Navigation – Plan du site

AccueilNuméros9Dossier“Platform Realism”. AI Image Synt...

Dossier

“Platform Realism”. AI Image Synthesis and the Rise of Generic Visual Content

Le « réalisme de plateforme ». L’intelligence artificielle générative et l’essor du contenu visuel générique
Roland Meyer
Traduction(s) :
Le « réalisme de plateforme ». L’intelligence artificielle générative et l’essor du contenu visuel générique [fr]

Résumés

À chaque mise à jour, les modèles de génération d’image par IA tels que DALL-E et Midjourney promettent des représentations toujours plus « réalistes ». Comme le montre toutefois cet article, ce soi-disant « réalisme » trahit des biais favorisant des valeurs esthétiques et des préférences idéologiques liées à un public majoritairement blanc, occidental, masculin et issu de la classe moyenne, tout en étant intrinsèquement générique. Les images générées par l’IA sont produites à partir de milliards d’images existantes, filtrées au moyen de concepts linguistiques, adaptées aux normes des sociétés et optimisées pour les attentes des consommateurs. On nommera cet effet le « réalisme de plateforme » : une esthétique de second ordre reposant sur des images génériques. Cet article développe ce concept en replaçant l’esthétique picturale de l’IA générative au sein d’une histoire plus vaste du contenu visuel générique, en examinant son infrastructure au sein du capitalisme de plateforme contemporain et en esquissant ses possibles conséquences pour la culture visuelle numérique.

Haut de page

Texte intégral

1In May 2023, artist Bjoern Karmann introduced a gadget called Paragraphica, a lensless “camera” based on AI text-to-image synthesis (fig. 1). For Karmann, Paragraphica gives us a glimpse into how generative AI “sees” our world—a “seeing” not based on the optical registration of light but on the collection of data. When you look through the viewfinder, instead of a visual representation of the scene before you, the camera gives you a textual description based on locational data, including time of day, weather, and nearby points of interest. Moreover, if you push the button, this description is turned into a prompt, given to a text-to-image generator, and thereby used to produce what looks like a photo, though it is entirely different. In Karmann’s words, “The resulting ‘photo’ is not just a snapshot, but a visual data visualization and reflection of the location you are at, and perhaps how the AI model ‘sees’ that place.”1

1. Bjørn Karmann, Paragraphica

1. Bjørn Karmann, Paragraphica

A lensless “camera” based on AI text-to-image generation, 2023.

© Bjørn Karmann

  • 2 On the notion of the “invisual,” see Jussi Parrika, Operational Images: From the Visual to the Invi (...)

2Not only does Paragraphica not produce photographs, it is also not a camera in any meaningful sense of the word, and it may not even be part of the history of “optical media,” as written, among others, by Friedrich Kittler. There is no opening to let the light of the outside world in, and no image is projected inside onto the back of the device. There are only flows of invisible contextual data channeled through open APIs and blind statistical processes based on mass quantities of training data scraped from the web, including millions of images produced by photographic cameras. The visible world thus enters Paragraphica not as a simulation but only as a faint and distorted echo, mediated by “invisual” data.2

3Karmann has adorned his gadget with the replica of a star-nose, the sensory organ that allows certain functionally blind moles to find their way in complete darkness. He also added reels that allow you to adjust the parameters of the AI model. However, as Paragraphica is not a camera but a black box, it is nearly impossible to know how these influence what happens inside. Indeed, it might even be that Karmann’s prototype is a fake. If so, it is pretty realistic—everything he describes could be realized with current technology.

  • 3 Shane Denson, Post-Cinematic Bodies (Lüneburg: Meson, 2023), 187.

4Today’s computational photography sometimes already works in a way that is quite similar to Paragraphica. If, for example, you take a picture of the full moon with your Samsung phone, the moon you will get might be a synthetic simulacrum. The cheap optics and sensor technology in most phones are not well equipped to work under difficult lighting conditions, which is why AI-based optimization algorithms are routinely used to produce hi-res images out of noisy data. As one Reddit user found out, such “optimization” even works when you re-photograph a pixelated photo of the moon; the smartphone’s AI filters recognize what you are trying to photograph, automatically add all the details no longer visible in the sensor input, and turn a blurry bright sphere into a hi-res image of the moon, with all the craters rendered in crisp shadows. In other words, the software knows what we expect from a photograph, as it has been trained on millions of images that have already been taken, and thus produces, in media scholar Shane Denson’s words, images that “anticipate their viewer’s gaze.”3

  • 4 Sianne Ngai, Theory of the Gimmick: Aesthetic Judgement and Capitalist Form (Cambridge, MA: Belknap (...)

5Paragraphica is not just a black box but also a “gimmick” in the sense that cultural theorist Sianne Ngai has given this term. For Ngai, the gimmick is both a cheap trick and a technological marvel, a device that promises to save labor while simultaneously appearing excessively laborious: “doing too much and yet also not enough work.”4 In the same sense, having AI produce synthetic images instead of taking pictures seems both lazy and overly complicated, at once a shortcut and a waste of resources. In more than one way, generative AI, with its promise to save and even abolish labor while consuming enormous resources, including often precarious and badly paid human labor, might represent the ultimate form of capitalist gimmick. By turning this gimmick into a tangible gadget, Karmann has created a philosophical toy in the truest sense: a physical model that allows us to grasp some of the underlying mechanisms of AI’s impact on our current digital visual culture.

  • 5 On the notion of “photorealism,” see Jens Schröter’s contribution to this issue. On realism and dig (...)
  • 6 I would like to sincerely thank Jacob Birken for letting me use this term, which he originally deve (...)

6Although the reality shown by Paragraphica is in no way a mirror of the reality around us, we are supposed to understand these images as somehow “realistic” representations. In general, “realism” is the holy grail of AI image synthesis. With each update of their models, companies like Midjourney and OpenAI strive for increasingly “realistic” renderings, which is also what many consumers expect from these tools. Judging from the companies’ promotional messages and users’ online discussions, this notion of realism conflates three dimensions of representation. First, synthetic images are expected to match a given textual description (or prompt) as closely as possible—in other words, they are supposed to be legible. Secondly, they are optimized to meet certain expectations of visual coherence, physical integrity, and spatial consistency. In other words, they are supposed to be plausible (and not feature, for example, hands with additional fingers). Finally, such consumer expectations of legibility and plausibility are profoundly shaped by other media—not only by photography but increasingly by “photorealistic” video games and digital renderings: synthetic images are expected to look like the images we already know.5 This set of expectations is what, to use a term coined by Jacob Birken, one might call “platform realism”: a second-order aesthetic of generic images optimized to match textual descriptions, consumer expectations, and, ultimately, images that already exist.6 In what follows, I want to discuss and further develop these three aspects of “platform realism”: the question of legibility (i.e., the primacy of language both in the production and reception of these images); the question of plausibility (i.e., the way they are statistically optimized to cater to our expectations); and finally, their structurally nostalgic relationship to past media.

Legibility: Guided by Concepts

  • 7 Antonio Somaini, “Algorithmic Images: Artificial Intelligence and Visual Culture,” Grey Room 93 (20 (...)

7Before the image, there is a text. When you “photograph” a scene with Paragraphica, you first get a textual description, which is then transformed into a prompt, and the image you get is presented as a visualization of that prompt (fig. 2). This apparent primacy of the sayable over the visible aptly illustrates an argument that film and visual culture scholar Antonio Somaini has made, namely that what can and will be visualized by text-image generation models such as Midjourney, Dall-E, and Stable Diffusion depends to a considerable extent on what can be formulated in textual form and written as a prompt. In Somaini’s words, these tools open up “a new visual landscape in which images and words are increasingly inseparable.”7

2. Bjørn Karmann, Paragraphica

2. Bjørn Karmann, Paragraphica

From left to right: photo of the real-life location of the “camera”; text description based on location data; and AI-generated image derived from this description, 2023.

© Bjørn Karmann

  • 8 Roland Meyer, “The New Value of the Archive: AI Image Generation and the Visual Economy of ‘Style,’ (...)

8I have written on the logic of the prompt in an earlier essay, and I want to summarize the main argument I have made there.8 In my view, prompts cannot be adequately understood as mere descriptions of images, which are then translated into visual form. Instead, prompts resemble search commands. Despite the popularity of the term “prompt engineering,” their function is not that of an exact program to be executed but that of fuzzy parameters narrowing down possible results and steering the model toward a specific region within a seemingly infinite statistical space of possibilities. In other words, the relationship between words and images in AI image synthesis is less one of translation and execution than one of matching and guidance: words determine which images are synthesized, but these images are not so much produced by words as found through them.

9From the user’s point of view, one of the main features of textual prompts is the possibility of combining words that are not commonly associated with one another. The logic of the prompt is the structural logic involved in recombining discrete elements. It is precisely the improbable, even counterfactual combinability of already known and nameable image objects, concepts, and modes of representation, such as “Pope” and “puffy Balenciaga jacket” or “Trump” and “police arrest,” which accounts for the viral appeal of many of the most popular AI images (fig. 3). The mash-up of semantic concepts triggers the mash-up of visual elements, and the result, at best, is an image that is both expected and unexpected, immediately legible and surprising at the same time.

3. Image generated by Elliott Higgins using the Midjourney program, published on Twitter/X, March 20, 2023

3. Image generated by Elliott Higgins using the Midjourney program, published on Twitter/X, March 20, 2023

© Elliott Higgins

  • 9 For a beautiful visualization of this process, see Eryk Salvaggio’s film “Flowers Blooming Backward (...)
  • 10 Andreas Ervik, “Generative AI and the Collective Imaginary: The Technology-Guided Social Imaginatio (...)

10On the face of it, prompt-based AI image synthesis functions much like an online image search—you type in a few words, and the interface presents you with a selection of images to match your query. These images are found, however, not in databases of existing images but in a field of noise. Having been trained with countless examples of how digital images lose all visual information and degrade into a state of entropy, so-called diffusion models are able to reverse this process, turning entropy into visual information. Once you have entered a prompt, these models begin to gradually denoise pure noise until a pattern begins to emerge that matches the input prompt, a pattern which can then be reinforced in every further iteration.9 As media scholar Andreas Ervik has put it, AI images “may seem like concrete solids; they may resemble photographs or some other products of traditional image production. However, they are, in fact, localized zones of coherence, drawn from a flux of potential intensities in a field of noise.”10 One could think of this process as a kind of reconstruction of information lost in a noisy channel. One could also see this as a series of iterative tests: the image is tested again and again in how well it already matches the prompt, and its legibility is further optimized in every iteration.

11This is made possible because these models have not only “learned” how images fade into noise but also how words relate to images, or rather, how text patterns correlate with visual patterns. To this end, billions of image-text pairs harvested from the web are used as training data, and the patterns found in this vast virtual archive of our current digital visual culture are encoded in a so-called latent space: a condensed statistical representation of possible relationships between words and images. Within this distribution of statistical possibilities, the prompt functions as a road map, guiding the model to a specific region and reducing the output variance. Without a prompt, the diffusion process would lead to a random image.

  • 11 Kate Eichhorn, Content (Cambridge, MA: MIT Press, 2022), 29.

12If words and images become inseparable in AI image synthesis, the main precondition for this is that they are already very closely linked in today’s online visual culture. The training of AI image-synthesis models fundamentally depends on a visual landscape in which most images are surrounded by clouds of visible and invisible text, including ALT descriptions, copyright notices, and all kinds of captions, labels, and snippets of articles. Typical of current digital image cultures, this virtual archive of image-text pairs is messy and, at least to a certain degree, flat: a “single and indistinguishable flow” of “content,”11 which knows no hierarchies between high and low, no sharp boundaries between media and genres and formats, and no clear distinction between real and fake. Within this endless flow, all images, whether holiday snapshots, works of art, memes, product advertisements, or documentary footage, are presented in the same formats, distributed via the same platforms, and subject to the same algorithmic operations. In the image-synthesis process, however, none of these training images are any longer available to the model; only the highly condensed patterns are derived from them. Therefore, AI image synthesis is completely different from a remix of existing image elements or a collage.

  • 12 Adrian MacKenzie and Anna Munster, “Platform Seeing: Image Ensembles and Their Invisualities,” Theo (...)
  • 13 See Mary L. Gray and Siddharth Suri, Ghost Work: How to Sop Silicon Valley from Building a New Glob (...)

13While training data is scraped from all over the web, the most important sources include platforms such as Wikipedia, Flickr, Pinterest, and Shopify, where masses of images have already been aggregated, preformatted, crowd-curated, and annotated. As Adrian McKenzie and Anna Munster have described, our online visual culture is one in which no image can be isolated from large amounts of visual and invisual data: “Today, we propose, it is the image ensemble—images, not simply quantified, but labelled, formatted and made ‘platform-ready’—that enables the emergence of a new mode of perception, and indeed a reformulation of visuality itself. We call this platform seeing.”12 In a sense, AI image generation constitutes a second-order “platform seeing”—as it presupposes the mass aggregation, processing, and labeling of image ensembles and opens them up for novel models of pattern extraction on an unprecedented scale. It is not only masses of copyright-protected content that are appropriated in the process but also the invisible and often precarious “ghost work” of all those who annotate, sort, and evaluate these image ensembles.13 Their work ensures, not least, that today no image comes as an isolated entity.

  • 14 “Bild-Begriffs-Kopplungen”: Matthias Bruhn, Bildwirtschaft: Verwaltung und Verwertung der Sichtbark (...)
  • 15 For an in-depth historical analysis of the visual economy of stock images, see Estelle Blaschke, Ba (...)
  • 16 Boris Groys, The Total Art of Stalinism: Avant-Garde, Aesthetic Dictatorship, and Beyond, trans. Ch (...)
  • 17 Jörg Colberg, Photography’s Neoliberal Realism (London: Mack, 2021).
  • 18 On the notion of platform capitalism, see Nick Srinicek, Platform Capitalism (London: Wiley, 2016).

14In a world of digital platforms of networked online content, every image comes with a text, but an online text only seems complete with at least one accompanying image. Most content management systems expect you to provide an image for each entry. Such images are not mere illustrations but attractors supposed to make content more visible, shareable, and likable, and more often than not, they are stock photographs. As art historian Matthias Bruhn has argued, stock images function as pre-produced “image-concept couplings.”14 While they are usually “realistic”—i.e., rendered in a style that conforms to our expectations of photographic “realism”—they depict not so much real-life situations but rather stage visualizations of general concepts, symbolic ideas, and abstract values: a legible text, dressed up as an image.15 Interestingly, this is also what the “socialist realism” art movement was about, at least according to cultural theorist Boris Groys: “More than a true ‘reflection’ of any reality, the work is a hieroglyphic text, an icon, or a prescriptive newspaper article.”16 Photography theorist Jörg Colberg has recently taken up Groys’s analysis to describe the “neoliberal realism” of photographers such as Annie Leibovitz, Gregory Crewdson, and Andreas Gursky. Looking like photographs but functioning more like paintings or film stills, where no detail is left to chance, these works for Colberg are embodiments of neoliberal ideology, in which no social change seems possible, and the power of capital reigns supreme.17 Accordingly, the products of AI image synthesis could be seen as manifestations of platform capitalism, in which there is nothing outside of the circulation of content, the mass aggregation of data, and the extraction of patterns.18

15In a basic sense, text-to-image synthesis is pattern recognition in reverse: instead of classifying images and labeling image elements, it converts labels back into images. In both cases, the prerequisite is that these models have been trained with large quantities of previously labeled images. Thus, platform realism shows us a world optimized for pattern recognition and instant legibility, in which everything (and everyone) matches a pre-formulated label. Let us once again take Paragraphica as an example: instead of showing you, for example, this sunny afternoon today in this specific park here, it shows you a generic park that existed nowhere and never. This time- and placeless simulacrum has no indexical relationship to any reality, not even a simulated one; it is synthesized from a combination of visualized concepts, visual patterns associated with “parkness,” “afternoonness,” and “sunniness,” into an instant stock image, guided by written concepts and optimized for legibility.

Plausibility: Expectable Expectations

  • 19 Hito Steyerl, “Mean Images,” New Left Review 140/141 (2023): 82.

16Unlike stock photography or socialist realist painting, however, the generic images of AI image synthesis are not so much the product of artistic choices and conventions as the result of statistical calculations: millions of images of sunny parks have found their way into the AI image of a sunny park, each contributing only a tiny fraction. In her essay on “Mean Images,” artist, author, and filmmaker Hito Steyerl has therefore aptly called AI images “statistical renderings”: “Visuals created by ML tools are statistical renderings, rather than images of actually existing objects. They shift the focus from photographic indexicality to stochastic discrimination. They no longer refer to facticity, let alone truth, but to probability.”19 In the case of commercial AI image-synthesis tools, however, probability is deeply intertwined with plausibility: these images are not simply the mechanical outcome of a blind statistical process but are catering to certain expectations. “Mean images,” in Steyerl’s sense, are shaped both by pattern recognition and constant user feedback—they are, as the double meaning of “mean” indicates, based not only on statistical averages but also on the often vulgar product of an incessant digital plebiscite, manifestations of an algorithmically powered aesthetic populism.

  • 20 Christo Buschek and Jer Thorp, “Models All the Way Down,” Knowing Machines, accessed August 16, 202 (...)

17Tellingly, plausibility and applause have the exact origin: plausibility is what deserves an audience’s applause and approval. In this sense, AI images are optimized for plausibility; they aim for the broadest possible approval, and this optimization begins with the selective filtering of training data. For example, LAION 5B is the open-source database of 5 billion image-text pairs used to train Stable Diffusion and Midjourney; it also offers a collection of subsets called LAION-Aesthetics, filtered for “high visual quality.” For this, a neural network was trained to predict which images people find particularly pleasing aesthetically—and the training data for this model, as Christo Buschek and Jer Thorp have pointed out, mainly came from two sources: synthetic images rated by the users of a Discord channel and photos scraped, together with the ratings, from a website called dpchallenge.com, which stands for “digital photography challenge.” In both cases, as far as can be reconstructed, the user base providing the ratings was rather small, predominantly male, and mostly from the US and Europe, with the tastes of only a handful of users significantly dominating the outcome.20

18Trained with the results of these online plebiscites, the neural network could then rate images seemingly autonomously. These algorithmically produced ratings, which are meant to predict how humans—or at least a very narrow and specific group of humans—would rate an image, were then used to filter training data scraped from the web in order to train and optimize image-synthesis models such as Stable Diffusion and Midjourney. The outcome is what one could call a recursive algorithmization of taste: machines predict the ratings of human users, which are then used to optimize other machines to please the aesthetic expectations of other human users. Significantly, the images that achieve the highest scores often appear representational but not photographic: watercolors of exotic streetscapes, tourist destinations in atmospheric lighting conditions, and portraits of conventionally attractive, mostly white young women are among those with the highest scores (fig. 4).

4. Assemblage of images with the best ratings in LAION-Aesthetics, published by Andy Baio on Waxy.org, August 30, 2022

4. Assemblage of images with the best ratings in LAION-Aesthetics, published by Andy Baio on Waxy.org, August 30, 2022

19The recursive algorithmization of taste does not stop with filtering training data. AI image-synthesis tools such as Midjourney produce masses of data about how customers react to their model and its output through their interface design—simply by quantifying which images are upscaled, shared, and liked by other users. Commercial AI tools are optimized to predict and adapt to their paying customers’ aesthetic and ideological preferences by constantly monitoring reactions to images. As it turns out, those customers do not expect unbiased representation. This became very clear when Midjourney, in its fifth version, released in March 2023, switched to a generic stock photography look as its default mode rather than the exaggerated illustration style it had been known for, while at the same time allowing for more ethnic diversity. Apparently, this was not a success with the user base, as the next release, version 5.1, made whiteness the norm again (fig. 5).

5. Images generated by the author using Midjourney in March 2023 (with Version 5, above) and June 2023 (with Version 5.2, below), based on the “family” prompt.

5. Images generated by the author using Midjourney in March 2023 (with Version 5, above) and June 2023 (with Version 5.2, below), based on the “family” prompt.
  • 21 See Fabian Offert and Thao Phan, “A Sign That Spells: DALL-E 2, Invisual Images and the Racial Poli (...)

20Midjourney’s competitor Open AI, however, handles things differently. In the latest version, you interact with Open AI’s Dall-E via ChatGPT, which means that instead of typing a prompt directly into the text-to-image generator, you ask ChatGPT for an image. The chatbot then prompts Dall-E with several prompts based on its interpretation of your requests and OpenAI’s general guidelines. If ChatGPT suspects that you want something inappropriate, be it pornography or a copyright violation, it will simply refuse or offer alternatives. In any case, however, it will generate requests that are much more specific than what you asked for. For example, if you ask ChatGPT for a picture of a “family,” it adds possible settings, such as a picnic and the notion of diversity. This technique, called “prompt transformation” or “shadow prompting,” was introduced by OpenAI in 2022 as a response to the criticism that its model promotes whiteness as the norm and reproduces racial stereotypes.21

  • 22 Thao Phan, “Black Nazis, Asian Vikings & the White Paranoia That Haunts Generative AI,” Australian (...)

21The flaws of this approach to mitigating bias were highlighted when Google’s AI image generator, Gemini, was launched in the spring of 2024. Many right-wing online accounts were outraged by what they called “woke indoctrination,” as the software deliberately attempted to produce more diverse results, even in cases where this did not match the expectations evoked by the prompt, resulting in female popes, Asian Vikings, and black Nazis (fig. 6). In fact, these results likely did match the prompts, albeit not those formulated by users but rather those augmented internally by Google’s software. Having been accused of promoting “anti-white bias,” Google finally had to suspend the image-generating features of its model. The whole affair was, as feminist technoscience researcher Thao Phan has pointed out, a study in “white paranoia”: “While they may not be accurate, these images are nevertheless instructive, showing us what and whose anxieties are shaping the conversations around AI and its potential harms.”22 Any idea of plausibility is already a biased one. Almost by design, the output of the image-synthesis model is likely to cater to the tastes of a mostly white, young, male, and Western demographic. Moreover, this demographic makes itself most loudly heard whenever its expectations are not met.

6. “America’s Founding Fathers, the Vikings, and the Pope according to Google AI,” posted on Twitter/X by EndWokeness, February 21, 2024.

6. “America’s Founding Fathers, the Vikings, and the Pope according to Google AI,” posted on Twitter/X by EndWokeness, February 21, 2024.
  • 23 See Estelle Blaschke, “The Excess of the Archive,” in Documenting the World: Film, Photography, and (...)

22The ways Midjourney, OpenAI, and Google handle the problem of bias indicate that there might be different variants of “platform realism” competing. However, it is precisely this underlying logic of competition that ultimately unites them: platform realism is the product of a digital visual culture in which images constantly compete for scores and ratings in the virtual arenas of online user interaction. Every AI image is a test image designed to be evaluated and compared with other images. Such large-scale evaluation of images is nothing new. As media scholar Estelle Blaschke notes, digital image libraries such as Corbis have long relied on sophisticated ranking systems to structure their ever-growing archives.23 In AI image synthesis, however, the rating and ranking of images via social media platforms, the training of generative models, and the optimization of their output form an almost perfect feedback loop of human reactions and machine predictions.

  • 24 I have developed this aspect in Roland Meyer, “Es schimmert, es glüht, es funkelt: Zur Ästhetik der (...)

23In turn, the specific qualities of synthetic images reveal something about the prevailing aesthetic preferences promoted by digital platforms, their recommendation algorithms, and filter aesthetics. The typical “glow” of synthetic images, for example, which often makes them appear as if illuminated from within, looks particularly appealing on mobile screens and is produced by exaggerating the effects of certain digital filters which have proven particularly successful on social media such as the “teal and orange” color scheme.24 However, it is not just the images that are tested. We, as users and consumers of these images, become test subjects as well: our reactions are recorded, our desires monitored, and our approval ratings fed into the ongoing optimization of the models. What is presented to us as images are concrete abstractions from the statistical evaluation of reactions to images, quantifiable as likes, shares, and upscales. Platform realism is an aesthetic of expectable expectations, an endless recursive feedback loop.

Pastness: Nostalgic Vibes

  • 25 Eva Cetinic, “The Myth of Culturally Agnostic AI Models”, arXiv, last revised November 29, 2022, ht (...)

24Returning once again to the example of Paragraphica, what these images make visible is not only a “reality” structured by language, as a visualization of concepts, and optimized for plausibility, statistically aligned to user expectations, it is also a “reality” that is very much the product of the past rather than the present. The vast aggregations of image ensembles these models are trained on have a historical index, a so-called cut-off date—nothing that happens after the training process enters the world of these models. As digital humanities scholar Eva Cetinic writes, “By encoding numerous associations which exist between data items collected at a certain point in time, those models therefore represent synchronic assemblages of cultural snapshots, embedded in a specific technological framework.”25 Thus, AI image synthesis has to rely on interpolating data from the past to produce an image of the present or even the future. AI image synthesis is a backward prediction: it makes plausible guesses on what could have been. Platform realism, therefore, is structurally conservative, even nostalgic.

  • 26 Benjamín Labatut, “A Wild Wild Wind: Bennett Miller’s AI-Generated Art,” Gagosian Quarterly (Summer (...)

25Such structural conservatism becomes visible in the dominant aesthetic of current AI art. In spring 2023, Hollywood director Bennet Miller had his first contemporary art exhibition at the renowned Gagosian Gallery in New York. In his show, visitors were presented with works that looked like photographs, undatable at first glance but vaguely reminiscent of the period around 1900 (fig. 7). Overall, there was a dreamlike atmosphere in these pictures: children with wavy hair staring at us, airplanes falling from the sky, bodies floating weightlessly in space, all rendered in soft sepia tones and a blurriness that made most of the details disappear into vagueness. Miller, who had been working on a documentary film about “artificial intelligence” for five years and therefore had exclusive early access to OpenAI’s software, had created all of these pictures with DALL-E and now presented them as limited edition pigment prints.26

7. Bennett Miller, Untitled (Picnic Fall), 2022–23

7. Bennett Miller, Untitled (Picnic Fall), 2022–23

Pigment print of an image generated using DALL-E, 85.7 × 85.7 cm.

© Bennett Miller, courtesy of the artist and Gagosian Gallery

26The artificial nostalgia of Miller’s work seems typical of many artistically more ambitious approaches to new image-making tools: artists and designers, in particular, tend to use the latest image-making techniques less to imagine possible futures than to explore alternative pasts.27 Often, this impulse goes together with the idea that AI image synthesis can show what never was but could have been. One well-known example is German photographer Boris Eldagsen, one of whose pictures from the series Pseudomnesia won a photography prize in 2023—until Eldagsen spectacularly withdrew it from the competition (fig. 8). Though Eldagsen does not use the term “nostalgia,” the conceptual statement on his website makes his approach to the past very clear: “Using the visual language of the 1940s, Boris Eldagsen produces his images as fake memories of a past, that never existed, that no-one photographed.”28

8. Boris Eldagsen, Pseudomnesia | The Electrician, 2022

8. Boris Eldagsen, Pseudomnesia | The Electrician, 2022

Promptography, 26 × 15.2 cm.

  • 29 Elif Batuman, “What Am I Doing Here?,” The Guardian, April 26, 2008, https://www.theguardian.com/bo (...)
  • 30 Peli Grietzer, “A Theory of Vibe,” in “Site 1. Logic Gate: The Politics of the Artifactual Mind,” G (...)

27What Eldagsen calls here “the visual language of the 1940s” is more than a style; it is rather a particular historical “vibe” that manifests itself in the photographic composition, in the costumes and hairstyles, the sharp black-and-white contrasts, as well as in the physiognomy and expression of human faces. A “vibe,” in this sense, is as much about what is represented as about how it is represented, an atmospheric quality shared by images (or, for that matter, sounds and texts) that seem to be part of the same aesthetic “world”—in the words of Guardian pop critic Elif Batuman: “What gives a vibe ‘authenticity’ is its ability to evoke . . . a certain nexus of historic, geographic and cultural forces.”29 In his “Theory of Vibe,” literary theorist Peli Grietzer elaborated on this world-building quality of aesthetic “vibes,” calling the vibe an “abstractum that cannot be separated from its concreta.”30 It is this historical concreteness, the rootedness in the specifics of a particular cultural moment in space and time, that makes a “vibe” distinctive.

  • 31 See Dominik Schrey, Analoge Nostalgie in der digitalen Medienkultur (Berlin: Kadmos, 2017).
  • 32 André Habib, “Rayures, Poussière, Grain. ‘L’effet-pellicule’ au temps du numérique ou la survivance (...)
  • 33 On the notion of “remediation,” see Jay David Bolter and Richard Grusin, Remediation: Understanding (...)

28Unlike styles, “vibes” are always mediated. They seem to be bound to historical media and their visual qualities, and it is these qualities—the blurriness of old VHS tapes, the bright colors of Kodachrome, or the graininess of 16 mm film stock—that AI image synthesis excels at replicating. As media scholar Dominik Schrey has shown, digital culture has long cultivated a nostalgic relationship with the look of the analog, not least in the form of filters simulating the specific noise of older media.31 Contemporary cinematography, in particular, has developed a range of post-production techniques aimed at reproducing “scratches, dust and grain” and thus simulating a “filmic” look as a digital effect, as film scholar André Habib has pointed out.32 In AI image synthesis, such earlier forms of digital “remediation” come to a kind of conclusion, as every AI image is by design made to look as if it were the product of an older visual medium, be it oil painting, photography, or game engines.33 In all these cases, what is remediated are not media-specific properties but aesthetic vibes, nameable and repeatable patterns extracted from mass quantities of digital visual content: a particular means of representing space and distributing light and shadow, a distinct color scheme, a specific form of visual noise, and, first and foremost, a certain way of giving visual coherence to an image.

  • 34 Fredric Jameson, Postmodernism, or, The Cultural Logic of Late Capitalism (London: Verso, 1991), 19

29By remediating the vibes of past images, AI image synthesis promises to fill the imaginary gaps in the virtual archive of past images. However, this virtual archive, scraped from the web, is not only limited and incomplete in many ways, leading to well-known racist and sexist biases, it is also structurally messy and opaque. The sources for what Eldagsen calls the “visual language of the 1940s” might include documentary photographs from that era as well as staged magazine illustrations, film noir stills, scenes from 1980s Hollywood movies, pop album covers, and images from video games that attempt to recreate the look of the era. Consequently, the memories evoked by AI image synthesis do not refer to a historical past of documented events but to what literary theorist Fredric Jameson called “pastness”—a pseudo-historical plausibility based on aesthetic styles loosely associated with a particular historical period.34

  • 35 On “mood boarding,” see Tom Holert, Regieren im Bildraum (Berlin: b-books, 2008), 38.

30With AI image synthesis, such visual markers of the past can now be transformed into endlessly reproducible patterns and curated to evoke affects and atmospheres associated with a generic “pastness.” This is not only reminiscent of a postmodern aesthetic of pastiche but can also be seen as a continuation of established commercial image practices such as “mood boarding,” used by commercial artists to convey the visual moods sought in a particular design or advertising campaign through the selection and juxtaposition of existing images.35 Like mood boarding, AI image synthesis is a technique for curating vibes—based not on assembling a manual montage of discrete pictures but on extracting and synthesizing patterns distilled from a seemingly freely accessible virtual archive. In platform realism, history dissolves into a series of aesthetic “vibes.”

Coda: Disposable by Design

  • 36 On the energy costs of generative AI in general, see Alexandra Sasha Luccioni, Yacine Jernite, and (...)

31While platform realism is structurally nostalgic, it is also inherently anti-nostalgic. Nostalgia is about loss, about longing for a past that is ultimately irretrievable. Therefore, the fetishized objects of nostalgia are those rare and valuable relics that keep some traces of this lost past alive. Image-synthesis tools, however, produce images that are neither rare nor valuable. On the contrary, AI image synthesis is inherently a wasteful process that consumes enormous energy resources and computing power for images that are, in most cases, completely worthless.36 To get an image that meets your expectations, you almost always have to produce dozens of variations first. Most synthetic images are never upscaled, shared, or liked but discarded immediately. Even those that do circulate widely have increasingly become a form of visual waste, contaminating the web with fake content, cheap clickbait, and reactionary propaganda. Platform realism is an aesthetic of wastefulness.

  • 37 Sina Alemohammad et al. “Self-Consuming Generative Models Go MAD”, arXiv, July 4, 2023, https://doi (...)
  • 38 Francis Hunger, “Spamming the Data Space: CLIP, GPT and Synthetic Data”, database cultures, Decembe (...)

32However, from the AI companies’ point of view, all these wasted images are not worthless. As a test image, they produce valuable data for further optimization of the models. Every image you do not upscale tells these companies something about your expectations. However, this optimization through excessive overproduction comes at a price. Synthetic content has already become part of the training data for future models, a process that is likely to intensify in the coming years. And this has dramatic consequences: typical AI artifacts, glitches, and errors, feeding back on themselves, become increasingly amplified, creating what one research team called an “autophagous (‘self-consuming’) loop.”37 Perhaps, as Francis Hunger argued, we have already seen the last days of the web as a largely uncontaminated data resource.38 Even if companies eventually find a way to filter out synthetic content for their future training data, the sheer scale of spam, scams, and disinformation flooding digital platforms today is likely to exacerbate the already palpable sense of distrust, resistance, and aversion that AI image synthesis provokes. Like all extractivist business models, platform realism ultimately threatens to destroy its own foundations.

Haut de page

Notes

1 Bjoern Karmann, “Paragraphica” (2023), accessed August 16, 2024, https://bjoernkarmann.dk/project/paragraphica.

2 On the notion of the “invisual,” see Jussi Parrika, Operational Images: From the Visual to the Invisual (Minneapolis: University of Minnessota Press, 2023).

3 Shane Denson, Post-Cinematic Bodies (Lüneburg: Meson, 2023), 187.

4 Sianne Ngai, Theory of the Gimmick: Aesthetic Judgement and Capitalist Form (Cambridge, MA: Belknap Press, 2020), 56.

5 On the notion of “photorealism,” see Jens Schröter’s contribution to this issue. On realism and digital renderings more broadly, see Jacob Birken, Vom Pixelrealismus: Takeshi Muratas Stillleben Cyborg (Berlin: Schlaufen, 2023), esp. 53–55, 82–83, 125–26.

6 I would like to sincerely thank Jacob Birken for letting me use this term, which he originally developed in a lecture on NVIDIA’s “digital twins.”

7 Antonio Somaini, “Algorithmic Images: Artificial Intelligence and Visual Culture,” Grey Room 93 (2023): 74–115, here: 107.

8 Roland Meyer, “The New Value of the Archive: AI Image Generation and the Visual Economy of ‘Style,’” IMAGE 37, no. 1 (2023): 100–111.

9 For a beautiful visualization of this process, see Eryk Salvaggio’s film “Flowers Blooming Backward into Noise” (2023), June 1, 2023, https://www.cyberneticforests.com/news/flowers-blooming-backward-into-noise-2023.

10 Andreas Ervik, “Generative AI and the Collective Imaginary: The Technology-Guided Social Imagination in AI-Imagenesis,” IMAGE 37, no. 1 (2023): 42–57, here: 46.

11 Kate Eichhorn, Content (Cambridge, MA: MIT Press, 2022), 29.

12 Adrian MacKenzie and Anna Munster, “Platform Seeing: Image Ensembles and Their Invisualities,” Theory, Culture & Society 36, no. 5 (2019): 3–22, here: 5.

13 See Mary L. Gray and Siddharth Suri, Ghost Work: How to Sop Silicon Valley from Building a New Global Underclass (Boston: Harper Business, 2019).

14 “Bild-Begriffs-Kopplungen”: Matthias Bruhn, Bildwirtschaft: Verwaltung und Verwertung der Sichtbarkeit (Weimar: VDG, 2003), 70.

15 For an in-depth historical analysis of the visual economy of stock images, see Estelle Blaschke, Banking on Images: The Bettmann Archive and Corbis (Leipzig: Spector 2016).

16 Boris Groys, The Total Art of Stalinism: Avant-Garde, Aesthetic Dictatorship, and Beyond, trans. Charles Rougle (Princeton, NJ: Princeton University Press, 1992), 56.

17 Jörg Colberg, Photography’s Neoliberal Realism (London: Mack, 2021).

18 On the notion of platform capitalism, see Nick Srinicek, Platform Capitalism (London: Wiley, 2016).

19 Hito Steyerl, “Mean Images,” New Left Review 140/141 (2023): 82.

20 Christo Buschek and Jer Thorp, “Models All the Way Down,” Knowing Machines, accessed August 16, 2024, https://knowingmachines.org.

21 See Fabian Offert and Thao Phan, “A Sign That Spells: DALL-E 2, Invisual Images and the Racial Politics of Feature Space,” arXiv, October 26, 2022, https://arxiv.org/abs/2211.06323.

22 Thao Phan, “Black Nazis, Asian Vikings & the White Paranoia That Haunts Generative AI,” Australian Academy of the Humanities (April 2024), accessed August 16, 2024, https://humanities.org.au/power-of-the-humanities/black-nazis-asian-vikings-and-other-problems-with-generative-ai/.

23 See Estelle Blaschke, “The Excess of the Archive,” in Documenting the World: Film, Photography, and the Scientific Record, ed. Gregg Mitman and Kelley Wilder (Chicago: University of Chicago Press, 2016), 224–53, here: 245.

24 I have developed this aspect in Roland Meyer, “Es schimmert, es glüht, es funkelt: Zur Ästhetik der KI-Bilder,” 54 Books, March 20, 2023, https://54books.de/es-schimmert-es-glueht-es-funkelt-zur-aesthetik-der-ki-bilder/. On the aesthetics of digital filters, see also Berit Glanz, Filter (Berlin: Wagenbach, 2023).

25 Eva Cetinic, “The Myth of Culturally Agnostic AI Models”, arXiv, last revised November 29, 2022, https://doi.org/10.48550/arXiv.2211.15271.

26 Benjamín Labatut, “A Wild Wild Wind: Bennett Miller’s AI-Generated Art,” Gagosian Quarterly (Summer 2023), https://gagosian.com/quarterly/2023/05/01/essay-a-wild-wild-wind-bennett-millers-ai-generated-art/.

27 One important example of this would be the work of Grégory Chatonsky, who uses diffusion models to explore counterfactual autobiographical narratives; see Somaini, “Algorithmic Images,” 105–6. Other examples would include Roope Rainisto, with his series Life in West America, or Philip Toledano, with his series Another America.

28 Boris Eldagsen, Pseudomnesia, accessed August 16, 2024, https://www.eldagsen.com/pseudomnesia/.

29 Elif Batuman, “What Am I Doing Here?,” The Guardian, April 26, 2008, https://www.theguardian.com/books/2008/apr/26/popandrock.

30 Peli Grietzer, “A Theory of Vibe,” in “Site 1. Logic Gate: The Politics of the Artifactual Mind,” Glass Bead (2017), www.glass-bead.org/article/what-is-it-to-think/.

31 See Dominik Schrey, Analoge Nostalgie in der digitalen Medienkultur (Berlin: Kadmos, 2017).

32 André Habib, “Rayures, Poussière, Grain. ‘L’effet-pellicule’ au temps du numérique ou la survivance simulée de l’involontaire,” in: La haute et la basse définition des images. Photographie, cinéma, art contemporain, culture visuelle, ed. Franco Cassetti and Antonio Somaini (Paris: Mimesis, 2021), pp. 125–141.

33 On the notion of “remediation,” see Jay David Bolter and Richard Grusin, Remediation: Understanding New Media (Cambridge, MA: MIT Press, 1999). See also Jay David Bolter, “AI Generative Art as Algorithmic Remediation,” IMAGE 37, no. 1 (2023), 195–207.

34 Fredric Jameson, Postmodernism, or, The Cultural Logic of Late Capitalism (London: Verso, 1991), 19.

35 On “mood boarding,” see Tom Holert, Regieren im Bildraum (Berlin: b-books, 2008), 38.

36 On the energy costs of generative AI in general, see Alexandra Sasha Luccioni, Yacine Jernite, and Emma Strubell, “Power Hungry Processing: Watts Driving the Cost of AI Deployment?,” arXiv, last revised May 23, 2024, https://doi.org/10.48550/arXiv.2311.16863.

37 Sina Alemohammad et al. “Self-Consuming Generative Models Go MAD”, arXiv, July 4, 2023, https://doi.org/10.48550/arXiv.2307.01850.

38 Francis Hunger, “Spamming the Data Space: CLIP, GPT and Synthetic Data”, database cultures, December 7, 2022, http://databasecultures.irmielin.org/spamming-the-data-space-clip-gpt-and-synthetic-data/.

Haut de page

Table des illustrations

Titre 1. Bjørn Karmann, Paragraphica
Légende A lensless “camera” based on AI text-to-image generation, 2023.
Crédits © Bjørn Karmann
URL http://journals.openedition.org/transbordeur/docannexe/image/2299/img-1.jpg
Fichier image/jpeg, 479k
Titre 2. Bjørn Karmann, Paragraphica
Légende From left to right: photo of the real-life location of the “camera”; text description based on location data; and AI-generated image derived from this description, 2023.
Crédits © Bjørn Karmann
URL http://journals.openedition.org/transbordeur/docannexe/image/2299/img-2.jpg
Fichier image/jpeg, 315k
Titre 3. Image generated by Elliott Higgins using the Midjourney program, published on Twitter/X, March 20, 2023
Crédits © Elliott Higgins
URL http://journals.openedition.org/transbordeur/docannexe/image/2299/img-3.jpg
Fichier image/jpeg, 525k
Titre 4. Assemblage of images with the best ratings in LAION-Aesthetics, published by Andy Baio on Waxy.org, August 30, 2022
URL http://journals.openedition.org/transbordeur/docannexe/image/2299/img-4.jpg
Fichier image/jpeg, 392k
Titre 5. Images generated by the author using Midjourney in March 2023 (with Version 5, above) and June 2023 (with Version 5.2, below), based on the “family” prompt.
URL http://journals.openedition.org/transbordeur/docannexe/image/2299/img-5.jpg
Fichier image/jpeg, 892k
Titre 6. “America’s Founding Fathers, the Vikings, and the Pope according to Google AI,” posted on Twitter/X by EndWokeness, February 21, 2024.
URL http://journals.openedition.org/transbordeur/docannexe/image/2299/img-6.jpg
Fichier image/jpeg, 368k
Titre 7. Bennett Miller, Untitled (Picnic Fall), 2022–23
Légende Pigment print of an image generated using DALL-E, 85.7 × 85.7 cm.
Crédits © Bennett Miller, courtesy of the artist and Gagosian Gallery
URL http://journals.openedition.org/transbordeur/docannexe/image/2299/img-7.jpg
Fichier image/jpeg, 385k
Titre 8. Boris Eldagsen, Pseudomnesia | The Electrician, 2022
Légende Promptography, 26 × 15.2 cm.
URL http://journals.openedition.org/transbordeur/docannexe/image/2299/img-8.jpg
Fichier image/jpeg, 367k
Haut de page

Pour citer cet article

Référence électronique

Roland Meyer, « “Platform Realism”. AI Image Synthesis and the Rise of Generic Visual Content »Transbordeur [En ligne], 9 | 2025, mis en ligne le 26 février 2025, consulté le 13 mai 2026. URL : http://journals.openedition.org/transbordeur/2299 ; DOI : https://doi.org/10.4000/13dwq

Haut de page

Auteur

Roland Meyer

Roland Meyer is a media and visual culture scholar and DIZH Bridge Professor for Digital Cultures and Arts at the University of Zurich and the Zurich University of the Arts (ZHdK). He has published on the history and theory of operative images, the cultural and social implications of automated facial recognition, forensic image practices, the aesthetics of AI image synthesis, and the media archaeology of augmented spaces.
Roland Meyer est chercheur spécialisé dans les médias et la culture visuelle et professeur DIZH Bridge de culture et d’arts numériques au sein de l’université de Zurich et de la Haute École d’art de Zurich (ZHdK). Il a notamment écrit sur l’histoire et la théorie des images opératoires, les implications socioculturelles de la reconnaissance faciale automatisée, les pratiques liées à l’imagerie médico-légale, l’esthétique des images générées par l’IA et l’archéologie des espaces augmentés dans les médias.

Haut de page

Droits d’auteur

CC-BY-NC-ND-4.0

Le texte seul est utilisable sous licence CC BY-NC-ND 4.0. Les autres éléments (illustrations, fichiers annexes importés) sont susceptibles d’être soumis à des autorisations d’usage spécifiques.

Haut de page
Rechercher dans OpenEdition Search

Vous allez être redirigé vers OpenEdition Search