I vaguely remember the first time I came across an artwork made with AI. It was a short film with impressive graphics. It had vibrant and captivating moving images that morphed from one scene to the next. I was in awe. I began to feel the shared sense of excitement in the air at the time, excited for the possibilities for making my own short films.
As someone who at the time never thought I could ever make these impressive and technically complex artworks myself, AI seemed like the perfect avenue for me to get closer to a world I had always wanted to be a part of.

Now about two years later, I have moved quite far onto the opposite side of the spectrum about the use of AI models in art. In fact, I often find myself feeling a negative reaction I can't quite articulate properly yet, when I come across these same kinds of artworks browsing online or on social media today.

The term ‘AI’ or Artificial Intelligence (Wikipedia 2026) has been used to describe ‘the capability of computational systems to perform tasks typically associated with human intelligence, such as learning, reasoning, problem-solving, perception, and decision-making.’ However in recent years, due to the rise of Generative AI models such as ChatGPT and Midjourney, AI is known for a different purpose - making art.
Upon its initial introduction to the masses, Generative AI models seemed to be the future. The newest breakthrough in human technology, the ‘inevitable’ turn in human history that was supposed to fix all our problems.
Suddenly it was everywhere, and as someone who was always more of a techno-optimist, I thought this was a great turn of events.

All too soon however, AI was everywhere – and not in a good way.
Almost every new tech startup launched was now an AI startup; and established companies that typically had nothing to do with technology were suddenly making ‘AI-driven strategy’ pivots; like they were trying to wedge AI assistants into any part of their business model they could.

On social media, there were many more ads pushing all the ways this shiny new technology made life so much better; with some even advertising a frankly dystopian product called ‘Friend’ (Brunell 2025). This is a pendant embedded with an AI ‘companion’ you could wear around your neck and served as a partner to converse with day-to-day. Needless to say, it was not very well received by a lot of people, judging from the vandalism and graffiti it received on its ads around the city of New York (Tangermann 2025).

Even in my personal life, the name ‘ChatGPT’ was everywhere, with people starting to refer to the LLM using anthropomorphising pronouns like ‘he/she’. Even more alarmingly, it was becoming more normalised to ask it for advice about things like where to visit on holiday, what to eat for dinner or even what to do about the issues you were having with a friend. With this in mind, I started to realise how embedded it was becoming in our day to day lives; and often not by choice.

Within the world of art, the ‘AI aesthetic’ emerged and along with it came the realisation for many artists that their works were being stolen and used to train these models to produce content without their consent.

In light of this, the wave of anti-AI sentiment became even more visible within the creative spaces I navigated. It was more common to label someone as ‘not a real artist’ if they had used AI in creating an artwork. Even the mention of AI in a creative context was regularly met with hostility and disdain.

The topic has become something very polarising today, with still no clear consensus about the legality of using models that have been trained with stolen or copyrighted artworks to create art. Without a doubt however, it’s safe to say the impact of AI on artists has been quite profound.

The loss of intellectual property is another outcome of the AI boom that must be discussed. Due to the aforementioned lack of legal clarity of using AI to produce art works, notions of authorship are often unclear and debatable.The paper titled ‘The GPT Judge’ by Grimm et al. explores this topic in further detail (Grimm et al. 2023).

Consequently, the phenomenon of cognitive offloading is another aspect of the AI boom that is increasingly becoming more and more evident.

The outcome is clear - AI is making us lazier; according to a paper published by Fayaz et al (Fayaz et al. 2023). For some artists, using AI sometimes means offloading the essential tasks of art making such as ideation and execution to these LLMs in the name of efficiency and optimisation.

One of the most important parts of being an artist is the ability to be able to develop ideas in a way that feels subversive and meaningful; by being able to question and break norms and potentially collaborate with others to do so. Art has largely been used as a vehicle for change in our society.

However it seems increasingly, some relatively renowned artists do not resonate with that, which to me was unexpected.
Upon listening to a talk by the artist called ‘niceaunties’ who almost predominantly uses AI tools to make their work, something seemed amiss to me. Hearing them talk about how much nicer it is to work with AI because it ‘doesn't get angry when you ask for changes’ or how ‘the line of production has shrunk and therefore become more efficient’ is something that feels out of place for a practice that seeks to challenge, explore and question things.

The artist here in question describes their own practice as intended to ‘rewrite the stories of women’ making them ‘unapologetically free’ (niceaunties 2026). However, it seemed ironic to me to use a generative AI tool like KlingAI or Seedance to ‘rewrite stories’ when a major problem with AI is that it is regressive in nature.

AI models are often trained to operate on probability - meaning they operate by comparing what ‘is’ to what has already been. Google itself mentions using the Bayesian model (Google 2026) as part of training its models, leading us to believe that common AI models are trained to always bring the output back to what exists and what has been done before. The outcome of this current model is that a lot of art ends up looking the same, because these models are drawing from the same probability distributions and end up simply producing the ‘average sum’ of a piece of artwork, based on all of the ones that exist in its dataset. Unfortunately, this also means that subtle (or unsubtle) biases rooted in misogyny, racism and homophobia are too often inherent, and need to be actively countered or overwritten.

In my opinion, this makes it quite static in nature. It easily becomes a restricting force for an artist wanting to try something new, because it is so heavily based on reading information from previously gathered datasets - from previously established norms. AI models - specifically ones trained by 'Big Tech', are not easily able to create new arguments for norms to be rewritten and deviated from. When artists use it, they often cannot break out from the norm and question things, which arguably is one of the core purposes for art creation.

A lot of modern artists increasingly rely on technology or even ‘non-human’ collaborators to produce ideas and make art which is not a problem in itself. As a computational artist myself, I believe combining art and technology is a wonderful way to allow artists to question things in ways they may have not been able to do before, and be more audacious with their work.

However, the problem arises when the artist becomes increasingly detached from the ideation, execution and implementation of an artwork. There, one begins to wonder whether the outcome is truly something we can consider to be ‘art’. If they are detached from questioning the outputs of the tools - which are often designed by non-artists for artists, they risk falling into and perpetuating the ideals and imaginations of these people, which unfortunately are increasingly resonating more and more with fascist fantasies of society. Art is a process through which artists can think critically and question things, even the tools they use themselves.

Jenna Sutela’s work ‘nimiia cétiï’(Sutela 2018) is a brilliant and imaginative example of this. Her work uses machine learning, neural networks and bacteria to open up a dialogue about communication that is rich and unique.
Holly Herndon’s spawning ceremonies (Herndon 2022) where people gather to willingly use their voices as training data show an interesting example of a collective and consensual collaboration between humans and an AI model.
Chipo Mapondera is another artist who uses an AI model in a subversive and thought-provoking way. She similarly operates in a collective; joining arms with members of different African indigenous cultures and a culturally rooted ‘AI climate oracle’ to help fill in the gaps of parts of their heritages that have been lost. Her works are developed closely with choreographers, archivists and ‘knowledge keepers’ whose main goal is ‘telling Indigenous stories’ (Iqoqo 2025).

In my point of view, we must first start by asking ‘what is art and is AI art actually art? Who is art for? Why does it need to be optimised?’.
Additionally, we must question what purpose the creation of art serves; is it solely for generating income or for questioning the norm and imagining the world in a different way?

Art is a way to express our thoughts and emotions, to communicate, and to connect with one another on the basis of our shared human experience. It doesn't seem right as an artist to create your own echo chamber of thoughts and narratives, completely disconnected from the reality of day to day life; devoid of the quality of human experience, just because it’s ‘more efficient’.

I strongly believe in the idea that there is no efficiency in creation.
In an ideal world, there should be very little place for trying to be efficient in the creative process because efficiency and producing should not be the goal. The toil of ideation and creation for an artist is a worthwhile process that should not be fast-tracked. If we ‘try to do everything efficiently, we are turning off the power of serendipity’ (Tenner 2019). As Dimacali mentions in their article on AI Art ‘just because we can, doesn't mean we should’ (Dimacali 2025).

AI is further perpetuating the capitalist mindset within art spaces that the purpose of creating art is for the market, or that production and scale are the end goal. The effect of this way of thinking on art and artists tries to rationalise it as a matter of fact.
Despite this, we need to keep reminding ourselves about the purpose that art is intended to serve, before allowing this mentality to further creep into spaces where they are actively causing damage.

Indeed, the impact of AI is a serious societal and needless to mention - environmental issue (MIT 2025). Despite this, we must all tackle it with a shared vision to protect the rights of artists and the general developmental and intellectual well-being of our society.
All things considered, the first major steps we need to take as a society are to create regulations to protect artists and put safety measures in place that actually have the best interest of the wider community at the center. Leaving these decisions at the hands of the CEOs of big tech companies who seem to have less and less interest in being humane and socially conscious is a guaranteed way for things to spin even more out of control.

Before we accept using AI in the artistic creation process, AI models as they are need a massive overhaul. No more should they be ‘Vampire Machines’ (Corral 2023) – these massive, extractive black holes of data, flattening identities and consuming any form of input to the extent of even cannibalising themselves (Vice 2025). Perhaps it might be better to create smaller, more intentional models whose data sets have been more ethically produced and gathered. Models created for a specific purpose, either by the artists themselves or in collaboration with others in their network. This helps to challenge the assumption about whether all things can truly be categorised into text labels in the first place.
The purpose of technology should not be trying to ‘solve’ art and make it repeatable or scalable; it should be to provide artists with a way of generating questions that they didn't have before.

Although AI currently seems to be advancing at a speed that feels overwhelming to imagine attempting to control and regulate, we all deserve to live, work and create in spaces that intrinsically support, rather than detract from our creativity and expression; and therefore our human experience itself.

Imagine that one day, you are told that an algorithmic risk system has placed you on a watchlist. Your past records, your place of residence, your social relations: this set of quantifiable features sits close enough to the features of some high-risk group. As a result, even though you may not have committed any unlawful act, you still receive a determination about yourself: a score, a risk level, a contact from the police. This system is Chicago's "Strategic Subject List", a system that ran for several years before being shut down (Saunders, Hunt, and Hollywood 2016).¹ So the question I want to ask is: why can a judgment directed at an individual appear before that individual's concrete action has taken place?

¹ The "Strategic Subject List" is used as the example here because it is a case that has already ended and can be reviewed in full. After the system was shut down, its logic did not end with it; this essay is concerned with the structure they share.

1. Judgment

To see clearly what this system does to judgment, we must first describe the old structure of judgment. According to Mireille Hildebrandt's account of the "informational" character of law: modern legal judgment is organised around actions that have already taken place (Hildebrandt 2016);

The concrete action is the object of judgment: a person can anticipate in advance that failing to perform a contract after signing it may bring liability, precisely because the connection between action and consequence allows the meaning of a behaviour to be understood before it unfolds,

Reasons are the medium through which an action can be interpreted; a transfer of money may be fraud, or it may be support for a dependant. The same external behaviour, under different reasons, points to different structures of responsibility.

Context is the background against which actions and reasons can be read. Together, these three hold up the entire structure.

At the same time, Hildebrandt adds a feature that is crucial for what follows: judgment can be modified. Rather than a result that closes the moment it is formed, judgment is a process continually reopened by new actions, facts, and interpretations (Hildebrandt 2015). The subject's later conduct flows back and may loosen the initial treatment. It is this capacity to be reopened that keeps judgment open to self-correction.

2. "Judgment"

So what changes in judgment once the system intervenes? Gallese discusses a case in which a traffic stop was initiated entirely because it took place in an area a predictive model had marked in advance as "high-risk" (Gallese 2026). The court pointed out that this mode of enforcement had shifted policing from judgment based on individualised suspicion towards a preventive logic rooted in prediction rather than concrete suspicion. So it is with the "Strategic Subject List". Its operating logic is the same as in the case above: inferring future risk from past data. A person placed on the list may never have committed any violent act; they may not even have done anything in the present scene that calls for scrutiny. They enter the system because of the data extracted from them. Here the "profiling" described by Antoinette Rouvroy and Thomas Berns comes into play: what the system touches is a statistical position recomposed from digital traces and everyday trajectories (Rouvroy and Berns 2013). Strictly speaking, these materials belong to the past, and often they do not belong to the subject's own actions: a high score may stem from records of their having appeared several times alongside someone already flagged. In attribution, it is thin. At this point the judgment is complete, yet the three conditions for completing a judgment are absent one by one:

The concrete action disappears: the basis for handling this person is a computed proximity to future risk, which has nothing to do with what they have done.

Reasons disappear: what the model delivers is correlation, the co-occurrence of data points, with no stage at which anyone, when asked, gives a reason.

Context is displaced: the background becomes past data compressed into features, and this data often belongs to a group, while the consequences fall on the individual.!

At the same time, that capacity "to be reopened" is absorbed in reverse. If a person marked as high-risk later commits no violence, this can be read as the risk never having existed, or as the risk having been preventively contained. Either way, the score is retained. The subject's later conduct, their appeals, even the fact that the predicted behaviour never occurred, are all absorbed as input for the model's next round of self-optimisation. This makes the system more precise and therefore harder to hold accountable; the channel of correction becomes a device that reinforces the initial treatment.

This is the structure I want to call the "empty judgment". Judgment still operates, complete and efficient: classification, contact, intervention, and consequences all land on the subject. Yet action, reasons, context, and the openness to modification no longer appear in the process of judgment. Technology opens a wormhole at the very place where judgment used to unfold. The stretch of process that should have been supported by action, that should have slowly taken shape in time and in waiting, is bypassed as a whole, hollowed out, flattened. What is hollowed out is the process itself, the road between action and verdict, the time of waiting; the subject stands intact at the exit of the wormhole and receives the result. Whenever an appeal is thrown at it, the wormhole draws it in and grows more solid. And what do these bypassed things stand for? What do these actions, reasons, and contexts stand for? — I take them to be precisely the conditions that depend on a stretch of time not yet arrived, on an event still waiting to unfold. They cannot be computed in advance, because they have not yet happened.

3. Separation

And these bypassed conditions are exactly where a person's standing to be treated as a subject resides; when judgment can take effect as usual while bypassing these conditions, what is suspended is the subject's position to speak within that judgment. In recent years, one line of research has sought to respond to this suspension through design: channels for appeal, response, and demanding explanation are seen as key to protecting the dignity and autonomy of those being judged, and this line studies how to build such channels into systems according to their needs (Yurrita and others 2025). But in my view, this optimisation brings further disregard of autonomy. Even if the system becomes accurate enough to absorb an individual's appeal and, on that basis, genuinely change the result it gives, delivering an adjusted "perfect judgment".

Why? In the old structure, correction passes through reinterpretation: the subject's statement is read as a reason, placed back into its context, connected to an action, and the verdict changes because the understanding of a person has changed. But in the responsive system, the appeal enters as data, and the verdict changes only because the computation has been run once more. The two results can look identical. A person is flagged; they appeal; the score is lowered; they are removed from the list. Their own conduct did change the outcome, and the weight of the consequences landing on them is no different from that of a verdict that has passed through the full process; from the result alone, they can no longer tell whether they have been seen or recomputed. The more perfect a predictive system, the more open it is to being influenced by appeal, the more it satisfies the requirement of contestability on the surface, while the appeal is absorbed as data throughout. A perfect judgment is a judgment in which even modifiability has been synthesised. But return to what came before: the action supporting this treatment was never examined, reasons were never given, and the context never belonged to them. Judgment in the sense of the old process did not happen; what happened was the entirety of its effects. Then what is this thing that lands on them, and where does it come from? It is produced by computation, and its origin contains no event that has already taken place.

As far as existing systems go, this artefact is not yet perfect. Appeals are absorbed as data, scores are retained, and the subject still has a chance to notice, in the seams of the results, that what they face is a computation run once more. The real question arises in the perfect system imagined above: there, the experience the system delivers passes, at the level of effect, every test the subject is able to perform. This is something new: an artificial experience of "being judged", detached from the event that should have generated it, and holding nonetheless. I have no intention of disparaging it; it does act on the subject, and it does respond to the subject's conduct; what can be described is that it did not arrive along the old process.

What can be established ends here: the hollowing-out has already taken place, empty judgments are already being produced, and the closer the system comes to perfection, the harder this output is to distinguish from judgment. When that perfect artefact arrives, whether what the subject receives is a judgment or an experience of "being judged", and how that experience should be defined, this essay hands over as an open question.

Every year, platforms such as Steam send me a year-in-review summary. The genres I spent the most time playing, the games I played the most, the distribution of my activity across the year — these visualisations, supposedly generated from “my” annual data, often make me feel as if data can truly read me. After all, these numbers are not invented; they are records of actual behaviour.

Yet the limited categories in Steam’s official recap never fully satisfied me. I wanted to produce a more complete reading of my own gaming history. Following a similar logic of data-driven summarisation, I developed a practice-based project titled Digital Gaming Identity, which explores a player’s long-term gaming preferences across their Steam library. At the time, I believed that I could create a fuller and more truthful self-summary. However, when I finally transformed the data into visual charts, I felt a strange sense of distance. The result neither confirmed my existing self-image nor replaced it with a more objective version. Instead, it placed different versions of myself side by side, creating a strong sense of misalignment. The person it described seemed to be me, but not entirely.

Why would a self-portrait built from real data make me feel both like myself and unlike myself? I cannot say with certainty whether the data had left out part of me, or whether my own memory of myself was already partial and unstable. However, this personal sense of misalignment drew me to investigate how such a gap came into being. It also points to a larger question: in an age where data has increasingly become an important medium of self-understanding, what kind of relationship should we have with the numbers that claim to represent us?

I therefore returned to my own project, retracing the process through which the data portrait was constructed step by step. This process, I hoped, might reveal how the gap between the data portrait and my lived experience had emerged.

For this project, I used Steam’s official Web API to retrieve player data (Valve Corporation n.d.). The tool was designed to allow users to input their Steam profile URL and generate a corresponding player portrait. The discussion in this essay draws primarily on my own Steam data as the main test sample, while also referring to observations from other test users.

The first decision I encountered was which data fields to retrieve. Driven by my curiosity about which games a player had invested the most time in, which genres dominated their library, and what kinds of play modes they seemed to prefer, I fetched the publicly available list of owned games and their playtime. I then connected the selected games to further metadata, including genres, categories, and recent activity. These seemingly straightforward choices had already begun to quietly define the identity they were supposed to reveal.

On the one hand, playing games is not merely a measurable action. It also involves attention, emotion, memory, skill, habit, and context. A game might be played intensively for one week and then abandoned for months; another might remain in the library for years as part of a personal collection. A player may spend hundreds of hours in a game because they genuinely enjoy it, but the same amount of time could also be accumulated through social obligations, repetitive grinding, collection habits, idle background running, or simply forgetting to close the application. Yet when these experiences enter a platform database, they are reduced to the fields that the system is able to record, store, and process. This translation is not neutral.

José van Dijck (2013) understands datafication as the process through which social behaviour is converted into quantifiable data, while also warning against mistaking data for a direct route to reality. Data is not a transparent window onto reality; it is the result of reality being recorded, filtered, and formatted by specific systems. Steam’s records appear objective, but they are still only fragments of actual behaviour.

On the other hand, in retrieving the data, I had already become a second layer of filtering. My own interests and assumptions shaped what I chose to include. For instance, I did not focus on achievement data. Achievements can often indicate how deeply a player engages with a game. However, in my own design judgement, they were not the dimension I wanted to examine most closely. This means that for a player whose engagement is primarily expressed through achievement systems, such as a dedicated trophy collector, the portrait generated by my project would miss an important dimension of their gaming identity. Their most representative behavioural trace had already been excluded at the level of data selection.

This trade-off further suggests that when I choose which fields to retrieve, I am not simply presenting the truth. Rather, I am using my own curiosity to draw a boundary around what counts as a “player” in the first place. This is not merely a subjective choice, but an epistemological operation. It determines what can be known about a player, and what is excluded from view from the very beginning.

After collecting the raw data, the second decision concerned how to filter and classify it. In the project, I attempted to summarise the distribution of a player’s time investment across different play modes. Through the Steam API, I retrieved each game’s category labels, such as Single-player, Multi-player, and Co-op. I then tried to connect these categories with playtime through a weighted association and visualised the results as comparative axes, showing the proportional distribution of time across different modes.

However, during testing, I found that the resulting percentages did not fully match my own experience. The root of the problem was that Steam’s categories are not mutually exclusive. A single game can be labelled with multiple categories at the same time.

This discrepancy was not caused by a technical bug. Steam categories are designed primarily for store organisation, browsing, and discovery. They describe a game’s affordances — the types of play that the game supports — rather than the factual behaviour of a specific player. In other words, categories can tell us what a game makes possible, but they cannot tell us how much time a player actually spent in each mode within that game.

This technical gap became another source of unfamiliarity in the player portrait. What the visualisation ultimately presented was not simply “what I did”, but rather “what kinds of games I own” and “what my games allow me to do”. I wanted to measure my actual footsteps, but the result looked more like a map of all possible routes. This also made me realise that I was not only dealing with “my own choices”. Steam’s system itself was also actively participating in the shaping of the portrait. Its data structure, its classification logic, and the non-exclusivity of its categories all contributed to the image being produced. These non-human technical elements can also be understood as actants within the process (Latour 2005).

The third decision was how to turn the processed data into a readable player portrait. The final output of the project consists of a set of charts and a textual player persona generated from those charts and data. Here, the persona can be understood as an interpretive identity constructed through data.

In designing this persona, I faced a central problem: platform data does not naturally produce a fixed identity. Different computational logics generate different readable versions of the same player. For example, if a player is analysed by the number of games they own, they might appear to be a Single-player / Indie / Action player. If the same player is analysed by playtime, they might instead appear to be a Multi-player / Online / Casual player. I therefore had to actively design an interpretive framework for the data.

For the persona generation, I created a specific set of translation rules. I selected games with more than two hours of playtime, ranked them by total playtime, and used the top five as representative titles. Their genres and categories were then weighted by playtime and mapped onto a set of player traits, including immersion, strategy, challenge, social interaction, competition, and variety. These mappings were manually designed. For instance, RPG and Adventure were more strongly associated with immersion, while Action and Fighting were more strongly associated with challenge.

The resulting persona is produced through a combination of data, designed rules, mapping systems, and language. It is not fabricated from nothing, but it does go beyond what the data directly states. It is a reasonable inference, but not the only possible interpretation. As Katherine Behar (2018) describes in ‘Personalities without People’, data systems can generate coherent and plausible personality descriptions that do not necessarily correspond to any real individual. This further suggests that data visualisation, in its final form, does not present facts alone. It presents a narrative constructed by the designer on the basis of facts. Even when the underlying data is accurate, the interpretation built upon it always contains an element of authorship. This authorship explains why data cannot automatically become meaningful information without design intervention. It is also where the creative potential of data practice lies.

Returning to the question at the beginning: why does a self-portrait built from real data make me feel both like myself and unlike myself?

The answer is not that the data is false, nor that my memory is simply wrong. Rather, a data portrait is never a replica of the real self. The apparent clarity of the visualisation conceals a long chain of decisions. Choosing which fields to retrieve, how to classify and weight them, and what kind of narrative to write around them — these seemingly technical decisions are in fact a series of interpretive choices. Every step translates “me” into another language, and every translation involves both preservation and loss. What these choices determine is not only whether the portrait appears more or less accurate, but the entire framework through which a player can be known and described.

Looking back at the project, I realised that this strangeness does not only come from being translated. It also comes from the double nature of data visualisation itself. On the one hand, it is a creative practice: it allows me to organise fragmented records into a readable story and produce an engaging, seemingly complete summary portrait. On the other hand, it also makes clear that this portrait is not “me”, but a version of me produced through a series of design decisions. The shaping of this identity involved not only myself as the designer, but also Steam’s system, the API’s data structure, and the visual grammar of the charts themselves. These non-human actants, together with my own interpretive decisions, jointly produced a narrative about a “player”. I am both the author of this narrative and the subject it attempts to describe.

It is precisely this doubled position that makes me more aware than anyone else of the gap between the portrait and myself. This awareness itself may be closer to lived experience than the portrait could ever be. It leaves me in a position from which I cannot fully withdraw. I cannot treat the data portrait as an honest mirror, but neither can I simply dismiss it. It does capture certain contours of me, but in a language that I cannot fully claim as my own. Faced with this portrait, I remain suspended between affirmation and refusal. Perhaps the relationship between humans and data is not ultimately about finding ever more precise measurements, but about learning, in each encounter between resemblance and estrangement, how to understand, question, and use these versions of ourselves.

To most passers-by, pigeons appear almost identical. They belong to the familiar background of the city: birds gathering on pavements, rooftops and window ledges, visible almost everywhere but rarely noticed as individuals. The_Unarchivable began by interrupting this habitual form of attention. What does it mean for a life to be individually recognised by a system built to classify and generalise? Is this a question of how notions of identity become embedded within recognition systems or how algorithmic abstraction reflects a human image of the subject? More simply, is it an attempt to test the limits of the algorithm itself? It centred on Huihui and Diandian, two homing pigeons raised by the artist. Although they lived under human care and had individual names and histories, their appearance closely resembled that of the anonymous pigeons encountered across the city. Their position therefore remained uncertain: they were intimate and familiar lives, yet they also belonged visually to a population often treated as an interchangeable urban background.

The project asked whether technologies normally associated with classification, surveillance and identification could be redirected towards overlooked non-human life. Instead of detecting only the general category “pigeon,” could a machine-vision system learn to recognise Huihui and Diandian as separate individuals? Could this process create a space in which non-human life became visible, recordable and open to broader discussion?

This gap reflects a hierarchy in the distribution of computational attention. The Unarchivable works precisely within it, redirecting a category detector towards the individuation of two specific lives.

Recognition is therefore produced through abstraction. Visual differences are transformed into measurable features and compared with the categories established during training. As Trevor Paglen argues, machine-readable images increasingly operate through classifications, training sets and mathematical abstractions rather than through forms of vision addressed primarily to human spectators (Paglen 2016). What appears through the interface as the recognition of a stable object is consequently the visible result of a series of computational operations.

This process reveals the double function of classification. Classification can reduce a complex life to a limited category that a technical system is able to process. Yet in The Unarchivable, it also created an opportunity. What the project pursued, in this sense, was the individuality of non-human life — a status rarely granted to animals, with the partial exception of pets. The model was not asked simply to detect “a pigeon.” Huihui and Diandian entered the dataset as separate classes, requiring their differences to become computationally meaningful. By separating them from the general category of pigeon, the project attempted to interrupt the anonymity imposed on urban non-human life and to grant them a form of individuality usually reserved for humans or the animals closest to human domestic life.

The trained model operated as a speculative form of digital identification. A technology commonly used to locate and categorise objects was redirected towards two lives usually excluded from individual records. In this sense, machine vision became not only a tool of detection but also a method of sustained attention.

Yet this attention remained conditional. Huihui and Diandian could be recognised only through the photographs selected for the dataset, the labels assigned to them and the visual patterns learned from those examples. The model did not discover their identities independently. It produced them computationally through the conditions constructed by the dataset. When their bodies no longer corresponded closely enough to these learned patterns—because of rapid movement, overlap, occlusion or changes in silhouette—the stability of recognition began to break down.

Learning to Fold developed from this contradiction rather than from a rejection of the earlier project. The two works form reciprocal moments within the same investigation. The Unarchivable asks what machine recognition can make visible; Learning to Fold examines the conditions under which that visibility becomes unstable. The technology that grants an individual position to the bird also reveals its limits when the body can nolonger be maintained as a coherent and readable object. During the first project, detection was repeatedly disrupted when the pigeons moved rapidly, overlapped with one another or entered visually complex environments. Bounding boxes flickered, shifted, multiplied or disappeared. Learning to Fold isolates one recurring source of this instability: the folding and unfolding of a pigeon’s wing. The same YOLO object-detection framework was used to examine how rapid transformation, self-occlusion and changing silhouettes affected computational recognition.

The project distinguishes embodied seeing from computational abstraction and classification. Embodied seeing involves an encounter situated within a perceptual and meaningful world. Computational abstraction, by contrast, translates visual variation into quantifiable features, while classification assigns those features to predefined categories. Machine recognition combines the latter operations, even though its interface often presents the outcome as though the machine had directly perceived a stable object (Paglen 2016). This contrast can be stated more directly. Embodied seeing unfolds as a temporal continuum — a duration in which perception and its object change together over time — whereas the computational paradigm identifies the living body with a static, fixed image, a discrete instant abstracted from the flow of time. The failure that follows is therefore not merely a matter of insufficient computing power. It is built into a paradigm that treats a life as though it were static, and it is for this reason that the paradigm can never fully capture a body that is, in its nature, continuous.

The project develops the term algorithmic disorientation to describe moments when this calculation cannot be maintained consistently across successive frames. “Disorientation” does not imply that the algorithm is conscious, confused or capable of subjective experience. Rather, it names an operational condition in which the system loses the spatial and temporal consistency necessary to produce the appearance of one coherent object moving through time. The calculations continue, but bounding boxes compete for position, confidence values fall, and the detected object may fragment into several uncertain presences.

Since it was above all the wing’s folding, a flexible, hinge-like skeletal movement that repeatedly triggered this disorientation, isolating that movement from the rest of the living bird offered a way to examine it directly. To investigate this condition, Learning to Fold moves away from the complete and alive pigeon and focuses on a custom-built skeletal wing structure. The structure abstracts the wing into hinges, extensions and folding relations, allowing selected movements to be repeated under more controlled conditions. It does not claim to reproduce pigeon anatomy. Instead, it functions as an experimental device for testing folding, vibration, acceleration, concealment and changes in outline. These episodes of algorithmic disorientation cannot be attributed to one cause. Within the experiments, folding produced overlap and concealed previously visible features. Speed introduced motion blur and greater visual differences between consecutive frames. Lighting, camera frame rate, object scale, processing latency and the limits of the training data also affected the model’s outputs. The project therefore does not claim that pigeon wings are intrinsically or permanently beyond computation. Improved cameras, larger datasets, faster hardware or different model architectures might reduce many of these errors. What matters is not the appearance of the wing, but its duration: the temporary interval in which the system used in the project cannot stabilise the moving body as a continuous computational identity.

Henri Bergson’s concept of duration helps to frame this interval. Bergson distinguishes lived duration from spatialised models of time that divide movement into separate and measurable units. In duration, successive moments do not exist as isolated points but flow into and transform one another (Bergson 2001, 100–06). Object detection, by contrast, samples movement through separate frames and reconstructs continuity from a sequence of individual predictions. Flickering and multiplying boxes expose moments when this reconstruction no longer appears seamless. The wing continues moving, but the model’s account of it becomes discontinuous. Bergson gives this partiality an optical figure. The mind, he suggests, refracts the continuous flow of duration into separate, spatialised units, pausing it, dividing it and giving each division a name (Bergson 1911, 242). Refraction is the bending of light as it passes between media of different densities, and what the bent ray delivers is not the object as it is, but a version of it, angled by the medium it has crossed, so that no single lens ever returns reality whole. The machine is one more such medium. It refracts the moving body into discrete, classifiable frames, and what it captures is therefore always a version of the pigeon, coherent within its own paradigm yet never continuous with the living duration from which it is cut.

Zach Blas extends this concern into computational culture through the concept of informatic opacity. Blas examines opacity in relation to systems that convert identity, difference and behaviour into readable and classifiable information (Blas 2018, 197–99). His account is particularly relevant to technologies whose operation depends on rendering bodies computationally legible.

Temporal opacity is not a term borrowed directly from either Glissant or Blas. Rather, it is developed through their work to describe unstable legibility across successive moments of object detection. The shift is deliberate: where Glissant claims opacity as an ethical right asserted against a colonial and capitalist drive towards total transparency, temporal opacity relocates opacity into the register of time, naming the mismatch between a body that endures continuously and a system that can only sample it in separate instants. Learning to Fold brings opacity into the temporal operation of machine vision. The wing is detected in one frame, lost in another and multiplied in the next. Its presence is not absent, but unevenly legible.

Temporal opacity therefore does not prove that life will always exceed technology, nor does it turn the pigeon into a conscious political resistor. It describes a provisional condition produced through the relationship between a moving body, a camera, a dataset and an object-detection model. While this instability lasts, the bird remains visibly present without being fully maintained as one stable computational object.

Together, the two projects position recognition and opacity not as simple opposites, but as two conditions within the same encounter between machine vision and non-human life. Recognition can interrupt indifference and bring an overlooked life into attention. Yet complete legibility may also reduce that life to the categories through which a system is able to recognise it.

More powerful hardware, faster cameras, larger datasets or different detection models may eventually reduce, or even remove, many of the failures observed in the project. Yet the time held inside

the computer is not the time in which the pigeon lives. The machine proceeds by cutting continuous movement into frames at a fixed rate, but between one frame and the next there is always an instant too fine to be divided, and in that uncountable moment the living body slips past capture, like a fine fold in time. The bird does not leave the technological system. Instead, its movement opens a fold within recognition itself: a moment in which it remains present, but not fully resolved.

沉默呵，沉默呵！不在沉默中爆發，就在沉默中滅亡。
Silence, silence! Unless we burst out in silence, we perish in it.

— Lu Xun, 'In Memory of Miss Liu Hezhen' (1926), translation by the author

I asked a text-to-image model to write my presence in Chinese characters, and it gave me something unreadable. That failure is what this text is about. This manifesto does not propose to fix the failure. It lets it speak.

The characters I asked for were six: 失語者宣言／我在此處—The Aphasic Manifesto / I Am Here. I gave the same prompt and the same seeds to three text-to-image models: SD 1.5 (Rombach and others 2022), 2022, West; FLUX.1 schnell (Black Forest Labs 2024), 2024, West; and GLM Image (Zhipu AI 2025), 2025, China, the last trained specifically for Chinese. I assumed the Western models would be mute and GLM fluent. What came back was stranger than that.

SD 1.5 produced no single character at all: a scroll of dense, illegible marks with no boundary between one character and the next. FLUX gave the shape of one character, but wrote all six as the same form, varying only in small ways. GLM, trained for Chinese, came closest. It separated the six and its strokes were more plausible, yet it still failed. 失 gained extra strokes, the radical of 語 loosened, the interior of 宣 broke down. Then 處 (place): four images, all resembling 慶 (celebration), without exception.

The six characters were not arbitrary. They are the manifesto's own declaration: 失語者宣言／我在此處. I asked the model to render the manifesto's own title as written characters, and it returned a presence that cannot be read. When the same model was given the six matched English words (loss, language, declare, speech, self, place), it rendered each one legibly. At the same unit of meaning, one word set against one character set, the alphabet survived and the hanzi did not.

It was only after the experiment that I understood why. A diffusion model does not write a character; it learns how pixels tend to distribute when a text description appears, then samples the cluster of pixels that best matches. Ted Chiang describes such models as a lossy compression of their training data, a blurry JPEG of the web (Chiang 2023). For Chinese characters, the compression is lossier still. They are rare in training data, and often an image contains a character while its label is only English, so the model is never told what the symbol means. It fills in strokes until the texture looks convincing. The rarer the character, the vaguer the impression: 處 is more broken than 我 because it is far rarer, and errors run toward too many strokes, the fingerprint of texture-matching rather than writing.

The translation, in fact, begins before the model is involved. Chinese writing already carries a text-to-image step inside it. When I type, I type in pinyin, a phonetic romanisation using Latin letters, and the input system must translate that sound into a visually unrelated character. The syllable chu, typed without tone marks as it is on any keyboard, maps to dozens of distinct shapes (處, 除, 儲, 礎). This sound-to-glyph conversion is a feature of pinyin input, not of the diffusion model. A pinyin input system navigates it by exact lookup: one keystroke sequence, one code point, one glyph, chosen by the user. A diffusion model has no such table; it reconstructs the character by sampling from a compressed distribution, and that is where rare shapes collapse into common ones. In pinyin I say I am here, and between the sound and the glyph the sentence is already being translated. That is the first point in the chain where meaning passes through translation, and where distinctions can begin to be lost. Chiang's lossy compression names the rest: when distinctions are compressed across layers of translation and then reconstructed, close shapes flatten into one. 處 (place) and 慶 (celebration) share an outline close enough that, under compression, the difference is lost. The model returns not the character I asked for but the nearest shape it can assemble.

One objection is obvious: why use an image model at all? A language model writes 處 without ever failing, because it does not write: it retrieves a fixed code point and a font renders it. There is nothing to see there. The failure I examine is not a failure of Chinese language itself, but a failure that occurs at the boundary between linguistic encoding and visual synthesis. I chose the system where writing must happen pixel by pixel, because aphasia can only appear where synthesis is required. Twenty-six letters cover most English; each of thousands of Chinese characters must be learned alone. Even a balanced dataset would not equalise this, because the architecture already treats the complexity of the alphabet as its default. Recent computer-vision research names this exact pattern long-tail text rendering: image quality and character accuracy both fall as a character grows rarer in the data (Zhuang and others 2025). My 處-into-慶 is one instance of a documented law.

My East-West binary collapsed. The variable is not which culture trained the model but frequency. The three models form a descending order (SD 1.5 heaviest, FLUX next, GLM lightest), without exception. As a native reader I see it at the stroke level: the outline survives, the stroke logic is gone. From a distance it looks like writing; up close it is misalignment, one stroke too many or too few. Readability dissolves with distance.

I set this out as a diptych. On one side, the six traditional characters, correct in form, alongside their generated counterparts with broken interiors. On the other, pseudo-English: a fine-tuned model was given the manifesto's English line and asked to generate it as collapsed marks, the same process applied in reverse. The script that is supposed to survive and the script that is supposed to break trade places.

This descent is neither random nor a mere technical flaw; it is a visible register of power. Crawford argues that machine classification is never neutral: it records who holds weight in the world of data (Crawford 2021). The claim that such classification is objective is what Haraway calls the god trick, a view from nowhere that hides the position it speaks from (Haraway 1988). To stage this at small scale, I trained a CNN on a deliberately skewed set, 95% Latin characters to 5% Chinese. The training curve records the result plainly: the Latin class reaches near-perfect accuracy in the first epoch and holds there; the Chinese class oscillates and recovers more slowly, never closing the gap. It is a conceptual reenactment, not forensic proof, but the shape of the gap is the argument.

Tsang Tsou-choi, the self-styled Kowloon Emperor, brushed his name and lineage across Hong Kong: utility boxes, road signs, walls (Tsang 1990s). The authorities cleared it as graffiti, refusing to read it as writing. My collapsed characters and his wall-names share a fate, unread by the system, one processed as graffiti, the other as noise. The algorithm does not fail to see 處; it meets an unfamiliar symbol and substitutes the nearest known one. In every classificatory system, marginal writing is not processed: it is overwritten.

What appears here is Glissant's right to opacity, but reversed. His right to opacity is claimed by a subject: the refusal to be made fully legible to colonial power. My characters are the inverse. Their opacity is not claimed but imposed; the statistics of the dataset, not the characters, made them unreadable. The model grants them visibility without readability: a shape appears, the stroke logic is gone. That is why their opacity is evidence of marginalisation, not resistance. The reversal happens in the act of naming. The characters do not resist; I resist on their behalf, by refusing to correct them and by calling their collapse a manifesto. The right to opacity is claimed by me, for them, after the fact.

One risk must be named. Xu Bing's Book from the Sky invented some four thousand non-existent characters, later absorbed by Western institutions as mysterious Eastern aesthetics (Xu 1987–91). This project shares the same fate, its collapsed characters read as intriguing Eastern signs, the aesthetic swallowing the argument. I state the political position plainly, rather than trusting the image to speak, precisely to resist that absorption.

These characters do not need to be fixed. Aphasia is not an error to correct; it is the trace a training-data power structure leaves on the image. I show the collapse not to improve the model, but because the collapse is testimony. Cultural marginality settles into the loss of readability across viewing distance. From afar it looks like characters; up close, it is ruins. The algorithm does not write. It samples. It forgets. We do not fix aphasia; we let it speak.

The year is 2025, three years since Stable Diffusion hit the market. A colleague and I were walking through the subway platform of Shanghai when a peculiar-looking poster entered our vicinity. The advert featured an anime-styled illustration of two people. The girl in front looked like she’d been meticulously rendered by an artist, whereas the boy in the background had smeared, half-fused fingers. The generic yet uncanny depiction immediately presented itself as "AI art". My colleague, who happened to be invested in this field, made snarky comments that the image was a product of an outdated "checkpoint model", and that he would have used "state-of-the-art" illustration models such as Illustrious-XL and NoobAI-XL. I jokingly asked him if he could reverse-engineer the prompt used to produce the image in front of us, and watched a long chain of AI runes eject from his mouth.

1girl, 1boy, looking at viewer, hands up, short hair, dark skin, school uniform, best quality, masterpiece.

AI and Danbooru

Both Illustrious-XL and NoobAI-XL have been trained on short, discrete tags as uttered by my colleague, and both point to the website "Danbooru" as the origin of this style of tagging. Launched in 2005, Danbooru is an image-hosting website intended to be a "high-quality archive for anime-styled images". The archive is co-curated by its users, as they submit images and are encouraged to contribute tags that describe the image. The large number of posts (11 million as of now), coupled with the refined pairing between images and tags, made Danbooru the ideal repository to train image models on. Its first interaction with AI happened in 2017 as writer Gwern scraped its data for classification research, producing the dataset danbooru2017 (Gwern 2015). In 2022, company NovelAI trained its proprietary text-to-image model using the site’s data, which was subsequently leaked in a security breach (NovelAI 2022). The leaked model, now known as animefull-latest.ckpt, showed the Internet that "AI can draw anime", thus spawning a gold rush at Danbooru where users bulk-scraped its data for model fine-tuning. The site’s traffic skyrocketed alongside its significance — its vocabulary of tags would soon become a standard prompting format adopted by many models, and has gained far more popularity than the natural-language captioning format associated with OpenAI's CLIP (Wagner and Cetinic 2025).

Looking at the stunning output of the aforementioned models, it becomes hard to dismiss Danbooru’s feat in the field of generative AI, whether that be its original intention or not. Moreover, Danbooru’s method of data curation stands in stark contrast to prior datasets such as ImageNet. While the latter outsourced the labour of labelling to data workers at Amazon Mechanical Turk, Danbooru relied completely on the volunteered co-curation of its user community, who collectively decided the tagging conventions. This is augmented by the site’s open-source codebase, which has allowed numerous "image-boorus" to emerge, each serving its own fandom (Britt 2019). One could argue that Danbooru represents an alternative approach to the making of an AI system — one that is federated, community-driven and non-profit, standing against the dominant players in the industry.

However, before one envisions a bright future of AI based upon Danbooru, the present criticisms of the site seem more glaring. The most heated debates are on the issue of "art theft", as the majority of Danbooru’s images are uploaded without the consent of their original artists. The taxonomy, on closer inspection, is also itself skewed, containing an expansive list of sexually explicit terms while remaining thin on domains such as nature or art history (Britt 2019). There is also asian (but not european), bad vulva (but not bad penis), a symptom of the far-right and gender-exclusionary pre-discourses that haunt the anime fandom (Petit 2022). Researchers have also traced how Danbooru-derived tagging systems normalised sexualised and misogynistic depictions in downstream generative pipelines, fostering an ecosystem where fantasy is misapplied to real-life individuals through deepfake models (Wagner and Cetinic 2025). What struck me in the subway, however, was that the site has furnished a system in which an artwork can be flattened into an array of tags, and the suggestion that this reduction can be run in reverse to produce endless combinations of art in the fashion of Fordist manufacture.

Media Mix, Worlds and Databases

Or perhaps that wasn’t Danbooru’s fault, but an established mode of consumption and production in the anime-manga industry. Culture critic Azuma Hiroki coined the term "database consumption" in his analysis of otaku culture in the 1990s. He proposed that each piece of character art, upon being received, is broken down into a collection of "Moe-elements" drawn from a "database". "Moe-elements" are parts of a character design that evoke an affectionate response from its audience, which may include "cat ears", "short hair", "maid uniform", or "glasses" — exactly what Danbooru's tags are about. On the other end, the industry has adopted a method of media production based on the database — "create character settings first, followed by the work or project, including stories" (Azuma 2001).

This method of media production is part of a broader phenomenon known as "Media Mix", a term often associated with the Japanese publishing giant Kadokawa. "Media Mix" originally denoted the practice of serial franchising of media across different platforms in post-war Japan (Steinberg 2023), though it has gained further complexity since Kadokawa Tsuguhiko took over the company in 1993. Inspired by the then-popular TRPGs (table-top role-playing games) such as Dungeons & Dragons, Tsuguhiko’s approach to storytelling involved designing a "world view" — a set of rules and character blueprints, which could then be used to enact stories through the distribution of creative agency (Otsuka 2004). This version of Media Mix is interested not in producing narratives but in building a mechanism that can generate an infinite number of narratives through the recombination of narrative elements. Characters are one of these elements, which Azuma sees as the "meta-narrative nodal point" that holds narratives together. "The know-how for enhancing the attractiveness (through the art of the Moe-element) of characters has rapidly accumulated… and the development of Moe-element databases has become a necessity" (Azuma 2001). Azuma’s insistence on the term "database" reveals the computational nature of the practice of Media Mix, and its similarity with "Generative AI" as we know it today. AI Dungeon, an early face of GPT-2, realised Tsuguhiko's TRPG-derived vision of storytelling almost exactly. Diffusion models, meanwhile, generate character art from tag-annotated datasets that are offsprings of Danbooru — the manifestation of the "Moe-element database" as it was conceptualised by Azuma.

The Empowerment Deceit

Another characteristic of the Media Mix is the mobilisation of consumers as producers. One should be familiar with the culture of "doujin" where fan-made derivative works circulate at conventions as well as online. The creativity amongst the fans is celebrated by Ian Condry in The Soul of Anime (2013), and resonates with Western studies of fandom such as the works on participatory culture by Henry Jenkins. Another who credited fan creativity is Otsuka Eiji, who worked for Kadokawa as one of the early practitioners of "Media Mix" (and later, a harsh critic of it). In his text World and Variation (1989), Otsuka saw the model of media production they had developed as a system that could empower the consumer of texts. Since Media Mix delivers a "world view", consumers who internalised the "world" are able to produce small narratives of their own using its building blocks (such as characters). This would end in the dissolution of the boundary between the "official" and the "derivative" (Otsuka 1989). A similar kind of empowerment discourse exists amongst the field of generative AI, which suggests the technology could empower the ordinary user by eliminating the skill required to create, as prompting replaces the artistic training required to produce art. Though, as Marc Steinberg points out in his analysis of Media Mix and its platform economy (Steinberg 2019), such production is being actively harvested as a resource. Kadokawa in particular published collections of fan-made derivative works of franchises it owns, and launched the web novel platform Kakuyomu in 2016, where the company’s editors scout the site’s metrics for amateur web novels valuable for publishing and adapting. Users not only contribute content to the site but also participate in the curation of content through their views and ratings. The generative AI industry replicates this arrangement at scale. An example would be aforementioned NovelAI which trained its proprietary model on Danbooru's fan-tagged database, whose images are in turn sourced from fan illustration sites such as Pixiv. Such is the fate of the participatory culture media mix cultivated: consumers are empowered just to the extent that they can be extracted. A seemingly democratic platform in fact institutionalises a system where users absorb the precarity of uncompensated labour.

The deception of empowerment does not stop at an economic level. In his later work, Otsuka traced Media Mix and its participatory model back to wartime propaganda employed by the Imperial Rule Assistance Association of Japan. The very techniques of distributing character and a "world view" for the consumers to elaborate on were pioneered in Yokusan Ikka (1940), a state-sanctioned manga project whose cast of characters were designed to be easy to draw and was deliberately released for amateurs and artists alike to produce their own stories with. What’s embedded within the characters and their "world view", however, is a fascist ideology that justified Japan's colonisation of the East Asia. Here, the participatory culture was turned towards mass mobilisation (Otsuka 2018). Through their fannish creation of small narratives, citizens voluntarily internalised the state's ideology all whilst perpetuating it as their comics circulated via mass media. The consumer’s creative agency within a system like Media Mix is bounded by its construct, as the "world view" (or "database") handed to the amateur artist predetermines what can be said within it, much like the dataset of a generative AI model delimits what is possible in its output. Juggling the tags of the text-to-image community yields illustrations that meet its codified standard of anatomical correctness and beauty — 4k, masterpiece, amazing quality, award-winning. The user is free to recombine these graphical clichés and circulate the way of seeing they encode, but never to redefine them.

Depature

I hesitate to close the case. Media Mix, the same participatory framework that once mobilised citizens for wartime propaganda, also cultivated the self-publishing scene of doujin, where fans sustained decades of genuine innovation and no shortage of subversion. Danbooru, despite everything it has fed into, remains as a vast archive built and governed by its own community. Perhaps the question is not what made this collaborative culture complicit, but whether any collaborative framework can escape the capture. Could a database belong to its builders, before it belongs to who runs the mechanism on top of it?

The train arrived at the platform. My colleague was still refining his prompt as we boarded, chasing a better render of the girl. The advert would be seen by millions of commuters before its contract ended. No one drew the picture, yet everyone did. The doors closed and the two protagonists slid out of view, looking at viewer, until the end.

Are seeing and knowing the same thing? The assumption that they are runs deep in Western thought. Since Descartes proposed that the mind mirrors an independent reality, and Locke imagined ideas as copies of external things, philosophy has treated knowledge as a question of correspondence — how accurately the inner representation matches the outer world. Richard Rorty's critique of this tradition argues that the mirror metaphor is not a discovery about the nature of mind but a historically contingent habit mistaken for necessity: to verify that the mirror is accurate, one would need to step outside it and see the world directly, which is precisely what cannot be done (Rorty 1979). This article begins from that premise. It argues that neither human recognition nor machine vision functions as a mirror of reality, and that the assumption that either could — what this article calls the mirror paradigm — naturalises the specific, partial, and constructed character of both. When two different kinds of seeing inhabit the same space and claim authority over the same object, the transparency of reality becomes difficult to sustain. The question is not which vision is more accurate. It is what the mirror paradigm makes appear natural about each. The first thing it naturalises is that both human and machine vision are forms of embodiment — different in kind, but equivalent in their partiality. Consider what the human visual system actually is: an eye sensitive to a specific range of wavelengths, connected to a brain that processes incoming signals through structures shaped by evolutionary history, cultural formation, and accumulated personal experience. Human vision does not see everything. It sees what it has been designed — by evolution, by culture, by lived experience — to see. Merleau-Ponty's account of embodied perception makes this precise: recognition does not happen only in the head but through the whole body (Merleau-Ponty 1962). Before any conscious calculation begins, the body already knows its place in the world — its orientation, its distances, its relations. This is what Merleau-Ponty calls the pre-reflective corporeal schema. To recognise a face — particularly one's own — is not to match a visual input against a stored template. It is to encounter the visible surface of a life that has been lived through a particular body. Inside that act of recognition are all the previous encounters, the feelings those encounters left behind, and the gradual revisions that experience has made to the image one holds of oneself. Machine vision is equally embodied, though differently so. Its body is composed of training data, hardware architecture, algorithms, and thresholds — each of which functions in specific, designed ways that are particular to it, just as the eye and brain function in specific, evolved ways that are particular to them. The training data carries cultural weight and historical decisions about whose faces count as normative. The hardware imposes physical constraints on what can be processed and at what resolution. The algorithms encode a point of view, prioritising certain patterns over others. The thresholds enact value judgements about what counts as a valid result. What the machine sees is what it has been taught to see — and in this respect, the parallel with human vision is structural: we also see what we have been designed to see, by the accumulated selections of evolution and experience. The machine is not looking from nowhere. It is looking from a very specific somewhere: a training distribution, a set of design decisions, a history of inclusions and exclusions that shaped what it learned to recognise (Crawford and Paglen 2019). The abstraction, then, is not the machine. It is the mirror paradigm — the mentality built into the presentation of machine vision as objective detection, as though the output were the world disclosing itself rather than the product of a particular, constructed way of seeing. This is what Haraway identifies as the god trick: the fantasy of a gaze that sees everything from no particular position (Haraway 1991). But a gaze requires eyes, eyes require a body, and a body is always somewhere. The god trick does not reside in the machine itself; it resides in the claim that the machine's output is neutral, in the institutional and cultural decision to present machine classification as a transparent window onto reality. Human vision operates within the same paradigm in its own way: it presents itself as simply natural, as just what seeing is, allowing the evolutionary, cultural, and affective conditions that shape it to recede into the background as though they were no conditions at all. Both visions are in the same relation to the mirror paradigm — the paradigm that naturalises each, making each appear as simply what seeing is rather than as a particular, constructed way of engaging with the world.

This is why the questions that matter are not questions about accuracy. They are the questions the mirror paradigm renders unremarkable: who built the training data? Which faces were included, and which were not? Whose features were encoded as normative, and whose were made marginal or invisible? Crawford and Paglen's analysis of large-scale image datasets demonstrates that these are political decisions, not technical ones — choices that encode assumptions about race, gender, and social value that are reproduced at scale in every act of classification the system subsequently performs (Crawford and Paglen 2019). When the mirror paradigm is set aside, what becomes visible is not a void but the machinery behind the result: the datasets, the design choices, the thresholds, and the consequences of error. The database is not a flawed copy of the world waiting to be improved. It is itself a way of seeing the world — one that determines in advance what is worth knowing, what counts as a face, and whose features are legible as normal (Sluis 2019). If both human and machine vision are partial and embodied, what distinguishes them? The answer lies not in the presence or absence of a body but in how each relates to time and to the act of recognition itself. Human recognition is what this article terms re-cognition: the act of knowing again, shaped by the living accumulation of everything that has come before. Bergson's account of memory as duration rather than storage captures this precisely (Bergson 1911). Memory is not a filing system from which fixed records are retrieved; it is a living process that continues to move in the present, colouring each new act of perception. This means that the same face encountered twice is not encountered in quite the same way. Between the first meeting and the next, time passes, experience accumulates, and what one knows of that person shifts. Each act of recognition carries that accumulated weight into the present moment, and in doing so, modifies what will be recognised next time. This is also why there is no such thing as a first sight in the strict sense. Even when a face is encountered for the first time, it is encountered through an already-formed perceptual system. Clark's predictive processing framework describes a brain that continuously generates expectations and reads incoming sensation as confirmation or correction of those expectations (Clark 2013). A new face is not processed from scratch. It is processed against the accumulated pattern of every face seen before, every cultural norm about what faces look like, every affective association the category of faces already carries. The entirely fresh perception is a philosophical fiction. Re-cognition precedes recognition. Machine classification relates to time differently. In a stable system, the same input produces the same output. The machine does not accumulate experience of a particular face across encounters; it does not revise its understanding of that face in light of what has happened since. Where human re-cognition is perpetually in motion — modifying itself through every new encounter — machine classification is fixed at the moment of training until the system is retrained. Even post-training feedback loops, the cybernetic corrections introduced through user interaction, constitute a separate learning process rather than a modification of the recognition act itself. The machine recognises, in a sense, only once: it applies a trained model to a new input without the act of application changing the model in the moment. This asymmetry matters because it defines what each system can and cannot carry. Human re-cognition carries history — the specific, irreplaceable deposit of every encounter, every emotion, every revision. It carries flesh, in Merleau-Ponty's sense: the experience of being a body that is also seen, that meets the world as both subject and object (Merleau-Ponty 1962). Machine classification carries data — the curated, annotated, politically inflected version of the world that was encoded in its training. Both are partial. Both are shaped by what came before. But they are shaped by different befores, and they revise themselves — or fail to revise themselves — in different ways.

The mirror paradigm naturalises this difference away. By presenting both as kinds of seeing that aspire toward the same end — an accurate reflection of an independent reality — it allows human vision to appear simply natural and machine vision to appear simply objective. The specific conditions on which each depends recede into the background. Setting the mirror aside does not leave a void. It leaves two genuinely different ways of knowing — each with its own body, its own history, its own partiality — and the political question of what happens when one of them is presented as though its specific conditions were no conditions at all: as though it were simply detecting the world, rather than seeing it from a very particular somewhere.

Technological capitalism claims that eliminating waiting, satisfying needs in advance, and channelling desire smoothly toward immediate return is a great enhancement of human agency. This claim presupposes that friction is, by default, an obstacle to the subject, so that to eliminate friction is to liberate the subject. This paper argues the opposite: friction is in fact a necessary condition for the constitution of the subject, and the technological apparatus that systematically eliminates it structurally destroys the very ground on which the subject takes shape. If the genesis of subjectivity depends on specific formal conditions, then a technology aimed at eliminating those conditions necessarily triggers the structural collapse of subjectivity.

I. The Subject Is Sustained by Friction

By friction I mean the ineliminable structural resistance between the subject and the objects of its desire. Lacan distinguishes need from desire: need has a concrete object and subsides once satisfied, whereas desire points to no concrete object and circulates perpetually around a structural lack that cannot be filled (Lacan 2006). As Žižek observes, this lack within the symbolic structure is the subject itself (Žižek 2009). Once the lack is filled, desire vanishes, and the subject with it. What sustains the subject is precisely the fact that the lack cannot be filled. This is the first source of friction.

The second source is temporal. In Cinema 1, Deleuze locates subjectivity in the interval between stimulus and response: inanimate matter moves the instant it is acted upon, whereas a living centre hesitates, and it is precisely this non-zero delay that opens the space of interiority and choice (Deleuze 1986). Once the interval falls to zero, response degenerates into reflex. What sustains the subject here is that the response cannot be given immediately.

The lack and the delay are both a gap: the distance between what the subject desires and what it already possesses. Friction is not this gap itself but the pressure the gap exerts; it is the result of the gap, and the gap in itself is not a force. If I am hungry with nothing to eat, the missing meal is the gap; the hunger that drives me to look for food is the pressure it exerts, and that driving force is the resistance. The absence of food does not, in itself, push me; it is the felt state of hunger that sets me in motion. No intentionality is implied here: nothing resists the subject on purpose. The resistance is merely the tension of a state that can never be closed — and it is precisely by pressing against this tension that the subject comes to feel that it exists.

Although these two sources belong respectively to a theory of desire and a philosophy of time, they point to the same structural feature: the subject depends on a resistance that cannot be reached directly. Lacan’s lack is spatial unreachability; Deleuze’s interval is temporal unreachability; the two are dimensions of one and the same resistance. This paper uses the word friction to name this feature, and concludes: the subject is sustained by friction, and to eliminate friction is to strip away the necessary condition for sustaining the subject.

II. Prestress and Load

Friction is heterogeneous, comprising two structurally distinct components. The first is a permanent, background resistance, which I name alterity: the way the world and others remain beyond the subject’s will, holding to an irreducible opacity — Lacan’s big Other (Lacan 2006). In the structural analogy, it corresponds to prestress: an internal tension that exists prior to any external load and holds the structure in form.

The second is a contingent resistance, arising in one concrete situation after another, which I name the frustration of desire: each instance of wanting and not obtaining, in which desire is continually reignited and structured. It corresponds to load: the external forces the structure bears, one after another.

This translation holds because both fields confront the same problem: how a thing maintains its own unity in the absence of a self-sufficient point of support. Lacan removed the Cartesian cogito as that point;thereafter, unity is maintained by a relational structure organized by lack, in which no privileged entity grounds the whole (Lacan 2006; Descartes 2008). Tensegrity confronts the very same problem when it refuses the foundation, and gives the same answer: distributed relational tension. The correspondence rests on this formal isomorphism.

III. Tensegrity

On the basis of this legitimate translation, the paper argues, by construction, that the structure able to sustain the subject must be, and can only be, tensegrity. A tensegrity structure consists of two kinds of member: a continuous network of cables in tension, and compression struts that do not touch one another and float within the tension network. Its core mechanical feature is prestress: even with no external load, the cable network is already taut and the struts already compressed, and the structure’s form and stiffness are shaped entirely by this internal opposition that precedes any external force (Fuller 1982; Calladine 1978).

These members find exact counterparts. The struts are the subject’s desires: each presses outward, and no desire touches another directly. The cables are friction, the tension that binds these outward pressures into one whole form. The empty spans between the struts — the non-contact the cables hold taut — are the unreachable itself, and it has two dimensions at once: a spatial gap (Lacan’s lack) and a temporal gap (Deleuze’s interval). A single tensioned band already carries both properties: it has a length in space and a duration over which it stays taut. Space and time are therefore not two different members but two projections of the one resistance the cables embody.















From this, three constraints follow. One: load-bearing relies on overall tension, with no foundation — Lacan’s subject can bear force only through a tension network spread across the whole. Two: the struts must float and never touch — physical non-contact is the structural equivalent of Deleuze’s temporal interval. Three: once tension is withdrawn, the structure disintegrates — losing the tension that maintains lack, it degenerates into loose parts. Tensegrity alone satisfies all three at once. Calladine’s theorem consolidates this: an underconstrained tensegrity acquires stiffness in a state of self-stress, and once the self-stress is withdrawn the structure collapses (Calladine 1978). The mode of failure matters: a wall stripped of its foundation crashes, whereas a tensegrity stripped of its tension does not crash but slackens — the cables go loose, the struts drift, and the form quietly sags out of existence.

IV. The Subject as a Living Space Pried Open by Tension

What truly holds this structure up is the relational field sustained by tension; the members are what that field holds in place. Mapped onto the subject, the subject itself is this space pried open by force. The void here is a living void — tautened and sustained by the tension around it, holding its form only through that tension’s continuous action. It is wholly distinct from a dead, silent emptiness: the living void has form and stiffness, deforming under force and rebounding when force is released. Once the tension disappears, this living space closes.

V. The Smoothness of the Algorithm and the Weightlessness of the Subject

The subject is a tensegrity supported jointly by the two components of friction; the smoothness of the algorithm is, at the structural level, the erasure of both at once. Norman defines the ideal interface as one that disappears completely (Norman 1998) — a philosophy that erases the interval on which subjectivity depends, and one enforced rigorously at the lowest layer of the data pipeline.

On one hand, algorithms erase prestress. Feature extraction and data scraping reduce the body and contingent events in dimensionality into structured feature vectors (Bradski 2000; Lugaresi and others 2019), and the alterity that cannot be captured parametrically is discarded as noise. The environment is remade into a compliant mirror, and unpredictable otherness dies away with it. Read together with Han’s account of the disappearance of the other (Han 2018), the permanent background tension that holds the subject up is withdrawn.

On the other hand, algorithms erase load. The loss function minimizes the error between predicted and actual values (Goodfellow, Bengio, and Courville 2016), while prediction pre-empts conflict before it occurs. As Zuboff argues, surveillance capitalism extracts profit by eliminating the uncertainty of behaviour in advance (Zuboff 2023). The external forces the subject would have borne are intercepted before frustration can occur: desire is filled in before it can be denied.

When both components are withdrawn, the tension field slackens and the subject falls into weightlessness: the internal opposing force that maintains its form disappears. And because pain was the felt result of friction, eliminating friction eliminates pain along with it, and the dismantling is undergone, from inside, as comfort. Opposition is precisely one of the elements by which a structure takes shape.

VI. Conclusion

The persistence of the subject relies on alterity as prestress and on the frustration of desire as load. Within this framework, the smoothness pursued by contemporary algorithmic systems is a systematic dismantling of the structure of the subject: recommendation flattens the world and withdraws the prestress, while prediction fills desire in before it can be frustrated and cancels the load. This dual absence dissolves the subject’s tension field and closes its living space, pried open by tension, in a frictionless environment. In seeking to abolish all resistance and delay, technological capitalism also destroys the very ground on which the subject takes shape. And the end does not announce itself: there is no crash, only a slackening that is felt, in so far as it is felt at all, as ease.

“The fantasy worlds stored inside a computer… bear the ontology of numbers: although not tangible objects, they are nevertheless manipulable entities and can be experienced as real things with resolute efficacy” – Timothy Binkley, The Computer is Not a Medium (1988)

We present a work-in-progress artistic practice. We explore a mystery and through this inquiry, enter and encounter a computational space. Mathematics and computation both generate this space, and also serve as the tools for exploration. At each step, the practice is subjective – taking into account our own perspective and subjective encounters or changes.

Crucially, there is no mathematical or technical prerequisite for this practice. Rather, one can work their way down, as we did, starting from whatever level, picking and choosing which layers of abstraction to personally struggle with, and which to leave.

Curiosity & Mystery

“Much of what I write is coming from the position that mathematics and the natural sciences are intrinsically interesting” – Greg Egan (2009)

We begin with curiosity about a mystery of how something works (or could work), usually related to maths, computing, biology. For example, we might ask “why do ferns curl?”, “how does a perceptron work?”, “how do limbs form?”

We try to set off from a place where we understand almost everything, but then step by step enter and inspect subsequent mysteries.

The mysteriousness is subjective. What mystifies us is often trivial to others. For example, both authors spent much time understanding the equations behind Gray-Scott reaction–diffusion, something well established in mathematics. This subjectivity separates our approach from institutional science practices that require new contributions to knowledge.

Artistic Computational Inquiry and Encounters

We then try to computationally or mathematically stoke the curiosity or mystery. For example, writing simulation code, or struggling to derive or understand a formula.

As part of the inquiry, we end up encountering something new. This could mean a changed understanding of how things work, a new view into ourselves, a new metaphor. It does not have to be technically advanced and can be humbler in practice, for example understanding binary search, or a for loop.

This process repeats. The loop is: curiosity ⟶ struggle to learn or inquiry ⟶ encounter (understanding) ⟶ repeat.

Here, an encounter is new in that it is subjective. A polar bear encounters a rain forest for the first time, while the rainforest is quotidian for a jaguar. This subjectivity comes from being individuals with interests and multi-headed, multiple simultaneous lines of inquiry and influence. We read theory, fiction, and do visual research. As if a novel unfolding through braided plots, we do not try to connect all of the threads, but rather allow them to diffuse or entangle.

Sometimes this inquiry leads to a solution, and other times the mystery stays. For example, the reaction–diffusion equations can be gradually understood, with some effort. But something like “how do limbs grow” stays mysterious.

The main tools of the inquiry are computation and maths.

By struggling with and understanding the maths of these equations, we can start to see how they work. The core idea of reaction–diffusion is that there are two processes (equations) happening at once. One can imagine it like rabbits and carrots. In one equation, rabbits are multiplying, eating carrots, living, dying. In another equation, carrots grow and are eaten. Each equation has parameters that determine the rates of growth, eating, multiplication, and death.

Computation makes the encounter interactive, in a particular way that pen and paper and maths do not. By simulating these equations with code, we can accelerate time to see how they play out over millions of computations, making the simulation feel like an entity that is responding back to us immediately.

Computation and Maths as Space

The mystery becomes a computational space where an artist can go, if they want.

For example, the set of all possible combinations of reaction–diffusion parameters becomes a computational space that we can explore. We can graph a range of values of carrot-growth-speed and plot it versus a range of values of rabbit-birth-speed, and see every combination.

We try to enter a computational space and see what tentative ideas unfold from our encounter, rather than having ideas and then using computation as a tool to illustrate them.

With complex mysteries, the computational space almost never is how one would expect, so the surprise becomes a further mystery for investigation and learning.

Entering a computational space, we must carry in our own sensitivities, interests, ways of thinking. We try to have, “One foot planted in the craft work of design and the other foot planted in the reflexive work of critique” (Agre 1997). The encounter with this space is, like the mystery, subjective and enters into relationship with whatever else we are reading, looking at, talking about. This creates emergent similes from osmoses.

Bringing something back

We do not primarily aim to create specific outputs, science papers, or art artefacts, although these things occur. We try to bring back souvenirs, traces, tools, stories, anything from the encounter to share with colleagues.

We try to log and discuss the inquiry, and continue to further mysteries, following a “gradient of interest” (Machine Learning Street Talk 2025).

For example, for a recent computational arts show, we both showed experimental logs. Savannah’s log of reaction–diffusion included engravings of various results onto metal. Mikhail showed a series of screenshots of notebooks and visualization tools, combined with notes on what changed.

Our logs are personal, including screenshots of experiments, statistical findings or tables, notes from conversations, code. It is important to include thoughts, connections, personal responses and questions. We like logs because they capture more of the inquiry. A common challenge in ALife is that researchers simulating creatures see directly how “their evolving algorithms and organisms creatively subvert their expectations or intentions,” but then this creative subversion does not make it to their final papers because it does not “fit into the standard scientific narrative” (Lehman and others 2020).

“Fitting In” as a Design Problem

Institutional science and art contexts both challenge this practice, since they are oriented around contributions or artistic outputs. However, it is often possible to treat institutional requirements as a design problem (where one must come up with a formal solution), to figure out how they can serve the inquiry. The alternative is to treat them as a practice problem which means reorienting all of the above to produce Contributions, or specific artistic outputs. For example, in the past we did specific art outputs like installations or games as part of an arts program. We later found handing in unfinished Python notebooks and things to discuss in small groups was more demonstrative of our practice than making work to specifically show in galleries.

In Practice

How do we know when our work is good? We are mixing many lineages and none of their criteria apply well. For example, science wants contributions, publishing; critical tech art wants a package of theory, technical decisions, related artists, concepts. For our practice, perhaps the best criterion is whether experiments are happening and if new encounters and changes emerge for the artist. Larger more psychedelic changes are as valid as smaller changes. For example, going from not-understanding to understanding how a sine wave works would also count as change in this practice. We have no criteria for change, other than that we see the mystery differently than when we started. Making no progress towards understanding is in itself interesting, and leads to questions around why.

This notion of change is somewhat related to ideas around artistic research. Artistic research suggests that artists produce different types of knowledge (related to our idea that the encounter is personal and need not be new to science) and practice-based research suggests that change (although not individual) might be a goal of practice. Pickering, Rheinberger, and others in Science and Technology studies overlap with us often regarding experimental methodology. There are also echoes of Bogost’s idea of Carpentry, where philosophical objects are built to test ideas (Candy 2006; Borgdorff, Peters, and Pinch 2020; Pickering 2010; Rheinberger 1997; Bogost 2012).

“The computer artist works with a hyperreal creation which stands behind the scenes as a kind of meta-physical object” – Tim Binkley (1988)

Computation history as support for this methodology

Our methodology is connected to computation's roots . Computation is not fundamentally about image-making – its history lies in the union of different disciplines. George Boole’s algebra was a combination of maths and logic. Logic, originally a branch of philosophy, has become an integral part of mathematics. Boolean algebra was applied to electronic circuits by Claude Shannon when mathematics and electricity were still distinct and separate practices. Through the fusion of these different disciplines, new ways of thinking emerged (Boole 1854; Shannon 1938).

The practice is also deeply inspired by people who consider themselves scientists who we consider to be artists: mathematicians like L Mahadevan (who looks at the mathematics of morphogenesis and other strange topics) and computer scientists like Alexander Mordvintsev (Deep Dream, Neural Cellular Automata) (Mahadevan Natural Philosophy n.d.; Mordvintsev, Olah, and Tyka 2015; Mordvintsev and others 2020).

Conclusion

This practice does not aspire to achieve a specific outcome; it drives curiosity and creative practice forward through mystery and investigation – valuing inquiry and the potential that unfolds from it. Developing this as a methodology helps us invite others to adapt these ideas to their own practices, and also helps us understand our work relative to other computational arts lineages.

We are interested to know, from readers, what do you think of this? What is uninteresting or does not apply to you? What would it mean to apply ideas from this to your practice? What obstacles do you see? What mystery would you be curious about? Where has computation or simulation changed you?

E-mail for correspondence: savannah.r.perry@gmail.com