Introduction
The rapid development of contemporary generative image models has transformed image-making in ways that remain difficult to fully account for. Systems such as DALL·E, Midjourney and Stable Diffusion can now generate visually convincing images within seconds, prompting widespread speculation about whether AI might replace artists. Yet this question already assumes that painting is primarily concerned with producing images. My concern is different. I ask what happens when an AI-generated image enters the painter’s studio, and what it means to paint from such an image rather than simply display it.
This distinction also requires a clearer account of what is specific to AI-generated imagery. Painters have long worked from photographs, printed reproductions, digital images, memory and collage. In every case, the work of painting transforms its source through interpretation, revision and material experimentation. The significance of AI therefore does not lie simply in providing another image to paint from. Unlike photography, which retains an indexical relationship to a moment in the world, AI-generated images are synthetic constructions assembled through probabilistic relationships learned across vast datasets. They frequently combine convincing visual coherence with spaces, objects and atmospheres that have never existed. For the painter, this produces a different kind of dialogue. Rather than translating an image of something that has been seen, painting negotiates with a proposition that has no prior existence outside computation. The distinctive question is therefore not whether painting from AI-generated images differs materially from painting from photographs, but how the synthetic character of AI-generated imagery alters the painter’s imaginative encounter with the image before material work even begins. Recent scholarship has begun to theorise this synthetic status directly: Manovich and Arielli (2024) argue that AI-generated images constitute a distinct ontological category, irreducible to either photographic indexicality or earlier forms of digital image-making.
Painting is not reducible to image production. It is an embodied, material practice, one that inscribes time, physical effort, and decisions made in physical conditions and lived experience. Its significance lies not only in the resulting image but also in the lived encounter between perception, memory, intention, matter, and duration. This is the sense in which painting differs from computational image generation, not because it lacks sophistication, but because it operates on a fundamentally different register.
That said, Generative AI has become an occasional presence in my studio, and it is not a simple or comfortable one. When I work from an AI-generated image, I am not just using a new kind of reference material. I am entering into a negotiation with a computational system that proposes images without lived human experience or existential investment in their outcome. Figuring out what to do with that, what it demands of me as a painter, is what this essay tries to work through.
Throughout this essay, ‘AI’ refers specifically to contemporary text-to-image and image-to-image generative systems used as painterly source material, a deliberately narrow definition given how loosely the term ‘AI art’ is often used to cover an increasingly heterogeneous field (Zylinska, 2020). I do not intend the discussion to encompass every form of AI-assisted artistic practice. Artists working with robotic painting, live AI performance, computational installation or autonomous software raise related but different questions; recent scholarship on Sougwen Chung’s human-robot collaborative paintings, for example, extends Hayles’s notion of the cognitive assemblage into precisely this different context (Wang, 2025). My concern is a particular studio practice in which computationally generated images become provisional propositions that are subsequently translated into painting through sustained material engagement. In this context, AI does not replace painting; it becomes one interlocutor within the painter’s process. The AI gives me something to push against. The painting is what happens in that pushing. In this text, I therefore refer to painting not simply as an artistic medium but as a way of thinking. The studio is approached as a cognitive ecology in which body, materials, memory, environment, and technology continually shape one another. Generative image models enter this ecology not as replacements for the painter but as new participants in an already complex dialogue between perception, imagination, and matter.
1. The Quasi-Personhood of Painting
The studio has always been a place where control and chance are in negotiation. Jackson Pollock’s drip paintings are often read as heroic expressions of individual gesture, but as Rosalind Krauss has pointed out, their real interest lies in how they hold accident and intention together (Krauss, 1985). Pollock let paint fall; gravity and viscosity were part of the work. Something similar is true of Generative AI image models, though the comparison only goes so far.
When an AI-generated image enters the studio, it does not introduce chance the way a brush slipping or paint pooling does. It introduces something more like a probabilistic field, a configuration shaped by statistical weighting across millions of training images. It does not fall. It does not resist. And when I try to translate it into paint, that difference becomes very apparent. The luminous flatness of the screen is immediately interrupted by the physical weight of oil or acrylic. The ‘smoothness’ of the digital image meets the drag of pigment on canvas, and something must give. That friction is where the actual work happens.
In my Ideal Studio paintings, which I began in the early 2000s, I used handmade collage to prepare source images for painting. My painting Fallingwater Seasons Remix (Figure 1), for example, was built from at least four different images, hand-collaged together. Working with Generative AI image models is, in some ways, a continuation of that process, assembling, selecting, and combining, but with crucial differences. With collage, I was handling physical materials, making cuts, and holding things together with my hands. With Generative AI image models, I am working at a remove, making choices about prompts and then waiting to see what comes back. The only real control I have is to change the prompt or to run the same prompt so many times that I eventually find something I want to work with. The painting illustrated in (Figure 2) is painted from an AI generated image which took almost a hundred generations to arrive at. That method, running a single prompt repeatedly until something usable emerges, is not one I have returned to since, even though I found the friction of the process and the reflection it afforded interesting. I have tested some AI that allows in-the-moment modifications of an image, such as Nano Banana. Still, at this point I prefer to use AI with existing material either as a prompt or in combination with collage work in Photoshop. I may return to text-to-image generation for inspiration should a conceptual project that suits it arise.
In The Love of Painting, Isabelle Graw articulates the provocative thesis that painting functions as a ‘quasi-person’ (Graw, 2018). Although paintings are inanimate objects, they circulate socially as if they bore the trace of an individual presence. A painting is rarely experienced as purely mechanical; its marks are read as indexes of labour, intention, attention, hesitation, and even psyche. Graw’s theory has two important implications. First, it highlights that paintings are not simply visual phenomena but social actors. They demand attention not merely as images but as residues of someone’s actions unfolding through time. Second, it underlines the irreplaceability of the brushstroke. A brushstroke is not simply a stylistic device but an inscription of embodied labour, a material record of a body’s encounter with resistance, gravity, rhythm, and duration. In this sense, the brushstroke functions not as a representation of subjectivity but as one of its physical traces.
The deeper stakes of Graw’s argument concern intentionality, but intentionality should not be understood here as conscious control over every mark. According to phenomenologists such as Edmund Husserl and Maurice Merleau-Ponty, intentionality describes the directedness of embodied consciousness, our fundamental orientation towards the world (Husserl, 1982; Merleau-Ponty, 2012). Every act is an act of or about something. Human painting emerges within this horizon of lived experience, memory, expectation, and perception. Contemporary generative image models operate differently. They generate images through statistical inference across vast datasets rather than through lived orientation towards the world. Their outputs may appear intentional, but they do not arise from existential involvement or embodied concern.
Graw’s account also resonates with Roland Barthes’s understanding of photographic indexicality in Camera Lucida (Barthes, 1981). Barthes argues that analogue photography possesses a unique ontological force because it records a physical encounter between light, body, and photosensitive surface, what he famously calls the ça a été [‘that-has-been’]. Writing, handwriting, photography, casting, and other analogue practices similarly preserve material traces of human presence. Painting therefore does not possess a monopoly on embodiment. Its distinctiveness lies elsewhere. Unlike photography, which indexes the presence of what stood before the lens, painting indexes the act of making itself. The brushstroke does not simply testify that something once existed; it testifies that someone acted. The painting records not only presence but the unfolding duration of embodied labour.
AI-generated images do not work this way. They are computationally derived. Their ‘marks’ do not point back to a body moving through time; they point to a distributed statistical process that is, by design, repeatable and impersonal. You can simulate the appearance of a brushstroke; you cannot reproduce the lived act of making that brushstroke. This is also what distinguishes AI painting, in my sense of that term, from simply displaying an AI output as art. When I paint from an AI image, I am re-embedding it in a process that carries bodily risk and temporal duration. I am making it mine, not by claiming authorship of the original image, but by insisting that the painting is something different from it.
A key aspect of my process in the studio is the number of choices I continuously have to make to make an artwork. Literally, many thousands of choices and physical gestures. This is one of the main arguments that sci-fi writer Ted Chiang gives as evidence of in his New Yorker article ‘Why AI Isn’t going to Make Art’ (Chiang, 2024). Writing about fiction writing, he states: ‘Art is something that results from making a lot of choices.’ And he goes on to say, ‘I think the same underlying principle applies to visual art, although it’s harder to quantify the choices that a painter might make. Real paintings bear the mark of an enormous number of decisions.’ But there is a limit to the number of choices made by AI due to the inherent structure of the models: ‘By comparison, a person using a text-to-image program like DALL-E enters a prompt … (The newest version of DALL-E accepts prompts of up to four thousand characters, hundreds of words, but not enough to describe every detail of a scene.) Most of the choices in the resulting image must be borrowed from similar paintings found online; the image might be exquisitely rendered, but the person entering the prompt can’t claim credit for that’ (Chiang, 2024).
Contemporary robotic painting systems push this question further. Ai-Da, for instance, (Lima, 2024) is a humanoid robot with camera-equipped eyes and a robotic arm that applies paint to canvas. The robot’s works have been shown at the United Nations and sold at Sotheby’s for over a million dollars. At first glance, it seems to bridge the gap: it sees, interprets, and executes. But its decisions emerge from calibrated statistical models, not from consciousness or embodied negotiation with material. It does not risk failure in the existential sense in which a human artist does. Harold Cohen’s earlier system AARON (Cohen, 1995), developed from the 1970s onward, raises similar questions. AARON could generate drawings autonomously, but Cohen himself resisted calling it an independent artist; for him, it remained an extension of his own creative inquiry. Both AARON and Ai-Da function as interesting provocations, but they highlight rather than erase the specificity of human embodiment in painting. These precedents are important when thinking about what is genuinely at stake. Machines can produce images. They cannot undergo the embodied human struggle of making a painting. Graw’s ‘quasi-person’ can therefore be understood less as a claim about individual genius than about the persistence of embodied presence within the painted object. The painting continues to circulate socially after the painter has left the studio because it retains the visible residue of an encounter between body and matter.
2. Deleuze, Sensation and Painting as Thought
If Graw foregrounds painting’s social ontology, Gilles Deleuze shifts attention to its phenomenology. Where Graw asks how paintings continue to act socially as ‘quasi-persons,’ Deleuze asks what painting does to perception itself. In Francis Bacon: The Logic of Sensation (2003), Deleuze argues that painting is not primarily representational but diagrammatic. It does not simply depict the world; it intervenes in perception, making forces visible that ordinarily remain unseen. The diagram is therefore less a compositional device than an event of discovery. It interrupts habitual seeing, requiring the painter to think through material encounter rather than prior representation.
When an AI-generated image surprises us, it is not because a body has struggled with sensation, but because a probabilistic model has produced an unlikely configuration. That is a genuinely different thing. Painting is embodied because it produces a different mode of cognition and perception than language or computational image generation. Computational image generation does not undergo this struggle in the Deleuzian sense. Novel configurations emerge through probabilistic inference rather than through the reciprocal negotiation between body and material resistance that characterises painterly practice. A generated image may surprise its user, but the surprise belongs to the human encountering the output rather than to the system producing it. The computational model does not experience uncertainty, hesitation, fatigue, or perceptual crisis about its output.
The Deleuzian account of sensation is difficult to transfer directly to Generative AI because the model does not inhabit sensation as a living body does (Deleuze and Guattari, 1987). It cannot reorganise sensation because it does not itself inhabit sensation. This distinction becomes especially important once AI-generated images enter the painter’s studio. At that point the image ceases to function as a finished output and instead becomes what Deleuze would recognise as a diagram, a proposition that interrupts habitual ways of seeing. Its significance lies not in its completion but in its capacity to provoke further material thinking. The painter does not reproduce the generated image. Rather, its computational order is repeatedly broken apart and reconstituted through the contingencies of pigment, gravity, touch, drying time, revision, and accident. The diagram therefore survives the transition from computation to painting not because it is copied, but because it is transformed through embodied labour.
In some ways, AI-prompts or AI-generated images could be seen to resemble the proposals of conceptual art. Joseph Kosuth’s Art as Idea as Idea series relocated the artwork from the object to the proposition (1969). Sol LeWitt insisted that the idea becomes the machine that makes the art, with execution delegated to others (1967). For both Kosuth and LeWitt, the artwork begins as a conceptual proposition rather than a materially resolved object. Execution remains important, but its role is transformed. The work exists first as a structure of thought. AI-generated images can be understood in similar terms. They function less as completed artworks than as propositions, conceptual scores, or speculative prompts that invite further interpretation. Their significance lies not in their completion but in their openness. Like LeWitt’s instructions, they establish a field of possibility rather than determining a final form. The crucial difference, however, is that LeWitt’s instructions remained grounded in human intentionality. They originated within an embodied artistic practice before being delegated to others. Generative image models produce propositions through computational inference rather than lived artistic intention. Their outputs therefore require a further act of interpretation if they are to become paintings. But of course they can still be considered artworks in themselves if the human artist decides that they are.
Conceptual art has often moved toward dematerialization; Lucy Lippard’s ‘dematerialization of the art object’ was a deliberate critical move away from the fetish of the physical (Lippard and Chandler, 1968). What I find myself doing with AI is almost the inverse. The AI output is already dematerialized: screen-bound, luminous, infinitely reproducible. When I paint from it, matter reasserts itself. The digital proposition becomes a physical object with weight, texture, and presence. The smoothness is interrupted by the density of the material. The act of painting therefore re-materialises the computational proposition, returning algorithmically generated hallucinations to surface, duration, and bodily encounter. For example, if an AI model used a Claude Monet painting as inspiration for a digital hallucination, it would be reduced to a smooth digital version, which would mean it loses its process referentiality, especially in relation to the surface and the duration of making the painting. If an artist paints from that same digital hallucination, it re-acquires those connections through another human interpreting the image through duration and application of surface. If conceptual art temporarily displaced the artwork into language, instruction, or proposition, painting after Generative AI performs a complementary movement. It returns propositions to matter. I understand this process as an inversion of the dematerialisation described by Lippard and Chandler and in doing so, it demonstrates that meaning does not reside solely in ideas but also in the embodied processes through which ideas become material events.
Like painting, writing emerges through the coordination of brain, hand, muscles, memory, and tools, whether pen, typewriter, or keyboard. A handwritten manuscript, like a painting, bears the traces of bodily movement and temporal duration. The distinction is therefore not that painting alone is embodied. Rather, painting and writing mobilise embodiment differently and consequently engage the perceiving subject through different cognitive and sensory pathways. Language is mediated through symbolic systems whose material support ultimately serves semantic meaning. Even when typography or handwriting becomes expressive, words are decoded linguistically before they are apprehended sensorially. Painting operates differently. Colour, scale, texture, viscosity, gloss, opacity, and material resistance are not vehicles for meaning; they are themselves constitutive of meaning. Before interpretation begins, the body has already responded. The eye adjusts to chromatic relationships, the nervous system registers texture and depth, and the viewer’s movement through space continually alters the work’s appearance. Painting therefore occupies a distinctive position among embodied practices because it exists as a materially present object in the world. It is encountered through proximity, scale, light, surface, and duration. Its meaning cannot be separated from its physical existence. In this sense, Deleuze’s ‘logic of sensation’ (Deleuze, 2003) extends beyond the painter’s body to encompass the embodied encounter between painting and viewer.
In my Greenhouse Archive Translations series (drawings and paintings, 2025–26, ongoing), I worked from a decade of photographs documenting abandoned Danish glasshouses. I used AI to generate speculative remix images based on their decay, impossible architectures fusing steel frames, broken glass, and invasive plants. On screen, these images hover between plausibility and fantasy. On canvas, their instability becomes explicit. Painting the work meant confronting the granular texture of rust, the opacity of shadow, and the scale of overgrowth. The AI image had smoothed all of this out. Painting it allows the viewer to question the object and the image.
The act of painting therefore does more than materialise a digital image; it translates a computational proposition into an embodied event. I use the term ‘translation’ deliberately. Unlike transmediation or remediation, which describe movement between media while preserving informational content, translation acknowledges transformation. Every translation alters meaning because each material possesses different affordances, resistances, temporalities, and sensory capacities. The movement from algorithmically generated image to painting is therefore not the transfer of information from one medium to another. It is the encounter between two fundamentally different modes of thought: one statistical and computational, the other embodied and materially situated. Translation, in this sense, resembles the work of a literary translator. A translated poem is neither a copy nor a replacement of its source but another work that emerges through negotiation with the possibilities and limits of a new language. Likewise, the painter does not reproduce the AI-generated image. The painter discovers what the image becomes when translated into pigment, scale, texture, gravity, drying time, and bodily gesture, once it enters the cognitive ecology of the studio. Seen in this light, AI-generated images become less significant as finished pictures than as participants within a broader ecology of painterly thinking. Their importance lies not in replacing embodied practice but in provoking new material encounters.
3. Embodiment and Nonconscious Cognition
If Deleuze understands painting as the production of sensation, N. Katherine Hayles extends this discussion by asking how cognition itself is distributed across bodies, environments, and technologies. Rather than opposing humans and machines outright, Hayles complicates the distinction by introducing the concept of the ‘cognitive nonconscious’. In Unthought, she writes that ‘cognition is a much broader capacity than consciousness; it extends to all living systems and some technical systems as well’ (Hayles, 2017: 11). This is not a mystical claim but a practical one; we know more than we know we know, and much of what shapes our actions happens beneath the level of deliberate thought. Hayles further argues that technical systems can engage in nonconscious cognition but notes that they lack features central to biological cognition, such as emotion and embodied affect (Hayles, 2017: 22, 30). Rather than replacing human cognition, technical systems become participants within wider cognitive assemblages. Cognition therefore emerges relationally, through continuous exchanges between bodies, environments, materials, and technologies rather than residing exclusively within individual minds.
Painting itself is a cognitive assemblage. When I am working on a canvas, I adjust pressure before I decide to; my eye recalibrates color relations through perceptual shifts I am not consciously directing; my body compensates for gravity and fatigue. These are not simply calculations. They are felt emotive negotiations with material conditions. The smell of oil, the drag of a brush across an uneven ground, the resistance of canvas these are not external to the cognitive process but constitutive of it. They shape what the work becomes.
Philip Guston described the studio as a haunted place full of presences, artists who have shaped your vision, histories you carry with you (2011). As work progresses, he suggested, these presences gradually fall away, until eventually, if you are lucky, even the self departs, and the painting emerges from somewhere beyond conscious control. This resonates with Hayles’s account of nonconscious cognition, but it extends it into something more existential. The painter is not only a cognitive system processing sensory data; she is someone who can be absorbed, who can lose herself in the act, and whose work carries that loss as part of its meaning.
Gertrude Stein captures something related when she writes that the minute you or anyone else knows what you are, you are not it (1937). The deepest moments in painting, for me, happen when self-consciousness falls away, when seeing, acting, and sensing unfold before reflection catches up. Generative AI models have no version of this. They do not experience the rupture between knowing and being that Stein describes. They do not have the capacity to lose themselves yet. Painting depends on that oscillation. It is one of the things that makes the practice irreducible.
Hayles’s distributed model of cognition finds an important parallel in Jane Bennett’s account of material agency. In Vibrant Matter, Bennett argues that matter should no longer be understood as passive substance awaiting human intention. Instead, she describes ‘the curious ability of inanimate things to animate, to act, to produce effects dramatic and subtle’ (Bennett 2010: 6). Her concept of ‘thing-power’ does not attribute consciousness to objects but recognises that agency is often distributed across human and nonhuman actors. Pigment, oil, wax, paper, humidity, gravity, drying time, studio light, and bodily movement all participate in shaping the work that eventually emerges. Taken together, Hayles and Bennett suggest a way of understanding the artist’s studio that extends beyond either phenomenology or traditional accounts of artistic intention. The studio can be understood as a cognitive ecology, a term Edwin Hutchins coined in his book Cognition in the Wild (1995), which describes an environment in which thought emerges through the continuous interaction of body, materials, memory, environment, tools, and time. Cognition does not occur first in the artist’s mind before being expressed through paint. Rather, thinking unfolds through the reciprocal negotiations between painter, surface, pigment, atmosphere, and the accumulated histories embedded within materials themselves. The studio is therefore not simply the location in which painting happens; it is one of the conditions through which painting thinks.
My own studio practice illustrates this ecology. It is located in a former commercial plant nursery, now a garden in the process of regeneration. This matters as the environment is not a backdrop; it participates. Seasonal light, humidity, changing vegetation, and the cyclical rhythms of growth and decay continually shape the paintings that emerge there. I work with materials that carry histories: charcoal, beeswax, honey, Guinness, coffee. Coffee baked into paper leaves unpredictable tonalities. Beeswax behaves differently in winter and summer. Guinness oxidizes into browns and violets I could not have calculated in advance. These are not metaphors for something else; they are inscriptions of place, time, and material agency. They push back against my intentions, and the work is shaped by that pushback. The painting that results is not simply something I made; it is something that happened through a field of forces in which I was one factor among several. The painting develops less as the execution of a predetermined image than as an evolving negotiation among human intention, material agency, and environmental condition.
Hayles nevertheless cautions against collapsing cognition into computation. While computational systems instantiate forms of nonconscious cognition, pattern recognition, adaptive feedback, probabilistic inference, they do not therefore become equivalent to human thought. As Hayles argues in How We Became Posthuman (1999), the danger lies in the ‘erasure of embodiment’ that occurs when information is imagined as independent of its material instantiation. Human cognition remains inseparable from metabolism, sensation, affect, mortality, memory, and lived history. These are not secondary additions to cognition but conditions of its possibility.
From this perspective, the distinction between material agency and algorithmic agency becomes clearer. The agency of paint does not consist in autonomous intention but in its capacity to resist, exceed, and redirect human plans through its own material behaviour. Pigment flows, wax contracts, paper buckles, gravity pulls, and solvents evaporate according to physical processes that neither the painter nor the material fully controls. Algorithmic systems likewise generate outcomes that often surprise their users, but this surprise arises through probabilistic opacity rather than material contingency. The painter therefore negotiates with two different kinds of agency: computational systems generate unforeseen propositions, while materials generate unforeseen events. Painting begins where these propositions encounter matter.
4. Foucault, Fried, and the Painting as Event
Michel Foucault’s reading of Édouard Manet helps us think further about what distinguishes painting from mere image production. Foucault argues that Manet exposed the canvas as an object rather than a window. By reconfiguring light, flattening perspective, and repositioning the viewer, Manet made the constructed quality of the pictorial act visible (Foucault, 2009). His canvases insist on their own surface. In Olympia (1863), the flatness of the figure against a stark background interrupts the expected depth; in A Bar at the Folies-Bergère (1882), the mirrored background unsettles the viewer’s position. In both, painting becomes an event of visibility; it stages its own conditions of appearance rather than transparently mediating a scene.
Generative AI image art is an output: the terminal resolution of a process that begins (though not always in pure code) with some form of human intention. The usual progression is one of translation and attenuation: a person chooses to upload an image, articulates a prompt, and in doing so converts experience and desire into language; language is then processed into code; code resolves into a generated image. At each stage, the human gesture recedes further from the result. One could, of course, write code that produces images autonomously, without any linguistic prompt at all, but even then, the human act has merely been displaced earlier in the chain, not abolished. In either case, the process is sequential and self-closing: it ends in a generated file, a finished thing, sealed at the moment of its making. Painting, by contrast, remains open; it is a becoming between body, material, and world. Like Manet’s canvases, it makes visible the act of making visible.
Michael Fried’s account of the tableau extends this. For Fried, the tableau is not simply a large-format picture but a specific mode of address. It creates a staged presence that confronts the viewer while resisting mere spectacle (Fried, 1980). In Courbet’s The Painter’s Studio (1855), for instance, multiple registers of looking are set in play simultaneously: the artist at work, the model, the assembled social types, and the viewer are all drawn into a complex choreography of attention (Fried, 1990). The painting is not a representation of an artist’s milieu; it stages painting itself, its sociality, its labor, its address. The viewer is not a passive consumer of the image, but someone caught in an event. The tableau therefore stages not only representation but the viewer’s own bodily encounter with the painting. It insists upon duration, attention, and physical co-presence in ways fundamentally different from the endlessly replaceable digital image.
My Museum Collection canvases (2005–26, ongoing) work toward a similar condition. Consider a large-scale painting from 2005, in which a monumental Louise Bourgeois spider stands outside the Guggenheim Museum in Bilbao (Figure 3) beneath a yellow sky, alongside a more recent work in progress from 2026, where another Bourgeois spider appears among the trees outside the Louisiana Museum of Modern Art in Denmark (Figure 4), with a Brancusi column rising in the distance. Despite their temporal distance, both paintings operate through a comparable logic. Painted at two meters across, the works envelop the viewer. It is not an image representation of a sculptural garden; it is a tableau that demands duration. The viewer has to negotiate their own memory of canonical Modernism or of visiting the museums against the strangeness of its metamorphosis, and that negotiation is not passive. What AI can generate is an image of a garden with sculptures. The generated image itself does not constitute the kind of embodied tableau I am describing. It does not slow it down in duration, nor can it make it confront the viewer’s body as a human-made physical object.
Giorgio Agamben’s distinction between chronos (quantitative, sequential time) and kairos (qualitative, intensified time) is relevant here as it offers a useful way of describing the contrast I experience (Agamben, 1993). Generative AI image models produce in chronos: rapid, serial, undifferentiated. Painting works in kairos: the weeks of sustained labor required to translate an AI-generated image into pigment represent a temporal inversion, a deliberate slowing of what computation produces in seconds. This is not inefficiency. It is the point.
Seen in this light, translation becomes the central creative act of contemporary painting after generative AI. The originality of the work no longer resides solely in inventing unprecedented images but in discovering how computational possibilities are transformed through material encounter. Painting remains irreducibly creative because every act of translation is also an act of interpretation, and every interpretation unfolds within the embodied ecology of the studio.
5. AI, Painting, and Narrative
The theoretical arguments developed thus far emerge directly from my own studio practice. Rather than applying philosophical concepts retrospectively to completed works, my practice has become the site through which these questions first arose. The encounter with contemporary generative image models did not replace drawing or painting; it altered and continues to alter the conditions under which they unfold. What follows is therefore not offered as illustration but as a reflection on how theory and practice continually inform one another within the studio.
One of the things that painting can do that AI image generation models cannot do so far is participate in narrative, not in the sense of telling a story, but in the sense that W. J. T. Mitchell calls ‘pictorial narrativity’: an image that unfolds temporally through the act of looking, carrying traces of decisions, revisions, and hesitations that themselves tell a story of making (Mitchell, 1994). A painting is a record of time spent.
AI-generated images are, in a way, saturated with latent narrative. They emerge from vast cultural datasets, carrying traces of millions of prior images. In that sense, they are full of history. But paradoxically, they have no backstory of their own; they are appearances without labor, images without the temporal sedimentation of process.
In my Hybrids and Museum Collection series, this tension becomes explicit. The AI image generation prompts I work from propose speculative spaces, museum gardens where canonical sculptures merge with botanical forms or hybridized artworks from art history. On screen, these read as fragments of an alternate art-historical archive: they suggest exhibitions that never happened, institutions that never existed. When I paint from them, the work shifts register. Layers of paint accumulate; the narrative of struggle inscribes itself in the surface. What began as a machine-generated hallucination becomes a story of making, a record of the encounter between painter, algorithm, and material. I was able to do this before AI of course, and still do, but with AI image generation, I have another tool to turn to in order to find different propositions and research ideas.
Walter Benjamin’s distinction between storytelling and information is useful here. In The Storyteller, Benjamin argues that storytelling unfolds through repetition, memory, and lived experience, while information is instantaneous, contextless, and already saturated with explanation (Benjamin, 1968). AI image generation outputs function like information: rapid, endless, self-presenting, arriving already smoothed and resolved, without visible seams or signs of having been made (at least in the current more sophisticated models and not the early AI image generation models where glitches and imperfections were much more prevalent). Their opacity runs in the opposite direction: not transparent, but internally inaccessible, concealing the unmappable paths through latent space that produced them. Painting converts them into something closer to storytelling. By laboring over the AI image, slowing it down, and embedding it in pigment, the painter turns an informational hallucination into a narrative inscription. This is not a nostalgic argument for handicraft. It is an argument about what visual narrative requires: inscription, risk, and the embodied labor of telling. AI can generate images that look like stories. Painting can turn them into stories whose narrative is carried not only by what they depict but by the visible history of their making.
Rather than treating artificial intelligence as an autonomous intelligence or creative entity, it may be more productive to understand it as a material and processual medium whose possibilities emerge through artistic engagement. In this sense, Generative AI functions less as an independent author than as a set of technical conditions that shape and constrain creative action, much as photography did in the twentieth century. Vilém Flusser’s concept of the apparatus is particularly useful here Flusser, V. (2000). For Flusser, the camera is not simply a neutral tool but a programmed system that structures what can be imagined and produced, while simultaneously requiring human intervention to activate and negotiate its possibilities. The photographer is therefore always implicated within the apparatus rather than standing outside it as a sovereign author. AI image generation can be understood in similar terms: as an apparatus composed of datasets, computational architectures, interfaces, and probabilistic operations that artists enter into and work through. Meaning does not arise from the machine itself but from the dynamic interaction between human intention, embodied judgement, iterative prompting, selection, and transformation within this technical system. Interpreting AI as material rather than intelligence shifts attention away from questions of machine agency towards the creative negotiations that occur between artist and apparatus, where authorship becomes distributed across a process of continuous exchange without diminishing the central role of human criticality and aesthetic decision-making.
This understanding is reinforced by Sofian Audry in Art in the Age of Machine Learning (2021), where he identifies four persistent myths surrounding machine learning in artistic practice. In Myth 3, ‘Machine Learning Can Create Art Without Artists’, Audry challenges the popular narrative that AI systems are autonomous creators capable of replacing human artists. He argues instead that machine learning is inseparable from the networks of human labour, technical design, and aesthetic judgement that make its outputs possible. Training datasets are selected and curated by people; algorithms are designed, tuned, and implemented by engineers and researchers; and artistic meaning emerges through the choices made by artists who formulate prompts, evaluate results, discard unsuccessful iterations, and integrate outputs into broader conceptual frameworks. Rather than eliminating the artist, machine learning redistributes creative agency across a complex assemblage of humans, computational systems, and cultural infrastructures. Seen in this light, AI does not operate as an independent artistic intelligence but as another apparatus through which artistic practice is mediated. Like Flusser’s photographic camera, it offers a programmed field of possibilities that only becomes meaningful through human engagement, interpretation, and critical decision-making. The apparent autonomy of the system is therefore less a property of the technology itself than an effect of obscuring the extensive human participation embedded within its operation. AI clearly participates in cognition and artistic production without sharing the existential structure of human intentionality.
6. Memory, Technology, and the Painter
As I am working on this text, I just heard on the news that the artist David Hockney has passed away. Hockney was a big influence on my father (who was also an artist) and then subsequently an influence on me. My father’s interest in Hockney also connected to the theme of new technology, and we often discussed Hockney’s experiments with Photoshop and then later with the iPad drawings.
My father passed away in 2022, and a large part of the grief involved in losing him has been connected to existential questions about meaning in life and the human condition. Looking back on my father’s artwork, career, and life, it is possible with the benefit of hindsight to see how everything fitted together but also how complex human existence is and how creative practice becomes a mirror for humankind to see itself. My father and I were deeply connected on the themes of history, education, and creativity and the potential of what technology could do in those areas. I have often wondered what he would think of the rise of AI in the years since his passing in 2022; would he have been excited by the potential or horrified by the consequences? My father was an educator, and he often brought home new technology from work for us children to experience and experiment with. I have strong memories of large sheets of paper on the drawing room floor and programming the turtle robot to draw abstract shapes. He brought home early computers; each upgrade brought new possibilities. As a teenager, I won a prize in a competition to design a magazine cover for the computer magazine Your Sinclair, where I drew a portrait of Indiana Jones using Mac Paint.
Reading and listening to interviews with Hockney now that he has passed, he didn’t seem that caught up with his legacy or so afraid of death. He seemed to live his life in the moment and be constantly amazed at everything around him. My father could also live in the moment and was fascinated by life and history but he was also very different in his outlook. He feared death, yes, but he feared even more the possibility of being forgotten. Since his passing, that has been the hardest part for me: the ache of knowing how much it would have hurt him to imagine his memory fading.
I wonder again about my father, about what he left behind, about how much of him I can keep alive through memory, through story, through words, through painting. My father and I also shared a love for Claude Monet. He first introduced me to those shimmering canvases while helping me assemble a portfolio. It wasn’t just the colours or the brushstrokes that moved him; it was the patience of them, the way they suggested not just beauty but duration, the embodiment of a landscape observed over time. I visited Monet’s garden and studios at Giverny for the first time in April 2025, and after getting over the fact that his large studio had been turned into a gift shop, I began to see how Monet had constructed the whole environment as a vehicle for making his paintings, so that he could control the colour, the light and composition in an embodied way.
The drawing illustrated in Figure 5 was made by my father when I was five years old. It remains one of my earliest memories, not because I remember every visual detail, but because I remember the experience of being drawn. I remember sitting still, waiting, aware of being observed. That sensation, the awareness that another person’s sustained attention was shaping an image of me, has stayed with me throughout my life. The drawing is therefore not simply an image of a child; it is also an index of a life, saturated with the accumulated habits, memories, values and experiences that formed the person who made it.
The image contains histories that neither a viewer nor even I fully understood as a child. Through the Georgian window lies the landscape of the Irish border, dividing Northern Ireland from the Republic. During the Troubles, that boundary shaped everyday life through checkpoints, soldiers and the constant presence of violence, even if these realities remained only partly comprehensible to me as a young child. The television on which I sit quietly echoes that wider political world, carrying daily images of conflict into our home. Yet what dominates the drawing is neither politics nor documentary description, but care. My father’s sustained attention transformed an ordinary afternoon into something that continues to exist long after both the moment itself and, now, the artist who created it have disappeared.
This is why drawing and painting matter to me from an existential perspective. Their significance does not lie simply in the fact that they record perception through the movements of a body. Cameras and increasingly sophisticated machines also produce records of the world. Rather, painting becomes meaningful because it is one of the ways human beings attempt to resist the transience of experience. We invest time, physical effort, emotion and attention into making something that might outlast us, even while knowing that its meaning can never be guaranteed. Every painting therefore carries the possibility of failure, but also the hope that another person, perhaps years or generations later, will encounter it and recognise something of another human life.
When I now live with my father’s work, I do not experience the drawings merely as objects but as continuing conversations. They remind me that creative practice is ultimately less concerned with producing images than with creating relationships across time. This, I believe, is where painting differs most profoundly from contemporary AI image generation. AI can generate convincing visual outcomes, but it does not know what it means to mourn, to remember, to hope, or to fear being forgotten. It cannot experience the uncertainty that gives artistic labour its existential weight. The meaning of my father’s drawing lies not only in what it depicts but in the fact that another human being chose to spend part of his finite life looking carefully at his son. Painting, in this sense, is not simply a technology of image-making but a fundamentally human act of making meaning through a distillation of observations from an embodied experience of the world.
Hockney’s large-scale paintings, particularly his Yorkshire landscapes, further insist on the irreducibility of embodiment. Their scale demands physical negotiation from both maker and viewer. The painted surface accumulates decisions over time, erasures, reworkings, adjustments, forming what might be understood as a temporal sedimentation of perception. These qualities cannot be compressed into instantaneous production. In this respect, Hockney’s work aligns with the argument developed earlier: painting is not merely an image but an event structured by duration, material resistance, and bodily presence. At the same time, Hockney’s openness to new technologies demonstrates that the relationship between painting and technological mediation is not oppositional. His experiments with photography, fax machines, and digital tablets explore how tools can reconfigure vision without abandoning embodiment. What persists across these shifts is a commitment to the temporality of seeing and the situatedness of the body within space.
Hockney’s practice offers a compelling counterpoint to Generative AI because it demonstrates that technological experimentation does not negate embodiment but can intensify it. Hockney has repeatedly emphasized that seeing is an active, embodied process. As he has remarked, ‘We see with memory’, suggesting that perception is not a neutral recording but a temporal and bodily construction (Hockney and Gayford, 2016). John Berger puts it in another way, which points even more to the individual human seeing something in the world from their own experience: ‘An image became a record of how X had seen Y’ (Berger, 1972). In relation to Hockney’s artwork, this insight underpins his sustained critique of single-point perspective and informs works such as the photographic joiners of the 1980s and his later digital drawings, where multiple viewpoints are assembled into a single image. These works do not present disembodied vision; rather, they register looking as a durational, physical act unfolding through time.
In Secret Knowledge (2001), Hockney investigates the historical use of optical devices by Old Masters, arguing that technologies such as lenses and mirrors were integral to pictorial construction. In one section, he uses a drawing device called a Camera Lucida (the same title used for Roland Barthes’ seminal book on photography). He is convinced that Ingres used such a device in making a series of portraits and wants to test his theory. While not all of the results are at the level of the Ingres drawings, he makes a strong argument that this could have been the way the drawings were made. Crucially, his claim is not that technology replaces skill, but that it becomes part of an expanded field of embodied practice. The artist’s hand, eye, and body remain central, even when mediated. Technology, in this sense, is not external to embodiment but folded into it, augmenting rather than displacing the act of seeing. This distinction becomes critical when considering contemporary AI systems. Unlike the camera obscura or the iPad, AI does not extend the body’s perceptual capacities; it produces images without perception. AI shares the apparatus’s dependence on human activation but breaks from it in lacking any perceptual register at all. Hockney’s use of the iPad, for example, retains the gestural and temporal qualities of drawing: pressure, velocity, rhythm, and revision. His digital works remain inscriptions of bodily movement and attention. By contrast, AI-generated images lack this indexical trace. They are not records of looking or making but statistical resolutions derived from datasets.
As noted in the introduction, this distinction also requires a more precise account of what is specific to AI-generated images as painterly source material. The friction encountered in painting from an AI-generated image is not fundamentally different from painting from a photograph. The resistance of paint, the unpredictability of materials, and the interpretive labour of translating an image into a painting remain central regardless of the origin of the source. What distinguishes AI-generated imagery is therefore not the painterly process itself, but the character of the images it produces. The image no longer records something that has been seen but proposes something that could never have existed, requiring the artist to negotiate between recognition and invention, between visual credibility and imaginative speculation. AI image generation therefore functions less as a substitute for artistic practice than as a new kind of material within it: an apparatus capable of generating synthetic propositions that become meaningful only through the embodied processes of painting.
Recognising this distinction also reframes the question of meaning. The significance of a painting cannot be understood as residing solely within the object or within the process of its making. Since Marcel Duchamp’s Fountain (1917), it has become clear that artistic meaning emerges through the relationships between object, maker, context, and viewer. AI image generation models complicate this ecology by generating images on demand that are capable of responding to our intentions, desires, and cultural references with extraordinary fluency. Its apparent creative power lies precisely in its ability to produce convincing visual propositions that resonate with human subjectivity despite having no lived experience of its own. Yet this does not diminish painting’s distinctive role. If AI can generate an endless stream of synthetic images, painting remains a primary language through which those images are subjected to bodily experience, reflection, duration, and material transformation. The contribution of painting is therefore not simply to translate AI images into another medium, but to slow them down, test them against embodied perception, and reconstitute them through the existential conditions of human making. It is this dialogue between synthetic image generation and lived experience, rather than any claim of technological opposition, that defines the distinctive contribution of painting in the age of generative AI.
Positioning AI image generation within this lineage clarifies its limitation. The issue is not that AI uses advanced technology, but that it lacks a body through which technology can be meaningfully experienced. Hockney’s practice shows that tools become artistically significant when they register and extend embodied perception. AI does not meet this condition; it simulates visual outcomes without undergoing perceptual or material processes. Embodiment, therefore, should not be framed as a nostalgic return to pre-digital authenticity, but as a critical framework for evaluating how images come into being. Hockney exemplifies a model in which technological innovation deepens, rather than displaces, the bodily conditions of art. In this light, painting’s continued relevance lies in its capacity to hold together perception, material, and time within lived experience. Generative AI remains outside this assemblage, producing images that can resemble, but that were never seen, and never envisioned.
Conclusion: Painting, AI, and the Future of the Tableau
Introducing AI image generation into the studio has not felt like a rupture. It has felt more like an extension of long-standing negotiations between intentionality and chance, control and material resistance, the image I plan and the painting that emerges. Working from its outputs, I have become more aware of what painting is by contrast: an event that is irreducibly bound to embodiment, duration, and the specific conditions of a body in a studio at a particular time. My own practice is shaped by this understanding. I occasionally use AI image generation as a foil, a source of propositions that I push against. The generated image suggests chromatic arrangements or spatial configurations I might not have arrived at otherwise. But when those suggestions meet oil, acrylic, or beeswax on canvas, they are transformed. Painting has always absorbed external forces, psychoanalysis, linguistics, photography, digital networks, and found ways to metabolize them into renewed forms of critical visibility. AI image generation is the latest of these. Just as Pollock’s drips engaged the unconscious and gesture, or Kosuth’s tautologies interrogated the linguistic frame, working with AI models makes the algorithmic conditions of contemporary image culture visible.
I hope that the central contribution of this text lies not in defending painting against artificial intelligence on technical or procedural grounds, but in articulating the distinctly human dimensions of painterly practice. At a moment when Generative AI increasingly challenges assumptions about what is distinctively grounded in human lived experience in creative production, painting offers a way of understanding artistic practice as an existential search for meaning rather than simply a process of image generation. The act of making a painting involves an investment of human experience in pursuit of something whose significance cannot be known in advance. This openness to uncertainty is inseparable from the possibility of failure, and it is precisely this risk that gives painting its human character. Rather than merely producing visual outcomes, painting constitutes an embodied process through which artists negotiate their relationship to the world and seek forms of connection with others. Unlike symbolic or computational systems that operate through predefined structures, painting functions as a primary language rooted in sensory experience, affect, and the holistic condition of being in the world. While I have acknowledged the importance of the physical mechanics of painting, vision, gesture, gravity, fatigue, and material resistance, I have tried to focus on a more compelling argument that emerges through questions of emotion, desire, embodiment, and meaning. These are not incidental qualities that accompany the act of painting; they are the conditions that make painting a uniquely human mode of inquiry and communication.
The dialogue between AI image generation and painting is not about surpassing or replacing. It is about reframing. It extends the history of the tableau into a moment where the object of painting is simultaneously material and virtual, tactile and computational. And in staging that dialogue, painting once again demonstrates what it has always been capable of: making visible the conditions of seeing and holding that visibility open for the time it takes to look. To engage AI image generation models is to continue painting’s philosophical inquiry into its own objecthood, and in doing so, to recognize that the tableau remains a vital stage on which the drama of visibility and perception is perpetually rehearsed.
Competing Interests
The author has no competing interests to declare.
References
Agamben, G 1993 Infancy and History: The Destruction of Experience. London: Verso.
Audry, S 2021 Art in the Age of Machine Learning. The MIT Press.
Barthes, R 1981 Camera Lucida: Reflections on Photography. Howard, R (trans.). New York: Hill and Wang.
Benjamin, W 1968 The storyteller: Reflections on the works of Nikolai Leskov. In: Arendt, H. (ed.) Illuminations. Zohn, H (trans.). New York: Schocken Books, pp. 83–110.
Bennett, J 2010 Vibrant Matter: A Political Ecology of Things. Durham, NC: Duke University Press.
Berger, J 1972 Ways of Seeing. London: Penguin.
Chiang, T 2024 Why A.I. Isn’t Going to Make Art. The New Yorker, 31 August. https:www.newyorker.com/culture/the-weekend-essay/why-ai-isnt-going-to-make-art. [Last accessed 21st May 2026].
Cohen, H 1995 The Further Exploits of AARON, Painter. Stanford Humanities Review, 4(2): 141–158.
Deleuze, G 2003 Francis Bacon: The Logic of Sensation. Smith, D W (trans.). London: Continuum.
Deleuze, G and Guattari, F 1987 A Thousand Plateaus: Capitalism and Schizophrenia. Massumi, B (trans.). Minneapolis: University of Minnesota Press.
Flusser, V 2000 Towards a Philosophy of Photography. Mathews, A (trans.). London: Reaktion Books.
Foucault, M 2009 Manet and the Object of Painting. Barr, M (trans.). London: Tate Publishing.
Fried, M 1980 Absorption and Theatricality: Painting and Beholder in the Age of Diderot. Berkeley: University of California Press.
Fried, M 1990 Courbet’s Realism. Chicago: University of Chicago Press.
Graw, I 2018 The Love of Painting: Genealogy of a Success Medium. Berlin: Sternberg Press.
Guston, P 2011 Philip Guston: Collected Writings, Lectures, and Conversations. Coolidge, C (ed.) Berkeley: University of California Press.
Hayles, N K 1999 How We Became Posthuman: Virtual Bodies in Cybernetics, Literature, and Informatics. Chicago: University of Chicago Press.
Hayles, N K 2017 Unthought: The Power of the Cognitive Nonconscious. Chicago: University of Chicago Press.
Hockney, D 2001 Secret Knowledge: Rediscovering the Lost Techniques of the Old Masters. New York: Viking Studio.
Hockney, D and Gayford, M 2016 A History of Pictures: From the Cave to the Computer Screen. London: Thames and Hudson.
Husserl, E 1982 Ideas Pertaining to a Pure Phenomenology and to a Phenomenological Philosophy. Kersten, F (trans.). The Hague: Martinus Nijhoff.
Hutchins, E 1995 Cognition in the Wild. Cambridge, MA: MIT Press.
Kosuth, J 1969 Art after philosophy. Studio International, 178(915): 134–137.
Krauss, R E 1985 The Originality of the Avant-Garde and Other Modernist Myths. Cambridge, MA: MIT Press.
LeWitt, S 1967 Paragraphs on conceptual art. Artforum, 5 (10): 79–83.
Lima, E 2024 AI art and public literacy: the miseducation of Ai-Da the robot. AI and Ethics. Berlin: Springer. http://doi.org/10.1007/s43681-024-00488-5
Lippard, L R and Chandler, J 1968 The dematerialization of art. Art International, 12 (2), pp. 31–36.
Manovich, L and Arielli, E 2024 Artificial Aesthetics: Generative AI, Art and Visual Media. https://manovich.net/index.php/projects/artificial-aesthetics [Last accessed 1 August 2026].
Merleau-Ponty, M 2012 Phenomenology of Perception. Landes, D A (trans.). London: Routledge.
Mitchell, W J T 1994 Picture Theory: Essays on Verbal and Visual Representation. Chicago: University of Chicago Press.
Stein, G 1937 Everybody’s Autobiography. New York: Random House.
Wang, G 2025 Curving spontaneously: Sougwen Chung’s collaborative paintings with algorithmic systems. Journal of Visual Art Practice. http://doi.org/10.1080/14702029.2025.2567148
Zylinska, J 2020 AI Art: Machine Visions and Warped Dreams. London: Open Humanities Press.




