Video Transcript
For most of my career, one of the big themes that I've been trying to understand is: what is our relationship to infrastructures and to technologies, and how do the things that we build change us. In other words how do we change the landscapes around us, and in doing so, how do we change who we are.
So I've done a lot of work looking at things like communication systems, surveillance systems -- you know looking at technologies that we use to interact with each other -- and I started thinking about computer vision and AI in about 2012 because I thought something really weird is going on here, which is that we're starting to build machines that see the world for us and do things in the world for us. As somebody who's spent many years thinking about vision, thinking about images, I know that there's always politics to the way that we see, I know that there's always a subjective experience that we bring to how we see.
There's nothing neutral about visual perception, and so I wanted to understand if we take that as a given, that seeing is never neutral, then what kinds of politics, what kinds of values, what vision of the world, for lack of a better word, is built into the computer systems that we're making to see the world for us, and how in turn is that going to change us.
I initially was working with computer vision labs, and over time those tools became more accessible. We started building computer vision systems in my studio and then there was this so-called machine learning revolution that happens in the late 2010s. So we started building early models and then started dissecting those models. I wanted to understand what is in this model, what way of organizing the world's information are has the model been trained on. So I spent a lot of time looking at what are called training sets and data sets; you know the information that you give to a machine learning system which then of course structures the way that that system works, and when you do that you find all sorts of strange and often terrible things.
The work here in the exhibition is called âFrom Apple to Abominationâ, and what it's looking at is one of the main data sets that's used in machine learning -- mostly research and development work -- it's a data set called ImageNet, and it's a data set that was created between 2009 and 2011. The people who made the data set intended it to be what they called âa map to the entire world of objectsâ; what they wanted to do is collect a database of pictures of everything in the world, and the idea was with this database, then you could train a computer vision or machine learning system to recognize everything in the world.
So what they did maybe makes sense theoretically? They took a dictionary, they took a special kind of a dictionary where all of the words that meant the same thing only had one entry, right, so there's a lot of words where there's two words that mean the same thing, so in this dictionary they there's just one entry for it; and what they did was then they took all of the nouns, so they threw away the adjectives and verbs, and they said well if you are a word, and you are a noun, then that is something that corresponds to something in the world, and therefore we should collect pictures of that thing in the world.
So an apple is a perfect example: an apple is a noun we can create pictures of apples and then maybe train a machine learning system how to see an apple. Now the problem is that they did not really take into consideration that all nouns are not equal. We have abstract nouns on one hand -- a noun like consciousness, like that's a noun but it doesn't - there's no image of what that is that's in the world. Now there's other nouns that are racist and sexist and misogynistic and horrible, and you can think of a lot of nouns that fit that category, and so the problem that they had was that they took all of these nouns and they scraped the internet for images that corresponded to these nouns, packaged it all together, and said okay everybody go use this now, and it became a standard, and they never bothered to look inside the data set to see what was actually there.
They thought âOh it's it's too big, how could anybody possibly do that?!â They had about 22,000 categories of objects in there, and about I think about 14 million images, and I thought âWell I can look at that in a dayâ, and so I started looking at it and starting find these horrible things. So this piece is really about that it's about that problem of classification, it's about the problem of trying to build a computer vision system or a machine learning system to interpret the world, because we might think of an object like an apple and we can all agree that maybe there's such a thing as an apple in the world and maybe that's a picture of it, but very quickly that starts to get much-much more complicated, so that's really what the piece is about is: telling a story about things that seem self-evident gradually and somewhat frighteningly becoming unhinged.
There's many-many problems that you see on this journey, and an obvious one is is what we call bias -- that if you have a category like CEO maybe it's all like white guys, and if there's a category like criminal then it's all black guys, and so you have this this racist thing going on, or ethnocentric kind of thing going on very often. But it gets even more complicated than that, you know -- if you think of a category like man or woman, you know, we might say okay, well in in our everyday lives we walk around and think, oh, well that's a man, or that's a woman, but that's not really how gender works, you know. I think nowadays we ask people how they identify -- maybe a man, maybe woman, maybe neither, maybe both, maybe you know, it's complex, and so that idea that we're just going to take a picture and then attribute a gender to an image, you know, that's a problem.
Now it gets even harder than that, like one might argue that men and women exist in the world and there's some kind of relationship between masculinity and the human condition or whatever. We take a word like criminal, which is a noun in the database -- the idea of criminality is a historically specific idea -- what even constitutes a criminal is constantly changing, so that is a noun but it's not fixed at all, it's entirely relational, and so how could you have an image of that - it makes no sense. That's the kinds of problems that you start to see, and they're complex problems â right? - and when you when you talk to people in the industry they'll think, âOh well, we'll fix this with more data,â and the argument that I make is: these are not solvable problems with computers because these are not quantitative questions, these are qualitative questions, they are cultural questions, they're historical questions, they're political questions.
So is AI going to take people's jobs? Yes, absolutely. Is AI going to in increase economic inequality? Yes, absolutely. So there's problems with both of those. In terms of how AI will influence our own subjectivity, that's a question where I think it gets really weird â right? We have the beginnings of that with what are essentially surveillance systems using AI. So for example, your car spies on you, sends information about how carefully you drive and how closely you adhere to traffic laws to insurance companies, who can then use that information to modulate - you know - how you pay your insurance, and what what they want to do in that industry is moving towards a model in which they call it âpay as you go insuranceâ. So as you drive your insurance rates go up and down based on - you know - how safely the AI thinks that you're driving, and how closely you follow its recommendations. So that's a very simple example but very obvious - like that's quite a transformation. Where I think it will get much much weirder is in what we call parasocial relationships -- so the relationships that we have to each other with you know online figures, whether that's a podcaster, or an influencer, what have you.
We look at a phenomena like Tik Tok, for example, famously destroying our brains because it's, you know, recommendation algorithms are very good, it's a technology that's gotten very good at keeping our attention by giving us things that it it thinks that we want to see; it works. Now there has of course been over the last, you know, really year or two, this turn towards generative AI, towards making images, towards making texts, towards making videos, obviously something like ChatGPT would be in this, Mid Journey, Stable Diffusion ... these sorts of tools. Now what is going to happen, is that increasingly the entities that we interact with online will be generated for us individually, right. So right now people have to make Tik Tok videos and the algorithm curates them in order to capture our attention. In the very near future, those videos or that media will just be individually generated for us â right - and the machine learning system will learn what we are reactive to most, and obviously this is not going to be done in a vacuum. The point of this kind of media is to extract value from you in one way or another, whether that is your attention, or whether that is - you know - surveillance data in order to monetize one way or another, or whether that is to influence you to buy something or vote a particular way, what have you -- and that's where I think this starts to get really weird, where we increasingly will be living in a world where: each of us - the media that we interact with - will be individually generated for each of us, and will contain dramatically different worldviews that are attempting to manipulate us in ways that are precisely tailored to our own individuality, and this of course will you know amplify a lot of trends that that we've been seeing already - you know - things like polarization, people having the the lack of a shared worldview, the political manipulation obviously - things like disinformation ... I think that's going to be far more dramatic than what people are imagining. My concern is that that although these types of media will create different worlds for each of us, that will have not that much overlap with one another, and when you do that ... I'm not sure how you have a concept like democracy.
When we're talking about this generative turn, I think we see the beginnings of the dynamics that-that will have when we look at things like a Tik Tok algorithm, or even Instagram, Twitter -- these are content platforms that have gotten very good at pulling on your emotional strings, your intellectual strings, your attention, and when with generative media that process of doing that will become much more efficient, for lack of a better word. In other words I'll be able to generate media for you and measure the reaction that you have that, and very quickly learn how to optimize something for you. âMyâ goal in generating media for you it is to sell you advertisements - right - and people think that's the end of it, but no-no-no-no that's the beginning of it -- I want to extract value from you, so your attention is a kind of value, but I can also try to get you to buy something, that's kind of value. I can try to get you to vote in one way or another, or have a political position. I can get you to try to adopt a particular identity that is going to have certain economic interests associated with it - you know - I can extract information from you that's useful far beyond advertising ... if I know about like what kind of food you eat, I can sell that to your health insurance company ... if I know you know how you drive I can sell that to your auto insurance company ... but there there's a lot to be done here. So this does create a world that is characterized by cultural political and social atomization, like that's a feature of it, that's the point - right - and why is that bad? I think we've seen the beginnings of why that's bad already, with - you know - certainly in the United States, where I'm from, youâre seeing the most bizarre forms of politics kind-of emerging from the fact that we're increasingly not living with the same worldview. That can go much-much further - you know - we're just seeing the beginnings of that.
So I think when we ask the big question why is that bad, I think ultimately two reasons: one, it is a process by which increasingly the market or the economic system is turning us against ourselves in order to extract value - right - so you can think about it as a kind of colonization or an extractive practice from places that previously were not subject to a market logic - right - so you can think about it as being analogous to like like a process of colonization or a process of occupation. These are the kinds of metaphors that I would use for that, but it's our everyday lives, it's our brains, it's our subjectivities ...
The other thing that is bad about that is weâll increasingly have a world in which consensus will be impossible, because there is no basis for that, and the consequence of that is: I don't know how you have a coherent political system when that is the culture that you're living in. And - you know - thirdly ... this will benefit very few people a great deal, but that will come at the expense of - you know - everybody else, and so again seeing this question of inequality is very much a part of that.
These are complicated dynamics; all of these - you know - the machines influence, us we influence machines ... what often gets left out of that kind of conversation is thinking about who owns the machines and what are they being optimized to do? Who owns the machines - like - people trying to make money - you know. What are they optimized to do? Make money - you know. So like, that's so we can frame the question differently than like -- oh there's these machines, and then there's the humans, and to what extent do they work in accord or in contradiction to one another ... I think it's much simpler to say something like - you know - the machines are engines of capitalism and what they're trying to do is extend that logic of capital to as many places as possible, and to make it as optimized and efficient as possible, and that comes at the expense of places that we've had, whether that's in our brains, or in our collective experience that were - you know - relatively insulated from from those sorts of quite brutal market logics.
So I think in the work that I do, very broadly, comes out of a landscape tradition, if you want to think about that way -- looking at the world, and also trying to understand how we are looking at the world. Obviously with machine vision and computer vision it becomes: How is the world looking at us, right? What I think art can do is help us to break apart different forms of common sense, whether that's things that we take for granted - without understanding that we take that for granted, or things that are around us that we don't think twice about. Art is a vehicle through which we can pay much closer attention to the environments around us, the technologies that we use - you know - the landscapes, for lack of a better word, and art can uniquely allow us to help us learn how to see them, and I say see very literally ... art can show you something that a newspaper article cannot show you, that a journal article cannot show you, that a book cannot show you ... can literally help you learn how to see. So in terms of how art can help create a - you know - more equitable or joyful, hopefully society, I think it can help you see the world around you differently, and when you do that, the hope is that helps you imagine what a better society could be.
