AI-Generated Aesthetic

In April 2022, OpenAI released DALL·E 2. In July, Midjourney opened its beta to public access through Discord. In August, Stability AI open-sourced Stable Diffusion. In a span of roughly four months, image generation by text-to-image neural network — a research curiosity for the previous five years — became a tool that anyone with an internet connection could use to produce, on demand, an arbitrary number of images of an arbitrary subject in an arbitrary style. The cultural impact of this transition has not yet been fully metabolized. What is already clear is that AI-generated imagery is the first major visual culture in human history that was not produced by humans drawing or photographing or designing, but by mathematical models trained on the entire historical archive of human visual output.

The technical details matter for understanding the resulting aesthetic. Modern text-to-image models work by learning, during training, a vast statistical model of the relationship between human-readable text descriptions and image patterns. When given a new text prompt, the model produces images that are, in some sense, the most probable image corresponding to that prompt according to its training distribution. This has several visual consequences. First, the models produce images that look like statistical averages of their training data — soft-edged, slightly plastic, often with a kind of dreamlike compositional logic that is not quite realism and not quite illustration. Second, the models excel at producing imagery in styles that are well-represented in their training data and struggle with styles that are not — a Midjourney request for "in the style of Caravaggio" produces something recognizably Caravaggesque, while "in the style of a contemporary Bengali street painter" produces something much less specific. Third, the models exhibit characteristic failures: extra fingers, melting backgrounds, illegible text, faces that don't quite cohere, perspective that drifts. These failures, which were severe in 2022 and have diminished in subsequent generations of models without disappearing, have become part of the recognizable aesthetic.

By late 2023, "Midjourney aesthetic" was a recognizable visual register. Highly saturated colors, an atmospheric haze over almost every scene, complex compositional layering, subjects rendered with a kind of plastic-doll smoothness that no human painter would produce, lighting that suggests cinematic post-production, and a dreamlike refusal of ordinary perspective. The aesthetic was so consistent across users that art directors learned to identify it instantly and, increasingly, to specifically request work that did not look AI-generated. By 2025, a counter-movement of designers had emerged explicitly working against AI-default styles, foregrounding handmade marks and visible human production decisions.

The relationship to existing visual movements is complex and important. AI-generated imagery has an obvious aesthetic kinship with Surrealism (entry: surrealism) — both are committed to dream-logic image compositions, juxtapositions of impossible scenes, hyperreal rendering of imagined subjects. A Midjourney landscape with two moons and a glass tree growing from a flooded plaza is, formally, a Magritte. The art critic Hito Steyerl has argued that AI image generation is, in some sense, the technological realization of what Surrealism aspired to be — the unconscious externalized into images. There is also a clear continuity with Vaporwave (entry: vaporwave) and broader internet-native aesthetics: AI image generation has absorbed those visual languages into its training distribution and produces them on demand. And there is a continuity with the long tradition of image-as-statistical-aggregation that runs through photography (especially the composite photography of Francis Galton in the 1880s), through Andy Warhol's silkscreen repetitions, through the early "deep dream" Google Vision experiments of 2015. AI-generated imagery is the latest, and currently most powerful, version of a tradition that has been developing for over a century.

Three honest critiques are worth naming, because the publication's premise is to take all movements seriously, including the ones we are inside. First: the training data problem. All major text-to-image models were trained on enormous datasets of human-produced images, largely scraped from the internet without consent, compensation, or attribution to the artists whose work formed the training set. The legal status of this practice remains contested in 2026, with ongoing lawsuits in multiple jurisdictions. Whatever the legal outcome, the ethical question is real: the visual culture these models produce is built on the uncompensated labor of millions of human artists, photographers, illustrators, and designers. To call this culture's products "AI art," without acknowledgment of the human work that made them possible, is at best naive and at worst a form of laundering.

Second: the homogenization problem. Because the models produce statistical averages of their training data, they tend to converge on a particular aesthetic — the "Midjourney style" — that bleeds into all output unless actively resisted. Visual culture risks becoming, at the margins, a fractal of the same model output, recombined endlessly. The early 2010s critique of Instagram filters — that they were producing a homogenized "Instagram aesthetic" across millions of photographs — was correct, but small in scale compared to what AI-generated imagery is doing across all of visual production.

Third: the displacement problem. AI image generation is, already, displacing real human work — stock photography, illustration, concept art, certain kinds of commercial photography, low-end editorial illustration. Whatever its aesthetic merits, the technology is reshaping the economic conditions of visual production in ways that are accelerating faster than design culture's ability to adapt.

But there are real possibilities, too, and a publication that pretends otherwise would be dishonest. AI image generation is, when used carefully, a tool that extends the reach of designers who already have visual judgment — letting them produce mockups in minutes, explore variations they would not otherwise have time to draw, communicate visual ideas across language barriers. It is also producing genuinely new image categories: the surreal, atmospheric, hyperreal compositions that have become identifiable in their own right, and that may, in time, be recognized as a distinct visual contribution rather than merely a technical artifact. The painter David Hockney, asked about AI image generation in 2023, said something like: "It is another tool. The question is what people will do with it." He was right. The medium itself does not determine the value of what is made with it; the human choices about what to ask, what to keep, what to throw away, and what to combine with other media continue to matter.

The longer-term consequences are unknowable in 2026. What is clear is that AI-generated imagery is the first visual movement on this map whose primary author is not a human. Every previous movement, from Impressionism through Vaporwave, was made by humans using tools. AI-generated imagery is made by models using prompts. The boundary between "made by a person" and "made by a machine that learned from people" is, at the present moment, the central question of contemporary visual culture, and it is not yet clear what the right answer is. This essay, in 2026, can only describe the moment and say honestly that the moment is not yet over.