At the workshops in the coming days, you will have experimental forms of exchange. There will be very different kinds of atmospheres, but this evening we're going to start with a very, very classic format of two keynotes from two absolutely wonderful speakers that we're very, very happy to have here. Are you, Selena, are you ready? Oh, yes, you are. You already set that up. Okay. The first speaker, sorry. First of all, we will have, Selena and Paris will be speaking one after another. First one keynote, then the second. So please make a note of your questions, your comments that you want to add. And we will have a discussion of both of the talks together afterward. All right? So that will be the format. So first of all, it is my very great pleasure to introduce Selene Savitsch. She is assistant professor of proto history of AI and machines in the arts at the University of Amsterdam. Her research and her research interests animate a practice at the intersection of computational processes and post-humanist and post-colonial critique of technology. I think that's such a wonderful sentence. It's on the website. You can all read it for yourself. And Celina is going to talk to us about the generative pasts of AI. Selina, thank you. Thank you. Thank you. Well, thank you for the wonderful introduction. And thank you all for being here. Indeed, I'm going to talk about this nearly invisible title that I made here for you. I'm going to start with, again, introducing myself, and with some self-reflexivity, as they like to say. So I want to say that I speak from the position and experience of an immigrant. I work in academic research, and I'm currently assistant professor in Amsterdam. I work in academic research and I'm currently assistant professor in Amsterdam. But I come from the European periphery, the city of Belgrade, which remains outside of the European Union. And then in terms of research, I also feel that somehow I had to make some borders crossable for myself. make some borders crossable for myself. So I made myself familiar with AI through different kind of parts of studying design and technology, let's say, beginning with studying architecture in Belgrade, then moving to study media design in Rotterdam, then doing a PhD in Lausanne, and later working in Vienna and Basel on studies of data and information architectonics. And now I work in Amsterdam. So I never formally studied computer science or such a field which produces AI, as we call it. such a field which produces AI, as we call it. But I did learn some basics of programming, thanks to some people present here as well. And then I also extensively read, and mainly in English, about media theory, STS history, philosophy, anthropology, and cultural studies related to computation. So in the context of all this and also of our meeting today, I use this term AI as an umbrella hype term that is inclusive of machine learning, artificial neural networks, statistical imitation models, data-driven applications, and also doubt its existence following Lucy Suchman's recent call to treat its existence as controversial. So let's go on. This talk is going to be about generative pasts of AI, This talk is going to be about generative pasts of AI, and what I mean by this has to do with what I would call temporal ambiguity of the root gene, which belongs to both generative and genealogy and generative. And the point is that this, the fact that the gene is at root of the past and of the future in this sense of this proto-European root, which I think is proto-European, proto-Indo-European, is that it's also characteristic of the way we use and make AI technologies. So in terms of it being generative, we speak most recently about generating images and text, but also it is kind of always generated from the past. It's really always based on what is the past data. It's never really about the future at all, at all. It doesn't even tell the future. It doesn't even exist in the future. It's always in the present. So there is this wonderful book by a Dutch feminist scholar, Iris van der Tuin, that writes about generative genealogies of feminism. And I take this term from this book and I run with it. And I then pay attention to the way present and historical mechanisms of dominance, such as patriarchalism, colonialism, imperialism, and capitalism maybe, and others, take part in these generations. So one thing to ask would be, for example, if you want to write a generative history or generative genealogy of AI, would be how does chat GPT come to be? What past has generated it? And how can we look at that past and see other things and see surrounding events and see, you know, other ways things could be today or things could have been? In this sense, this talk is generative because I hope that with discussing these ideas and these pasts, we can think also together about other options and think about what we want. So I have been reading this book, a wonderful book called Wrong Way by John McNeill. And this book is, let's say, about self-driving cars. But it's not. It's about much more. And then I found, I was looking for an illustration of this book because I was wondering if I got the description right. If this car, if the person is inside in this or that way, horizontal, looking forward or backwards, I did not understand it. So I was looking for an image of it. And I found this wonderful illustration by an illustrator, Tim Einthoven, whom I contacted to use his image in this key moment of this keynote because I was interested in actually how do so since we speak a lot about the AI generating stuff and generating images I wonder you know how do we imagine AI at all how do we really how do we imagine AI? And let's see how professional illustrators do commissioned work that illustrates AI. So this is one example. And we will come back to it. So don't worry. It's really wonderful. Now, these are some other imaginaries. This is from a magazine called Elements or Elementi, published in Serbia. And this one is from April 2023. It's about chat GPT. And I think this one is as well about chat GPT. It's by the same author, Nikola Koric. Thank him as well for these images. And so it's depicting, yeah, this kind of natural language processing technology aids that help us to write. Then there is this wonderful other image of Tim Einhoven about self-quantified work, and again, the others that are part of it. And one more image that I want to show by Shivan Sven Wang for a magazine, It's Nice That, where it was published as a banner for the article by Julian Posture called Why Creative Labor Isn't Always Seen as Real Work. And this title makes this good arc to answer the first question, how do we imagine AI and how does AI come to be as they show labor, they show work, sometimes also pleasure, sometimes play, and most importantly, all these images picture relations. So there are all these relations between people and technology, relations that are enabled by technology, or relations that are enabled by technology or relations that are theirs regardless of technology. And so let's come back to this first image, which I promised to show you again. How would you describe this image? Can anybody please tell me? Yeah. yeah anything more yes can somehow, from some point of view, organize this space of realization of desire. Yeah. But you can see in the picture the labor behind this. Yeah. And you see also the desire that it is so in this image, that it exists. Yeah? Yeah? Do you have more suggestions? I think it's beautiful to collect them because we will look into that further. It evokes a sense of missing privacy somehow, because these hidden actors and agents somehow stuff the way it should be. Maybe that the people are fitting into the design of the car. It's not the design for people, but that it's sort of shaping the car. Yeah, that's very important. Yeah, I was thinking immediately when I saw this image, I want to process it through an image classifier. I want to see what it sees in it. So I did that. And that made me wonder how I would describe this image because it's so failing. So here, the first try is Image Recognize, a website, a free image classifier that uses CNNs for object detection and facial recognition. So you have different options here. You can look for celebrities, you can analyze faces and so on. You can extract text depending on the image and so with the basic functionality of object detection, it detects person, art, machine, wheel, painting, face, head, but it does not detect a car, curiously, because, I mean, car is clearly present here for us. And you mentioned that people are in the shape of a car, and yet the AI does not see, I mean, this basic AI does not see a car here. Then another tool by a company called Pally, like a pal, Pally, that apparently focuses on social media management. So you can post, like it enables multi-platform posting. And what you really need when you multi-platform post apparently is also object description so it is an image caption generator tool now I first asked it here without context on the left and the description is the image is a cartoon of a person riding a skateboard. Because I guess this guy with the casket tricked it. Then I tried, because it has this possibility to include additional info, I thought, OK, I describe it here. I have a very long description, which maybe you can't see, so I'm going to read it from here. People kissing in a car. Instead of mechanical and electronic parts, the car is working on human force of four persons hiding inside car niches. Car model reminds of a Volkswagen Beetle. I thought any of these words will catch its attention that there is a car here. And no, again, the description says the image is a cartoon of a person riding a skateboard. It's very, maybe it's not a very good tool. So the best tool was this tab caption app, which even though it has a very low rating, has a very low rating, 2.8 stars in the app store, did some good results. So it says, just another day at the garage with my amazing team. The work we do may seem complicated, but together we make it look easy. This is quite accurate. It looks easy, but it's not. There's such a lot of work going into self-driving car. Working on cars is not just a job, it's a passion. You're right about the desires. The illustration captures the spirit of collaboration and dedication that fuels our work every day, literally. Then there are more, which I will read from my screen. Ever wonder what goes on inside an auto repair shop? Well, this colorful artwork gives you a fun idea of the organized chaos that happens behind the scenes. We might not wear capes, but as mechanics, we still manage to be superheroes for our cars. Check out this cool graphic representation of our daily hustle in the workshop. Anyway, these are very nice descriptions, and they really match well with the industry of self-driving cars. So to conclude this examination of the generative tools or image recognition tools, obviously those ones that I try here for you are not the most cutting-edge products of the industry. We could have found better ones which will do a better job at classification and so on. But the interesting thing is that these are very accessible ones. They are free, they are online, so a lot of people can use them, and they get massively used. Therefore, these networks that are typically like the neural nets of these classifiers, for instance, for object detection, which are typically used in research and industry developing autonomous robot vehicles that are also fueling self-driving cars, if we may say so, is very differently able to recognize a car in a drawing of a car. Like it doesn't see a car, but it's supposed to do at least that. And so image classification is a lot of fun. But let's get serious. So these illustrations show us how designers think and feel about artificial intelligence. People are always in these images, they're both in the center and in the scenes, behind the scenes of these technologies, they're measuring, labeling, selecting, supervising, optimizing. And on the other hand, artists' interpretations of AI are differently classified and interpreted by state of the art artificial neural networks and their specific applications, such as object detection and image to text conversion. So somehow there is a big mismatch between the interpretations that we get from people about what is really going on and the way even get from people about what is really going on, and the way even these interpretations are able to be summarized to these supposedly artificially intelligent algorithms or machines. And an important question to ask in terms of the generative genealogies like how is how are these machines able to do this and the simple question is with using a lot of data so there's lots of data everybody speaks about it since maybe 20 years we speak about this big data glut, about lots of data, too much data, big data, and so on. But the question is, what are these data exactly? And how do they become part of this system? So the question is, how somehow we have all this data. And now one complicated answer to that would be to look at the work that is obscured to public, which is going on continuously and feeding these machines. But let's start from what we can see here. And what we can see here are images of cats from the Lion B dataset. These are photographs of digital drawings or digitized drawings of cats. Suppose you are all familiar with the Lion B project and on the Lion project in their Lion B5G or something dataset which got then pulled down because of problematic content. But initially, it was put up with the motivation to democratize research, experimentation, and also suggesting that it will give something back to the public, which is not often mentioned explicitly and not very well articulated. What is interesting is to notice how easily it's taken for granted to take things from the public such as assuming the access to web crawling which has happened previously and is contained in the common crawll web index, which is also one of the foundations for the accumulation of images in the LION datasets. So when we are thinking about images of cats, this is a famous photographer, Walter Chandoha. He was the cat photographer, and he worked before the internet. And he produced lots of photographs of cats, like really, really a lot. And he was very, very well known and very well liked. And his photographs were featured in magazines and covers of journals and advertisement campaigns, so a lot of places. And he was really able to stage these cats because he was close to them and he was able to make them pose in specific poses and get a cat which is waking up or, you know, you could order a specific cat pose from him and he would give it to you. And some of his images are probably in the Lion Bee data set. But Chendoa said that cats are his favorite subject because they're expressive, playful, and cute. While it's important to note that you see this camera he works with. It's a very old camera. It's an analog camera. And interestingly, cameras became so flexible and nice and portable around the time when he was born. So we're here in this timeline of Walter Trandoha getting into the world in which cameras are suddenly portable and you can take them with you and also have films. So you can take multiple shots and create images of whatever you want. Also much more affordable, like cheaper. And also, I think when he was maybe 10 years old, color film became available. Also, it's interesting to say, or kind of stupid to say, it became available. People were able suddenly to do it. It didn't happen on its own. And now, to stay a little bit longer with photography, without showing any other images, but on a much less joyful note, and speaking of subjects in a different context, I would like to mention that we should recognize also how life is converted into data through various technologies of capture, including photography. So to think about that, I like to read and mention the book called Potential Histories by Ariella Aisha Azoulay, where she discusses the operation of camera shutter in early photography as a materialization of imperial technology of dispossession. So it's a very big claim, but it's also very well explained. So, on one hand, there is the action of the camera that produces the photograph, like the shutter will close and open, and then also close. On the other hand, it institutes this reproduction of imperial divisions. In the moment when the shutter is closing, it sort of distances the person taking the photograph from the other side. So on one side you have this imperial power, because indeed photographs of photo cameras tended to be more accessible to some people than to others. And then you have the dispossessed people and objects on the other side. And it speaks of the big, big work of archiving that went along these processes of colonization where you would actually document the subjects, their objects, take them to the museums, actually document the subjects, their objects, take them to the museums, make them more and more appropriated by archives and other knowledge infrastructures. And so this is an important thing to keep in mind when we think about the lion bee data set and how it came to be. Another important thing to think about of what photography enabled, which I guess you're all very familiar with, but let's briefly mention it, the standardized capturing of faces of people at arrest. So this is a technique called bertillonage, which was popularized in Paris at the end of the 19th century by the French criminologist Alphonse Bertillon, who developed it so that you can systematically measure and index arrest cards, which should aid in identification of past offenders. So the point was to be able to recognize if you are again seeing the same face, so you would take always the same position and you would put the person always in the same situation so that you could compare and say yes this is again this guy who stole this apple at the market or not. Therefore, this is a very specific application. So it serves to narrow down the search. But it shares history with attempts of famous eugenicist Francis Galton, who was interested in sort of the opposite in kind of predicting behavior based on what you can actually measure out of the physiognomy of a face so interestingly there is this photograph of him by Bertillon in the proper setup because because they were meeting. And as far as I was able to read up on this, were interested in these techniques for different reasons. And in the end, Bertillon disagreed with ideas of actually extrapolating measures of physiognomy into social behavior, which is what Galton was into. But indeed, having all these appropriate faces of people into photographs enabled also systematic measurements and typologization and taxonomization of facial features. and taxonomization of facial features. And therefore, yeah, they sort of enabled eugenics in this way. But people are not necessary for this at all as subjects. So we see here images of irises. These are, of course, contemporary images. They are taken probably very recently. But they are a kind of a pair image data set to the original data set, which is behind. This data set consists of four value measurements of 150 flowers, of four value measurements of 150 flowers so that the so that you could describe three different species of iris flowers and they are measured according so what is measured are the sepal lengths and widths and then petal lengths and widths. So there is like two features with two measurements for each flower. So this data set was collected by a person called Edgar Anderson, who wanted to quantify the morphologic variation of species of iris flowers. And it gained prominence in the research of Ronald Fisher, who was a British biologist and statistician, who published this paper on the use of measurement in taxonomic problems. And based on the combinations of these four measures, he was able to develop the linear discriminant analysis model to distinguish the species one from each other. So he was able to, purely facing only these four values, able to purely facing only these four values, write such a routine to be able to determine which bag out of the three possible bags this iris flower would belong to. This paper was published in the Annals of the Eugenics because it was contributing to the same idea of identifying nature through measurement and finding the logic of morphology, physiognomy, those things that are to do with shape, with the outcomes of growth into something that could be systematically classified into different categories, but also predicted, like what is this person or this iris going to do? that could be systematically classified into different categories, but then also predicted, like what is this person or this iris going to do? Is this going to be a criminal? Is this, you know? Yeah, so why I like to dwell a lot on this data set is because it's one of the first data sets you can encounter when you want to try a basic machine learning classifier. Like if you want to try out something, you will most likely fall onto the IRIS data set and it will be able to show you whether your algorithm is well trained, whether you understood how to train an algorithm, whether you understood anything about machine learning. You're going to get to work with this iris data set. And if you're successful, you will get correct associations of irises, if you would show it new irises. It's a beginner data set for learning, and it has this history. So in this way, still today, physiognomical and morphological variations prove to be a stimulating framework for producing data, while the assumptions and conditions that produce them continue to inform contemporary practices of learning and developing machine learning. contemporary practices of learning and developing machine learning. Another small arc I will make into the past before we deal with the future is to also mention the thing which I, in the beginning, said we cannot see, so what is actually happening in terms of how AI is able to do that, the work of annotating, labelling and caring for the images. So on the right here, you can see an image from a paper by Antonio Casilli, which is a sociological research into basically a lot of aspects of labor that goes into AI. And I think this particular image fascinated me because he talks about the roots of contracting that goes between the origin of images, for instance, then where they get labeled, then who this work gets bought by, and then the kind of requests for data sets and where this gets produced and how this gets sold back. So it's obviously very complicated, but it somehow tends to mean that U.S. companies will subcontract a European company to collect some, or like, I don't know, a European company that will subcontract an Indian company that will subcontract a company in Indonesia and will then produce the more and more granular levels of work and also pay and then feed back that chain. So I invite you to actually read that paper rather than take this description as the explanation of what happens. Another interesting look at the datasets is by Sasha Luciani. That's the two screenshots you see on the left. So this particular talk was partly about the ImageNet data set and how it's flawed, of course. So the problem is that when all this labor happens, there is lots of errors as well. And these errors are nobody's really responsibility because they're in the chain of subcontracting. Obviously, we miss a lot of verifications. So here on the top, you see incorrect labeling of different animals. But then the more striking thing is that something that I was also able to observe in data sets of bird songs, which is that bird songs tend to be recorded and posted in Europe and the US, but bird diversity is not mainly concentrated there, but it's where people have recording equipment. And then when you have a bird data set from the world, you will have a lot of European and US birds, which is quite a bias, and it's quite a lack of data. And what also Sasha Luccioni was able to discern from ImageNet is this striking part on the bottom. A lot of the wild animals had people who killed them on the images. And the thing is that these images are then labeled as fish, or I think this is an antelope. And it turns out that in the ImageNet collection of fish images, most of them are held by people. It's okay once, but if every image of a fish is containing a person, then that AI will think, apparently, that images always come with people. That's another thing, and that's not correct. And, yeah. So, I hope I'm not too much over time. No? Okay. No? OK. Now, we have looked at the way past the generated AI is playing out today, and these kind of mechanisms of dominance, and how they make their way into contemporary technologies. The question is, what other presence could emerge from these pasts and how in general can we deal with this? So one thing that I really like or one interesting proposal that I like to consider is coming from Wendy Chun's relatively recent book. I still always think, oh, it came out yesterday, but it's from 2021, and it's already 2024. So Vendee Chun proposed to counter harmful technological disruptions by using existing AI technologies to diagnose inequalities and treat discriminatory predictions as evidence of past discriminations. So I would say, if I can reinterpret it, as to kind of turn the automated predictions upside down on their heads, onto themselves. And one short vignette that I would like to offer here into how this might work is my first interaction with this game called Semantle. So there is this website. You can play a game, a word-finding game every day, another word. Has anybody played Semantle here? Yes? One of you? I learned about it from, yeah, anyway, I played it for the first time once, and should I say first the solution and give away everything? And it's not a very pleasant one, so maybe I should kind of warn you. The first word that I was supposed to guess was the word grave. I was not expecting that. And it works like this. You put a word, and it tells you how close it is to the word that you should guess. So with that, with already two, you can figure out that you should guess. So with that, you know, already two, you can figure out where you should go in which direction, whether it's an adjective or it's a verb or is it about... And so the person who recommended this to me also said that you should try men, women or objects or, you know, there can be differences in these that could be guiding you towards the solution. So I've tried men and women, and a woman was much higher than men, and I thought, okay, this is something about women, and there was something about it being a terrible word, something about disaster, something that was also high, and I was like, okay, a terrible and a woman, what could it be? There was something about death, too, but I did not get the word at all. So I was like, okay, women, crime. And at the time, of course, and always, there is lots of talk about criminalization of abortion. So I was like, okay, abortion, miscarriage, all these words, terrible words. And they were all really, really high. And it was not the solution. And then, yeah, then I tried the hints, and we tried also giving up. And finally, the solution was a grave. What you can actually conclude from this interaction is that women are closer to the grave than men. than men, you know, and this is true in, apparently, in that model because this game is working with a word-to-vector model of Google News until 2021. So all the text that is in Google News, there is this file, you can download it, it's a zipped file of a word-to-vector model. So this word-to-vector model is an interesting thing that I really like. It gives you a representation of words in many, many, many dimensions. So it's kind of spatial, but it's very complicated. And so there is kind of the assumption that the meaning is contained in these measurements because it will understand or it will contain information about the words' relationship to other words, and words are always defined through the words that surround them and how often that happens in the text on which it trains. So it's very interesting then to think how come that women are closer to grave in Google News. So I was like okay let's download this model and see what is close, you know, can I really see it with my eyes uh so here is what are the 34 i mean any number of closest words to grave in that model um and actually women are not there and like the words that are close are quite logically there there is symmetry um graves also, but Sirius is high, I don't know. So I thought, oh, I don't see women. I listed hundred closest words. I listed thousand closest words, and women were still not there. So finally I found a woman. That's, I think, the first. I mean, I really could not do this too detailed, but a woman is the first instance that's 73,678 distance, or close word to the word grave. So that was quite strange, and made me think, you know, it's really important to have other ways of accessing this than, you know, creating these lists that will not really tell you anything because this interaction was much more valuable and it gave me an insight which I think is, we cannot really, I cannot forget it. And also it's true, like it really is true. When you ask it what is closer, woman is closer. So, yeah. I would like to now conclude with two examples of practice that counter these discriminatory predictions of database inequalities, as Wendy Chun suggested. So this is the first one I want to mention is the work of artist Netris Gaskins, who explored generative AI algorithms to improve the way black and brown people are rendered and give them more positive ways to view themselves. So she wrote this great field review of artistic practices that I will not go into details but I invite you to read and appreciate her work online there is this text called Interrogating AI Bias Through Digital Art it's appearing in Just Tech and there are other texts by Netris Gaskins that really give you an overview of that perspective. It's interesting to note, in the context of previous discussions on photography, that the way photography, colored film renders color, has also been quite scrutinized for the fact that it has worked with this assumption of global whiteness. So there's many texts that show that the Shirley card was the model and because of that, certain colors, like tones were excluded and more like less precise. And so it's unadjusted for darker skin color. So it poorly renders it. It's underexposed. And that shows that these camera-based systems have defaults that disproportionately affect marginalized populations or they disproportionately affect simply lives of people that are not imagined as the kind of normal, the norm, and that is significant to observe what is considered the norm. And this extends to digital technology and CGI in film, as well as in surveillance. We have also seen many instances of that. But the result is the lack of representation of people in data sets, and then lack of material visual heritage for people with darker skin, or like that is of a lower quality and so to counter that there are tools such as the old if I where you can actually color photographs of people that were taken in black and white which is interesting to do if you know your family did not have access to color photography in the 60s, for instance, when some families had access to this. Then also image style transfer. Here, Nedrys Gaskins talks about the kind of two options of making skin shiny with AI filters in Photoshop, and that one technique is based on one kind of calculation and the other on another kind of calculation, and the one that is default is the one that makes white skin shine but it actually does not work for darker skin and so um to kind of emphasize this uh she made this gilded series in which she used shine effects from vintage gold and bronze tapestries to create a mural-sized portrait of, for instance, here, Greg Tate, a preeminent black American cultural critique, author and musician, and to kind of glorify him with these other skins that are not part of the human skin repertoire because, anyway, this repertoire for AI tools is lacking. So that's kind of how I would transpose that message. And then I also want to mention the art of negative prompting. It's my favorite. This is the artwork by Stafford Swanson, a super composite, and it's about a character that emerges from a data set, we don't know which one, but it doesn't matter, it could be Lion B, it could be some version of I don't know what. Anyway, she asked this model, what is the opposite of Marlon Brando? And I don't know why it's written wrong. It's my error. Anyway, I think the prompt was like Brando minus one. And what you get was, in this model, a strange abstract image of a kind of castle with some text. And the text apparently was digitapnit, digitapnitics, something meaningless. Then she put that text and description of that image as a prompt, a negative prompt, and got this person. And so this person, well, you know, it's an interesting person, but it appeared that this person. And so this person, well, it's an interesting person. But it appeared that this person inhabits this kind of coherent space of negativity in that model. Because on different prompts and different negative prompts, this person was reappearing and was like the opposite of stuff. So it's interesting to think about the way we can explore the models, given that it took this big Stanford study to understand that there are these not safe for work images in Lion B, although you could open that data set and just see them, but that's not possible actually. actually know what is in the data. How can I know that woman is closer to the grave than men, even if this stupid word-to-vector file is very browsable and there are a limited number of lines? I don't have patience to actually go through them in order to learn something. And learning something, you know, it is temporal also. Like, our attention has a span. So I find this a very important example of how you can actually learn what is in the data or characterize a data set and characterize a statistical imitation model, such as a generative AI, image generative AI, such as a generative AI, image generative AI, in practices that will have things that are unexpected emerge. Yeah, that's all I have to show you. And I would like to end with a note of, yeah, opening this question up on how artists can prompt and model and generate different realities, where the models show how we can offset this generality and this assumption that these models are general, that skin can be made glowing with this algorithm. And so there is these defaults, the technological defaults. And I thank you a lot for your attention. I look forward to talking to you again. Thank you.