What Is Not an Image?

This chapter is on the operations of invisibility and invisuality. Of those two terms, the first seems to prescribe an optical situation of (not) seeing, while the second pertains to mathematical functions, data, and platforms. Considering how often seeing, vision, and other metaphors are used when discussing observations relating to and based on data, this division is not entirely clear-cut. What is clear, though, is that speaking of images in such terms helps to avoid a binary division between analog and digital images. Some of the terminology discussed in this chapter is also meant to open up to questions of data. There is a wonderfully useful phrase from Georges Didi-Huberman: “The entire history of images can thus be told as an effort to visually transcend the trivial contrasts between the visible and the invisible.”1 We can add that this history also needs to transcend trivial contrasts between data and the visible too.

Of course, knowing what is trivial and what is not demands a significant ability to evaluate and describe changing material, semiotic, and social contexts of images. As far as photographic theory is concerned, this question was central to how the material existence of an image becomes invisible when the image as representation emerges: to have an image is to forget there is a technical mediation that is the apparatus of capture and visualization. In Roland Barthes’s words, it is the pesky referent that gets in the way of the photograph itself: “Whatever it grants to vision and whatever its manner, a photograph is always invisible: it is not it that we see.”2 The same idea was voiced by Erwin Panofsky, too, in relation to painting since the Renaissance: “The material surface upon which the individual figures or objects are drawn or painted or carved is thus negated, and instead reinterpreted as a mere ‘picture plane.’”3 The material mediation, then, is what enables seeing and becomes unseen.

Furthermore, the history of photography—and the history of theories of photography—has been, for the most part, about what sort of visible or invisible rays, what wavelengths of radiation, are perceived as relevant and valid. What sort of light counts and literally also counts, as calculational images, as programmed images? Tell me what spectrum of electromagnetic radiation you are concerned with, and I will tell you what your discipline is. Is it the visible spectrum that registers people, objects, and landscapes? Or is it the invisible spectra observed by other remote sensing systems on satellites and the ground that register people, objects, landscapes—but also a multitude of other layers of reality from temperature to humidity, and from mineral resources to the chemical composition of vegetal surfaces (see Figure 3)?

These themes condense many of the approaches in visual theory as it is reflected through measurement instruments that sometimes are cameras, sometimes hyperspectral remote sensing devices, sometimes telescopes capturing the night sky as photographs of luminosity filtered through the spectroscopic lens. Some of the questions of the history of technical sensing and imaging resurface in contexts of data. In other words, if photographic theory has often been premised on the idea that it is a system of signs (and thus information), we can propose that images are systems of signals and noise and thus susceptible to interpretation and pattern recognition. A statistical distribution of characteristics becomes a base unit of an “aesthetic,” a recognizable trait for machine/computer vision and other systems that mobilize such an aesthetic as epistemically relevant. This also applies to remote sensing: classification and perception of different material characteristics from vegetation to soil and from sand to asphalt, snow, clouds, and water is often perceived as a spectral signature that actually is a spectral response pattern; not necessarily so much an “absolute and unique”4 representational entity as much a varying characteristic.

These might indeed be quite technical dilemmas, but they relate to the broader sense in which questions of sensing and images are integrated into a different epistemology than that of unique objects in representation. This starts to trouble what is visible and invisible as it becomes a question of masking/unmasking reality. It also includes how to reframe these questions in the age of cloud analytics, a term used by Louise Amoore to refer to the “capacity to extract patterns and features from data to open onto targets of opportunity, commercial and governmental.”5

In such descriptions that shift the frame of knowledge from optics to software and data, from pictures to models and patterns, we are not so far away from what Wendy Hui Kyong Chun argues is one essential feature of visibility/invisibility in the computational culture of software and interface design. Complexifying notions of transparency, visibility, and opacity, Chun dissects the relation of seeing as knowing, while arguing that we need to understand the (computational) system as “a conduit that also amplifies and selects what is at once real and unreal, true and untrue, visible and invisible.”6 This, one can argue, is also where data systems—operationalization of images as data in platforms—involve a production of what is counted as real and unreal, true and untrue, visible and invisible. Of course, many would be quick to add that the power of withdrawing from view and becoming invisible—or worse, natural—is the work of ideology.7 But it works the other way round, too, as per Foucault’s quip that “visibility is a trap” that is materialized as architectures; this trap is designed as observational events bootstrapped in existing racialized and gendered systems of power.8

This chapter outlines how the concept of operational images helps to map different regimes of invisibility and invisuality of images by providing three partly overlapping angles to the issue: a platform, a dataset, and a model. These offer insights into the current discussions about the transformation of visual culture (and, for example, issues related to labor of AI images) and the aesthetic and epistemic context in which images become something other than pictures. In simplest terms, the focus of the chapter is how visual culture studies (broadly understood) deals with invisuality as it is articulated through questions of data and machine learning techniques. Trevor Paglen on “invisible images”9 and Hito Steyerl’s “How Not to Be Seen” video essay (2013) are exemplary of the practices through which invisibility is not just unseen, but an entity produced, framed, and regulated as a cultural technique. These approaches also articulate the transformation of optics of vision and images in relation to a digital culture defined by machine vision. The previous chapter’s points about operative ontologies also help us understand how invisibility and invisuality are produced in specific media techniques. These terms, these three sections, are also ways to engage with the material epistemology of operational images in contemporary digital culture.

Transparency and opacity, intelligibility and noise entangle on surfaces of images. Some can also become an affordance, switching the focus from the value of visibility to the primacy of opacity. Informatic opacity is coined as a technique of resistance of biometric regimes of imaging the (vulnerable) body in Zach Blas’s practice-led work.10 In Stephanie Syjuco’s 2019 installation Dodge and Burn (Visible Storage), digital photographic techniques of a chroma-key green and Photoshop transparency checkerboard pattern entangle with imperial techniques of making invisible traditional patterns and garments, histories, and bodies. Grids appear at multiple scales of space, image, and bodies. Many other examples could be quoted, especially the question of camouflage, invisibility, and opacity that feature heavily in surveillance-driven discourse that also makes use of the terminology of the operational image.11 Much of the use of invisibility and opacity as terms are part of the methods of figuring out what forms of perception relate to operations of power in digital culture and how transparency and opacity are also designed (in the sense Chun underlines as the function of a medium that is the conduit of the visible and the invisible).

None of this was invented as late as digital culture, even if digital images are very specific kinds of nonrepresentational, rendered objects—even models instead of representations.12 The longer photographic history of the invisible focuses on scientific imaging in physics, biology, and astronomy: of objects that are too fast to be sensed (separating sensing and imaging) or too far (such as the astronomic photograph, like the opening examples—the Frauenhofer line mentioned in the preface and the astronomic analysis of luminosity at the Harvard observatory—of this book). Microphotography at Louis Pasteur’s laboratory picturing minuscule atmospheric dust (see Figure 4), Ernst Mach’s photograph of a bullet in motion (1884), August Strindberg’s celestographs (1893), and Ottomar Volkmer’s statement (1894) that “the progress of the natural sciences, and more specifically, the progress that has been made in the fields of photographic technology and electricity in the past decade, is so marvelous that even today it allows us to fix a number of invisible facts in nature in photographic images”13 testifies to this theme that runs through several decades of epistemic and aesthetic experimentation.

Peter Geimer offers a detailed glimpse into this history, from such examples of scientific use to the emerging artistic practices in the first half of the twentieth century. Furthermore, in another context, he proposes a closely related question. Geimer’s suggestion to switch the question from what is an image to what is not an image is particularly useful for a book focused on describing and conceiving such images that are primarily not seen (perhaps not even meant to be seen, but watched?), but which take an important instrumental part in an operation. His term for the nonimage or even the unimage narrates entities that cannot be seen as having been destroyed through archival deterioration or some other related process affecting old daguerreotype plates and other photochemical images.14 Such images at the threshold of their disappearance become part of the opening note on the materiality of the medium as its invisible support, although in this case, the ontological point becomes part of a temporal process.

For Geimer, the genealogy of unimages includes photographs that have vanished or eroded, that are defective and glitched. Besides errors that expose the photograph itself, this genealogy also includes operations where the photographic plate intersects with other practices of light (visible or not): chemical effects on surfaces and skin, X-rays, radioactivity, electricity, infrared, and so forth. Such an example is René Colson’s late-nineteenth-century physical experiments, related to the broader question of sensibility with the photographic as both an instrument and object of knowledge that prescribes realities that do not have an image of what they should look like but are diagrammatically traced into existence.15 Such operational ontologies of invisibility launch experimental probes into what is not visible, but that becomes a model for what can be known through the operations involved. They bring this invisible, even sometimes seemingly impossible view back as an epistemic and aesthetic object of manipulation into an apparatus of images.16 The function of images and instruments as experimental systems and models will be featured later in this chapter.

While the question of what is not an image, in Geimer’s take, refers to archival and historical situations, photochemical affordances, or liminal cases of photographic images, it is through this diversion that a broader question also emerges: it also implies all the cases of other (than) pictorial images. Unimages come to the fore, and unseeing becomes practiced in a version of Paul Virilio’s “sightless vision.”17 The operations of photographic chemicals and signs implies thus a broad field that is also the concern of this book: worlds that are not merely seen, as Paglen notes, but not necessarily only invisible either. The one that becomes a diagram, even if not exactly an image; the one that measures the world and produces a table or some other form of data media.

Invisibility is thus not merely a sphere of existence that remains unseen; it concerns a praxis. Before it becomes ontological, it is operative. Here, Geimer’s points are useful to quote in full as they concern the institutional and technical operations of knowledge in what becomes a discussion close to the operative ontology that underpins aesthetics of knowledge.

First, photographic visualization was not the transfer into the image of a phenomenon that had always already existed, though it had been latent; it was, first and foremost, the fabrication of an image. Second, as a general rule, it was not evident what this image showed; it needed to be decoded, stabilized, and isolated from an amalgam of facts and artifacts. And third, the “invisibility” that preceded it was not nothing but a sort of structured darkness that, despite its persistent obscurity, was interspersed with experiences, expectations, or imaginations.18

In something that resembles a response to Didi-Huberman’s note about transcending triviality, Geimer proposes to see invisibility as an interplay of material operations and cultural imaginaries instead of an ontological reality outside the human senses. Furthermore, as a discussion of images as entry points to knowledge objects for manipulation (an operationalization of invisibility), it brings to light the sort of extended angle to images that this book is engaged in through the notion of operational image. Visible or not, it is operations—platforms, datasets, and models—all the way to the bottom.

Platform Operations

First, dear reader, a thank you for your patience: you have made it this far, and I am moving deeper into the concept of the invisual. After all, the term was promised to be central for the developing argument, and so far, I have given only brief glimpses into the term’s use. Consider the postponement of this discussion to here, the second chapter of this book, as a kind of a watershed moment, for the argument of the book oscillates back and forth between visual and invisual culture. There is additional potential in the term “invisual,” which might be more useful in the long run than keeping with the invisible. It is a peculiar neologism, as prepositional twisting often can be: the shift from the visual to its negation also includes the metaphorization of space that becomes helpful, too.19 It is both a negation (prefix) and a spatialization (preposition) of the visual, and it includes an expansion from optics of vision to logistics of data, but entirely “in” space—literally. Not that this shift from invisible to invisual is one of linear historicization: “the invisual” becomes a heuristic term, and like “operational image,” helps to reevaluate historical sources in a different light. This resonates with some media archaeological methods that read recursively in and across historical periods.

The term “invisual” developed and used by Adrian Mackenzie and Anna Munster describes “platform seeing,” which offers a useful, interesting, and smart way of dealing with the dilemma of images in computational platforms. At what level can we claim that we are addressing images even when they do not cater to traditional forms of human seeing or observing? This is a question that is implied in the legacy of operational images as well: the computational is one threshold for what defines visuality, but even more specifically, digital platforms as environments of aggregation and operationalization of images become a more specific place where this focus condenses in relation to the economy and ecology of power. It is also on platforms where images turn into diagrams that can be operationalized in the technical sense and legal and other ways. If operational images always already were images defined by operations of measure, scaling, quantification, comparison, and analysis, in which ways does the platform become one key site that continues, as well as transforms, that legacy?

Mackenzie and Munster use the term “invisual” in their discussion of assembly and operations of images in corporate platforms from Facebook to Nvidia. This view to platforms includes multiple components from software and services to hardware and image processing. It is not restricted to platforms in the business or legal sense (although there are clear implications at stake). Their approach underlines an important lesson about images in technical and logistical terms. Images always come as series, that in Sekula’s words, constitute a traffic in images since the emergence of photography: quantity and value in circulation.20

Mackenzie and Munster are focused on the formatting and preparation of images as part of the operational status of the platforms, not on what happens or does not happen to the photographic image. This pragmatic, even methodological stance asks: What is being done to the image, and how does this logistical operation define its usefulness and function? The image as an aggregated series is a central starting point to what the image is and how it is being captured and observed: “Their operativity cannot be seen by an observing ‘subject’ but rather is enacted via observation events distributed throughout and across devices, hardware, human agents and artificial networked architectures such as deep learning networks.”21 Observation and sensing become the central tropes for this operation that shifts the stakes further from visual culture while also addressing the mass-image as defined by Cubitt. Observation is sensing in the broadest sense (a smart city of sensors and other data capture points, for example) and includes mathematical functions as (partial) observations of data.22

Besides the accumulating quantity of images captured and mobilized in platforms, the invisual mass-image is an example of administrative data media that is found in relational databases with operations of locating, identifying, recognizing, and measuring among the attributes of interest:

Relational databases like these are more interested in metadata—geolocation, personal identifiers, device identifiers, date stamps, facial recognition, distances and dimensions recorded by autofocus functions, upload addresses and the kinds of measurement prized by astronomical and meteorological databases. Commercially operated databases combine such data with data from searches and likes, swipes, shares, tags and other interactions, to construe from the relations between images a mass image of the world in constantly evolving organisational diagrams that only machines can read.23

When anything can be made into an image—a series of patterns for analysis or an organized set of pixel relations presented on a digital screen—the constitution of the image as “soft” has been argued to be a central aspect of this transformation that also includes operational images. The claim that we are dealing with algorithmic processing (to paraphrase Hoelzl and Marie)24—and not geometric projection—as the image’s primary quality is radically developed further to include the preparation of material that can be processed as if an image but for various ends other than human perception. Images can have other autonomous functions regulating, ordering, controlling, and relaying. Instances of data visualization since the eighteenth century could also be considered as part of this lineage of “proto data media”:25 the emergence of charts, graphs, tables, and diagrams as central features of modern statistics (consider William Playfair’s Commercial and Political Atlas and Statistical Breviary of 1786). Mackenzie and Munster write that invisual perception goes further than recapping machine vision. What is meant by observation and sensing is more than a substitution of human agency with machines or data:

Such a mode suggests that while visual techniques and practices continue to proliferate—from data visualization through to LIDAR technologies for capturing nonoptical images—the visual itself as a paradigm for how to see and observe is being evacuated, and that space is now occupied by a different kind of perception. This is not simply “machine vision,” we argue, but a making operative of the visual by platforms themselves.26

This making operative of the visual is where the relation to the main theme of this book emerges: it is not the image but the platform where (distributed) agency lies in this shifting regime of (in)visuality. What is recognized as an image is also an operation of data, commerce, law, and so on, defined by the complex ecology of platforms. There’s no individual observer or one individual image as the process of pattern recognition is automated through different techniques such as convolutional neural networks that are central for image analysis en masse.27 The fantasy space of pattern recognition resides in a particular assumption about the epistemic world of data: there’s always more; to work beyond the intelligible is the multiscalar promise that, instead of vast, invisible worlds of a micro- or macroscopic optical regime, is about the endless permutations inherent in data, training datasets for machine learning, relational databases, and their operations across platforms that reformat their surroundings.28

Image ensembles feed forward to the creation of (statistical, mathematical) models for various bespoke tasks. The example used by Mackenzie and Munster of the DeepMind company’s AI model for the game of Go shows the operationalization of and through the invisual domain: a world that is approached as patterns, diagrams, series, and data. The AlphaGo program was widely discussed in popular discourse between 2015 and 2017 as it beat several leading Go players, demonstrating how effective the neural network had become in playing the infamously complex board game. Using the Monte Carlo tree search in mapping and identifying the effective decision patterns functioning at the back of the neural network also implies interesting points about the centrality of cognition as perception, as observation. In provocative terms, Vladan Joler and Matteo Pasquinelli describe “the history of AI as the automation of perception,” where the automation of perception is understood as “a visual montage of pixels along a computational assembly line.”29 Here too, the (so-called) machine intelligence that goes into playing Go has a special version of pattern recognition as a new kind of cultural technique30 that sorts invisual observations (from images to remote sensing data) into operational actions.

Go game squares are also a visual pixel grid. Such grids can be operationalized in a logistics of images that can be adjusted into multilayered neural networks and thus inserted into training sets: game boards like Go can thus become a 19 × 19 observational matrix. Here, cognition is observation and is further processed as images integrated as heterogeneous forms of data into platform operations. This does not imply an image in the traditional sense but the logistics of platforms that create, organize, and mobilize data and datasets. Furthermore, from a 19 × 19 pixel space, the scaling up to millions of pixels across millions of images becomes the space where these problems are computationally challenging and wonderfully emblematic of an operative ontology of multiscalar data. As Mackenzie and Munster write:

We draw attention here especially to the primacy of image ensembles; the model trained via the DeepMind platform “learns” by observing many images. AlphaGo acts in the world to the extent that local spatial correlations can be associated with actions and rewards for those actions. The development of these systems centres on many cycles of observation followed by action. This cycling through observation and action constitutes the “training” of the model; a training that seemingly requires very little “prior knowledge” on the part of the model since it only receives pixels and game scores as input.31

Pixels in, models out. So-called images in, so-called statistical models out. The transformation enabled by such invisual platforms seems to produce one form of operationalization that describes not only a technical but also a political reformatting of the world that is one version of what Benjamin Bratton coins as platform sovereignty operating “within territories that are composed of intersecting lines, some physical and some virtual.”32 This new political geography of data rearranges planetary visibility according to a different logic than traditional cartographic projection.

Go could be claimed to be just another abstract and formal example that the history of (symbolic) AI has mobilized so many times, but in fact, platform invisuality adjusts to multiple kinds of live situations too. Many examples of research projects with cloud robotics, machine learning, and other experiments with making the real world into an observed “image” testify to this: consider, for example, the work on Dex-Net cloud robotics and point cloud data of 3D object models. Moreover, the perceptual capacity of machine vision, learning, and platform invisuality is not restricted to “ready-made” pixel arrangements: platforms are involved in different dynamics, from urban life to environmental earth observation to agricultural innovation—even “platform ruralism.”33 Such powers of operationalization define its particular grip on different material sites of reality: the smartphone, the image sensor inside the smartphone, urban and nonurban sensors from access cards to facial recognition systems to thermostats to driverless cars (see chapter 5), a long list of systems that are often covered under “surveillance” but include much more specific cultural techniques that are nested with other cultural techniques (such as pattern recognition). Platforms thus both distribute (image-events of observation and capture across various dynamic situations of people and things) and integrate (synchronizing, synthesizing these events into series, into mass-images in datasets). It’s a push and pull of images and data, of sensor capture and projective modeling that drives these operative ontologies of the visual/invisual.34

In other words, invisual images demonstrate one updated version of the pairing of the visible/invisible as a shift to the observation of data relations and patterns. The operational image is integrated into platforms as well as machine learning procedures that take images as training sets and models. The refashioning of surfaces that can be read as images can be seen to work as one form of abstraction, but it is in all forms and ways real. These abstractions and invisual events effectively establish interfaces to those material worlds. This platform-based way of observing and modeling is also about integrating material worlds into data operations and producing handles that can change those worlds: “Unlike other geographic projections, the interface is not only a visual representation of an aspirational totality; it is an image of a totality that when acted on also instrumentally affects the world.”35 In addition, operations of statistical modeling are actively involved in real-world situations, thus becoming one form of a technique of intervention. Abstractions are not unreal, and they are not separated from material surfaces.

As you can see, this meshwork of terms maps how a transformed notion of the image relates to questions of abstraction regarding platform operations. But in our case, the main focus is still on the coupling of operations and images; what sort of an understanding of images and operations is adequate in discussions of data and digital culture. As acknowledged throughout my discussion, the questions are not entirely new, and there are many answers already that relate to software, interfaces, and interactive screens. The argument about images as part of the broader programmable, executable culture was recognized earlier in software studies with different emphases by Alex Galloway, Wendy Hui Kyong Chun, and others. For Galloway, like in games, computational images are primarily geared toward execution and action.36 Chun opens up the question of software as execution to the genealogy of what is narrated as visible and invisible in computer history from aspects of gender and programming. Furthermore, following on from discourses of computers as information machines, this relates to how transparency is being sustained as the ideological backdrop for a machine that is primarily meant to make visible.37 This theme features in contemporary big data analytics and data visualization, too; how to draw patterns and make visible the otherwise unseen from datasets substantially bigger than the usual corpus of, for instance, humanities research.38

So while we could continue describing the various technical formulations of platforms and images, it is clear that the stakes of those techniques and humanities-focused discussions about data and (operational) images are not the only context. Some of it concerns methodology and how we address images in contemporary technical humanities; some concern “images in the wild,” which implies their political economy as much as their political ecology. Here, the platform as a central feature of capitalism, political geography, and digital culture39 is not a mere extension of the operational image but a site where it is executed with additional force, adding to the early 2000s context of discussions about software and visibility. While we have inherited a rich set of ideas about technical images as mobilizing an ontology of the invisible and bringing it to play a key part in various institutional forms of knowledge (and coercion) over the past 150 years, the platform describes a more recent way images are formatted into platforms and how they format the world in patterns of value, knowledge, control, and more.40

This can be discussed in terms of a “political economy of digital data.” As defined by cinema scholars Ruggero Eugeni and Patricia Pisters, it pushes us to analyze “the logics of production, circulation, and transformation of light (both from an optical-geometric and a physiological perspective); also, it drives us to evaluate the intertwining of these logics with that of the production, circulation, and algorithmic transformation of digital data—without forgetting the principles of sovereignty that govern these dynamics and determine their trends.”41 In short, and of particular interest for visual and invisual studies, the question of the image is what mediates this shift between the political economy of light and data.

As a more specific unit for a “political economy of digital data,” platforms are a condensation of economic transactions and, as such, set the stage for what Sekula articulated as the link between photographs and money. Now, though, it’s images beyond photography, and data, that are linked in this way. Here, the platform is not merely a replacement of the market as a meeting place of providers and customers topped up with various access-for-data arrangements but an intensive apparatus for extending logistics of data to a variety of urban and nonurban situations. In short, it is not only facial recognition as surveillance, but agricultural solutions, landscape surveys, spectral signatures, and many other things that are aggregated and, in that process, turned into a particular kind of a mass-image.

Furthermore, the platform turns the world into its own image. Like a recent collection of speculative texts probing the transformation of cities into platforms asked: What is the city as Uber, the city as Instagram, the city as Palantir, as GroundTruth, as Amazon?42 Here, the question of images as money is not merely about the circulation of photographs but the platforms that enable property and labor regimes that can be characterized by their operational invisuality. In other words, in the bundle of images, things, and people, platforms are also legal arrangements that assign positions of visibility while being operationally invisual: sellers, buyers, (gig) work, images, copyright, and other regimes of intellectual property through which the world is shaped in a particular anamorphic fashion, to follow Matteo Pasquinelli’s use of terms.43 But whatever we might refer to as (anamorphic) distortion, we can also refer to in terms of an expanded vocabulary concerning design in platform capitalism.

Engaging platform invisuality as posthuman property is an insightful way to focus on the materiality of data as an operative force in concrete spatial contexts.44 This term, mobilized by Jannice Käll, helps us to understand operational platforms in relation to broader spatial contexts of digital technologies. Autonomous vehicles and their way of (in)visualizing urban space is one example that I will follow up in chapter 5. But already, I should add that such an approach to spatial justice refers to how space is formatted by platforms, continuing the earlier points by, for example, Sarah Keenan on the law as production of space and place “from nation-states to public parks to eruvs.”45 Such a merger of law and invisuality is useful in helping us understand how platforms reorganize the world.

The operationalization of images is effective in contemporary contexts of data and where visual studies are not the only route to such an operational reality of observations.46 The operational image can be seen working in various contexts outside military targeting and vision; it is integrated into various other sites of execution where images do not anymore look like images. Or, even if they do, their primarily operational value might be as property, for example, one form of the continuation of war (a more frequently recurring reference point for operational images).47

Data Shots

Invisual culture is introduced as a core thematic and conceptual mediator that grounds much of the discussion in this book. The invisual is not to be mistaken as synonymous with the invisible, but they do stem from a shared concern about the transformation of images in different knowledge and aesthetic practices. Mackenzie and Munster’s analysis of the platform becomes a helpful scaffolding for many of the other ways I want to engage with operational images beyond the earlier input of Farocki and others. Investigations of invisuality—such a peculiar term that holds on to the legacy of visuality while denouncing it—are helpful as insights into the continued question about images and operations. But the answers as to where to address this question are somewhat different: some concern platforms (and thus legal and economic forms of the operationalization of the image), and some concern diagrams (thus inserting themselves into a long history of visual practices that have a particular role to play in mathematical, statistical, and relational forms of knowledge). Thus the concept of the invisual becomes a central way of understanding what operational images are in the contemporary context. In this way, they add an important qualification to the book’s argument: a shift from visual to invisual practices where images are not merely bundled up as datasets but operationalized across different institutional uses through platforms.

Two implications come to the fore: First, as per Hito Steyerl’s already quoted words: “Not seeing anything intelligible is the new normal.”48 Put differently: images don’t look like images at all, not that they are (anymore) invisible either. It is not even seeing, perhaps. Images are, rather, invisual. And second, images in and as data (sets) focus on patterns and models operationalized across platforms of aggregation, ordering, and training. It’s a complex set of nuanced transformations where “images” are sometimes anachronistic terms used for data but are still, in some cases, also a process of operationalization of the history and archives of existing photographs and other images.

Images are treated not as individual pictures but as entities on a different scale that does not fit in easily with a traditional vocabulary that assumes a human seeing subject who recognizes an image as an image.49 Images come in masses, as a mass-image that is increasingly the site of verification and validation of the world as it concerns various practices of power and knowledge.50 The image might be part of a logic of a target (war and administration, as in addressing), but it is also a statistical distribution governed by one (as is the case in digital photography and computation of images already in-camera). Abstraction and photography have a long history already. The data-intensive operations upon images continue some of that legacy,51 which could be said to include not only various practices lumped under surveillance, extraction of data, and commodification of that data, but also other techniques that take their aim at the face, the expressions, gestures, and actions: datasets such as Moments (MIT-IBM Watson AI Lab) of three-second videos of people, animals, but also objects and natural movement, Kinetics (consisting of YouTube clips), and AVA for “spatio-temporal localization of atomic visual actions.”52

Many of these then fulfill several earlier dreams of archives of cinematic gestures and bodies that were of interest to many twentieth-century thinkers and cinematographers and also act as media datasets of techniques of the body. Picking up the thread from the previous chapter on cultural techniques, here is a good moment to remember the line from Marcel Mauss describing how bodies walk, climb, swim, wash, soap, eat, and more. Eschewing the binary of nature and culture, Mauss also observed what cinema had done to bodies (with a particular gendered bias in his observation):

A kind of revelation came to me in hospital. I was ill in New York. I wondered where previously I had seen girls walking as my nurses walked. I had the time to think about it. At last I realised that it was at the cinema. Returning to France, I noticed how common this gait was, especially in Paris; the girls were French and they too were walking in this way. In fact, American walking fashions had begun to arrive over here, thanks to the cinema.53

While cinema, according to Mauss, was one form of archiving and then educating (forming and regulating) the body in a transnational fashion—enacting a particular form of a cybernetic feedback loop before cybernetics54—the contemporary datasets as categorized and categorizing snippets of bodies in movement constitute a second-order continuation, often extracted from existing media archives, whether from Hollywood or YouTube’s eclectic collections. Many archives-cum-datasets thus continue a particular form of institutionalization of the body that recircuits forms of micropolitics of the body. We can also call this the operationalization of the body registered and labeled in particular typologies, statistically reproduced in such models that return to format how the world of bodies is perceived.55 The operation is situated at one instance of perception-action and across a chain of operational links in an infrastructural system.

Any singular body, image, action, and reaction seems to give way to the aggregated mass image that fuels platform invisuality. This pertains to both the operations of datasets and the ensuing ways those are mobilized. As in, for example, the broad field of computer vision where “seeing is superseded by calculating probabilities” and where “vision loses importance and is replaced by filtering, decrypting, and pattern recognition,”56 as Steyerl argues. Once again, the image is decisively seen as something beyond represented or pictorial while it is tightly tied together with a statistical distribution of patterns regarding objects, people, and events. What a paradox this seems at first. We see things that refer not to things but statistical distribution. We observe people, but we do so based on statistically conditioned fine-tuning on the level of digital images, or in the case of AI, statistical models. We might still watch pictures, even moving images that feed back to a gestural repertoire available, but our gestures are already integrated into datasets that feed into training sets and model what gestures become meaningful.

Hence the shift from a discussion of digital or networked images57 to datasets of images is relevant in terms of our topic and the site where operational images function. The operational image is located not necessarily in a particular shift of ontology of image that is often specified as “digital” but in how images are operated upon and become operationalized through aggregation, algorithmic analysis, and the ensuing questions of data-driven mobilization of the mass-image. This chapter revolves around platforms, datasets (such as ImageNet), neural nets (and other machine learning techniques), and models in order to understand operational infrastructure.58 And although they are not really images as such, they are at the center of questions of operational images.

To continue focusing on datasets and the mass image, the ImageNet, also famous for the ImageNet Large Scale Visual Recognition Challenge competition, has become a go-to place for discussing photography and visual culture changes. In addition, it has become a recurring example in discussions on the role of images in the pipeline of machine learning techniques and computer vision. But many other image datasets could be mentioned too, corporate and noncorporate, urban and rural, human and nonhuman. Such examples range from Microsoft Common Objects in Context (or COCO) to Open Images (CC BY 2.0 license), from Fashion-MNIST (of Zalando online shop items) to FieldSAFE (for agricultural object detection), from LILA BC (dedicated to biology and conservation, such as datasets about glaciers and other land cover) to National Agriculture Imagery Program (NAIP) datasets. These examples bootstrap different approaches to the use of images—and what constitutes an image in the first place.

Of course, ImageNet is, for many good reasons, a central dataset practice of the past ten to fifteen years where changes in (in)visual culture become visible. Its extensive work in the compiling and naming of images consists of a vast logistical operation where the image—and the model processed from images—is created. The tens of thousands of Amazon Mechanical Turkers doing this work are but one component in this logistical operation, alongside other elements that speak to the centrality of taxonomic categories as taking place where images used to be. At the back of a longer history of principles of machine learning and computer vision, which is outside the scope of this little book and my argument,59 it is fair to say that ImageNet’s status in the 2010s helped to understand some aspects of the changes in photography through computer vision.60 This especially concerns the evolution of datasets, as Fei-Fei Li, the leading figure behind ImageNet’s work at Stanford University, argues. Furthermore, as Nicolas Malevé points out, this shift points to the move from the algorithm-centric world of image processing and computer vision to the way “modeling functions in contemporary machine learning systems”61 through large datasets compiled from a variety of sources—like in the case of ImageNet from the internet (Flickr).

Access to useful and usable large datasets and the availability of efficient machine learning techniques such as convolutional neural networks become the technical operation that starts to define a particular invisuality of images. Also, the ways of looking at them no longer allow us in any meaningful way to refer to them as “visible” or “visual” only.62 The automation of seeing, vision, and invisual operations incorporates a different sense of a subject and how knowledge about images is handled.

Analyzing what an image is and what is in an image is not simply about the decomposition of the constitutive elements, even if this is often how feature extraction is presented. To quote Malevé for a useful explanation about machine-learned images and models:

The developer assembles a data set reflecting the variations of the domain under study and utilises automated means to calculate an optimal function that treats the features of the data as parameters. In computer vision, this technique, at its most simple level, uses large visual databases in which discrete units such as pixels can be considered as data points. Common techniques of machine learning in computer vision are said to be “supervised,” which means that the data is curated . . . to provide examples from which the machine learning algorithm extracts regularities: the software “learns by example.” To come back to the case of the cat, in the data-oriented paradigm, the developer does not try to decompose the animal in distinct shapes and explicitly summarise their relations. Instead, she curates a large series of photographs, where the cat is displayed in various positions, and lets the algorithm detect the regularities traversing the various samples. Through this phase of “learning,” the algorithm produces the model of the cat.63

This is both an analytical and, significantly, a synthetic operation as it concerns the automation of modeling as an assembly of large (training) sets. Through these series of aggregated datasets for training, models are produced for use and reuse in different applications. This process is specific to how machine learning works but think of it in relation to Mauss’s techniques of the body and cinema as a circuit of capturing, registering, educating, and reinforcing.

Moreover, not only a generic category is observed: a cat, a body, even an identifiable movement. The aim is to see the mass-image as not about a taxonomic thing but as consisting of experimental processes of weighting perceptual parameters for particular features. This amounts to finding the right operational balance between any cat and a group of cats struggling in the wind at five in the evening.64 The latter is not so much a quantified example of a general group but a singularity in its own right. Regularities are extracted, which is very much the stuff of the promise of AI, but the complex meshwork of this perceptual apparatus is not evacuated due to this promise of standardization.65 Any mass-image in a dataset or the ensuing trained model is always situated in a world of singularities, and observation goes beyond recognition of a thing in the world. Even the statistical model that reproduces and reformats the world is a singularity in its onto-epistemological being: one technique with the operative ontological force that shapes the world.

Photographs—already transformed from a picture to a compression format (such as JPEG) optimized for digital circulation66—are thus part of the infrastructure of datasets that become the more effective and operational end of images in this regime of (in)visuality. Yet, as has been argued by Malevé, Kate Crawford, and Paglen, but also by Daniel Chávez Heras and Tobias Blanke, the photographic persists as remediated as well as a black-boxed element in machine learning datasets. The photograph is structured as a historical reference point of (computer) vision while the complexity of the photographic beyond a representational collection of things, people, and situations is easily ignored. Indeed, this mediation—often voiced as naive realism of photographs as if they were just pictures copying reality—is one place where the rich historical reality of diverse practices and materials becomes invisible and is seen primarily through the mechanisms through which an image is made relevant. (Cue Barthes and Panofsky from the beginning of the chapter, but updated into the question of the medium of machine learning and datasets.) The photograph is thus constantly redefined in relation to contemporary practices, which more or less bundled it and nested it inside their own mechanisms of analytic definitions, thus engaging in a technical and discursive redefinition of what an image is.67

Chávez Heras and Blanke argue that a lot of this work black boxes the photograph itself as a historical array of materials, lenses, and visual practices of taking pictures. In addition to research that argues that the labeling of training data images is a central site of politics (including politics of distributed global labor of Amazon Turkers), Chávez Heras and Blanke point out how we need further “archaeologies” of images too: to excavate what forms of mediations are already bundled up in (photographic) images. Unpacking machine learning procedures must also involve unpacking the assumptions about images. Hence, computer vision techniques as packaging particular kinds of image-theories—as Leonardo Impett has demonstrated68—can be extended to alternative, even conflicting, genealogies of photography where debates about representation, nonrepresentation, measurement, calculation, and materiality are implicitly part of the landscape of images as data. Finally, Chávez Heras and Blanke argue that forms of optical computation—“glass” computers as they coin it—are embedded in the camera lens, which is already a form of technical mediation and automated synthesis.69

I am not after accounts that explain machine learning through optical processes, but it helps to understand that different forms of computation are already bundled up in the operational scaffolding of visuality. Thus, it is useful to include such details if one starts to examine the assumptions about images in machine learning. Furthermore, Chávez Heras and Blanke’s methodology for doing this is immanent to the techniques concerned: their alternative engagement thus consists of assembling “a dataset with which to analyse photographic practice, and then use it to train a bespoke focal length classifier as a proof-of-concept for a system designed to investigate the optical perspectives implied in MbM [“Made by Machine: When AI met the Archive” dataset].”70

Thus to rephrase the earlier points: machine learning is a technical purification and standardization of what an image is that also models the image in ways that exclude much of what it has been as a material practice. As a process of layered selections, decisions, and training, this is instrumental to how the pipeline of AI takes place. This concerns datasets specifically constituted about images, such as photographs, and/or employs what is assumed as a so-called photographic image. The operationalization of photographic images includes both making invisible their rich nuances and historical development (let alone the variety of experimentation of what constituted the photographic in the first place) and the praxis of working with images without looking at images (or hiring cheap labor to do the looking instead).

This is not to promote a kind of nostalgic take that we should merely critically look at images as individual solitary pairs of eyes, like a particular stance historically assumed, as museum objects, or as other forms of (an)aesthetics. Instead, I propose that this approach runs through a longer genealogy of images as administrative media where images are instruments involved in bureaucratic ordering procedures. Sekula’s work is one early example of recognizing this aspect in the “instrumental potential in photography,”71 where even without a full-fledged review, one can pick up the cues in his discursive takes through the focus on photography as it shifts parameters of social ordering, and presents a “new juridical photographic realism.”72 This is where the linking of photography and “archival rationalization” was “designed to contribute directly or indirectly to the practical transformation or manipulation of their referent.”73 This manipulative, formatting power of operations persists as our reference point throughout the attention paid to distinct practices of ordering that present visual reality but not necessarily for the aesthetic gaze. Paglen’s famous coinage that an “overwhelming majority of images are now made by machines for other machines, with humans rarely in the loop,”74 is likely to be correct, but to be fair, most images were never meant to be looked at in the sense of aesthetic consumption. Instead, they were part of administering data: about people, nature, territories (like geological, geographical surveys) as well as in laboratories of chemistry and physics involving images with a very different eye to aesthetics.75

The same applies to contemporary practices of manipulating worlds through data and computational operations. Images take place, but they also disappear. As Malevé points out, from Flickr to a dataset such as ImageNet, a lot happens between, with all sorts of exclusions already taking place at that stage. Again, this could be construed as a version of Barthes’s argument about the disappearance of images in the process of fabrication of datasets. For Malevé, though, this is the paradox resulting from the photograph’s dual existence and disappearance in relation to operations of data and machine learning. “The paradox is that to resolve a photograph as data, all the heterogeneity that pertains to the medium, its apparatus, and its circulation needs to be repressed. The photograph needs to be made a transparent vehicle. And to make it transparent, computer scientists need to engage in the production of their own apparatuses and produce their own alignments.”76 But only one kind of image disappears, while another image surfaces. What was called “photography” might disappear, all sorts of data relations and scales might appear. This disappearance is not a romanticization of historical modes of hermeneutic understanding. Instead, it acts to point to the centrality of the image as it becomes a measured surface flatness, a function that was already present in much earlier uses of photogrammetry (see chapter 3). This is, in and of itself, incredibly impactful when it comes to images taking a role in intervening in the world, which is to say that it is also part of a long nonlinear history of images taking action.

Obviously, there are many direct political stakes at play. Many of them also include images as their central hinge of operations that hit the ground, hit the dataset, hit the bodies involved. Multiple forms of bias have become a central reference point for evaluating the social impact of AI in this sense of the mass-image. Representational forms of analysis of visual culture are clearly insufficient in dealing with the pattern recognition and data-operations at play that still affect how and what kind of bodies are at play, visible in which ways and at what scales, as a target or as labor, as a collective pattern of statistical data.77 Algorithmic techniques take a central role in those operations. Different infrastructures (classification mechanisms, datasets) become a site of activism. Joy Buolamwini shows us so in her work Gender Shades, which maps existing skin-type distribution in datasets used for facial recognition training in relation to the Fitzpatrick Skin Phototype Classification (FSPC). But besides highlighting the underrepresentation of black individuals, she also created a new bespoke dataset, “the Pilot Parliaments Benchmark (PPB), to achieve better intersectional representation on the basis of gender and skin type.”78 Cross-reading the Buolamwini project in relation to work on machine learning and politics of statistical distribution of patterns, we arrive at points articulated by Ruha Benjamin, Matteo Pasquinelli, and others: building on existing “world bias” where “datasets [might] reinforce race, gender and class inequalities”79 normative power can further continue the institutional bias in statistical and algorithmic forms that can significantly amplify such distortions, to return to Pasquinelli’s point mentioned earlier.80

The question at hand is not so much a perspectival distortion observed by a gazing subject but a distortion at the level of invisual statistical distribution: it concerns collective distributions, data, and forms of operationalization in and out of images. Furthermore, for a consistent politics of statistical distribution of patterns, we should acknowledge that “art and media are fundamentally about the deception of sensory organs,”81 which would just put data technologies in that infamous lineage of evil media:82 persuasion instead of negotiation, trickery instead of representation, fooling instead of communicating. Or, to put it differently: instead of saddling ourselves with “bias or no bias” as the two alternatives to choose from, the stakes are more clearly formulated in the question of what forms of justice are built into the various computational, institutional, and statistical techniques.83

Laboring people, too, disappear or are (re)morphed in operations of datasets under capitalism.84 The (racialized) arrangement of bodies and operations of extraction, logistics, and finance are central engines in the circulation of what this book discusses in and through and as operational images. Read thus, the dissection by Malevé and others regarding the Amazon Mechanical Turk work in those terms: what was once in the early nineteenth-century phases of photography described as the new medium’s defining characteristic—objective accuracy—is being transferred onto a laboring body. It is the body that needs to be accurate and coordinated and to see according to preset labels and taxonomic categories.85

To look at masses of images at high speed in order to annotate them, the mechanisms for reaching consensus among the laboring Turkers about annotations is a peculiar assemblage of eye, hand, and cognitive coordination in relation to the platform environment. The mass-image is constructed in and through platforms that anonymize labor as a version of the micropayment gig economy. In Malevé’s words, which describe in fascinating ways the amount of cognition—embodied and platformed—going into the image process:

To annotate at speed does not consist of a mechanical response issued from a passive subject. To understand the annotator’s contribution, it is fundamental to understand the process as one of elaboration which goes beyond rational choice and explicit judgement. The epistemic contribution consists in embodying a scale, figuring out rhythms and levels, understanding and refraining involvement. To attend to the process of elaboration means to avoid concentrating exclusively on the semantic decision. The elaboration is not limited to a pivotal moment, where the annotators assert the meaning of a photograph. It includes the complex methods through which they synchronise within an alignment and embody a scale.86

An active image. An activated body. An image that might be programmable. But so are bodies in their rhythms and scales. The earlier assemblage of chronophotographic data visualizations of bodies, work, and movement defined the connections from pioneers of images as recording instruments such as Marey to pioneers of management as intervention such as Taylor. This pair of recording-intervention and scientific measurement-management/administration recurs in many instances ever since, like a media archaeological topos of sorts.87 Many of the themes in those feedback loops of labor and operative images return in the distributed set of microlabor. The labor of annotation exemplifies thus also the algorithmicizing of the body or, in other words: it is not only the images that might or might not be operational, but the bodies that are instructed and operated upon accordingly to clean, organize, and produce the (mass) image. This point relates closely to Joanna Zylinska’s term “undigital photography” that aims to probe the long tail of practices, including human labor, that defines technical images in the age of AI,88 but one can also point to various avant-garde art historical precedents of programming the performer, and instructing the operator.89

Beyond avant-garde arts, the work of technicians of the image and laborers paid to look at images is at least as important as the usual histories of innovators and inventors. The history of operational images is, in this sense, the history of anonymous images and anonymous labor, too, as they are essential in the preparation of what becomes visible. If one wants a historical counterpoint to contemporary preparation of images into datasets into models, one example that can be quoted is the meticulous work that went into the preparation of photographic astronomy images from glass plates into reproducible copies in the late nineteenth century: a technique of mechanical accuracy was underpinned by a long infrastructure of trained hands making halftone prints to ensure the invisible became visible. The little story about operations that format the astronomic observation of celestial objects of light into an image is one where the political economy of light and labor, of scientific work and objects of knowledge, intermesh. While it shows that the technical, even scientific, images had relied for a long time on various kinds of hand-eye coordination in preparation of such images—in the case of scientific accuracy in depicting nebulae and stars, this meant paying attention to brightness, contrast, and other values as they would stand out against the pitch blackness of empty space background—there are of course differences that furthermore emphasize the point. The relation to and respect of such manual labor was different, a point that is somewhat summarized in this anecdotal quote from 1911 by astronomer Charles Perrine, who “reminded his colleagues dealing with slow engravers of ‘the advisability of not forcing them to rush work on fine subjects.’”90 Similarly, the work of female computers at Harvard Observatory is one strand in the gendered history of operational images, as noted earlier.91

Intermezzo: Labor on a Single Platform

The (gendered) labor that prepares and calculates images is often invisible and outside of the image itself. What would these scenes, described above, look like if they were a Farocki film? Consider describing the image being labored into a dataset as a version of Farocki and Antje Ehmann’s project Labour in a Single Shot. Started in 2011 and featured in workshops and as exhibitions, the project is organized as what sounds like a simple instruction of filming a work sequence in a one- to two-minute shot in ways that capture aspects of the sequence and the gestures of the body. A cinematic reference point is Workers Leaving the Lumière Factory (Louis Lumière, 1895), but now these short scenes feature contemporary work:

cobblers, cooks, waiters, window cleaners, nurses, tattoo-artists, or garbage workers. But most of the work activities happen behind closed doors. Often labour is not only invisible but also unimaginable. Therefore it is vital to undertake research, to open one’s eyes and to set oneself into motion. Where can we see which kinds of labour? What is hidden? What happens in the centre of a city, what occurs at the periphery? What is characteristic and what is unusual with regard to each city? What kinds of labour processes set interesting cinematographic challenges?92

The above is a workshop description of the brief for Labor in a Single Shot. We can follow up with what might sound a peculiar question: What would be the single shot of a process that is distributed through the Amazon platform, coordinated across multiple workers doing the same task, and resulting in images being interpreted so that they disappear in the process?93 Can this shot be taken? Does it have a specific relation to cities or some other geography, territory, or architecture? Perhaps it is a collection of single shots, of distributed works, coordinated into a collection of multiple images about multiple images being worked on and annotated. The image being operated upon—through labor and as a technical process—disappears twice, but the shot itself would frame this disappearance through proxies that inscribe and trace the reframing of the images (mass-image) as they are readied to become models.94

This could be called a model image or an image-model.95 In either case, the stakes are clear: the images are prepared to feed into a model that feeds back to what is captured in perception. Such models are not necessarily visual nor tangible in the architectural sense of “modeling” but are statistical and mathematical functions in machine learning. This also means that the image becomes a statistical distribution while being delivered and maintained by labor and the platforms that organize that labor. While platforms synchronize logistics as bodies, software, routines, data, and production of images (such as in the example of the ImageNet and other uses of the Amazon Mechanical Turk), this “labor on a single platform” (to misquote Farocki and Ehmann’s project) is one aspect of what is being referred to as the invisual.

Model Images

Images are compiled from datasets to platforms in various institutional, para-institutional, and cross-institutional uses. As Estelle Blaschke has shown, photographs are carriers of images. As images, they are also carriers of data (which admittedly can also be metadata such as image texts, annotations, etc.): standardized, scalable, and “driven by the desire to make images ‘more’ exploitable for a variety of uses—from business and industry to scientific research and knowledge transmission.”96 Images include various levels of taxonomic and other data that help to format their usability for a multitude of purposes, and where the image itself becomes an entry point to data aggregation that leads into the synthetic modeling of the world. The work put into perceptual AI and robotics indicates the challenges where sensor-based realities of action and movement are entangled with questions of operational images: the gap between sensing and images becomes minimal.

In addressing the administration of operational images, I want to underline that significant techniques, operative ontologies, and practices of invisuality help us understand some of the stakes in data platforms and computational contexts such as machine learning. “The model,” “the platform,” and “the dataset” are some terms that speak to the operational image without a primary reference point in optics or the pictorial. As you might have seen by now, this is not to ignore historical discussions but to argue that those are also to be reevaluated in light of data and instrumental imaging questions that stretch the notion of “image” toward some nonimages like charts, tables, and diagrams.97 To address, even briefly, the “model” in this context brings together many aspects at play, from images as instruments of intervention to the mobilization of data in and beyond the sensorial.

Images have acted as instruments in the history of various laboratories in experimental psychology, plant physiology, physics, and biology, as well as in the pipeline of production of models in contemporary machine learning. In scientific practices, photography of the invisible shifted the ground of verification of what is (and is in) the image.98 The status of evidence could be questioned when there was no way of referring to a “ground truth” outside the apparatus. Synthetic ground truths emerge much earlier than in recent AI-training data practices, and reality had to be approached through models of reality. While photography of nonphenomenological physical or astronomical worlds might have triggered this decentering of ground (truth), it became more widespread in the context of twentieth-century knowledge practices: the quantum measurement apparatus as elucidated by Barad was already quoted (in the previous chapter) as an example of this.

Arguably, one can say the same of earlier, much simpler optical visual technologies such as Galileo’s telescope in the early seventeenth century. In a way, that apparatus concerned an early example of a calculated data practice of visuality: “The telescope creates the senses anew: it defines the meaning of vision and sensory perception, turning any and all visible facts into constructed and calculated data.”99 According to Joseph Vogl, the telescope established its aesthetic–epistemic argument not only by empirically seeing more but in how this apparatus operated upon its own conditions of what counts as seeing and knowing. The assemblage that brings together sensing, images, heavens, skies, planets, observation, and data is also a creation of a particular world of mediations and, in Vogl’s words, also of limitations. It produces a bespoke model of what perception is, its own model image of what counts and can be coined as anesthetics:

The critical point of the historical analysis of media is not to be found in what a medium makes visible, tangible, audible, readable or perceptible; it is not so much located in the aesthetic of the data and in-formation provided by a medium but rather in the anesthetic side of a media process. Again, what does Galileo Galilei see when he turns his telescope toward the sky? What exactly are the visibilities that Galileo observes, then captures in his texts and drawings—the lunar surface, unknown fixed stars, the Milky Way, the moons of Jupiter? Sidereus nuncius leaves no doubt: Galileo sees, newly perceptible in his telescope, not just sun, moon, and stars but the difference between the visible and the invisible.

In other words, this operation upon the invisible also produces the main limitation of the epistemic sweep brought about by the telescope: to realize that because of this assemblage, “every visibility now bears a stigma of provisionality; every visibility is surrounded by an ocean of invisibility. Everything visible remains contingent, forever encompassed by the imperceptible and the unknown.”100

Later, (photographic) images take a place as models of invisibility. As Geimer points out, late nineteenth-century and early twentieth-century scientific photographs function as models in the fundamental sense that gathers more epistemic force gradually over the twentieth century in relation to physical, mathematical, architectural, and computational models.101 Here again, the question of what counts as proof and evidence—and what this implies about the invisible, inconceivable, even invisual—stands out:

The pictures may figure as evidentiary material or indispensable illustration in a demonstration, but their existence as such is not already proof “that something invisible may really exist.” Their function more closely resembles that of a model. “An experimental model system has always something of the character of a supplement in the sense Derrida confers on the notion. It stands for something only the absence of which allows it to become effective.” . . . The model never fully coincides with what it represents, and this systematic misprision is the source of its productivity. Applied to the case of photography, this means that the visualization of an “invisible fact” did not straightforwardly undo its “invisibility.” What the photographic image presented to the eye was not the spoils of a foray into the darkness of the invisible world, hauled back unscathed into the world of visibility.102

Instead of merely capturing the visible and the invisible, the focus on models helps to consider images as experimental systems. This concerns photographs and other kinds of images closer to data visualization too: “Camera obscura tracings, photographs, and the inscriptions of self-registering instruments”103 can be named as part of what Daston and Galison coined as the regime of mechanical objectivity of (roughly) the nineteenth century. More recent experimental systems are likely to include machine learning and cloud analytic practices.104 But to address telescopes and photographic (and related) techniques also includes the awareness of them as instruments that produce epistemic reality, which is why they became integrated into the scientific practices and institutions that have a vested interest in images as invisual operators. Invisuality starts early. Think of Marey’s chronophotographs as one instance where such expanded visual instruments are essential to capturing data105 (see chapter 3), but consider also the above-voiced suggestion to see photographs as models in their own right. Will this pairing up of experimental and instrumental uses of images also help to figure out the scope of the operational image in contemporary data culture as a particular model of what an image is?

A jump from telescopes to images and contemporary datasets, and machine learning training and other techniques, is not a linear one though. But the point about the provisional epistemic conditions that work through seeing/images is relevant. To follow this sketch about models also implies that every dataset is provisional, every pattern recognition implies a different possible pattern, every training set implies an alternative statistical model that could be created. In this operative ontology of the image, the roles are reversed. If anything that exhibits a pattern to be read and reconstructed is an image, it is the technique that produces the world as an image. It is not that images are being analyzed, but that the analysis produces potentially anything as an image—a never-ending cascade of patterns, images, perceptions.

In this sense, part of the book’s argument is that operational images are scattered across models, diagrams, and graphs functioning in various assemblages of knowledge. Indeed, in the contemporary context, we are likely to find such image operations outside the laboratory and the military in all sorts of assemblages of machine learning and image (training) datasets. They are also found in the multitude of sensors that capture faces, body postures, gestures, and more: environmental sensing and subsequent imaging, corporate practices for urban and nonurban uses from policing to agriculture, gaming and other forms of motion capture and interaction design, and a set of other things that propose a radical detachment from what images in the photographic context were.

The aim of producing images as standardized, scalable, and ready-made for operational (administrative) uses underpins images as models also beyond scientific institutions. Images as models (and data that help build models) emerge gradually in the history of technical media, even if this aspect becomes much more obvious later in the computational and digital knowledge practices. In other words, a model in the historical sense outlined by Geimer or Vogl does not suggest the same thing as a model that emerges as a core feature of deep learning in contemporary computational practices where “the model is the statistical representation of a large and diverse training dataset into one file.”106 Nor is it that what’s a model in the photographic sense is the model that emerges in the quest to map the hitherto invisible scales of climate change as a mass aggregation of data.107 But the question of the model becomes a site of investigation of the change of images, as it becomes a site of investigation of data and knowledge, which is why it warrants placement alongside the discussions of platforms and datasets. Model, as such, also becomes an epistemic device to investigate the operations of images and operations of formatting the world in their own invisual image. Such models are not visualizations of the world but operational interfaces through which changes are enacted, whether in the contexts of urban planning, military operations, or earth observation of land use and land cover, vegetation, agricultural crops, erosion, and so on.

In many ways, this focus on the model turns the tables: the contemporary practices help to see images as models, diagrams, platforms, machine vision. Such experimental systems have more than a rhetorical relation to invisible worlds. In other words, the invisible is not merely what is not seen, but it is also what is produced as part of the practices of establishing what is knowable. This also features in contemporary practices of data and machine learning from cultural techniques of pattern recognition108 to techniques of probabilistic modeling and the centrality of datasets as they are mobilized in the operationalization of the image. Knowing and seeing become probabilistic, and the models that are produced are rehearsed as rhetorical “seeing” from the identification or misidentification of terrorist subjects to border control to other situations with specific geographical and spatial characteristics. This kind of operational image is not merely an image that targets; it is a statistical model that builds an invisual environment as a model of the world through which actions can take place.109 The world is formatted for such epistemic conditions of actions; platforms are one example of this capacity to format.

Gradually, our understanding of image, visuality, and perception changes as those terms become part of machine learning experimental systems. For example, statistical models can become a multidimensional vector space110 representing a very different link to “visual” than one of “images.” An investigation of graphical notation systems in mathematics and visual operations of diagrams as part of a media archaeology of technical images would be one way to build a longer historical argument about depictions of contemporary machine learning and data culture as an interface of statistics, images, and material territories.111

We are far from Galileo’s telescope in historical time and technological logic. Still, one point of similarity remains: a world is created anew as it is being observed, and not one of heavens and planets only, but across a different kind of a vast field of data operations. Here, models have a mediating function while becoming, recursively, one example of operational images as something else than “just” images.

Still Not Just Images

This chapter addresses different variations on the themes of invisible and invisual images and, in the process, ends up in machine learning and datasets. Part of this has meant mapping the discussions on operational images in relation to contemporary AI and data culture work. In so doing, the chapter has focused on images as data operations in a world that is urban, dynamic, modeled, simulated, and layered in metaphorical and technical ways (such as neural nets producing models based on training data). Images do not feature as images but as platforms, datasets, and models.

Building on the earlier chapter, I proposed that we look at the productive processes and valorizations, the operative ontologies, of what is at stake in this invisual visual culture.112 Some of the stakes of those operations concern value creation. Some are in existence in the institutional arrangement of knowledge creation. Some are emblematic of the forms of power (including law) that distribute, as Foucault would have it, what is visible and sayable in the first place.113 Much more could be said—and has been said by other scholars—on all of these issues. For example, in chapter 4, I will continue these arguments in relation to work on the scientific data platforms of the Chinese Digital Belt and Road initiative and how the artist duo Geocinema has engaged with that aspect of remote sensing and imaging. In chapter 5 the points are continued in relation to the urban traffic of data as one example of corporate data practices of operational images of mobility and lidar scanning.

As this chapter comes to a close, it is tempting to ask: So what is an image if it is lost somewhere inside the machinations of photography and datasets and machine learning? Some decades ago, scholars could still respond by saying that images are like language. For example, W. J. T. Mitchell suggested this in the mid-1980s: “Instead of providing a transparent window on the world, images are now regarded as the sort of sign that presents a deceptive appearance of naturalness and transparency concealing an opaque, distorting, arbitrary mechanism of representation, a process of ideological mystification.”114 This move was meant to help discuss the constructed and even political nature of images, moving away from the naive realism that images depict things. In some ways, critical insights to software and execution articulated an updated route: programming languages are a special case of visual knowledge, as Chun pointed out, where politics act not only through natural languages.

What would be an even more updated version? What is an image if it is not just an image but also data? What is an image if it is not just an image but a model? What is an image if it is not just an image but also a platform operation of distributed machine vision across multiple devices and observations that can aggregate those observations under a platform logic (and business model) and execute those as real changes in the world (e.g., in legal terms such as property and control)?

Such contemporary answers are no less political than Mitchell’s suggestion about the language of images.115 But, likely, many of the answers are now implied in the work of scientists and engineers dealing with images. “Images began to function at least as much as a tweezer, hammer, or anvil of nature: a tool to make and change things.”116 This point by Daston and Galison about the history of scientific practice is apt in the broader sense, too: much before and after poststructuralism, images were something else than just images or language.

Many of these responses, though, are at first sight pragmatic: Which datasets of images can be mobilized in particular convolutional neural network environments, and how to find computational power for dealing with training sets of high-resolution images? For example, the deep learning neural network from machine perception to image (dataset) to statistical model-pipeline is practically processed as GPU processing power that incorporates invisual operations inside the computational architecture.117 Ranjodh Singh Dhaliwal refers to this moment of (in)visuality in terms of rendering as computation, underlying the centrality of parallel processing (GPUs) and platforms such as NVIDIA as part of a refashioning of contemporary digital capitalism. In short: it is not only that the invisible worlds are being rendered visible for any perceiving subject but that rendering is considered a central part of the computational apparatus—a political ecology as much as a political economy—in which operational images also take a central role. Jacob Gaboury’s work on the archaeology of computer graphics makes this point in detail: such digital images are far from images in the representational sense and are instead complex computational events of rendering that draw upon invisual data to produce visible surfaces.118

Observation and seeing are peppered across a heterogeneous set of events in the world which are integrated onto a platform, which feeds on the anamorphic perspective to images themselves: twisted and transformed, designed and designing, they model the world according to their own distortion (or programmed rendering).