00:00:09.440And whether we like it or not, we're in bed with the right.
00:00:13.540So, Adrienne, today we are talking about the aesthetics of AI and their weird, creepy,
00:00:20.900very uncanny gender politics that they have inserted into our culture.
00:00:25.500So AI imagery has become a real fixture in like certain kinds of social media. I see it mostly on Elon Musk's Twitter, which I guess now we are all calling X. And while it's fair to say that AI slop, this like particular kind of imagery generated by these machineries is pretty pervasive.
00:00:44.620I think it's also clear that there are certain populations online that gravitate more towards it to find it persuasive or they find it to be like a useful tool for their communications and political projects. Right. And there are others who pretty much reject it. And there are some places online where like using AI imagery is in fact like taboo. Right.
00:01:04.760So I'd go so far as to say that the use of AI imagery is becoming itself a kind of signaling tool.
00:01:10.940So whether you use it or whether you don't, it says something to your audience about you and about where you stand.
00:01:16.820And some camps are really partaking of this technology very liberally in creating images to promote their project.
00:01:25.100Yeah, and it's very telling that, like, it's very specifically tied up also with certain platforms.
00:01:30.300That is to say, Elon Musk loves AI and AI-generated imagery and seems to really want to center it on, as you say, the platform he now calls X.
00:01:40.580But, I mean, I'm guessing that the Venn diagram of embed with the right listeners and truth social users is two separate circles.
00:03:19.040Yeah, there's like a rosiness and hyper reality on like Donald Trump's face in this image. Like what I have said about AI before is that a lot of the aesthetic references that this imagery creates seems to be a cross between like the nostalgia and sort of unreality of like a Norman Rockwell painting and the like vulgar hyper reality of like pornography, right?
00:03:43.720It's like drawing from very ideologically specific image pools for its references.
00:03:51.360There's a famous line from Walter Benjamin that I keep coming back to when I look at images like this.
00:03:56.160Benjamin said that fascism consists of the introduction of aesthetics into politics, where communism responds by politicizing art.
00:04:02.800And yeah, there's a very specific aesthetics, I think, at work here.
00:04:06.900And I think this line offers a clue as to why AI slop specifically is becoming so ubiquitous during this particular historic moment.
00:04:14.720And to discuss that, we have a guest today, which is very exciting.
00:04:18.020Our guest today is Roland Maia, who is professor of digital cultures and arts at both the University of Zurich and the University of the Arts in Switzerland.
00:04:26.520And he's been doing amazing work on the aesthetics of AI generated images and their uses.
00:04:31.240And I found his work on this absolutely essential in understanding what the hell I'm even looking at here.
00:04:37.940Now, one thing I should say, Roland Bothe teaches at a University of the Arts, meaning he teaches practicing artists, and is partly trained as an art historian.
00:04:45.700So today, we're not going to be talking so much about the economics of AI slop.
00:05:19.140Adrian is like leading our domination of the German language media market, I think.
00:05:24.060Like he's just going to be like our like crusading spearhead and to like convert all German speakers into feminist radicals, for which I'm very happy.
00:05:34.880But it also means that it brings us these wonderful people like you, Roland, who I'm so excited to talk to.
00:05:54.860What are we looking at when we look at AI images?
00:05:57.140I think not necessarily every AI produced image is slop.
00:06:00.140Obviously, what you see on social media platforms, very quickly generated meme-like AI images, clickbaity images.
00:06:10.920So there are, of course, artists who are using generative AI, different kinds of models, different kinds of tools in very ambitious way.
00:06:19.120There is a whole discourse on what that means and whether it's correct or not.
00:06:23.240But there are different uses of so-called generative AI, but slop is kind of the cheapest way of using that for very quick reactions most of the time.
00:06:34.060So these images are both used as reactions to current events, like people producing image of
00:06:42.060the Hollywood sign in flames because they have the urge to somehow visualize what is happening and
00:06:48.060are not content for some reason with the images that are existing en masse. And these images are
00:06:54.540also meant to, yeah, solicit reactions from online audiences. So to have clicks, to have people
00:07:00.380sharing them um that's also a business model in parts that has been described for the kind of
00:07:06.620facebook ai slop the the trim jesus and so on so these are actually produced mainly in the global
00:07:14.780south by people who make a living of producing ai images that people then um both people and
00:07:20.860bots react to but the people are uh who actually make the money for them because they are then
00:07:25.980sent to phishing websites or shown advertisements and so on.
00:08:14.120Have they been circulating around far-right message boards for weeks by the time I see them?
00:08:18.580Or have they been, as you say, crapped out by a content farm in like Albania 12 hours ago by the time I get them in my feed?
00:08:27.720I think, but that's a bit of speculation on my part because I haven't done empirical research on that.
00:08:34.300I think the life cycle is, as you say, a bit shorter than like memes who have this kind of ongoing attractiveness because they are a kind of template that can be used to produce endless variations.
00:08:48.160But you have the kind of same thing also with AI-generated images.
00:08:53.240So I have all these kind of meme cycles where people react to images that have been shared widely and then they produce endless variations of that.
00:09:01.360also very much with right-wing or neo-fascist imagery. So there was a whole wave in Britain
00:09:06.920last year of people posting this kind of content that would show how a Britain that never was in1.00
00:09:14.820some kind of nostalgic past is now under attack, of course, by foreigners, migrants, Muslims.1.00
00:09:22.160And then people use this kind of template and use the hashtag Remember England to post1.00
00:09:26.340ever more absurd versions of an England that actually never existed anywhere with Britons
00:09:32.860on the moon and people having the Union Jack on everything you can imagine and having this kind
00:09:39.640of nostalgic, patriotic imagery running wild. So there was a kind of meme-like image reaction
00:09:45.760chain that I think went on for quite a while. But it's not like these now classic memes that
00:09:52.320are with us for decades and give us a template to express certain ideas still, although we all
00:09:58.760know them. Also, if you think of the Balenciaga Pope, that was a big thing for a few weeks,
00:10:04.280but now nobody would make another Pope AI meme, I guess, or hardly anybody.
00:10:10.100Right. So one thing I think that this is driving towards, you know, AI isn't just about a technology.
00:10:17.660It's about many things, right? Like it's not about the possibility of generating these kinds
00:10:21.580of images. As you're saying, this technology is, well, I guess I don't even know how old it is,
00:10:26.580but it's not that old. And yet already we have a pretty good kind of implicit taxonomy of like
00:10:32.700the kinds of things that it is used for, the kinds of people it appeals to, the kinds of stories it's
00:10:38.460used to tell. And I think that's key, right? Like when you look at AI, you're not looking
00:10:43.340just at a technology and saying, well, this can be used for anything. This actually helps certain
00:10:48.360kinds of content proliferate, right? I would say so, but I think it's important also to look at
00:10:53.280the technology on a very kind of basic conceptual level and think about, okay, what does it actually
00:10:57.980do and how does it function and what it is based on? And for me, thinking about AI-generated images
00:11:04.280is one way is to think as a kind of pattern recognition in reverse. So in pattern recognition,
00:11:09.880object recognition, facial recognition, you kind of label things in images, right? You train these
00:11:15.080systems to label every cat image out of millions of images of cats and dogs and whatever. So you
00:11:21.440have to train them with images that are already labeled as cat images. And now you can turn around
00:11:26.160this process and say, okay, give me a cat image. I give you a million cat images, produce me another
00:11:31.560cat image that kind of looks like all the cat images that you've been trained on. And this,
00:11:36.880I think, then also explains what makes this attractive, what it can accomplish and where it
00:11:43.740fails because what it does is it basically starts with noise with an image where you can see nothing
00:11:50.700and then it tries to find patterns that it has learned from billions and billions of images
00:11:56.640scraped from the web many people say stolen from the web whatever but it tries to find the pattern
00:12:02.260that are already associated with a certain label with a certain text with a certain description
00:12:06.160with a prompt and then tries to amplify that and that is kind of a process that goes step by step
00:12:12.560iteratively and it makes the image ever more readable and ever more legible in every step
00:12:17.540it's becoming more of a cat image it's become more of a image that you can read as the visualization
00:12:24.200of that concept and of course these concepts come from all over the web they come from social media
00:12:30.740they come from our whole visual culture from the whole archive of digital visual culture and all
00:12:38.000the stereotypes, all the cliches are very much baked into that. And not only are they baked into
00:12:43.120that, they are amplified within the process because the whole process is a kind of optimization of the
00:12:49.460image to become ever more like what you prompted. So you get a visualization of these formulated
00:12:55.300written concepts in the form of a visual cliche that is drawn from billions of images. That makes
00:13:02.420that attractive for some purposes, but there is already a kind of ideological bias in that very
00:13:09.520much. And I think we have to talk about that. Yeah. There's a really interesting point that
00:13:14.180an AI researcher at Stanford once made that I keep flashing back to. He said, look, if you train
00:13:20.700something on the past, it will repeat the patterns of the past. If you think about the patterns of
00:13:26.240our past, he's like, it's not shocking that this thing is pretty racist because you fed it on
00:13:31.720what's available and what's especially freely accessible, which tends to be older things.
00:13:36.040And so like, yeah, it repeats biases that are baked into the digital record that we've
00:13:40.600assembled over the last over 30 years.
00:13:43.160And the generative, quote unquote, AI promise is basically, as you say, pattern recognition
00:14:51.020And it also means that the kind of imagery that is uploaded most is the kind that's going to be reproduced most, right?
00:14:58.200Like just in terms of sheer quantity of what sort of pictures are online into this training data set that is like most of the Internet now for a lot of these technologies.
00:15:09.000Like that is what's going to feed into this style of imagery and inform it in the future.
00:15:14.980So like whatever dominates our Internet now or whatever dominates our visual space now, that is what is going to be reproduced in the future.
00:15:23.260Absolutely. But in a way, it's even worse because these mid-journey stable diffusion, they are kind of fine-tuned to a certain aesthetics.
00:15:31.860There was a great paper by Cher Thorpe and Krzysztof Buschek who really got into these data sets.
00:15:36.940And there are special data sets for the aesthetic refinement or fine-tuning of these models.
00:15:42.600These are much smaller data sets with images that have a very high aesthetic score.
00:15:47.460And that is a prediction by an AI of which images are most attractive to people.
00:15:54.120But people in that case are the people who produce the training data for that kind of predictive AI, people who already rated images online.
00:16:02.360And they could show, yeah, well, these people and these ratings, that's a very small demographic of mostly white, young, male, North American, middle class, very online guys.
00:16:13.120And their aesthetic expectations are baked into these models as a kind of standard aesthetic of what makes not only a cat image then, but a beautiful cat image.
00:16:24.180And that is not only kind of the statistical mean of what's on social media, but a very kind of specific aesthetics.
00:16:30.540And that, I think, also accounts for this specific kind of glow and this kind of filter aesthetic that you see a lot, the shininess, the game-like and fantasy-like image worlds that are predominant in this kind of image production.
00:16:45.060So there is like the whole web in all its aesthetic forms that informs this, but also a very small subset curated by the preferences of a very small group.
00:16:58.120Yeah, which the internet has been for a long time, but which was never, I think, quite as visually spectacular, right? This was true for Yelp, or even Google Maps initially mapped places that Google programmers tended to like, right? Like you could find your fancy coffee shop, you couldn't find the handicapped accessible soup kitchen or whatever, that wouldn't be on it, because they didn't care.
00:17:18.380Part of what we see in AI slop, I think, is exactly the fact that the democratizing functions of the Internet were never quite what they were cracked up to be.
00:17:26.460This was always the playground for the kinds of people that, yeah, as you say, look at the Elon Musk Roman centurion picture and like, oh, isn't that nice?
00:17:35.200I want to print this one out for myself.
00:17:40.920So I just wanted to ask you personally how you came to study these.
00:17:44.900I mean, it is a funny, so I'm a literature professor and I try not to engage with large language models because it's basically taking the thing I love and I spend most of my time analyzing and giving me a bizarro world, sloppified version of it.
00:17:59.720Some colleagues find it deeply interesting and I'm like, no, it's literally like if I had a stroke and then I wrote a bunch of texts, it would look kind of like this.
00:18:08.760And, you know, I don't need to explore that.
00:18:11.160How did you, as someone who's, as you say, partly trained in art history, gravitate to these images?
00:18:17.760Yeah, basically, my background is in German called Bildwissenschaft.
00:18:21.360So that would be like a parallel project to visual cultural studies.
00:18:26.000So opening up the field of art history towards popular images, scientific images, like a whole range of media images and so on.
00:18:33.880But with a kind of very specific twist in German discourse that they focus very much on what is an image like in the singular, what does the image do? What's the power of the image? And what interested me always more is what, okay, but what do many images do? How do people operate with images? What kind of functions do images have, especially in a context like surveillance or identification?
00:19:00.980So that's actually what I wrote my PhD on.
00:19:04.300And then, of course, you get into a history of facial recognition and its early beginnings.
00:19:10.840And for me, what then interested me is how with social media, you get large amounts of images, huge image populations that become a kind of resource of information for the training of facial recognition systems.
00:19:25.240So Facebook was among the first companies developing a quite efficient facial recognition software because they already had these masses of labeled facial images, of course, of all their users.
00:19:37.440And they also provided a whole environment of surveillance that these algorithms could then be used upon.
00:19:44.700And so that interested me, how do images become this kind of resource of information?
00:19:50.700And then when these things became popular, like three years ago, for me, it was the next step of that.
00:19:56.300Like you have the surveillance capitalist business model that extracts information from large populations of images.
00:20:03.200And now it's turned into the production of ever more images.
00:20:09.040And that kind of fascinates me, not so much the single images and what it means, but really these kind of operations that go on with images in a networked environment.
00:20:19.240I mean, it's interesting that you mentioned scientific images. I've been fascinated with those myself, especially when it comes to things like phrenology, craniometry, and all these 19th century pseudosciences, or the way evolutionary science in the 19th century operated with these images that were the facts underlying them were frequently true, but the visual presentation made some suggestions that just are not scientific.
00:20:42.620The whole thing is about plausibility, about how images make things plausible to people.
00:20:47.140And that seems to me, I can see how you get from that to AI, because part of what AI is doing is sort of pushing the frame or moving the Overton window on what seems plausible, doesn't it?
00:20:57.420Yeah, and that is also a kind of more direct link even, because in like 19th century phrenological, physiognomic image making, Francis Gelton and his composite portraiture, there is this fascination for photography.
00:21:12.100as a way of capturing data that can then be used for statistical purposes
00:21:18.240to then visualize, for example, the mean criminal, the kind of ideal face of the criminal.
00:21:24.780And you have this completely hallucinatory but very influential idea
00:21:29.500that you can use photography as a statistical practice.
00:21:32.900And that is what basically pattern recognition AI and also generative AI
00:21:37.280is a kind of statistical view of images that then becomes productive.
00:21:42.500And of course, it's also used for the same kind of phrenological,
00:22:05.600Yeah. So I think maybe to make explicit something we've been sort of alluding to a bunch,
00:22:11.080what makes AI slop hard to talk about is that AI generated visuals indeed look like most of our
00:22:16.360images, right? They in some way highlight the conventionality. They make conventions obvious
00:22:21.600by overdoing them, right? You write in one of your wonderful threads on Blue Sky that
00:22:26.160AI images match but overall fulfill our stylistic expectations. And I think it seems to me that
00:22:31.620our time just generates these insane amount of images and we've become on the whole basically
00:22:36.220less sensitive to how standardized they really are, right? And that they discipline our gaze
00:22:41.640in certain ways, that they get us to expect certain things, supply certain things. And I
00:22:45.380think what you were saying about phrenological images, what you were saying about physiognomic
00:22:48.520images is exactly that. These were used basically to train people in scrutinizing others in real
00:22:56.300life in certain ways. And AI, it seems to me, derives a lot of its visual plausibility from
00:23:01.960the fact that so many of our images were normed to begin with. That is to say, this is not really
00:23:09.180mid-journey screwing with our perceptive apparatus. It's mid-journey exploiting the fact
00:23:14.920that the previous 10 years of internet pictures have become so normed, so statistically
00:23:21.740graphable, basically. Have you guys ever heard of Norma and Norman?
00:23:26.300the statues no oh my god they're great norma and norman were these two statues of the statistically
00:23:34.320perfectly average man and woman that were brought around the country oh to like county fairs in the
00:23:41.560early 20th century in the u.s as part of like eugenics programming basically and like the idea
00:23:47.240was that the average was in fact aspirational right and fairgoers at these county fairs would
00:23:54.080have these like eugenics pavilions would be encouraged to model themselves after norma and
00:24:01.320norman and of course those statistical data sets that were used to create the average of norma and0.54
00:24:06.460norman in fact excluded like everybody black for instance right they were like statistically very
00:24:11.500tricky but there's something similar happening with ai right in which this homogenizing force
00:24:17.160is trained on a lot of exclusions and i think there's something about our real world even now
00:24:22.940that's furnishing this kind of homogeneity of images that AI then like feeds off and mimics,
00:24:28.040like how every actress in Hollywood has the same face now, because they're all undergoing these
00:24:32.720standardizing cosmetic procedures that homogenize their facial structures across what would otherwise
00:24:38.920be a lot of like just natural or ethnic or familial difference, right? So like AI stylistic
00:24:44.500repetitiveness, I think we can see that as just a continuation of the same meta trend.
00:24:49.620But I also think this might be a function of the maturing technology, right? Because one of the earliest observations that I had about AI imagery was like the uncanniness of things that got wrong and it's like tendency to fail at its attempts to imitate.
00:25:06.220Yeah, like the extra fingers, the extra teeth, the weird like misplaced shadow of an eye like next to the nose or something like it's excess of the features that it was unable to standardize.
00:25:19.460Right. And that was something where the technology's attempt to like be very loyal to repetition and reproduction was also ironically what created this monstrosity or these monstrous images.
00:25:31.040Right. Yeah. I mean, I always think of the fact that is this true or is this apocryphal that that when he cast Peter Lorre in the film M, the director Fritz Lang apparently looked at Cesare Lombroso's compendium of criminal faces and he never told Lorre.
00:26:19.040This is from a Twitter blue check user, no shock there, with a reply from Elon Musk from November
00:26:26.0402024. Ironically, it is a picture of Elon Musk with very floppy hair that does not look transplanted
00:26:32.640at all in front of a NATO flag and in a, I would say, Roman centurion uniform, although I think
00:26:39.080we'll be nuancing that later on, with the caption, thank you, Elon, for making the West great again,
00:26:45.180cross sword emoji, fire emoji. And Elon Musk replied, I have a centurion. Cool. Yeah. What
00:26:52.580do we say about this? Let's do some art history here, people. No, what's fascinating about both
00:26:57.840Musk and Trump, how they love AI. I think that's because AI loves them also, because it's so easy
00:27:04.020to make a Musk image, because there are so many images of him already in the training data set.
00:27:08.900It's quite easy to produce a Musk-like face, much easier than, I don't know, a portrait of Adorno, for example, which you don't really get.
00:27:18.280You get a kind of, I don't know, old, middle-aged philosopher guy who maybe has some likeness to Adorno, but he's not Adorno.
00:27:26.040But Musk, you always got Musk, you always get the Pope, you always get Trump, of course.
00:27:30.100But you get him here in an extremely exaggerated masculine form that obviously also flatters his idea of how he should look.
00:27:40.780So it's him becoming his own cliche mixed with these kind of gender cliches that are baked into the technology that are here performed or shown in the image.
00:27:50.780And then combined with extremely readable, legible, very obvious kind of symbols that are combined in a way that you can almost read the prompt from the image.
00:27:59.680I mean, there is no ambiguity in this image.
00:28:03.340It is kind of the visualization of Musk in a centurion uniform standing in front of a NATO flag.
00:28:11.700One thing that I think is worth talking about is the breastplate, which I think betrays another thing.
00:28:16.400And as you say, you can kind of read what this AI was trained on, which is it's a superhero costume.
00:28:22.320His pecs seem enormous on this thing, which is true of the Batsuit, I think, historically.
00:28:27.420And again, I'm not a specialist on Roman armor, but my guess is this is trained on Marvel.
00:28:32.740This has real Thor and or Captain America vibes rather than, you know, Russell Crowe in gladiator vibes, let's say.
00:28:40.280Or maybe both. I think the point is that it's a kind of synthesis, a combination of the two, and that for these models, all these kind of visuals exist in the same space.
00:28:50.580There is no categorical difference between a 19th century history painting or a film still from some Marvel blockbuster.
00:28:59.920They all are sources of repeatable visual patterns.
00:29:03.520And if they are similar enough and if they are tagged with similar enough text, then they form the space of possibility in which these technologies operate.
00:29:12.760So I think an important point to make oneself clear about these images is that they are based on this kind of huge archive of visual culture that is completely messy and completely flat.
00:29:24.420Everything is kind of in the neighborhood to everything else without any kind of categories or high and low or whatever, or historic, authentic versus pop culture.
00:29:35.660That doesn't play a role in the logic in which these models operate.
00:29:39.700Yeah. So my guess is whoever made this, as you say, we can read the prompt. We can imagine what the prompt was. It was probably Elon Musk, Roman centurion's uniform in front of gold embroidered native flag for some reason. But my guess is that this person ran this a bunch of times and picked the ones that they like because it's very noticeable.
00:29:59.860I mean, again, like from a physiognomic standpoint, that this feels it was designed for exactly the thing that happened, namely that Elon Musk would react to it, right?
00:30:07.780The man famously has a thing about his hair.
00:30:10.180His hair is gorgeously floppy in a way that pure Elon Musk's hair hasn't been since his mid-20s.
00:30:15.640His chiseled jawline is absolutely what Elon Musk wants himself to look like.
00:30:20.820Very, very obviously, if you watch his repeated surgical intervention, that's the face he's aspiring to.
00:30:27.120Meaning it almost seems like whoever generated this image basically kept hitting return until they got an image that's like, oh, Elon's going to love this, right?
00:30:36.700As opposed to one where he might look the way a 50-something-year-old ketamine user really does look who hasn't had a good night's sleep in 10 years and whose body is starting to catch up with his rotten soul.
00:30:50.040It does feel like it's both, as you say, completely statistical, where everything is the same.
00:30:56.360But then it's ultimately, it's almost like a dating site image, right?
00:30:59.960This image was put out there in order to ensnare exactly one guy, right?
00:31:06.300It wanted that reply, which is probably all that user needed to monetize their blue check, right?
00:31:12.580So they get money for this engagement to get 100,000 retweets.
00:31:16.180And the 20 seconds they spent making this thing paid off, right?
00:31:20.360So, like, independent of the politics and the, like, instrumentalization of these images, like, just for a second, something I've noticed about, like, this new genre or, like, generation of AI imagery is that they look realistic and unrealistic at the same time.
00:31:37.860And Roland, you have talked about this as a kind of, like, platform realism.
00:31:42.640Yeah, that's actually a term I borrowed from a colleague of mine, Jakob Birken, and then tried to run with it.
00:31:48.380And it's an attempt to synthesize a couple of observations.
00:31:53.000So first of all, these images are made for being shared on online platforms.
00:31:58.780And that's a lot of what their purpose is.
00:32:01.240They're very much based on online platforms, on the content that is produced there and that is then fed into the training of these models.
00:32:08.340And as I already said, the kind of aesthetics of these images also relies on feedback mechanisms that come from platforms.
00:32:17.780So every image that is clicked, that is shared, that is liked, that is upscaled, tells these companies what users expect, what kind of imagery, what kind of aesthetics.
00:32:27.620And that can be then turned into these kind of recursive feedback loops that produce a very certain kind of aesthetics.
00:32:35.420And the realism part, I think there is a couple of things there.
00:32:40.320The one is that most of these images are neither photographic nor painterly, but somewhere in between.
00:32:48.940So they imitate a painterly practice that is already imitating photography in a way.
00:34:43.500That whole package called platform realism, this kind of generic aesthetic of a second order that is already based on images that are already generic in a way, like stock imagery on social media platform content.
00:34:57.720This like slippage between the imaginary and the sort of like recourse of what's already been produced brings us, I think, to A.I.'s use by like proponents of the trad movement, like various social and gender traditionalists.
00:35:18.180And this is really like our meat and potatoes, Adrienne. This is the shit we eat for dinner.1.00
00:35:21.920Yeah, we made it through half an hour without mentioning the trad wives. I was getting worried for us. But yes, that's where my mind went to.1.00
00:35:27.720Because there's like this deeply traditional visual vocabulary, right? And like, duh, that's because this tech is trained on these homogenized, highly conventional sets of images. But then it's also a strange and uncanny because it's stripped of context, it's stripped of actual human meaning making practices, and it's stripped of like, what we might think of as the elements that add authenticity to history, right?
00:35:52.860So, Adrian, you and I talked about this one.
00:35:56.280It's actually a short video that was produced on Twitter.
00:35:59.260We talked about it in our very fun conversation with Matt Bernstein for his podcast, A Bit Fruity.
00:36:05.240And it's a video I saw originally posted on X.
00:36:09.380It's about six seconds long by a Twitter user named Elijah Schaefer, who says,
00:36:16.600this type of content awakens in a man something so primal that not even an OF model in lingerie
00:36:22.520could compete. And then Adrian, what is this video depicting? So it's a, I'll go back to
00:36:29.140your description of it, which was fantastic. It is a young woman slash girl, more shading into
00:36:35.420girl, I would say, in traditional garb, kind of grays and a little bit of purple. Yeah, there's
00:36:41.420an apron and puff sleeves involved yeah it's like a pinafore situation yeah demurely holding a bowl1.00
00:36:47.760of four eggs with a bunch of chickens behind her and i think your joke was well yeah because she
00:36:53.040is a only fans model in an apron this is ai right i'm almost certain that it is i believe it's either
00:36:59.880ai or it's partaking of so many of the like conventions and aims of ai imagery that it
00:37:06.360almost doesn't matter. But it looks like AI to me.
00:37:09.040And movements are very strange. Either the camera did something odd or this is, in fact,
00:37:14.340an AI image. It's not a way you would actually move if you were in a body that you inhabited
00:37:18.900for your own purposes. Maybe if you are what looks like a trafficked Slovenian teenager who0.99
00:37:24.400has been slapped into an apron and put in a yard with a bunch of chickens, maybe you move in that0.96
00:37:28.840stiff and self-conscious and unnatural way. Yeah, if you're in this image, reach out to us and be
00:38:02.080And so this like Elon Centurion image, which Musk clearly read as Roman, the armor actually looks more like something you'd see in like a Warhammer video game.
00:38:12.120And Adrienne, you and I have talked about this concept from Alexandra Minna Stern, who talks about right-wing aesthetics as partaking of something she calls chaos futurism.
00:38:22.080So like these people like Musk, this like apron girl with the eggs, they don't look like actually period.
00:38:29.420They don't look like archaic or like historically accurate.
00:38:33.140They look like old timey and like futuristic at the same time.
00:38:36.160Yeah, there is this really interesting kind of meeting of the aesthetics.
00:38:40.700And I think that, Roland, you would probably say this is about the fact that video games are overrepresented in the image databases on which these things are trained, is my guess from what we've been saying so far.
00:38:53.360There is this very strange combination.
00:38:56.320It's not that the AI doesn't care about, period.
00:38:59.340It seems to compulsively combine the futuristic and the archaic.
00:43:10.180it looks like it's very video gamey, right?
00:43:12.880It looks like a bug carapace made of metal.
00:43:16.580And like these are categorizations that don't just like misattribute their distinctions, but actually like in practice eliminate them.
00:43:24.840Right. Because like for Musk and his purposes with this photo, that like is centurion armor.
00:43:31.960And so the relevance of the distinction becomes a bit like hard to pit down, at least in the discourse that these images like generate around themselves.
00:43:39.060So it's like doing for visuals what being on Twitter all day does for words, right? Like if you're not careful, the context collapse can erode you like intellectually and psychically so that you start saying things like you do not under any circumstances have to hand it to ISIL to like your grandma when you finally log off.
00:43:57.900So it's not just the mistake of the categories, it's actually their conceptual collapse, like the elimination of these distinctions altogether, while also like the eradication from the visual vocabulary, anything that doesn't fit into this like homogenized ideal.
00:44:10.760An interesting point about these models is how like every historical aesthetic and historical can mean also very recent, becomes some kind of nameable and repeatable style.
00:44:23.140And you can produce everything in the style of, and if you're more ambitious, you can combine lots and lots of styles, both styles of kind of individual creators, artists, and so on.
00:44:34.780But also a Polaroid photograph is also kind of a style.
00:44:38.680Every visual appearance, every look becomes a style and then can be recombined in prompting these kind of images.
00:44:46.280And that's extremely interesting because it's, for me, as a kind of art historian, how this category of style becomes extremely expanded and completely flattened, as well as the idea of history.
00:44:58.220Art history style was very much bound on history, and now it's like a whole resource of visual patterns that can be freely combined.
00:45:07.640I think it has a lot to do also with stock photography, I already mentioned, but also with mood boarding.
00:45:12.860So you have these aesthetic practices of recombining certain vibes, certain moods and fusing them together.
00:45:21.620So in mood boarding, you have them spread out and now you can curate vibes and moods and synthesize them into one image that looks like a single image, but actually it's a synthesis of untraceable influences and images that come before that.
00:45:37.340So maybe we should switch lenses very briefly and talk about it sociologically.
00:45:41.760That is to say, did something like mid-journey sort of migrate to the right over time?
00:45:47.360Or were far-right fora and sort of return accounts earlier adopters than others?
00:45:53.480I think that's a great research question, actually, for some kind of media sociological recent history.
00:45:59.720So I can only give an anecdotal kind of impression.
00:46:02.980And I have the impression that quite early on in 2023, when it started, it was at least not dominant.
00:46:09.840The far right was not the kind of dominant user group of these tools yet.
00:46:15.760I think that started very much around early 2024, at least then I noticed it.
00:46:22.380So first it was a kind of nerdy, people trying that out, very much debating also that there
00:46:28.900were big clashes of people online between people coming from all kinds of artistic backgrounds,
00:46:34.160very much hating it and others very much finding joy and fun in it. But I think it was not
00:46:42.140obviously politicized yet. But I think that started like some months or a year after that.
00:46:50.140And now it's very obvious that AI slop, as we say, is kind of the aesthetic of digital fascism. But
00:46:55.720I think that is a development. But it would be interesting to trace how that actually evolved.
00:47:01.200Well, I mean, in some way, we're having two creator economies sort of playing off against each other. One is the creator economy that generates a Roman centurion image and hopes Elon Musk retweets it and then makes a little bit of money off of it.
00:47:12.460And then there are these genuine creative communities, I would say, on places like Tumblr, who are deathly afraid of this stuff because it's going to destroy what little income they still get by drawing fantasy RPG character art or by illustrating a Kickstarter or making pornographic images of people's favorite cartoons.
00:47:35.000I mean, like all respectable forms of work, but all threatened by AI.
00:47:39.580And I do think that there is a kind of, there are two different kinds of class politics,
00:47:43.140I think, also smashing into each other here.
00:47:44.940And I think as people have gotten more triggered by AI images, just on a purely visceral level,
00:47:50.100being like, this is offensive to me as a creator, it's become one more way to trigger the libs, hasn't it?1.00
00:47:54.820It's just like, there's a good reason why Elon Musk loves this shit.0.99
00:47:58.600Part of it is that he looks great in it and he doesn't look great in real life.0.99
00:48:01.760but part of it is clearly also that like he knows we're going to hate it right it's there for the0.99
00:48:07.300anti-fans as well and we can just be like can you believe what this fucking guy just shared0.95
00:48:11.520i think that's a good point and a very fitting observation because i shared0.74
00:48:15.860i generated images and i don't know 2023 and i more or less stopped it and now only share other
00:48:23.460people's ai generated images and comment upon them in part because of course it triggers a lot
00:48:28.700of people. And rightfully so, there is a lot to hate about that. And that made it more and more
00:48:33.200attractive for the right. I guess that's exactly what's happening. I mean, on Luska, you can see
00:48:39.060that I'm blocked by people because I shared AI-generated images. But of course, right-wing
00:48:44.740accounts are not afraid of being on the block list of some creators or fantasy artists or
00:48:50.440illustrators or whatnot. That's kind of an honor bet for them. This is a dynamic where it is clear.
00:48:56.280Now it's a statement if you use it, more or less.
00:48:59.220Although people still try to use it for, let's say, progressive purposes or for making fun of the right.
00:49:07.380But it's the question of whether that actually works and it's a minority.
00:49:11.340Yeah, like for instance, there is the video of Volodymyr Zelensky punching Donald Trump in the face.
00:49:16.380It's not all in one location, but you're right.
00:49:19.060It's a good question whether or not that actually works.
00:49:21.720visually contesting this kind of neo-fascist rhetoric that really has become coextensive
00:49:27.600with AI slop. Can I ask you briefly about the, you mentioned in the beginning, these kind of
00:49:31.740hashtag remember England pictures, and you put them in our planning doc here. Is the second one
00:49:36.960supposed to be a parody of the first? Maybe I'll describe what I see. I see the Houses of Parliament
00:49:40.940with a British soldier returning from what looks to be World War I. Well, it's supposed to be0.99
00:49:45.700Dunkirk, I guess, 1940, facing a bunch of Muslim women, right? And it's very clearly playing into1.00
00:49:54.080all kinds of racist, you know, great replacement myths. But then on the right, we get one. My
00:50:00.000daughter asked me if I remember when English buffins put a man on the moon. Of course I do.1.00
00:50:03.960We all do. And then there's a picture of a cross of St. George on the moon, which has a moon in it
00:50:10.260for some reason. Two lions, a girl on a bicycle. What's happening there? Is this person making
00:50:15.840fun of the first or is this, do people not care what they're putting up anymore?
00:50:20.120It's making fun, as far as I understood it, of a whole wave of kind of right-wing
00:50:25.220nationalist imagery that played upon this idea of, oh, remember the old England before
00:50:31.960the foreigners came? And that also already had these lions in there sometimes, also with a
00:50:37.600strange kind of racist animal kingdom-like imagery.
00:50:43.180And they took these elements and kind of recombined them
00:50:46.800in a way that it got more and more absurd
00:51:30.920So I have wondered about whether or not they just lost a lion somewhere along the way.
00:51:35.680Yeah, I think this also like touches on or brings us to AI's very tricky depictions of race, because a lot of these models don't do very well with depictions of people of color, right? They either tend to homogenize non-white people's features into those of like an idealized white person, so that often has like skin bleaching effect, or they will depict them as just like flat out racist caricatures, right?
00:52:01.740So I think in a lot of cases for the users of AI, this is like a feature, not a bug, right?0.78
00:52:07.080Because like if AI functions for the right as a kind of like visual wish fulfillment, it's pretty clear that one of those wishes is for an all-white world, or at least for a world in which non-white people are in clear subordinate positions.0.70
00:52:21.320And that's worth pointing out that there's the famous example of the, I believe, hand dryers that would only react to white skin, right?0.81
00:52:29.120This is a classic thing that like how these models are trained and who they're trained by does tend to encode very real biases in their outputs and tend to reiterate invisibilities and lacunae in whatever the record and whatever data this thing is trained on.
00:52:44.280And we'll reproduce those and sort of make them our future, make the past mistakes, basically our future biases.
00:52:49.740Right. But I think that's in some sense like by design for a lot of the people who are using this technology like most enthusiastically.
00:52:55.900And that might bring us around to like the elephant in the room in basically all discussions of AI imagery, which is AI's use in generating pornography, particularly like non-consensual pornography or deep fake porn.
00:53:09.480And when we talk about AI, we're basically like mostly talking about porn, especially when we're discussing like videos and moving images.
00:53:18.340So like researchers at the AI monitoring company Sensity estimated that 90 percent, 90 percent of deep fake videos, so not the still images, but the videos that are posted online are of pornography and that of those 95 percent feature images of real non-consenting girls and women.
00:53:39.480So this is like something that is just now a new tool of like very older forms of sexual abuse, right?
00:53:49.340Like this technology has become very easy to use, very cheap or free to use, very easy to find online.
00:53:57.020And for a man or really a boy to make a pornographic video of somebody he wants to target or humiliate,
00:54:07.120He really only needs like a couple of photos of a woman or girl's face. You can do this with like three or four still images can be used to make like not a perfect, but like a fairly convincing and certainly very like uncanny, disturbing pornographic image and video of her. Right. So this is like a pretty standard part of the job of like any woman who has like any kind of public role. Right. So it famously happened to Taylor Swift. It happens to AOC. But you don't need to be like really famous.
00:54:35.380Like this has also happened to me, for instance.
00:55:04.340And that is like a form of misogynist harassment and coercion that this technology has just enabled.
00:55:11.700It's just basically like a high-tech manifestation of what are like basically low-tech, more conventional forms of sexual harassment.
00:55:18.800So AI imagery becomes wish fulfillment, but it's not only a way to like gratify the solitary impulse of the mind, right?
00:55:27.960It's also a way to affect somebody else's status in the real world, be that if you want to flatter Elon Musk by making him look a lot more muscular than he really is, or if you want to humiliate your classmate in the sixth grade by showing her going down on some random man, right?
00:55:48.560It is a way that these other kinds of images can be used to encompass people who would not be in that category of image, right?
00:55:58.640Because like the AI porn videos that they make are trained on real porn videos, which are ubiquitous, but they are kind of repetitive.
00:56:05.300So it's very, very easy to create generative AI of that image and just put on another face.
00:56:12.160So then there's this other situation where AI is being used to create images of women who do not exist. And those women that it summons into being are like very specific, right? They all kind of look the same. They look young. They are like somewhat like uncannily clear skin. They have like an almost Pixar, like plastic-y quality. They always have long hair.
00:56:37.860They always have very big eyes that to me look like just a little too close together, like a predator eyes, like a cat that are like right on the front.
00:56:46.900And I'd like to take you guys through a strange little artifact I found, which is an AI generated article featuring AI generated images called the most beautiful person from every country.
00:56:59.940And we'll note that all of these AI generated fake persons are very young and that they are all women.
00:57:07.860So you can see like what AI thinks every country looks like, like the most beautiful, famous, like creepy AI person in Denmark is wearing like a parka with like a fur lined hood because it's cold in Denmark, right?
00:57:24.300The Australia, this is Austria. Sorry, why is the Australian one wearing a dirndl? No, okay, that makes more sense.
00:57:31.320Like the Ireland one is, it has red hair. It's stuff like that. It's like ethnic stereotypes.
00:57:37.220Is the United States one just Taylor Swift?
01:06:18.240like the real Rafa does not have any high alpine peaks right in the background and the tents are
01:06:23.820not arranged to spell out the slogan, all eyes on Rafa. And you're like, yeah, I think people know
01:06:28.620that. Like it's a little weird. So there's something about, there is a kind of fixation
01:06:33.300on AI that can also obscure what people are doing with images. I do think it might be fun to talk
01:06:40.480about this example because this was a kind of, for all intents and purposes, a use of AI for a
01:06:45.760very different kind of political message. Yeah, I mean, it is so strange that this image should
01:06:50.180be debunked because its whole message is obviously literally readable in the image. That's what it
01:06:56.480says. The whole purpose is to spread the image as a carrier of actually a text. And it's not about
01:07:03.720what it shows, but really only about what it says. And then debunking, I think, has become
01:07:09.420standard reaction to viral content. There are kind of two standard reactions online if something
01:07:15.080gets really ubiquitous and you see it everywhere and people are always urged to react towards
01:07:21.620images also with images and either they produce memes and variations and make fun of it or they
01:07:27.320get this kind of forensic gaze and try to find clues that it's somehow manipulated that there's
01:07:33.820something wrong with the image and they find it also in images where it's so obvious that it's
01:07:39.340a synthetic image, but there is this kind of fun in finding clues of manipulation that
01:07:44.860AI images very much lend themselves to because they have these kind of strange, weird little
01:07:49.940details if you look longer. In this case, the details are very obvious, but I think it's a
01:07:54.720standard reaction mode, this debunking, and I don't think it leads anywhere.
01:07:58.940Yeah, exactly. The question, what is this image trying to do and do I agree with it or not,
01:08:03.900is a much more interesting one in some way than saying, oh, this is AI, right?
01:08:08.520And as you mentioned, what does the spread and the global distribution of this image tell you about the global distribution networks of images today and how matter is moderating content and which kind of content becomes visible and which kind of content becomes invisible?
01:08:25.960That's an interesting thing about this image.
01:08:27.940Not a question whether this actually shows a realistic scenery of Gaza.
01:10:26.660And then a whole genre of online quizzes sprung out of that, where you can guess if an image is AI-generated or not, which is in itself a very interesting kind of genre, because it always tells you, yeah, the non-AI-generated images are really true and authentic, and only the AI-generated images are the fake ones.
01:10:47.300And now we all have to learn how to spot these little clues and details.
01:10:52.120And also from the Pope AI image, a whole wave of other AI-generated Popes flooded the web.
01:10:58.660So there was this mimetic kind of reaction change to what that image also.
01:11:03.120It's now a classic, I think, of the genre.
01:11:10.180And I do think that, yeah, anything that involves a Pope doing fashionable things is exciting.0.97
01:11:14.780I always think of the ads for the, remember when Jude Law was the young pope and everyone was like, oh, the tagline better be this pope fucks.0.92
01:12:01.700I mean, for me, it's the best example of what the reality these right wing accounts want to see in these images actually is.
01:12:10.940It's the reality, as we spoke, of gender cliches, of a world where everyone kind of matches a certain already established pattern and formula in the most stereotypical way.
01:12:22.820So I think they mean that in a way serious.0.96
01:12:26.840That's how real women are supposed to look like for them.
01:12:30.540And that, of course, means that everyone who doesn't look that way is less real.
01:12:35.420And I think there is a threat in this image very much.
01:12:40.940Yeah, I also think that there's something interesting about, right, like, where does the reality lie, right?
01:12:45.500Is the reality of what a woman looks like every woman you see, or is it the category of woman in some, right?
01:12:52.520This was the appeal of phrenologists and of physiognomists in the 19th century.
01:12:57.800They're like, well, no, the individual face is not that telling, but the human face as, as you say, statistical composite tells us what we are really like, right?
01:13:08.100In the aggregate, the real portrait of the human species, or in this case, the German woman, emerges, right?
01:13:15.320Like, where does reality lie, in the concrete or in these kind of abstractions?
01:13:20.780Yeah, the abstraction from online content in that case, abstraction from even fantasies kind of hypercharged through these technologies, but presented as something with the aura of statistical objectivity in a way.