Tracking Circulation
Where is the meaning of an artwork located under current conditions? Most of how the art world talks about itself assumes meaning sits in the work, in the conditions of its making, in the artist’s intention, in the institution that places it, in the market that prices it, or in some combination of these. When art is challenged or assessed, the defence is usually given in one of four registers, each of which assumes that meaning is somewhere recoverable. It had a political effect. It was placed in a recognised institution. It sold for a recognisable amount. It expresses something authentic about the person who made it. Each names something real. Object, intention, institution, and market each carry meaning. Meaning also exceeds them. It is also carried by the circulation systems through which works move, and shaped by that movement. The four registers describe where meaning settles when art is held still long enough to be assessed. Circulation describes what happens to meaning when it moves, which is most of the time. Tracking the movement is what this practice attends to, running next to the registers and picking up what they leave out. Circulation operates on a work in two distinct relations. For works that are made under conditions where production and circulation are temporally separate, a painting completed in a closed studio, a sculpture cast and then transported, a manuscript finished and then sent, circulation is what happens after the work has been made. For works that are made under conditions where production passes through circulation infrastructures from the start, AI generated images whose prompts route through training distributions, video uploaded and reshaped by platform compression as it is being edited, networked works whose form depends on what the network delivers at runtime, circulation is one of the conditions of production. A third case sits between these two and has become increasingly common: works that are made before they circulate but whose first encounter with audiences is through documentation, press release, and social media rather than through direct presentation. The work and its mediation arrive together. The reading often cannot easily separate them, and asking what the work is, separate from how it first reached us, is no longer always the right question. Tracking circulation attends to all three, and treats the relation between production and circulation as itself describable. Whether the work was finished before it circulated, met its audience through mediation, or was constituted in the act of circulating, is one of the things a careful reading registers. Tracking circulation is the practice of attending to what these infrastructures do to the work, and to what the work does back to them. The framework here comes from working inside the art world as it actually operates, sitting on a state arts panel, running an exhibition space, teaching, applying for grants. The setting is compromised and less than it claims to be. Reading from inside it is the only kind of reading available to me. The four registers, the three forms of residue, and the loop are not unique to art. Versions of them organise how research, design, journalism, and other forms of cultural production justify themselves and circulate. I describe the art world version because that is where I work and where I can read carefully. Whether the framework extends cleanly to other domains is a question for readers situated in those domains. The claims here come from a loop of proposing, making, and reading. The maker proposes a concept or a description, makes something under its terms (a work, a system, an image, a script, a piece of hardware assembled from recycled parts), reads what the making actually produced, and lets that reading reshape the next proposal. Theory, practice, and analysis each contribute to the loop. None has final authority over the others. The reading is geared toward learning something new. Toward challenging what was thought the moment before, incrementally falsifying the assumptions the proposal rested on, prompting new questions rather than affirming what the maker already believed. A reading that comes back without having cost the reader anything has not done its work. Falsification here is a soft notion. The reading can show that the proposal was incomplete, mistaken in a specific respect, or built on an assumption the work itself did not support. This is what makes the loop a loop rather than an illustration cycle. If the reading only confirms the proposal, the loop has stopped working. Falsification shows up in concrete forms. A caption the maker has to abandon because the work will not bear it. A prompt the maker abandons after the generations show what was actually being asked for. A hardware constraint that turns out to be productive rather than limiting. The cost is the loss of the proposal as it stood. The threshold between noise and signal is not formal: it is whether the divergence asks for a revision the maker can actually make. I. Art often gets defended in one of four registers, in the panels I sit on and the conversations I have with colleagues, in the applications I write and the reviews I respond to. When a work is challenged or a practice is asked to justify itself, when it is assessed by critics, curators, collectors, or grant committees, the defence tends to take one of four shapes. This work has political effect. It addresses a specific issue, advances a specific cause, gives voice to specific groups, raises awareness, provokes thought. This work has institutional legitimacy. It is included in the canon, the museum, the major exhibition, the grant programme, the curated catalogue. Recognised institutions have placed it inside themselves. This work has financial value. It sold, it was acquired, it commands a price, the market accepted it. This work is authentic. It comes from a real personal vision, a particular biographical experience, an unguarded interior, a true voice. These are not the only registers in use. There are also affective registers, what the work felt like, what it moved, what it cost the viewer. Phenomenological registers too, what the work organised in the room, how it shaped attention. Religious, devotional, therapeutic, and pedagogic registers exist and are invoked in particular settings. The four named here are the ones that dominate when art has to account for itself in public discourse, before panels, in reviews, in institutional defence. They are the registers that have to hold up under pressure. The others are present everywhere but rarely have to stand alone as justification. These registers often overlap in practice. Institutional placement can produce financial value, personal authenticity can be institutionalised, financial success can retroactively legitimise. They function as distinct grammars of justification. The first three name real things. Political effect happens, institutions legitimise, markets pay. The fourth is more complicated. Personal authenticity is invoked constantly. The underlying claim that meaning originates in the unguarded interior of an individual artist, and leaves traces only that person could produce, is more contested cultural myth than fact. Art is more collective than the fourth register admits. Ideas are shared, circulated, absorbed, recombined. Styles and gestures pass through studios, schools, and scenes. What appears as the artist’s unique interior expression is almost always a sedimentation of other people’s work, conversations, and inherited forms. The fourth register is a working fiction that shapes how art gets funded and presented, more than an accurate account of how art gets made. It belongs in the set of four because it is what people actually say. Aesthetic judgement is one of these alternative registers, and it deserves a closer look because of how often it gets folded into the four rather than standing on its own. It speaks to the work’s accomplishment rather than to its warrant. The four registers describe how art justifies itself in public discourse when challenged. They are grammars of justification, distinct from locations of meaning. The distinction matters because grammars and locations get confused. The four registers describe what a work is taken to be. Circulation describes what a work does to the conditions that carry it. A work defended in register 2, institutional legitimacy, is taken to mean what its institutional placement says it means, and a work that does not yield to such a defence is taken to be empty of meaning. Both inferences are mistaken. Justification grammars are what get presented to panels, critics, and markets when meaning has to stand up to assessment. Circulation describes the much wider field in which meaning is at work most of the time, including when it is not being defended at all. A different question is also worth distinguishing, because the four registers leave it out as well. What about pleasure, beauty, formal accomplishment, craft, the sensory experience of a work? These are present everywhere in art and account for much of why people look, make, and care. They are doing different work from the four registers. Aesthetic judgement is what works invite from viewers. It operates partly inside institutional infrastructure, with critics, curators, jurors, schools, and partly outside it, in studios, audiences, scenes. Institutional legitimacy houses much aesthetic judgement without exhausting it. Tracking circulation rubs up against aesthetic judgement on a different axis. A work can be beautiful and carry circulation features worth reading, formally accomplished and offer little, carry strong features and be aesthetically uneven. The two questions are independent in principle, related in practice. Aesthetic accomplishment often amplifies what circulation does with a work, since well made images circulate more and settle into visual vocabulary more reliably. Pleasure can carry circulation effects precisely because it disarms critical resistance. Formal accomplishment can also obscure circulation features when a work is so resolved that viewers stop looking. The two readings cross constantly, in both directions. A video uploaded to a platform changes the moment after it has been compressed by the upload pipeline, indexed by the recommendation system, surfaced in some users' feeds and not others, embedded in some thumbnails and not others, scraped into some training corpora and not others. The four registers ask what is this work? There is also a fifth posture, more recent and more diffuse, that asks instead whether the work can be trusted, whether it was captured or generated, whether it deserves attention at all. The forensic posture differs from the four defence grammars. It is performed by the viewer rather than offered by the maker. The forensic posture asks what the work is; tracking circulation asks what it becomes. The former seeks verification; the latter tracks transformation. Tracking circulation asks what does this work become as it moves, and what does the movement leave? The failure of the four registers is most visible where there is no work as object in the conventional sense at all. AI generated imagery sits at a particular point in this argument. It breaks the four registers more visibly than older image cultures do, because the makers, the works, and the conditions of circulation all sit outside the categories the registers can recognise. It also makes the algorithmic component of circulation more legible, because the mechanisms (training, compression, generation) can be inspected to a limited extent. AI image culture is not the only case the framework reads, but it is the case where the framework’s necessity is hardest to deny. Consider the genre of AI generated images circulating across Facebook groups, Reddit boards, and image archives. Cursed AI and its cousins, the steady production of generated images of Donald Trump as king or baby, of Greta Thunberg as saint or martyr, of impossible architectures, dream logic creatures, villages that never existed, futures and pasts the historical record does not contain. Many of the makers have few of the classical image making skills. They are not placed in galleries, not collected, not paid, and not trying to express a personal interior. The four registers cannot place this image culture, and in the conversations I have had about it, in studios and at openings, it tends to be dismissed as kitsch or visual junk. This is one place where this kind of reading becomes useful. In the overtly political imagery, political figures rendered as exaggerated archetypes, in some cases these contribute to the visual vocabulary in which those figures can be pictured, reaching populations the official political imagery reaches differently or not at all. In the nominally apolitical imagery, impossible architectures, creatures, alternate scenes, the contribution is to what configurations of life can be pictured at all, often in worlds pointedly other than the present. The escapism, read carefully, has political features the maker may or may not articulate. The reading does not claim the work does politics regardless of intention. It says specific features of the work and its circulation can be described as politically charged when this kind of attention is brought to bear, whether or not the maker would describe them that way. Articulation is not the test. Precise description is. The work for Frihed, lighed og hiphop was produced and posted on social media as I developed it. The first captions came from a need to make available my reading of the images and give them some context that might explain why I found them interesting enough to share. The captions were therefore wholly unlike the prompts that made the images. They were formulas after the fact rather than part of the iterative process of generation. They came from a need to frame the images and make a possible reading available as they went into circulation. A note on lineage The recognition that art happens in circulation rather than in the object is not new, and the lineage runs deeper than the twentieth century practices most often cited. Long traditions of pattern, ornament, weaving, calligraphy and architectural surface have always understood images as fields that accumulate meaning through repetition, transmission and reuse rather than as single objects to be assessed. Several twentieth century practices made the circulation explicit and treated it as the substance of the work. Mail art from the 1960s onward used postal infrastructure as a medium, with works existing in their distribution and accumulating annotations as they passed through hands. Fluxus, from 1962, treated event scores, multiples, and the exchange of objects between practitioners as the substance of practice. Net art in the 1990s used the early web as both medium and circulation, producing works that could only be encountered in the act of being passed along, embedded, broken, mirrored. Each was attentive to the circulation conditions of its moment. What is different now is that the circulation has become hybrid. The social and institutional systems these earlier practices worked with are still present, and they now run alongside, and through, computational systems. Training corpora, the bodies of images and text used to teach a model. Recommendation engines, the systems that decide what surfaces in whose feed. Embedding spaces, the high dimensional fields in which a model represents the patterns it has learned. Generative models, the systems that produce new images from those patterns. Each earlier practice was scoped to the infrastructure of its moment, and each developed vocabulary that does not extend cleanly to the moment after. Mail art does not handle the moment when an image scraped from a mail art archive becomes part of a training distribution and re-emerges as statistical residue in unrelated generations. Fluxus does not address the algorithmic recommendation layer that decides which documentation surfaces in whose feed. Net art predates the platform and AI hybrid consolidated since the early 2010s, in which the work’s circulation includes statistical absorption into systems that then generate new images and reshape the imaginary the work was contributing to. The lineage is real. The vocabularies are partial. Tracking circulation extends the same critical attention into the layer the earlier practices did not have to address. II. If meaning is approached through circulation rather than through the object alone, reading requires vocabulary for the components of circulation. The word I have been using for what disciplined reading picks out, across these components, is residue. It names a reading effect, what becomes describable when this kind of attention is brought to a work, rather than something lying in the work waiting to be discovered. Residue is not whatever the method notices. It is what persists across careful readings by readers practising similar commitments. A feature that registers on one reading and not on another is not residue, it is noise, or it is something specific to that one encounter the framework has not yet learned to describe. The work is treated as a surface that the reading moves across, returns to, finds different things in on different passes. Reading is returning rather than penetrating. The surface is where the work is done, not a layer to look through on the way to a depth the reading is supposed to recover. Residue is to tracking circulation roughly what theme is to literary close reading or condition is to conservation science. A working term for what the discipline produces as describable. Three components of circulation are worth distinguishing, because each operates by a different mechanism and each requires a different kind of attention. The framework most centrally holds all three in view at once, because the most useful readings come from divergence between the three. Its second methodological commitment is to recognise when residue is thin or absent. A reading that finds circulation features in every work it examines has stopped reading and started rubber-stamping. The framework gains discriminatory power from being able to say honestly that the residue is weak, the reading collapses, the work does not sustain this kind of attention. Material residue. What the conditions of production and material infrastructure leave in the work. The projector manufactured under forced labour that runs the video. The sustainable paint in the canvas. The recycled hardware in the installation. The white walls of the gallery that house everything. Material residue is causal. A work is made under specific conditions and those conditions are carried into the work in ways that can in principle be tracked. It is the most stable of the three forms because the mechanisms are familiar. Production studies, infrastructure analysis, supply chain tracing, conservation science. A gallery’s white walls carry the history of connoisseurship, bourgeois display, and a neutrality friendly to the market into whatever is shown on them. The walls do this before and regardless of the work hung on them. The damage reduction approach to computers leaves me with last-generation hardware and open-weight models that produce less photorealism than mainstream ones. The server warms my studio in winter and I hear its fans spin up. To me the cloud is not an abstraction but present and infrastructural, made local. Imaginary residue. What circulation makes seem natural, available, or thinkable to those who have moved through it. The patterns of perception, memory, and expectation that circulation produces over time. Take a familiar example. The raised fist as a posture of solidarity appears to enter popular democratic use with the international labour movement in the early twentieth century, picked up by trade unions, socialist and communist parties, then by antifascist movements between the wars, then by pan-African, anti-colonial, and anti-apartheid movements. The exact origin is not fully clear. What is clear is that the gesture has accumulated, across more than a century of photographs, posters, films, and album covers, into something legible at a glance to viewers who have never been told what it means. No single image taught the gesture. Many images, across decades and contexts, settled it into the visual vocabulary of what political solidarity looks like. The raised fist is what circulation has made describable as imaginary residue. A feature of how seeing is now organised, traceable to no single source, accumulated through repetition and circulation rather than through any individual deposit. The same mechanism operates at longer time scales. The single point perspective that organises so much of what counts as a coherent picture space took several centuries to settle into the apparently natural way images are read, and is sustained by the continued repetition of buildings, photographs, screens and frames that lay it down again. The temporal range of imaginary residue runs from the recent to the very old. Imaginary residue is the most contested of the three. Several scholarly traditions have already described pieces of what the framework calls imaginary residue. Mere exposure effects in cognitive psychology, the tendency for repeated stimuli to be evaluated more positively. Cultivation theory in media studies, on how heavy viewers come to share the worldview of what they view. Maurice Halbwachs and Aleida Assmann on collective memory. Aby Warburg on recurrent visual formulas. Saidiya Hartman on critical fabulation as a way of recovering circulation patterns the archive does not contain. The framework treats these as partial, supplementary accounts and draws on them as such. Imaginary residue is real, can be tracked in specific cases through iterative work, and its tracking requires holding the descriptive question, what feature of perception has settled, carefully separate from the normative question, does it matter. The methods for tracking it are familiar ones. Comparative formal analysis of a feature across a scene over a decade, attending to when the feature first appears and how it spreads. Reviews and writing about works, looked at across time for when a form stops being called new and starts being called obvious. Studio genealogies that trace specific formal moves between teachers, students, collaborators, and scenes. Conversations with practitioners over years. With the raised fist, all four would yield something. Each gives a partial account. They are starting points. The flags and slogans of the labour movement remain potent, if somewhat docile, inside the current narrative of the labour movement. By repurposing them as hand-painted signs made from AI sketches, for example Frihed, strangeness og hiphop, the imagined freedoms are reactivated. A collective remembrance of collective struggle couples with the less than obvious hiphop, exposing the connection between cottage industry and mixing spray paint, taking over the means of production and using the record player to make new music instead of relying on the recordings of others. Algorithmic residue. What the work and its circulation leave in the computational systems that now index, rank, recommend, and generate images, distributed across training data, recommendation engines, and the weights of generative models. The mechanism is partly old. Images have long been accumulated, filtered, and redistributed by archives, catalogues, schools, canons. What is new is more legibility than purely social and cultural circulation has historically had, and a sharp acceleration in the rate of accumulation. Prompts can be run, outputs compared across many prompts, behaviour mapped systematically, and where models are open weight or training corpora public, the infrastructures can be inspected to a limited extent. This remains mostly observation of effects rather than direct examination of the infrastructure, because most contemporary models are proprietary. Their training corpora are not published, their weights are not released, their embedding spaces are not accessible, and the trace left by any specific image is statistically entangled with billions of others, not straightforwardly readable as the contribution of that image in isolation. Even with open weight models, examining the weights directly produces little that reads as cultural or political content. What one actually reads is the model’s behaviour. Algorithmic residue is more legible than what came before. Reading it is still a practice rather than a direct inspection, and the legibility is a matter of degree. In the Blue Light Diffusion project I trained my own diffusion model as best I could, using a script I wrote myself and a dataset of mostly blue objects I own. The resulting images were non-figurative but still looked photographic in texture, with a blue mass at the centre of the image, lit from a low angle as the February light in our flat is, and surrounded by a warm brown picked up from our floorboards. No recognisable objects remained, and yet there was a clear trace of the original. A relationship across compression, and something new at the same time. The three forms operate by different mechanisms. Material residue is causal trace. Imaginary residue is cultural sediment. Algorithmic residue is statistical imprint. These are different kinds of trace. The word residue holds them together as a heuristic. They share a structural pattern, in which something is deposited or transformed by a work or its circulation that can be read for political content. The three operate by different mechanisms, accumulate at different speeds, and fail in different ways. The three are held together as a working set, not as a single model. The framework distributes attention across them without proposing a unified account. The reason to hold the three together is that each becomes more legible when looked at through the others. Using the generative model as a lens on cultural memory sharpens what is meant by sedimentation, distribution, and survival. Using cultural memory as a lens on the generative model makes describable what kinds of particular get absorbed into a generalisation, and what kinds drop out. Using the material conditions of a work as a lens on both keeps in view that neither the imaginary nor the algorithmic is ever immaterial. Coordinating attention across the three is what tracking circulation does in practice. The most specific thing this framework picks out is the coupling of imaginary sedimentation with algorithmic accumulation. Older theories of cultural memory described the slow formation of what becomes thinkable through circulation, institutional reinforcement, and repetition. What has happened in the last two decades is that a second accumulation mechanism has come into operation at industrial scale, running in parallel with the first. Images are not only seen, remembered, and passed on. They are also statistically absorbed into systems that will reshape future seeing. Imaginary residue accumulates through human attention and cultural repetition. Algorithmic residue accumulates through statistical compression and the optimisation of model outputs. The two overlap where human seeing trains models and where model outputs reshape what becomes thinkable. They differ in how fast they accumulate, in how stable the accumulation is, and in what kinds of particular survive the process. The two mechanisms run next to each other, feed into each other, and are increasingly impossible to describe separately. Training local models on recycled hardware, prompting them and reading the outputs, and comparing what shows up in the generations to what one has seen circulate has been the empirical setting in which the vocabulary of three forms of residue took shape. The methods are most useful where they register conflict rather than confirm a single story. A divergence case is more telling than a convergence case, because divergence is positive evidence that the reading is registering something specific. This is the methodological payoff of holding the three components together. Divergence between them is the kind of finding that single mechanism approaches cannot easily produce, because they are not tracking two things at once. Take an example. Certain poses and aesthetic configurations from early twentieth century European modernism, the geometric portrait, the constructivist composition, the angular industrial body, show up heavily in the outputs of contemporary generative models, in the sense that the models reproduce them readily when prompted in adjacent directions. The imaginary tracking shows that these configurations have largely settled out of ordinary visual vocabulary. They appear citationally, in deliberately retro contexts, rather than as configurations a contemporary viewer encounters as just the way things look. The model has absorbed the imagery. The shared imaginary has moved on. The divergence is real and worth reading. The reading itself runs into a question it cannot fully resolve. Such a divergence could indicate genuine cultural lag, dataset bias, prompt distribution artifacts, or compression effects specific to particular model architectures. The framework cannot always distinguish between them on the basis of the divergence alone. Treating model behaviour as a clean observational apparatus for cultural reception is exactly the kind of move the framework should resist. A different divergence runs in the opposite direction. The imaginary residue carried by centuries of perspectival picture-making leads viewers to expect a particular kind of coherence from any image that looks photographic. The generated image enters under those credentials and is read as a photograph until the substrate declares itself, in the form of the procedure’s own returns: the face that has become a family resemblance, the crowd that tiles itself, the architecture that converges on two vanishing points. The divergence between what the algorithmic substrate is doing and what the imaginary leads the viewer to expect is one of the most consequential cases the framework registers. The dominant response is to read the divergence as failure. The framework reads it as a finding about how three different accumulation mechanisms operate at different speeds and to different ends. Two cases from working with models illustrate what this looks like in practice. Trying to generate multi-ethnic groups of people with the Flux model is difficult. The model tends toward uniformity, and the attempt makes describable how homogeneous commercial imagery has been for long enough that the training distribution has absorbed it as the norm. The model does not invent the bias. It amplifies a bias the imaginary has been carrying for decades, even where the imaginary itself has begun, unevenly, to move past it. The divergence between the model’s confident output and what an attentive viewer now expects to see in a contemporary image is the kind of finding the framework is built to register. A second case operates differently. Faces in Wan 2.1 video generation are unstable. They change while the back is turned, drift across frames, settle into a different person between one shot and the next. The instability is a compression artifact, and reading it that way is incomplete. It also makes describable something about the contingent and provisional character of identity in image space, which is harder to articulate when faces stay where the viewer expects them. What the model fails to produce becomes a way of describing what it would mean for an image to hold a face still in the first place. Within algorithmic residue there is a narrower moment. What I have elsewhere called survival residue. This is what survives of the particular inside the generalisation performed by a model during training. The trace that remains of a specific image once it has been compressed, distributed across weights, and averaged with millions of others. Most images that enter a training set are absorbed into the distribution without a distinctive trace. Survival residue is the specific case where the particular manages to remain legibly present inside the general. It sits at a point where the political stakes of tracking circulation become especially clear. What is carried forward, and what is lost. One way to put what generative systems do for the framework is that they make circulation partially observable. The training distribution, the compression, the procedure’s own returns, the divergence between what the model produces and what the imaginary expects, all of these are visible in ways the social and cultural circulation of older image systems was not. The framework does not depend on generative systems being special. It depends on them making visible aspects of circulation that older image systems concealed. The same mechanisms have been operating on photographic and pre-photographic image cultures for centuries. What is new is that they can now be inspected, however partially, in the layer where statistical absorption happens. III. The discipline of attention this requires is comparable to close reading, formal analysis, conservation science, or institutional critique. Each of those disciplines produces specific perceptions through its training and its commitments. Tracking circulation produces a different set of perceptions, with different commitments. Each of them, including this one, looks at a work through some framework. The framework here is one among several. What it makes describable is what the four registers leave out. The reading depends on conditions on both sides. Works without the relevant material and historical conditions will yield only weak and uneven readings. Readers without the relevant commitments will miss what is there to be read, or will read it poorly. What this discipline picks out is real in the sense that different readers practising similar commitments can converge on similar readings. It depends on the reading itself, in the way close reading and institutional critique do. What holds the practice to falsification rather than to comfortable iteration is partly personal discipline, partly external pressure. Left to itself, the loop has well known ways of going stylistic, where the reading selects what survives the proposal rather than what costs it. The intermittent pressure of external readers and the discomfort that builds up when a practice produces increasingly stylised work without genuine revision both help. These pressures are local and partial. Together they hold the loop to learning. Some of what the reading is doing, before any vocabulary is brought to bear, can be put in concrete terms as questions one finds oneself asking while looking. Does the image, object, or action suspend disbelief enough to warrant engagement? Does it hold long enough for a reading to begin at all, or does attention slide off it. Does it allow consideration of itself as an image, an object, a body, or only as a carrier of predetermined meaning? Is the work autonomous enough to be read, or has it dissolved into illustration of its own thesis. Is it reassuring, or does it leave me wondering and thinking afterwards? Does it close down or open up. Does it cost me something or does it confirm what I already thought? Does the encounter revise the proposal that brought me to the work, or does it return me to where I started. These are working questions rather than criteria. They overlap with aesthetic ones. A work that holds, that is autonomous, that opens up rather than closes down is also describable in aesthetic terms, and the overlap is one of the places aesthetic and proto political reading meet. A work that fails the first two questions, that does not hold attention, that resolves into pure thesis illustration, is a work the reading will struggle with. The conditions for circulation features to register are not in place. A work that fails the third and fourth, that reassures, that confirms, is a work the loop cannot learn from, however well the work performs in the four registers. Falsification in artistic practice is largely tacit. It happens across several channels, visual, material, embodied, infrastructural, and the verbal reading often comes after, catching up to what the practice has already registered through other channels. A generated image looks wrong in a way that takes a moment to articulate. A prompt produces something the proposal had not anticipated and the maker recognises this without being able yet to say what it is. A hardware build behaves differently than expected and the unexpected behaviour itself reorganises the proposal before any reading is written. We know more than we can tell, in Polanyi’s phrase, and in artistic practice the tacit channels of falsification are usually the leading ones. The verbal reading remains important. It is what makes the practice articulable to others and to oneself in the longer term. The practice does not wait for it. Recognising the different channels is part of taking the loop seriously. An instance from a project on Danish church architecture. The proposal was to generate a medieval Danish church in the background of a scene. The model produced a copper-spired church recognisable as German rather than Danish, every time, with small variations. The first reading was practical: rewrite the prompt, find the descriptors that would pull the model toward the Danish typology. The second reading was the one that mattered. The model’s confident routing of Danish medieval church architecture into German visual vocabulary is itself a finding about how the training distribution carries northern European architectural history. The question this opened was whether the work could absorb the divergence, perhaps by relocating the narrative to Germany, or by reading the slippage itself as part of what the work was about, the entanglement of Danish and German Protestant architectural lineages the historical record actually contains. In the end I judged the divergence too far from the narrative framing of that particular project and removed the image. The decision was a reading. The loop registered the falsification, considered what it might mean, and resolved it by cutting rather than rewriting. The relation between the bodily registration and the analytical vocabulary is worth being explicit about. The questions above, and the multimodal falsification described earlier, are preverbal. They happen as a tug, a discomfort, a noticing, a sense that something here is worth attending to or that something here is dissolving on contact. The five commitments below, and the three forms of residue, work on the articulation that follows. The commitments tell the maker how to read what has already registered. Materialist attention gets directed at a work because something has already pulled at the maker’s looking. A quality of the surface, a detail of how the work was placed, a recognition that the conditions of production matter here in this specific way. The verbal description of those conditions comes after. The tacit registration is what the commitments organise. The commitments are what makes the registration communicable, defensible, and buildable. The two layers operate together, and reading practice is partly the discipline of letting them do so. Five commitments specify what the reading itself consists in. Materialist. Attend to what the work is made of, how it was produced, where it is shown, and how it circulates, including how it circulates through algorithmic systems. These are central to meaning. They are where material and algorithmic residue become describable. A video made on recycled hardware in a shared studio and shown on a laptop screen in an artist run space is a different work, materially and therefore politically, than the same video shown at a biennial on calibrated monitors. Both are part of what disciplined attention can pick out, and both bear on what the work contributes to circulation. An instance from my practice. In Blue Light Diffusion I trained a small model on a salvaged graphics card using a dataset of blue objects from my flat. The constraint was not abstract. The card limited the resolution and the training time, which set what the model could learn. The recycled hardware kept the work tied to its production conditions in a way cloud generation does not. The materialist reading then picks out what the limitations produced: a model that returns blue masses and February light from a Copenhagen flat rather than a generic latent-space blue. Critical. Take the work’s stated intentions as one part of what can be read in it. What the artist says, what the gallery text says, what the title suggests, these matter. Careful reading often picks out something more, or something different. The work can be read as doing more than it states. The critical task is to attend to that surplus. An instance from my practice. The captions for Frihed, lighed og hiphop propose a grassroots labour aesthetic. A critical reading of the work has to attend to what the captions say and what the images do alongside them. The images route through training distributions that have absorbed mostly American urban settings, and getting them to look local requires prompting them as if they were in the US first and pulling them back. The captions name a position the images can only approximate, and the gap between the two is part of what the critical reading registers. Generous. A stance toward the work. Go willingly. Assume the work is trying to do something interesting until the work itself makes you abandon the assumption. Going willingly is a disciplined choice about how to meet a work in the first place. Looking for what might be useful in a work is the rarer act, because looking for wrong has become the default. An instance from my practice. In a sequence of generated portraits, several figures arrived with hands the model could not resolve, fingers running into each other, the wrong number of digits, knuckles in places knuckles do not go. Many actual people have hands and limbs that look nothing like the model’s idealised version. Visual culture has trained viewers not to expect to see them, and the glitched hand registers as failure only against the smooth and symmetrical hand the training distribution has settled on. A generous reading holds the unresolved hand as a body the field has not learned to picture, kept rather than corrected. Rigorous. A constraint on the claim. Rigour is what the reading owes the work once the generous stance has been adopted. Generosity is a form of discipline, with its own demands. The reading has to be accurate, defensible, and specific to what a careful reader can describe in the work. Rigour also requires handling the very common case where a work produces contradictory readings across the three components. A work that is materially progressive but imaginatively conservative, or materially compromised but imaginatively generative, or algorithmically active but materially and imaginatively thin. The response is to describe the contradiction precisely and let it stand. Rigour also includes the honest negative reading, where the framework recognises that the residue is thin or absent and says so without forcing the work into the vocabulary. An instance from my practice. States of Diffusion is materially refused (recycled hardware, open weights, plain text captions) and imaginarily generative (a non-corporate technical culture that current visual vocabulary does not carry). It is also algorithmically compromised because the open-weight model is downstream of the same corporate training pipelines the work is critical of. A rigorous reading names this contradiction precisely and does not collapse it into a single political verdict. The work is doing something useful in two registers and is implicated in what it critiques in the third. Rescuing. The reading attends to what in a work might survive being moved into different circulation conditions and still hold. A posture, a configuration, a structural move, a way of placing materials. The test is transferability across circulation contexts. The conditions that produced the work are mostly not coming back. Most of what was made under those conditions is bound tightly enough to those conditions that it cannot be carried at all. Some of what was made carries elements that retain force in new circulation environments, and the discipline of rescuing is the practice of telling these apart. The term has a lineage in Walter Benjamin’s Rettung, redeeming a fragment of the past from the continuum of history, and is closer in spirit to Adorno’s later reading of the same concept in Negative Dialectics, which holds onto the particular without the messianic apparatus. The use here is more modest and more practical. Rescuing recognises what is worth carrying without pretending the original conditions still apply. It is not a retreat into nostalgia but a tactical de-optimisation. Placing the slower, resistant material of manual craft, salvaged hardware, or social forms the system treats as obsolete next to the fast returns of the algorithmic layer produces a divergence the system does not smooth over easily. The divergence makes the infrastructure of the present visible. The hand-painted sign or the analog gesture is rescued not because it is better, but because its presence registers as a divergence rather than as input the system can absorb. The outdated and the optimised, placed in the same work, expose each other. An instance from my practice. For Frihed, lighed og hiphop I made hand-painted signs from AI sketches, the slogans of the labour movement and of hiphop rendered as signs you might see on an allotment shed. The hand-painted sign is the rescued form. Placing it next to the AI-generated source material puts the slow, resistant material of manual craft into the same work as the fast returns of the algorithmic layer. The juxtaposition makes the infrastructure of the present visible without retreating from it. These five commitments describe what tracking circulation involves, done slowly and repeatedly, over the course of a practice or a body of work. They describe how practice based art research at its best already operates, once it has stopped assessing itself in the four registers that do not describe its object. A reading typically proceeds by identifying candidate residues across the three components, testing whether they persist across readings, mapping where the components align and where they pull apart, and describing the direction of the pull without resolving it. These are aspects of one act of attention, not stages in a sequence. The reader holds them at once and lets the work guide which of them currently most demands articulation. A small concrete instance, from my own practice. States of Diffusion is a series of generated images and written captions that proposes a counterfactual technological history in which graffiti culture and phone phreaking merged in the late twentieth century around the discovery of conductive spray paint, giving rise to a grassroots ternary computing movement spread across cities including Copenhagen, Boston, Lexington, and Vancouver. The images were made by iterating prompts against a generative model on local hardware. An initial prompt produces an image that exceeds and resists what was asked. The prompt is revised toward what the image is actually showing rather than what was asked for. The next round begins. Bringing this kind of attention to the work produces a mixed result rather than a tidy synthesis. Materially, the project is made on recycled hardware running open weight models, in conditions that refuse cloud based generation, and the captions are written in plain text. This carries something of the work’s politics into its mode of production. The open weight model itself sits downstream of the same corporate training pipelines the work is critical of, and the material refusal does not undo that origin. Algorithmically, the project is a deliberate use of a system trained on biased data. The model has seen mostly corporate technology imagery, mostly American urban settings, mostly normative bodies. The iterative prompt revision pushes the generations toward configurations the training data does not directly contain but can be assembled out of. Whether this counts as working against the model’s biases or as displaying them in slightly altered form is left open by the reading. Imaginarily, the images contribute to the visual vocabulary in which a non-corporate, community maintained technical culture can be pictured at all. Ternary computing terminals beside spray painted infrastructure, gatherings around triangular monuments in open fields, the texture of an alternative path. What registers as imaginary contribution depends on whether the work is seen as plausible but different. The work is mostly seen by an art world audience that reads it as counterfactual fiction. What can be picked out is therefore probably less than what one might hope, because the work does not circulate in the popular technical channels where the imaginary it proposes would actually accumulate. The three components point in partly different directions. The reading does not resolve them into a single account of what the work is doing. That non-resolution is part of what makes the reading useful. It sends me back to the next round of making with specific things to revise. Reading can produce a negative result, and the practice is only useful if the difference can be specified. A negative result is part of what the framework should be able to produce. When the reader cannot establish specific features of circulation in the work beyond what the four registers already describe, this is itself a finding. When different readers practising the five commitments produce readings that do not converge in any specifiable way, that too is a finding. The reading collapses when it is more coherent as a report on the reader’s prior commitments than as a description of the work. A concrete instance, from my own attempts. A competent abstract painting in a mid career artist’s gallery solo, made in oil on canvas, exhibited in a white cube and sold to a private collector. Examined carefully across the three components, the material residue produces standard observations about studio production, the gallery’s market position, and the canvas’s industrial supply chain. Observations a competent political economy reading would already make. The imaginary residue is thin. The painting fits comfortably inside an established formal vocabulary it neither extends nor unsettles, and what registers as imaginary contribution is essentially that the existing vocabulary is reproduced. The algorithmic residue is sparse. The work circulates only through gallery documentation and a small social media footprint, and there is no specific feature of the work that I can describe as being statistically absorbed into image systems in a way another work would not. A reading attempted across the three components produces very little that the four registers do not already say. Some works carry circulation features and some do not, and the framework should be able to recognise the difference. The practice is not a licence to find features wherever one looks. Some works, examined carefully, resist the reading or yield only thin versions of it. Saying so when it happens is part of the practice. A closer instance, from working at scale. Most images generated for Frihed, lighed og hiphop were not used. They failed in describable ways. Some looked unrealistic in a way that broke the proposal’s footing rather than productively. Some carried implicit politics the proposal could not absorb, the homogeneity of race the model defaults to, a setting clearly not local. Some were too recognisable as generated, through image texture, contrast, or colour, in a way that pulled attention away from the work and toward the apparatus. Some were not conducive to the metapolitics of a grassroots position, the longer political work of building a stance through such imagery over time. Some were too literal, leaving nothing to think about as images. Most of the discards were sorted on these grounds and never used. A smaller number got through despite these objections, because they balanced out other arguments in the overall body of work or filled gaps in the narrative in ways I judged more important than their lack of conceptual residue. Some failed the opposite test. They were too close to my own prior thinking, registered nothing new, and were therefore deemed inert. The framework here did not always produce a clean decision. It produced a sorting practice across thousands of generations, with conceptual residue as one criterion among several. Not everything that is kept carries residue, and not everything that carries residue is kept. The framework names what it can. The work decides what it can carry. IV. A few limits worth being clear about. There are places where tracking circulation is the wrong tool. Formalism does better with what is going on inside the picture plane, with the specific decisions a work makes about composition, colour, surface, scale. Political economy analysis does better with the structural conditions of art’s funding, distribution, and labour, where reading can register effects but does not have purchase on causes. Conservation science does better with the actual material life of works over time, where reading describes what conditions of production deposit into the work but stops well short of how the work changes once it exists. Each framework has territory. A direct question worth answering. What is the external constraint that decides reading failure? Tracking circulation has roughly the constraints of disciplined close reading and institutional critique. Convergence between careful readers, disciplinary review, comparison against other readings of the same work, and the loop with practice. A reading is strong if other readers attending carefully arrive at compatible accounts, if it registers things subsequent readings recover, if it survives being held against a work over time. It is weak if it depends on its own elaboration, if other readers practising similar discipline produce incompatible accounts, if it cannot be carried into a discussion that does not already share its terms. These constraints are real and soft, and they are roughly the constraints any interpretive framework operates under. The framework’s epistemic standing is closer to that of close reading than to that of measurement. The framework can fail in two opposing directions. Over-description, where the theory absorbs every work uniformly and produces readings that say the same thing about everything. Under-description, where the reading retreats into anecdote and personal recognition and cannot be tested or argued with. Both are live risks at all times. The framework slides between three modes that operate differently. Diagnostic, describing what becomes available to disciplined attention regarding circulation. Selective, ranking works by what their readings produce, which is normative whatever language softens it. Generative, shaping what gets made next, because readings of one round inform proposals for the next. Sliding between them without flagging the shift is where descriptive vocabularies become evaluative criteria. The framework is most useful diagnostically, most openly normative when used selectively, and most vulnerable generatively. Readers who use it should know which mode they are in. The argument does imply a direction. From disciplined attention to specific perceptions to whatever those perceptions might inform later. Naming this openly is more honest than pretending the framework is neutral about where it points. Some direction is built into any framework with political stakes. A serious objection to this whole line of thinking is that any new vocabulary, any new practice of reading, any new criterion of value, will itself be absorbed by markets, by institutions, by grant bodies, by fashion, by trained habit and rehearsed norm. Tracking circulation will become a performance. Critics will write circulation analyses of works that merely look the part. The vocabulary will become a buzzword. The objection holds. It is also the condition. The history of cultural and artistic vocabularies has been a history of partial absorption, where a vocabulary used carefully by some practitioners becomes used carelessly by many, and over time cycles back into something usable again. The work is to keep going, all the time, opening up the terminology, the architecture, the materiality, the funding structures, the pedagogy, the modes of assessment, the criticism. There is no lost Eden to recover and no future Utopia to wait for. There is only the ongoing work of loosening, of making cracks, and of paying attention when something interesting happens in a crack. These limits are not failures of the framework but conditions of its use. The framework operates inside them rather than from outside, and a few of those conditions are worth stating directly. The data is inaccessible. The training corpora, the weights, the embedding spaces of the proprietary models are not going to be released. The framework does not work toward eventual access. It works toward the agency that remains possible without access, the agency of observation, inference, reading for divergence, and disciplined inspection of what the systems do at their visible surfaces. This is a permanent condition, not a temporary obstacle. The framework is reparative, in a precise sense. Strategic optimism and willed agency under conditions of catastrophe: the climate cost of compute, the political agenda of the ultra-rich who own the infrastructure, the speed at which the conditions of cultural production are being reorganised by systems most makers cannot see into. The reparative posture is a choice to look for what might still be carried forward rather than to surrender to the conditions or to retreat into purity. There is no purity. There is no way to work with these technologies that resolves what is wrong with them. The position is to work with the technology to better hate it, not to escape implication. Damage reduction is the working term: not redemption, not refusal, but disciplined engagement under conditions one did not choose. The framework will not last. It is one move in a longer cat and mouse game of capture and rupture, where each critical vocabulary becomes legible, then absorbed, then forgotten or critiqued into the next position. The work is to keep moving, and to be useful for the moment one is operating in. V. This is a discipline of attention oriented toward where meaning can be located under current conditions: in circulation, distributed across hybrid social and computational systems, traceable through its material, imaginary, and algorithmic components. That has been the work of these pages. Two other things depend on it but lie outside. The preliminary work done here is geared toward what comes after it. The proto-political reading, what tracking circulation actually picks out as politically charged, the specific features that read as carrying political force when this kind of attention is brought to bear, depends on a vocabulary like residue, divergence, and the loop being in place first, or it collapses into the politics-of-effect that the first register names. The longer cultural-political work sometimes called metapolitics, what is built from such readings over time, in turn depends on the proto-political reading. These are larger undertakings with their own demands. The work here is preliminary to both. The aesthetic is not apolitical. It carries a residue of its origin and a potential rekindling. Under surveillance and totalising infrastructure, the aesthetic is one of the few places where the conditions of perception can be worked on at all. What was named earlier as a parallel grammar running on a different axis is also, under these conditions, where the political reading has to be done. The framework treats it as such, and the work that follows takes this up. The practice is offered for use, not held as property. Residue and its three forms, divergence between them as the framework’s central operational tool, the loop, the four diagnostic questions, the five commitments: not finished, not meant to be. Practitioners who find any of it useful are welcome to use it, extend it, modify it for their own circulation contexts, and break it where it does not hold. A working description of a working practice. This will not save us from ourselves. It is not supposed to. It proposes a way of looking, offered from inside the conditions it describes.
