Harold Cohen’s AARON

What a prescient work of computer art tells us about AI

Joanne McNeil

Image Content Callouts

  • Cohen used the nose as a “device to establish the head’s orientation,” as he explained in a 1994 paper for Stanford Humanities Review. He would program AARON with rules to follow, body part by body part, including how each shape would be drawn on an array of coordinates.
  • The lush tomato red within the image set against a cherry red frame offers the sort of curiously alluring and non-intuitive color juxtaposition that is characteristic of AARON. In later iterations of the software, Cohen programmed AARON to create objects and assign specific hues to them, resulting in plants that are typically green in its images and human skin that is toned according to the range of skin colors. The lightness and saturation values are left up to chance. “To the degree that color is also an attribute of a face, there are a limited number of colors it can use. It would never decide to paint a face green because it doesn’t believe that faces can be green. However, there is no such limitation on the assignment of colors to things like sweaters or backgrounds,” Cohen said in an interview with The Computer Museum in 1995.
  • Notice how the woman in AARON KCAT has one hand—including five fingers—on her chest and another hand with four fingers at her hip, indicating that her thumb is there, but out of view. These are part of the library of shapes and relationships it has been programmed to produce. Cohen trained AARON to recognize that there are usually five fingers on a human hand, which are visible unless the person is posed in certain angles. In contrast, bizarre finger arrangements, including as many as a dozen digits per hand, are a signature oddity in images created with LLMs. This happens because applications like Dall-E are trained on an enormous corpus, encountering photographs in which people might have fingers obscured by an object or otherwise not in view. When you see a mutant hand in an AI art image, what you are looking at are the assumptions this program has made, generating forms without having specific rules to follow.
  • This is a screenshot from the screensaver version of AARON that Harold Cohen developed with Ray Kurzweil, the computer scientist known for his pioneering research in pattern recognition technologies and theory in what he called “the singularity”—a future point at which machine intelligence will surpass human talents. Kurzweil was such a fan of AARON that he kept a version of it running on display in his workplace lobby.
  • Cohen first intended for AARON’s output to look “indistinguishable” from the art of a human. But in his 1995 interview with the Computer History Museum, the artist explained how he came to appreciate its uncanniness. “I am much happier now when I see the program produce an image that looks as if it had been made by somebody who is seeing the world for the first time: seeing the world from a different point of view from someone who grew up human.”

Harold Cohen, AARON KCAT, 2001. Screenshot. Artificial intelligence software. Dimensions variable. Whitney Museum of American Art, New York; purchase, with funds from the Digital Art Committee 2023.20. © Harold Cohen Trust

In A Closer Look, a writer annotates a piece of art or an archival object.

The British artist Harold Cohen was a successful painter in the 1950s and 1960s, exhibiting widely across Europe at major exhibitions including the Venice Biennale and Documenta. Then, in 1968 he moved to California and started tinkering with computers. He had gone west to join the visual arts department at UC San Diego and while he was there he found inspiration in the campus computer lab. Cohen merged his training as a painter and amateur programming skills with a project he launched in 1973 called AARON, the work he is best known for today. Cohen’s aim, he once explained, was to create a “program that could function as an artist; an autonomous entity capable of generating original artworks.” This quirky and prescient work of computer art was recently on view at the Whitney Museum of American Art in New York.

Cohen often referred to the AARON software as his “surrogate self.” He programmed it to draw variations of lines and shapes with connected pen plotters, a machine that can guide pens across a paper surface according to instructions. AARON is not an acronym: Cohen picked the name, also his given Hebrew name, because he initially imagined a whole alphabet’s worth of computer projects. But it ended up being AARON that commanded his attention. The artist continued to tweak the project over decades, adding new forms for AARON to draw and methods for the program to autonomously select paint colors and control paint brushes.

This “surrogate self” to the painter was a highly prolific artist in its own right. Cohen sometimes left AARON running overnight and in the morning, he said, AARON would have made more than a hundred images for him to review.

The resulting images feature jagged lines and angular shapes, reminiscent, almost, of Egon Schiele, such as the one annotated here, AARON KCAT (2001). It comes from the screensaver version of AARON that Cohen produced in 2001 with the American transhumanist and computer scientist Ray Kurzweil. (KCAT stands for Kurzweil CyberArt Technologies.)

It’s tempting to describe AARON as a precursor to contemporary AI art, but if anything, it highlights where AI programs such as Midjourney and Dall-E fall short. AARON functions because of the meticulous research and detailed instructions it was programmed by Cohen to execute. To draw a human figure, as in this image, Cohen developed a set of rules for AARON, telling it how to create each part of the body, mapping out how the feet or head or torso should be drawn on an array of points that the plotter hits along an x and y axis. There is room for the machine to experiment in posing and coloring forms, but the fundamental shapes have been pre-programmmed. In contrast, LLM (large language model) applications scan an enormous corpus and use pattern recognition-based probability decisions to generate an image. You won’t see a person with nine fingers on each hand in an AARON image, as you might in an image made with Midjourney, because Cohen already programmed AARON to draw five of them. Cohen also doesn’t dismiss the consent of fellow artists. Unlike AI art programs which learn from images made by others, the only training data AARON processes is Cohen’s own work.

Cohen passed away in 2016, before LLMs began to enter broad public consciousness. At the Whitney, recreations of Cohen’s drawing machines were set up, for the public to watch AARON in action. He had once bragged that “I’ll be the first artist in history to have a posthumous exhibition of new work.” However less than humble Cohen was to forecast his own legacy, he is producing new work posthumously after all.

Originally published:
May 20, 2024


Louise Glück’s Late Style

The fabular turn in the poet’s last three books
Teju Cole

The Critic as Friend

The challenge of reading generously
Merve Emre

Rachel Cusk

The novelist on the “feminine non-state of non-being”
Merve Emre

You Might Also Like

Candy Darling’s “Newspaper” Centerfold

Encountering the infamous collage of a downtown legend
Lucy Sante

Ed Ruscha's "Every Building on the Sunset Strip"

Why the artists' book is a reminder of freedom
Jim Lewis

William Blake’s "Laocoön"

Why the poet's engraving reads like a protest poster
Anahid Nersessian


Sign up for The Yale Review newsletter and keep up with news, events, and more.