When AI Generates an Image, Where Does the Artist Create?

A Review of Min Jin Kwon’s “A Study on the Artist’s Intention and Structural Choice in AI Image Generation”

Since the emergence of generative AI, debates about art have often gravitated toward questions such as “Can AI be an artist?” or “Can entering a prompt really be considered a creative act?” Min Jin Kwon’s A Study on the Artist’s Intention and Structural Choice in AI Image Generation shifts the starting point of this discussion. Rather than first asking whether AI can create in the same sense as a human, the study looks back at a more fundamental question: How has human image-making actually been structured throughout history? By tracing photography, silkscreen, conceptual art, collective production, and algorithmic art, the author identifies a recurring structure in the production of images: condition-setting, mediation, and selection.

One of the most interesting aspects of the paper is that it does not treat AI as an exception within art history. Even before AI, artists did not always make every part of an artwork directly with their own hands. Cameras transformed light into images, silkscreen plates and inks transferred originals, Sol LeWitt’s works were executed by others according to written instructions, and assistants physically produced works in the studios of Jeff Koons and Takashi Murakami. In the algorithmic works of Harold Cohen and Vera Molnár, programs generated images. When these practices are placed side by side, a structure more persistent than the question of whether the artist “made it by hand” becomes visible.

Seeing Creation as a Chain of Judgments Rather Than Mere Production

In the condition-setting–mediation–selection structure proposed by the author, the first stage involves defining the range of possibilities within which an image can emerge. A photographer determines shutter speed, aperture, and ISO. A silkscreen artist selects the source image, colors, and stencil. A conceptual artist formulates linguistic instructions, while an algorithmic artist defines rules and parameters. Although their methods of execution differ, each artist first organizes the conditions under which a result can appear.

Next comes mediation, a process that is never entirely under the artist’s control. In photography, the camera and photographic system serve as mediators. In silkscreen, the plate and ink do so. In conceptual art, another person executing the instructions becomes the mediator; in collective production, assistants perform this role; and in algorithmic art, the program does. A system positioned between the artist and the result has therefore been part of artistic production for a long time.

Finally, there is selection. The photographer chooses one image from many exposures. The silkscreen artist fixes a particular result among variations in color and arrangement. The artist approves or modifies a work produced by others, while the algorithmic artist selects among computational outputs. The paper emphasizes that this selection is not merely an expression of preference. It involves judgments about relations across the entire image—balance, tension, density, center, and periphery.

This is where Rudolf Arnheim and John Dewey become important. Arnheim understands visual perception not as passive sensory reception but as an act of thought that organizes formal relations. When an artist looks at an image and judges that “this works” or “something is excessive,” the act of seeing itself functions as visual thinking. Dewey, by contrast, understands creation as a process in which action and response are repeated until an experience reaches consummation. The artist observes a result, modifies the conditions, examines another result, and continues adjusting until reaching the point where no further revision feels necessary.

By bringing these two theories together, the paper understands creation not as a single act of production but as a temporal sequence of judgment → execution → observation of the result → renewed judgment. From this perspective, AI image generation begins to look less unfamiliar than it first appears.

From Photography to Sol LeWitt, the Artist’s Hand Had Already Been Moving Away

Conceptual art and collective production are especially important examples in the paper. In Sol LeWitt’s work, what determines the form of the artwork is not the artist’s manual gesture but a set of linguistic instructions. Directions concerning lines and their arrangements are written down and then executed on a wall by someone else. Physical execution is transferred to another person, while the decisions concerning the structure of the work and the judgment of its final result remain with the artist.

The studio practices of Jeff Koons and Takashi Murakami make this separation even clearer. Even when an artist does not physically produce every part of the work, the author argues that the central creative structure remains intact as long as the artist determines elements such as color, form, scale, and overall direction, then reviews the result and decides whether to modify or approve it. Here, the artist gradually shifts from being “the person who makes” toward being the person who organizes and judges the process of generation.

Extending this historical trajectory to algorithmic art makes its connection with AI even clearer. Harold Cohen’s AARON generated images according to predefined rules, while Vera Molnár explored visual changes by repeatedly modifying mathematical rules and parameters. Cohen tended to establish relatively stable rules and then choose among outputs, whereas Molnár altered the rules themselves as part of the exploration. The distinction is directly recognizable in contemporary AI image-making. An artist can hold a prompt relatively constant and generate multiple outputs, or modify the prompt in response to each result and continually explore the generative space itself.

What emerged with AI, then, is not the first form of art in which the artist does not directly make the final object. Such structures already existed in photography, conceptual art, studio production, and algorithmic art. What changes with AI is that the relationship between the stages becomes far looser and less predictable.

What Distinguishes AI Is Not Mediation Itself, but the Opacity of Mediation

The paper does not merely argue for continuity between AI and earlier media. Its more consequential question is where the structure begins to change despite that continuity.

In photography, opening the aperture produces a relatively predictable change in depth of field, while slowing the shutter speed increases motion blur. There is a comparatively intelligible causal relationship between condition and result. Something similar applies to silkscreen: the chosen colors and plates are directly related to the resulting image.

With diffusion models, however, the same prompt can generate different images each time. A prompt provides direction but does not directly determine the result. Another layer also intervenes: the training data. Before an artist establishes any conditions, the model has already formed a latent space through its training data, and the user cannot directly inspect or control the internal organization of that space. The paper sees this as a decisive difference between AI and earlier media.

Seen from another position, an interesting issue emerges. Artistic control has not disappeared; the location at which control operates has shifted. With a camera, manipulating a condition tends to produce relatively predictable consequences. In generative AI, even carefully specified conditions cannot fully determine the output. As a result, artistic judgment moves away from being concentrated before generation and becomes increasingly important after generation, through comparison, modification, and selection.

For this reason, it is difficult to say that the artist’s role simply increases in proportion to the precision of the prompt. Actual practice involves repeated cycles of prompting, generation, comparison, selection, revision, and regeneration. The artist’s intention does not necessarily exist in complete form at the beginning. It becomes more specific as it encounters and responds to generated results.

In Midjourney, Intention Continues to Form After the Image Appears

The paper demonstrates this through an actual Midjourney generation process. The researcher enters conditions such as “Korean traditional ceramics, blue-and-white porcelain, moonlight, minimalism, high resolution” and receives four images. Although they share the same prompt, the images differ in composition, color, texture, and direction of light. The user compares them by evaluating negative space, visual density, and compositional balance.

If none is satisfactory, the prompt is revised with conditions such as “moonlight emphasis,” “delicate glaze texture,” or “increased negative space.” New images are then generated and judged again. Through repetition, demands that could not initially be articulated with precision gradually become more concrete.

The important point here is that artistic intention is not treated as a fixed starting point. The artist is not someone who begins with a perfectly complete image in mind and then instructs AI to reproduce it. Instead, the artist discovers more precisely what was wanted by responding to what the system produces, and then returns that judgment to the system as a new condition. Intention does not exist only before input. It also takes shape through the time spent observing and judging results.

This is precisely where Dewey’s theory of experience becomes useful. Generation and selection do not form a one-way relationship of command and execution. They constitute a cycle of action and response. A prompt shapes an output, but the output also reshapes the next prompt. Over time, conditions and selections affect one another until the artist eventually decides that further modification is unnecessary and the work reaches completion.

The Table Reveals a Shift in the Artist’s Role Across Media

Read horizontally, the table reveals differences among media. Read vertically, however, another trajectory appears. Moving from photography toward AI, the artist gradually becomes more distant from direct physical production. At the same time, the role of designing conditions and judging results becomes increasingly explicit. The artist’s role does not simply diminish; the point at which artistic intervention occurs changes.

This is one of the most important shifts in perspective proposed by the paper. If the question is “How can a human be the artist if AI made the image?”, the debate remains centered on who physically produced the result. But if the question becomes “Who organizes the generative conditions, judges the results, and determines when the process should stop?”, creative agency can be reconsidered from a different position.

In the Age of AI, the Artist Designs the Generative Process

The paper’s conclusion can be summarized in three propositions. First, from photography to AI, image-making repeatedly exhibits a structure of condition-setting, mediation, and selection. Second, the center of artistic creation need not lie in direct physical production but can reside in determining the direction of form and judging results. Third, AI inherits this historical structure while making the relationship between conditions and results considerably more complex and fluid because of probabilistic generation, the prior intervention of training data, and the invisibility of latent space.

Following this argument, the transformation brought by generative AI cannot be adequately described simply as the “loss of the hand.” The artist may have moved away from directly determining every visible form, but still determines the direction in which generation proceeds, interprets unexpected outputs, modifies conditions, and eventually fixes one result among many possibilities.

What therefore becomes increasingly important in AI-era artistic practice is not merely the ability to formulate precise instructions. It is the ability to perceive differences among generated results, judge how those differences relate to one’s intentions, and translate those judgments back into the next set of generative conditions. As the range of possibilities produced by generative AI expands, the work left to the artist does not necessarily decrease. It shifts to another position: away from the center of physical production and toward the orchestration of the generative process as a whole.