock's original film using only 2.78% of the film's original frames. As was just said, my name is Adam Cole. I'm a researcher and lecturer at the Creative Computing Institute in London. And my work does explore desire in the shadow of artificial representation. So a little interactive experience. Raise your hand if you've seen Vertigo before. Okay. Nice. Well, spoiler alert, I'm going to give you a little bit of the plot. It's been out for about 70 years, so you've had your time. In this film, a retired detective becomes obsessed with a woman he's hired to follow. And after her apparent death, he attempts to remold this new acquaintance into the exact image of this lost love. But it turns out this image was entirely artificial to begin with. In one of the final scenes, he demands this woman perform his fantasy for him. He says, I need you to be Madeline for a while. This fixation on constructing a desired ideal finds a modern parallel in our relationship to prompt-based AI tools, right? Where we might literally type into a text box, I need you to be. into a text box I need you to beat. So Vertigo is itself a film about an obsessive reconstruction of an artificial ideal. Vertigo Vertigo extends this logic to the material of the film itself, treating the canonical text as a probe to understand the normative conventions of classical cinema encoded in the model. So why did we choose Vertigo, right? We could have chosen many other films for this experiment. There's three key reasons. The first has to do with this idea of the cinematic ideal. Vertigo is considered by many to be the absolute best movie ever made. This is not necessarily my opinion. It was just ranked number one and number two in the side and sound polls. It also is very self-aware within the film about this idea of desire and this attempt to achieve this artificial ideal that is totally untethered to reality. The second reason it was chosen is its relevance to this idea of normative gender representation and the fact that it was famously attacked by Laura Mulvey in her initial articulation of the male gaze. And the third reason it was chosen is the title sequence developed by the artist John Whitney which is one of the absolute first uses of computer-generated imagery in a feature film. So the triangulation of these three kind of ideas, the cinematic ideal, normative representation, the computational image, forms the pillars that make the foundation of this project. So what did we actually do? We extract keyframes from the original film at every hard cut or every five to six seconds for very long takes. We then use a large open source model, 14 billion parameter WAN image to video model, to do first, last frame interpolation between those extracted keyframes. The reason we use this kind of process is that it isolates the model's predictive tendencies. It doesn't require semantic injection, there's no scene specific prompt that we use from a close-up compared to a long shot, we kind of just are interested in what happens in the spaces in between. So we reconstructed the entire film, and then we had to create some way to evaluate what did we make here? And there's three ways that we did that. The first is a computational analysis of the structural difference, the perceptual difference, temporal difference between the original and the reconstruction. We also evaluate this new film kind of qualitatively through a close reading to understand where the model fails, where it succeeds, and what happens in between. And our final form of analysis was enlisting sort of expert media theorists to get their opinion on what this reconstruction really means. So quantitatively, what did we find? We kind of created five tiers based on both pixel-based and perceptual measures. And we saw the following results. About 16% of the frames are nearly identical. Here you kind of see the difference is brighter and the darker parts are identical. And about 3.6% are severe failure. But what's interesting is that the majority of frames are somewhere in the middle, right? It's this plausible middle that is perceptibly like Vertigo, but not exactly the same as the original. But this aggregate number doesn't really tell us what it feels like to watch it. And so we wanted to understand where the model fails and where the model succeeds. So one of the places the model fails is that these AI reconstructions, they don't talk. And they also don't necessarily emote in the same way that their original character does. We also find these big spikes of divergences in scenes that deviate from conventional narrative film. So the opening sequence created by John Whitney, this sort of psychedelic dream sequence that the main character has, these have big spikes in their divergence from the prediction. What was surprising is the scenes that also succeeded very well. And these include scenes of incredible complexity, like this great tracking zoom-in shot of tons of characters in this really ornate kitchen with waiters crossing the screen. And the model was actually surprisingly capable of matching the original really closely. So what we find is that there's no correlation between interpolation distance and divergence depth. So it doesn't matter how many frames you're interpolating necessarily. We find that it's really content that governs the fidelity. Certain images are easier to predict than others. Aesthetically, the most fascinating scenes are not the ones that are catastrophic failures and it's not the ones that are perfect successes. It's these moments that are near perfect. This plausible middle that feels like it could have been a possible take that simply didn't make it into the film. In these moments, it is really hard to tell which is the original, which is the reconstruction, and it produces this legitimate doubt in the viewer between the two, this felt discorrelation that creates the aesthetic core of the work. So I said we'll come back to our three pillars. Let's talk a little bit about this cinematic ideal. After consulting several media theorists, they were both, they were all kind of in agreement that the intervention was legible, and they were all somewhat surprised about how neat the reconstruction was, how closely it tied to the original. There's two readings that we have as to what might be happening here, and they're not necessarily mutually exclusive. The first is that perhaps Vertigo is well represented in the training set, right? Perhaps it is compressed in the the model itself and this is a form of lossy retrieval for the original film. Another potential reading is that maybe Hollywood is statistically predictable, right? Maybe this film is not so original and that it's easy to predict it based on a consistent grammar that the model is trained on from a larger corpus of large video samples. Regardless of which one of these is true, it's important to recognize that if such a classic canonical film can be approximated from such a small sample set, then these questions that have arisen already from the 2023 SAG-AFTRA strike about likeness, about ownership, about consent, are dramatically acute and certainly getting more relevant by the day. Let's talk about our second pillar, which has to do with normative representations. While Vertigo is often considered this paragon of classical cinema, it's also one of the defining examples that Laura Mulvey uses to articulate her ideas about the male gaze, which is a really compelling formulation to reconsider in the age of machine vision. This is, of course, because machines do not actually have a gaze, right? There's no meaningful subject from which an AI model is looking. But what it does have is a statistical disposition towards certain images. And those images that it is most likely to produce are the ones that have been repeated by cinema time and time again. Whether it's the close-up of the blonde woman in soft focus, the heroic male lead, the damsel in distress, of course all these cliches play out in Vertigo quite beautifully. This reconstruction's pattern of successes and failures therefore suggests some inheritance, right? The point is not that the model was intentionally taught to look in a particular way. It's that certain ways of looking are simply more statistically likely to emerge. Ways of looking that align with this tradition of the male gaze. This is interesting to think about because the rhetoric that surrounds these models is this promise that we can produce any image that we desire. If we consider this inheritance, we have to consider the possibility that we will require some form of active resistance to flow against this normative current. All right. And finally, let's talk a little bit about the computational image. The paper claims about the present are expressed most effectively through the experience of actually watching the work. And here it's in its sort of difference mode where brighter pixels represent bigger divergences between the original and the reconstruction. The original and its double rhythmically diverge and then harmonize at every keyframe without the viewer necessarily being able to say in the moment which is which. And here's where someone's memory of the film and the prediction of the film become difficult to separate. But it's important for me to stress that I don't think we are entering some new world of unreality, right? We have been inundated with fictional images from when this movie came out in 1958 to 50 years before that when cinema got started, really all the way back to the arrival of mechanically reproducible images. What is new is this rise of uniquely predictive environments where content is algorithmically generated, personalized, and served at speeds beyond perception. One of the media theorists I spoke to said that cinema has always had this idea of compression, right? Cell animation, MPEG compression. Vertigo Vertigo fits that lineage not as a paradigm shift but as an acceleration, a confluence of speed, desire, and doubt that will reshape contemporary experience. Okay, in conclusion, Vertigo Vertigo shows that classical cinema's visual and structural norms are, in our opinion, remarkably encoded within contemporary generative systems. Quantitatively, we found that about 73% of frames were recognizably tethered to the original film, with failures really kind of clustered around scenes that depart from cinematic convention. Qualitatively, we explored this plausible middle between a copy and a hallucination. We could say that the experience of watching it produces its own 21st century vertigo, a felt discorrelation between the canonical original and its algorithmically predicted shadow, an experience that I think we're all getting quite familiar with as we're inundated with these AI doubles more and more. Rather than marking a rupture, the project argues that generative media extends and accelerates the historical logic of cinema itself, its voyeuristic gaze and its desire of perfecting an artificial ideal. The problem is not whether these images are real or artificial, but how these inherited structures are being produced, scaled and naturalized within predictive media environments at a rate beyond the grasp of our perception. Thank you very much. We have time for a couple of questions. Just one question or so maybe like a comment on the cemented decision you make while your research because you use the word like total failure when addressing frames that are highly, let's say, differentiated. And I was just wondering what's your view on that, because it might be that this differentiation created by the degenerative system are both on the one hand like the creativity, let's say, of those systems, and also as you from the practitioner who built those systems. Like, don't you feel that the more it diverges, it should be giving a more positive notation to that? Yeah, that's a really good question. I would say that our intention was to measure the difference between the original and the reconstruction. So we set up a system where failure really just meant more divergent and identical meant more identical, more similar. That's not a value judgment about which one is better or which one is preferred. And you're exactly right that it's predictability, it's tetheredness to convention is actually a negative if you're trying to make something that is not just repeating the past. So I wouldn't say my goal is to make Vertigo perfectly. My dream is not to get a model that is 100% perfect. We found this to be an effective tool to understand to what degree are these conventions really embedded in the model and what can we understand about how concrete they are? Thanks for the presentation. It's very interesting. I'm curious. I'm not really interested in which model specifically it is. But the fact that it's not specified presents it as if it's kind of like the models do that. And so I was wondering if you experimented with other models and had other outcomes, or if to some extent you could train or retrain some models to requalify some observations you had as to, let's say, the protagonist's reaction in some certain scenes? Yeah, so for this experiment, all the numbers are based on one single model, and that model is WAN 2.2, which came out probably about a year ago. We didn't do a formal test with other models, but I did try earlier models to see, like, has something changed? And there is a remark, it was remarkable how well the reconstruction worked. I mean, I didn't, the way this project started is I just tried one scene, obviously. I didn't do the full film to begin, and I really was quite shocked about how well it worked, and I was like, has this always been the case? And it's not true. So we don't know. It could just be the models got bigger and better. It could be that these open source models are training on more copyrighted material that aligns with classical cinema from the mid-century. And it would 100% be interesting to try it with other models to see if they behave very differently, and especially with commercial models, where bigger companies are very opaque about where their training data comes from. I mean, even open source models are very opaque about where their training model data comes from. I mean, even open source models are very opaque about where their training model data comes from. So it could be an interesting study to see if they behave differently or not. Hi, thanks for that. So I have two questions for you. One is, so you showed us the differentiated images overlaid, but you made this decision not to show us the two scenes by themselves to compare. So I'm curious how the generated scene holds up in terms of how natural it looks, the movements, the facial expressions, etc. And then also, you know, are you considering trying this out on a less canonical film to see how that works outside? Yeah, so for the first one, the way that we kind of shared it with external researchers to get their feedback was one of them side-by-side, one of them as the overlay, and one of them as the difference because I think you're right they all give a different experience of what you can take away. I would say the side-by-side being more analytical and the overlay kind of being more aesthetic. So in our research, we absolutely looked at them side-by-side as well. I think trying it on another film would be really, really rewarding. I think one of the researchers recommended trying it with Cassavetes films. What was the other one he mentioned? Basically trying it with more idiosyncratic filmmakers to see if there is a divergence or if it is sort of just, you know, if you have two keyframes, there's a predictable trend from one to the next. So that definitely would be an interesting follow-up project. Thanks. Yeah, just speaking of, like, trying another film, this idea struck me that maybe, like, in that model, well, definitely, like, vertigo is definitely one of like training materials for that specific model maybe when you are trying to recreate vertigo with that model it's like retracing the past in its memory so i would be really curious to see how it will behave or perform uh to recreate a film that's totally original. It's definitely not in the training set. That would be something I really want to see. Yeah, I think that's absolutely true. I think there's something interesting that a lot of these models are advertising themselves as increasingly cinematic. That's sort of when a new model comes out, that's the metric that they are interested in. But I think that the convention that they are interested in but I think that the convention that they are implying by cinematic is actually quite limited so I think being able to prove that sort of in a creative way would be a worthwhile endeavor Our next speaker unfortunately is not able to be physically here. It's Tomas Lorenzo. He's an associate professor at the University of Colorado Boulder, and the talk is called Speculative Pasts, Generative AI and the Political Imagination of History. Hello, thank you very much for coming to this talk that is titled Speculative Pasts, Generative AI and the Political Imagination of History. And very sorry I cannot be there in person. My name is Tomas Laurenzo. I am an associate professor at the Department of Critical Media Practices at the University of Colorado Boulder. And what I'm presenting here is Speculative Past, which is an ongoing research and artistic program that uses generative artificial intelligence to construct images and videos of histories that did not, could not, or cannot fully appear in the archive. The program approaches the past, not as a stable object to be recovered, but as a contested field of effect, ideology, omission, and technical mediation. AI-generated media can operate as a critical apparatus for the production of speculative memory, not the falsification of history, nor the simulation of documentary evidence, but the construction of plausible images that make visible the instability through which memory, city and political imagination are produced. General systems within this program are seen as instruments of estrangement, whose errors, excesses, recognizable biases and capacity for historical pastiche are not simply problems to be corrected but material conditions through which the politics of representation can be examined. Speculative systems therefore asked how synthetic images may help articulate memories marked by, among other things, dictatorship, interrupted modernization, artistic inheritance, and the exhaustion of official narratives, while remaining ethically distinct from claims of historical verification. So I'm gonna now present four artworks created within the program, and the first one is from 2024, and it's titled Abandoned Fictions. An aim of Uruguayan on abandoned fictions reflects on Uruguay's mid-20th century industrialisation, and specifically on the import substitution model preserved under Luis Valleveros, president in the mid-50s, which produced real social progress before its industries declined under structural and global market pressure. So it approaches that trajectory through speculation rather than documentary representation. Large format images generated with stable diffusion depict fictional abandoned factories drawn from the architectural language of mid-century Uruguayan industry. And the locations clearly do not exist. The model assembles composites from learned patterns, algorithmic interpretations of an unrealized future, a modernity that once appeared attainable and remained incomplete. Abandoned Fictions earned, was awarded the first prize at the inaugural, the first Colorado AI Art Competition in 2024. And then I continued exploring this collection or this perspective of fictional factories on a title that Abandoned Features in 2025, which extended our work into a serial meditation on the same industrial project with large format images being generated. So these are examples of the series of images. series of images. The second artwork I want to present is titled Montevideo, Montevideo 1983. Montevideo is the capital city of Uruguay. It was commissioned by the Sote Museum that belongs to the city. And this work reconstructs the city of my childhood during the final years of the Uruguayan civic military dictatorship that went from 1973 to 1985. It works not from documentation but from fragments, childhood perception, inherited fears, street atmospheres, family stories, political silence. For a child, such a context is not political analysis. It is tone, prohibition, adult silence, faces on television, the sense that some things cannot be said. And in 1983, I was six years old and everything was terrible. So the art world approaches the past as a field of political and effective reconstruction. And generative AI matters here, because the world concerns the impossibility of direct access. I cannot photograph my childhood. I can only reconstruct an image of the distance between it and the present. Childhood memory is not a stable recording, but a constantly reconstructed narrative. And generative systems offer a deliberately imperfect analogy for that reconstruction. And it's imperfect because they do not only infer. These models have been shown to memorize stream images and emit them at generational time. So what returns is neither reliably invented, nor reliably remembered. And the difference cannot be established from outside. So I'm gonna show a portion of this video. And this is of course not how the city looked in 1983. This is an effective reconstruction. It's an artwork. so I'm going to fast forward a little bit, for the sake of time. An important part of the artwork is the cabezudos. The cabezudos are like large-headed fierce characters from the local carnival. They stopped being part of the carnival in 2021, but they were part of my childhood's carnival. And myself and everybody my age, we were all absolutely terrified by these figures. And I decided to insert that memory into this reconstruction and it's not a representation of how they did actually look like but it's an effective again recollection on here they are The video ends with a reconstruction of a building that looks a little bit like one very famous building in the city and also a little bit like the Senate building and it makes it explicit that this is not a photographic reconstruction by adding this generated image. The second artwork I want to show is titled Hommage Numerique, that's digital homage in French, and it's a work that engages the legacy of early 20th century avant-garde cinemen, Rotman, Egeling, Len Lai and very especially Hans Richter. It revisits that tradition through contemporary generative AI. And the film does not claim to escape the recognizable aesthetic of current models, but it works against the assumption that this aesthetic is the medium's natural or only form. And it treats it as a historically specific platform style, something with a history and therefore an outside. I'm again cutting short. The videos are all available online. The third artwork is titled La Masa, the mouse, but it's also a boardplay with NASA, the American space agency. is also a board play with NASA, the American Space Agency. And La Masa is a surrealist short in black and white, blending science fiction, historical imagination, and invented ritual. It recounts the only successful interplanetary voyage of a fictional Mexican space agency in the 1960s, said to have combined mid-century technological experimentation with a ceremonial apparatus the film invents rather than documents. And it's narrated by Sol's surviving astronaut. It follows its recruitment and training until the boundary between technical and ceremonial knowledge dissolves. And it's clearly an invented mythology it does not claim any kind of knowledge about the traditions of Mesoamerica and it's Ayam Uruguayan and this was clearly a foreign construction which was also an exploration of how these imageries can represent perspectives that are created from the outside but again the the narration makes it clear that this is kind of like a dreamt story Cumplí 18 años en 1960. Toda mi vida había estado soñando con el espacio. Las galaxias, el ejército, el presidente de la UPABOM, el ejército de la UPABOM, el ejército de la UPABOM, el ejército de la UPABOM, el ejército de la UPABOM, el ejército de la UPABOM, el ejército de la UPABOM, el ejército de la UPABOM, el ejército de la UPABOM, Again, I'm going to forward. He is seeking troubles, smiling, his identity dissolves. So, Speculative Paths offers a model for engaging AI-generated media beyond the opposition between technical celebration and moral panic, treating the model neither as a creative subject nor as a neutral tool, but as a historically situated apparatus that can be appropriated, redirected, and made visible. The value of AI-generated media in intrinsic and artistic research lies not only in its capacity to produce convincing images, but in the opposite, in its capacity to make conviction itself unstable. What the project cannot yet claim is that this instability is what audiences experience, and that gap is the next problem rather than a concluding flourish. The past is never simply behind us. It is organized, narrated, forgotten, weaponized, and sometimes invented. So to speculate on the past is not to abandon history, but to recognize that history has always been mediated and that the present demands forms capable of making that mediation visible. So thank you very much. This are all available on my website. Thank you so much. It's fascinating to see these two in juxtaposition to each other, and if we have time, maybe at the end we can have some more conversation. The next talk is called Steering Through the Inner Residue, Feedback as an Artistic Strategy Against Generative Images. Our speakers are Ting-Chen Liu and Leon Etienne-Kerr. Ting-Chen Liu and Leon Etienne-Kerr are artistic associate and lecturer and research associate and co-director of the AI Lab at Bauhaus University Weimar and Offenbach University of Art and Design. Hello, hello. Ah, yeah, no. Yeah, thanks for coming to our paper presentation for steering through the inner residue, feedback as an artistic strategy against artificial generative images. That was too long, that's why we titled the talk, Will the Future of the Internet Be Purple? It will become clear what that means later. So my name is Leonid Yankeur, like I already said, I'm a scientific associate and co-lead of the AI Lab in HFG Offenbach. My name is Ting-Chung Liu. I teach at Bauhaus Uni Weimar. And so we work together as a duo where we start to do a lot of interaction from the inside to understand how generative system functions and today we're presenting the project steering through the inner residue which is a video installation that we work on trying to understand what the mechanism of AI generated image but before that we want to give some bit more context into where we reach this point does anyone of you familiar with the piece from Alvin Dussire, I'm Sitting in the Room? So, ah, I see hands. So basically what happened was that he had two tape recorders and he speaks a long sentence, or more like a passage, including the sentence I'm sitting in the room. He recorded it, play it back, use another tape recorder, record it, play it back, record it, play it back, record it, play it back. use another tape recorder, record it, play it back, record it, play it back, record it, play it back. As you might have guessed, the vibration and the reverberation of the room sort of records also into the final result. In the end, you hear those humming sounds coming out from the tape recorder. We try to implement the same mechanism on AI-generated image. Another context that sort of informed us is the idea of cybernetic, how systems self-regulate. At the same time, there is a concept that sort of pops out in 24, which is called model collapse. I think it becomes sort of mainstream nowadays about how when AI generated content being released into society and then being trained back again is sort of sabotage or damage the generated outcomes. And this is what we try to anticipate. Of course, we're not trying to rebuild the whole training process, but we try to speculate what could have happened in the image domain? Okay, so before we start with, don't worry, you don't have to read the code. There's not much code there. We want to start with a toy example before we start talking about generative AI coming from feedback and complex systems. So the idea is consider an image and an operation in this sense a blur, it doesn't matter which one, and you just repeatedly apply this blur to an image and an operation in this sense a blur, it doesn't matter which one, and you just repeatedly apply this blur to an image. I mean we can all guess what happens, right? The image becomes progressively blurrier, but there's this very interesting phenomenon if you add a second operation in this case a sharpen filter over the image and you do this progressively. I mean you could now think what happens, right? Do they cancel out? Does nothing happen? But this is actually what happens, right? There are these patterns that appear called Turing patterns named after Turing's paper on morphogenesis. And the interesting part becomes, and that also is the point that Turing made in his paper, is that and that also is the point that Turing made in his paper, is that the instructions for the patterns are not in the code themselves, but they emerge from the instructions that, in his case, he was talking about DNA, but we talk about code. So we can't really disentangle the patterns from the instructions that make these patterns, which is very important when talking about AI, because if you ever opened up a generative AI system, that make these patterns, which is very important when talking about AI, because if you ever opened up a generative AI system, you will see that there's lots of recursion and feedback and different paths interacting with each other, right? Autoregressive next token prediction, right, is feedback in itself. And diffusion models that generate images or videos themselves also have this iterative feedback structure. But we wanted to apply this to to image models, but not to the inner workings of image models, but from the outside, so a system of systems. So what we did, it's pretty basic. It's like the loop I showed you before. We took image models and we gave them an initial starting image, and we just let it reproduce this image again, and we did this over and over again. What happens you will see soon. So this is what it looks like, right? It's four lines of code, right? There's some boilerplate hidden there, but it really is not much. And important is that we remove the prompt, right? Prompts are very deceptive because they encode also our expectation of what we expect. So it's a lot of deception. We self-deceive when we write, I want to see a cat wearing a hat. And so I will also interpret the image in this sense. So we removed this from the equation and we left out the prompt and only did it on images recursively. So taking the output and putting it back in again. So what you see here is a continuation of this operation. We have this input, we use the output as an input again, output as an input again. And we try to speed it up, and you see an interesting pattern sort of emerged. That it became purple. At the same time, because of those little tiny deviation, it's sort of also animated. Just because there's tiny little difference from individual frames. At the same time, we start to question where does this purple comes from? Like what does it mean? Is it the same as the Turing pattern that we just saw? We try to have a different initial conditions as the input, but all of them sort of collapse into the same purple. So, our question then became where does this come from, right? Is this something that comes from the training data or does it come from the models itself? So we tried different models that kind of have similar architectures but not exactly and similar training data but not exactly. And what we could see very quickly is that the actual steering part that steers the image into patterns, into special patterns is the model assemblage themselves. So you can see here, we tried, let's just let it run a little bit. So it's very interesting to see that different models have different attractors that they kind of converge to different patterns and this is very interesting because I think when we talk about the steering that happens as a semantic steering that happens in generative AI we think a lot about things like how hegemonial structures right represent so it's if you would generate a flag it's always an American flag or something like this but there's also something that's not in the training data, there's something that is coming from the procedure itself, which you can see here. And something very interesting is also that sometimes if you just tune the parameters just right, you get kind of stability and attractors, conceptual attractors that repeat over and over again. And you can see it's very exactly what you would expect, it's very mundane patterns of advertising images and stock imagery, right? And in our experiments also, without any prompt, we also, there came a series of images of politicians and even Donald Trump came there without ever prompting it, right? Just because these conceptual attractors are so strong that they kind of steer towards kind of a mean of concepts. What seems to be also really interesting for us is the idea to coming back to when we're trying, when we're referencing Aving du Ciel's I'm Sitting in the Room, that in the end, what's being recorded as an outcome is not just this humming voice, it was the sort of attribute of the room. If that performance was done in a different room, the end result would be different just because how room sounds reflect as how I'm holding this microphone and speaking to you. It might be different when I'm speaking in the toilet. So it's the same situation happening with different models. Individual models also carry specific attributes. There is maybe different floating pool calculations, there may be different architecture how the feedback was run, but using this sort of outside of the black box mechanism that we're trying to apply here, we also see the tendency of different models, whether it behaves. Is there something specific that it really likes? So maybe that's also why the question about will the internet be purple was that a lot of people talk about model collapse in the sense of polluting the training data with synthetic data and then where will this lead, right? But our question was also just about influence, so just about running the models, right we take a picture with our phone, it's already run through an AI model. If we post it to Instagram, it's probably run through an AI model. YouTube is using AI now to compress our videos. So kind of what happens, what's the steering that happens also algorithmically just by AI's interpreting these images? And so here, this is the final installation. It's a continuous loop of like, I think it's 16,000 frames. And the only thing that we do is that we change out the model. So we have six starting images. And the only thing that we change is not the parameters or anything. We just change out the model. And you can see how it kind of iterates through this possibility space of possible images just through changing out the model. That's it. I think that's it. Thank you. I believe this is our last talk for today, so we have some time for questions. There is one more, I think. Are they here? Oh, you are here. Okay, I think. Are they here? Oh, you are here. Okay. I'm sorry. Okay. So we don't have time for questions. Well, maybe one or two questions and then we'll move on. Sorry. Apologies. Can this be related to certain data being more prevalent in connection to some topics? Like, for instance, it was mentioned that Trump essentially generated himself out of nothing because he was essentially hidden inside the model. was essentially hidden inside the model. Yeah, I think the point we want to make is that we kind of, and this is kind of new with generative AI I would say, you can't really differentiate anymore between training data and also algorithm, because it kind of fuses together here, so you can't really get one or the other. And I think a lot of research was done about the training data right but in the last years it kind of it's not so prevalent anymore because talking about training data it's really hard to predict what a model will output even if you know all of the training data so the algorithm plays an important part that was also kind of the the point of this experiment and so it's it's impossible to differentiate the two. So we have to interact with them. We have to run them actually to know what comes out and what the biases are. And that's one way of using feedback to kind of not prompt qualitatively but quantitatively interact with it. Okay. Do we have any other questions? Thank you. It's interesting. It's not so much... I mean, there's a question in it, but it's a quick FYI. I'm Philippe Passy. I'm a professor of creative AI in Vancouver. Technically speaking, this is very well known, and I could have predicted that before you started doing it. So it's called fixed point theory, and there's two types of fixed points you can attain, you know, stable fixed point and unstable fixed point. You get a great example of both. Did you know about the fact that this is all known? I mean, we had, like, references also from chaos, and complex systems, and kind of this idea. But what the fixed points are, that still was unknown to us. So that this can happen, that it loops or it gets into stable homeostasis or something like this. Yes, we were kind of predicting or we were kind of expecting something like this, but what these points are here, because here they have semantic meaning, or semantic meaning, but they say something if it is Trump that over-represents themselves. That is what we didn't know, right? So like, yeah, in complex systems or something like this, sure you have divergence or convergence. And I think we were expecting divergence in the beginning. And then when it converged, and then also into this pink noise, that was kind of interesting to us. So really looking at what comes out, not only that it converges, but what it converges to was the question we were interested in. Thank you. Our final talk is called Double Blind, Sensibilities of Extinction, Inducing Commemorative Practices of Countermapping and Distributed Archiving in an Extended Reality XR Environment. Our speaker is Margarita Kroll. She is from the head of department in design and lecturer at Vorarlberg University of Applied Sciences. Margarita Kroll is a transdisciplinary researcher and artist working at the intersection of arts, eco-social, as well as science and technology studies, recently focusing on the role of affect in more than human relationships. Thank you for the introduction. I'm Margaretha Köhler. I'm presenting an ongoing artistic research project which I developed together with my colleague David Altweger, who is also here, but actually we cannot see him yet. Just a moment. Yeah. No? No? Okay. Because he is there in the online call at the moment. But yeah, I'm going to start my talk now. In this research project, we focus on biodiversity. This is from an overall point of view. It tackles the loss of biodiversity, which is also integrated into the title. So, double-blind sensibilities of extinction. And I want to start with the title. So, double-blind, what does it refer to? While the project may initially appear to address the loss of biodiversity in the Austrian region of Vorarlberg, its deeper concern lies in the interrelationship between hierarchies among species and the ways in which humans situate themselves in relation to technologies and biological forms of life. So double-blind, we are all familiar with this term from academia. Double-blind is kind of gold standard of experimental method and it also guarantees in a way what we call objectivity. So we borrow this term for an artistic research project to question this notion of objectivity, because we use this term to address the ambivalence between safety and uncertainty, and also to question methods of mapping and archiving, which connects to the talk we heard before the last talk and also this question of how do we represent biodiversity loss and also this process of extinction. So I would like to start with the loss of biodiversity, which has been... Approximately one million animal and plant species are at risk of extinction in the coming decades. And there's an unprecedented acceleration of biodiversity loss in human history, according to the IPBSE Global Assessment Report on Biodiversity and Ecosystem Services. on biodiversity and ecosystem services. So 75% of Earth's land surface and 66% of the marine environment has already been significantly altered by human activity. This is not only a story about tropical hot spots, which is always framed like this in the media. It's also unfolding in the Alpine region. This project is made in Vorarlberg, seemingly stable central European cultural landscape. Around 68,000 species have been recorded and more than 1,200 plant species are already listed as threatened according to the Umweltbundesamt. So we work together intensively with a nature experience space in Vorarlberg, which is called Innatura, and they also provide us with a lot of data. And the typical farmland birds, for example, have also lost roughly 40% of their population. What is a pivotal point of this talk is that not all species are treated equal in the recording of extinction. Attention, funding, mourning, and also media coverage cluster around a small set of charismatic species. So there's a kind of inequality. And it's mainly the large mammals and animals that are visibly striking and emotionally legible to us. So mosses, fungi and microbiological organisms, for example, as well as insects and also countless unglamorous species or plants disappear barely noticed. Despite comparable or even greater ecological impact. or even greater ecological impact. So that's why we focus on these invisible or underrepresented species in the collective history or collective memory and would like to give them greater ecological visibility and significance. So that unevenness is one of the main points of this talk. And that's why we turn to Achille Membe's notion of necropolitics. of necropolitics. So what is happening is that this biological crisis is not only a biological one, but also a socio-economical and cultural historical one or cultural epistemic one. And this is also what is called semiocide because if the diversity of species is extinct, also our realm of memory and the way we can imagine a variety of species is diminished. That's why we turned to creating a hybrid or living archive. So the archive is something that is not a stable storage for us. It's a living archive that is more something like a cultural memory, as Asman calls it, a productive repository, which is living and where people or also other species can contribute to and create new meaning. So kind of counteracting this notion of the dead storage. So let's have a look in our hybrid systems lab which we built at the university in Vorarlberg, which is a kind of XR setting where you can enter this living archive by accessing it via distributed maps. So I think David will show it afterwards. We also have got a video where you can see this map where you can enter the space of collected specimens, but also non-existent species, which are created using AI models, which also links up to a project we presented two years ago here, Sphagnum Denses, and this is kind of the continuation of this project. So we can see this map here, which also is an open source archive. We work with local data from an open source repository called Vogis, where you can access all the environmental data, and you can also transform it into, for example, 3D models. into, for example, 3D models. And you can also contribute to this map with an accessible interface. So maybe, David, you want to jump in now how it's technically realized after I showed the video. So for us, this map is a tool for counter mapping, as Rebecca Solnit calls it, where not only the so-called neutral data or platform data is accessible or stored, but where also effects and emotions or other multisensory perceptive objects can be stored so it gets this more than human dimensions. Yeah so let's have a look, I don't know if the video works, otherwise David will show us the setting. I don't know. Es hängt. David, hello. Hello. Hello. Not yet, just a moment. No, unfortunately it doesn't work. Sorry, we cannot hear you. Okay. Sorry. I'm going to continue. Wo ist der Kollege? Sonst machen wir das jetzt einfach. Ja, das sieht gut aus. Okay. Sorry. Ja. I don't know why it doesn't work. Okay. Sorry, I don't know. Okay, sorry for the interruption. Yeah, I mean, we can also have a look into the lab where you can see the map in the background and yeah, you can enter this space with a head-up display or it's also accessible via touch. And then you can have a look at the different hybrid creatures where we also work with AI models which are combined, two AI models are combined to have these stereoscopic creatures where people can also interact with in the room so actually it was created with a specific touch designer patch where the two models are combined. And if people enter this space, they can really interact with those speculative creatures which are generated based on this historical data about extinct species, but also still existing creatures or species and also speculative species. So actually we use play as kind of critical methods where people can kind of wander through this space, which draws on this notion of the Deriv, which is also a technique developed by the situationists where you explore your environment and focus on the emotion you have during these Derivs. So it's a kind of ludic assembling of species, audience, tests, and it also reconfigures the rules by which spaces, memories, and futures are organized. So in a way, the main focus of the project was also to kind of cultivate something that is called speculative literacy that enables participants to imagine other ways of living with loss and uncertainty, ways that acknowledge our implications in extinction and in colonial and also platform-driven histories of mapping and collecting. So, yeah, I think I just got an insight into this project which is still developing. And thank you for your attention and your patience also. Thank you. Do you want to answer any questions? Yeah? Do we have time for maybe a question or two? Yes? The phenomenon that less visually appealing species are less cared about when they are that can also be seen in species that are far away. Like, for instance, the blobfish is known for being also endangered, but people don't care about that either because it's not visually appealing. And my question is is there a way to do something against that? I can't get where you are. Sorry. Okay. So actually, this background of necropolitics mainly refers to racial inequality, that some groups of society are overrepresented, and also their extinction is not even mentioned. So if we view ecological systems with this perspective, I think it's very obvious that there are a lot of crucial species which are kind of not discussed in any journal, for example, or any nature archive, for example, the museum we work with. And our approach is to kind of get people to reflect on the different types of extinction and also the different species that are extinct. So that's kind of the way to approach it with this artistic living archive. I don't know if it answered your question but I tried to. But we don't see David. David, can we try, will it work? No, now we can't hear him anymore. Hello?............................. Hi. It won't let us do it. Drag this. Let's try this again. Say it again. No, no, they turned it off completely. They don't like our idea. Okay, what's up? Do you want to do it again? No, no, no. I don't want to do it again. Okay, what's up? Do you want to do it again? No, no, no. I don't want to do it again. Okay, what's up? Do you want to do it again? No, no, no. I don't want to do it again. Okay, what's up? Do you want to do it again? No, no, no. I don't want to do it again. Okay, what's up? Do you want to do it again? No, no, no turned it off completely. They don't like our idea. Okay, what's up? Do you want to add anything so I just can? Yeah, then. Can you hear me now? A little bit. Not really. Da, da, da. Da, da, da, da, da, da, da, da, da, da. No, it doesn't work. Okay. That's a change. I think we should skip this experiment. So, I think we should just go back to the beginning. Okay. So, I think we should just go back to the beginning. Okay. So, I think we should skip this experiment. Can we turn the game all the way up? Hello, hello, hello? Yes? Yeah, you hear me? Yeah. Yeah, what's to add? I mean, in general, I would also like to add to what Margarita just said, in terms of, you know, there's hierarchies in all of these kind of like experiences of experiencing this loss of diversity and this loss of culture and this loss of communities. And what I would like to add is a political perspective that is really, really important and a crucial part of our work is that what we recognize a lot and what we see a lot is that a post-humanist perspective and the projects that come out of this way of thinking often engage with different species and different plants and different ecosystems and why why we see this absolutely central to not only developing empathy for other species but also never forget that this is about the destruction of cultures and bodies and and and systems that are also ecosystems and and I feel like it's so important you know in our post humanist perspective to include the solidarity with other communities and to not only focus on the nature part. And I feel like cultivating a practice and teaching people how to use tools and artificial intelligence in order to decolonize their workflows and their practices is really important. And to develop tools that are open source and to go away from software packages. For instance, OpenStreetMaps has been around for so long, and there's so much open source data, but there are so few applications that allow a broader public to experiment with what a map can be. Can it be a repository for poetry, for grievances? And we feel like AI and the hybrid systems at the moment really enables us to connect all of these technologies. And we feel like this is a really, really important time to reclaim all of these technologies that have been colonized by huge, huge technical conglomerates and use those technologies in order to make meaning and to make meaning out of this huge loss of richness that is happening around us and to be dogged about, you know, like holding up the practice of collecting and of archiving and of making memories and making meaning. And that's at the core of it. I say one more example, sorry, I'm probably talking way too long, but one more example is like by decolonizing our workflows i mean for instance we like ai enables us a lot in design to to automate certain things and like for students to be less reliable on software packages in order to create stuff and whenever we create some sort of freedom and space through the use of AI and AI-assisted workflows, we then take that space and fill it up with other practices that have been lost. So we reintroduced screen printing. We reintroduced, like, physical art making and design making. And then, so we work on both ends of the spectrum. And I feel like that's a really, really important thing to do, to not only automate, but also like, what are we going to do with that space that is being made by AI-assisted workflows? Thank you. I'm glad we were able to get you into the room, more or less. Thank you so much, and thank you for your patience. Of course, of course. Well, I think we're out of time, but please join me in thanking all of our panelists for their presentations today. We resume in 30 minutes for the next session. Maybe 29 minutes now. Thank you.