Creating images with artificial intelligence goes beyond simply running a command in software. It is an experimental space where human intent and the opaque nature of algorithms clash, creating friction. The user provides a prompt—an instruction in natural language—activates the system, and then waits for a result that is never entirely predictable. This cycle repeats: a new prompt, a new image, and another effort to realize the original idea. The true engagement happens in the gap between what is asked for and what is received. This space embodies the addictive mechanic Natasha Dow Schüll describes in Addiction by Design: a system of intermittent rewards, driven by small surprises, that keeps the user engaged endlessly.

The technical explanation itself reveals this playful nature. AI image platforms train neural networks using extensive datasets of visual and textual data. The model learns to connect words with shapes, colors, textures, and compositions—not through genuine understanding, but via statistical patterns. When a user inputs a prompt, the system translates it into a map of numerical relationships. It explores a range of possibilities, creating an image from noise—initially a blurry form that gradually sharpens and aligns with the instruction through iterations. This first stage, where the screen displays only a misty blur that slowly becomes clearer and more organized, resembles the experience of developing an analog photograph: the ghostly image appearing in the developer tray, filled with anticipation and uncertainty. Both the curiosity for the final image and the anxiety of the process feed into this experience; the technical delay itself becomes an integral part of the enjoyment.

Walter Benjamin referenced this apparition as a stain: a shapeless presence filled with potential, existing before any interpretation. Maria Filomena Molder revisits this idea to emphasize that the stain is a field of forces, not a void; it can originate from within, like moisture seeping through a wall or a flush emerging on the skin. This analogy links it to medical imaging — X-rays, ultrasounds, MRIs — where the stain signals something hidden, waiting to be uncovered. The viewer learns to interpret the indeterminate as evidence, turning the act of deciphering into a game of anticipation and reward. In AI, the shift from stain to sign mirrors this process of revelation, echoing the same rhythm of discovery.

The issue of authorship arises here: can an image be considered solely mine? Many view AI-generated images as a challenge to traditional ideas of authorship, yet art history has long been based on influence and appropriation networks. For instance, Francis Bacon drew inspiration from magazines, medical manuals, and nature documentaries, focusing on moments of flesh in tension, with bodies often distorted by speed or violence. Gilles Deleuze, in Francis Bacon: The Logic of Sensation, described Bacon’s figure as a condensation of intensities rather than an exact replica. Even when sourced from banal visuals, these images aimed to stimulate the pictorial gesture, never to substitute it.

Bacon drew inspiration from a private visual archive amassed over decades, which included newspaper clippings, medical journal pages with X-rays and photos, film stills, fitness manuals, and wildlife documentaries. He was captivated especially by images capturing moments of maximum tension — such as a rugby player’s open mouth, faces distorted by impact, or a bull at the moment of collision. These images, often stained, torn, or trampled in the studio, served not as models but as detonators of intensity. Bacon believed in subjecting them to brutal transformations — painting over, folding, or even turning the canvas around to work on the reverse — to strip them of their documentary role and turn them into sensory matter. This act of removing their referential context echoes the logic of generative AI: training data are displaced fragments, deprived of their original use, reassembled into a new figure that no longer belongs to their initial world.

Francis Bacon employed a unique technique: he painted on the back of his canvases. He valued the tactile resistance of the rough surface, which challenged the precise control of his gestures. This resistance compelled him to embrace imperfections and unforeseen accidents that he couldn’t predict or fully fix. It served as a counterbalance to the skill gained through repetition, allowing imperfections to surface. The artist saw this approach as a way to break free from the automatic movements of a trained hand—similar to Henri Cartier-Bresson’s idea of the “decisive moment” in photography, where a deliberate surrender to chance enables unplanned elements to appear within the image.

Andy Warhol extended this idea by mechanizing reproduction, turning the creative act into a form of coordination, and blurring the line between the original and the copy. Similarly, working with AI involves organizing, manipulating, and enhancing a pre-existing archive instead of creating from nothing. The human role is more like a curator than a divine creator, but it remains equally creative.

Bacon openly admitted to his addiction to roulette. He gambled to experience a randomness beyond his control — unlike painting, where even accidents are guided by the hand and body. This ongoing tension between control and surrender also exists in AI creation: the prompt is the bet, the image’s reveal is the spin, and the outcome always contains an element of unpredictability. Johan Huizinga, in Homo Ludens, notes that all games are governed by rules, and in this context, chance is limited by the model’s grammar, the allowed words, the styles within its training data, and the ethical restrictions built into its design. Yet, this limited randomness is enough to create the excitement that encourages repeated play.

The creator working with AI might use a method similar to Bacon’s reverse painting: the prompt acts as a verbal stroke, while the algorithm represents a rough, resistant surface—producing resistance and unpredictability. Rather than responding linearly, it moves through the request with noise, displacements, and interpretations that stray from the original intent. Embracing this resistance means collaborating with the machine as an improvisational partner, much like John Cage did in Silence, where chance influences the process. The gesture shifts from being just an instruction to becoming a choreography between control and randomness—similar to roulette, where each spin is both a decision and a surrender.

From this perspective, thinking about publishing futures goes beyond just imagining new formats; it involves creating ecosystems of production that acknowledge and investigate the materiality of images, focusing on their political and economic impacts. Hito Steyerl, in Duty Free Art, emphasizes that no digital image is neutral; its aesthetics and value are shaped by infrastructures, circulation, and filtering networks that determine what we see and what we overlook. Engaging critically with AI requires addressing this dual aspect — both the creative act and the system that frames it — expressing both in a practice that is aware of its technical and symbolic embedding.

Tools like Brian Eno’s Oblique Strategies, Italo Calvino’s ideas of lightness and multiplicity in Six Memos for the Next Millennium, and Vilém Flusser’s gesture theory all suggest ways to expand this practice. Each emphasizes that creation is fundamentally linked to the medium supporting the gesture. Therefore, addiction can be seen as a method — not just passive consumption but a force for sustained attention, where each image generated acts as a visual spin of the roulette wheel, a moment of insight that fuels curiosity and rekindles thought.

AI platform unpredictability is both structural and strategic. These platforms thrive on repetition: every unexpected variation prompts another try, more credits bought, and extended user engagement. Similar to the slot machines analyzed by Schüll, AI provides intermittent rewards and a steady flow of near-misses. This pattern fuels addiction and sustains a rhythm of expectation and discovery.

The temporality of these platforms introduces another complexity. Each model is swiftly replaced or ‘improved,’ destroying remnants of previous versions. Obsolescence is built into the system and fuels rapid consumption, similar to the fashion or hardware sectors, while traditional art forms tend to remain relevant for decades. In this context, visual memory is maintained through constant updates: archives are flexible, but the past rapidly vanishes, and the game restarts repeatedly.

The environmental issue remains equally pressing. Each generated image triggers data centers that use considerable energy and water. Schmidt et al. (2024) estimate that creating a single AI image uses about 1.7 Wh — roughly half a smartphone charge. Data centers already make up 1–1.5% of global electricity use (IEA, 2023), with CO₂ emissions reaching hundreds of millions of tons annually. In comparison, analogue photography involves toxic chemicals and consumes more energy per image—a 36-exposure roll can use up to 268 Wh (Grunwald et al., 2010). Although these impacts are spread across large-scale production, digital photography has less chemical waste but increases demands on production and disposal. AI, while more energy-efficient per image, could lose this benefit if energy comes from non-renewable sources and use keeps growing.

Ethical considerations are also built into the structure. Tools like Midjourney and DALL·E automatically filter out sensitive content, such as explicit nudity, graphic violence, images of public figures or politicians, and racially or religiously charged themes. This filtering occurs at two levels — during training data selection and through active prompt blocking. Hito Steyerl, in Duty Free Art, describes this as an architecture of visibility: what we perceive is linked to what becomes invisible through prior choices. AI, by shaping what can be represented, influences the collective imagination itself.

Ultimately, creating with AI is fundamentally a form of play. It is a game against the house — and the house always wins. Each image results from a negotiation between human input, algorithms, and economic interests. Chance, here, is more than a method; it is a diagnosis. It reveals as much about the image produced as about the system that insists on showing us another — and then another, and then another.

References

Warhol, A. (1975). The Philosophy of Andy Warhol (From A to B and Back Again). New York: Harcourt Brace Jovanovich.

Benjamin, W. (1982). Das Passagen-Werk. Frankfurt am Main: Suhrkamp.

Deleuze, G. (2011). Francis Bacon: Lógica da Sensação (José Miranda Justo, Trad.). Lisboa: Orfeu Negro.

Grunwald, M., de Groot, H., & Schaller, J. (2010). Environmental Impact of Photographic Processes. Journal of Cleaner Production, 18(1), 76–85.

Huizinga, J. (1949). Homo Ludens: A Study of the Play-Element in Culture. London: Routledge.

International Energy Agency (IEA). (2023). Data Centres and Energy Consumption. Paris: IEA.

Molder, M. F. (1999). O Absoluto que Pertence à Terra. Lisboa: Relógio d’Água.

Schmidt, L., Pereira, F., & Zhao, T. (2024). Energy Footprint of Generative AI Models. Nature Energy, 9(2), 134–142.

Schüll, N. D. (2012). Addiction by Design: Machine Gambling in Las Vegas. Princeton: Princeton University Press.