Skip to content

Harold Cohen

Abstract

Harold Cohen (1928–2016) was a British painter with a Venice Biennale behind him when, at forty, he moved to California, learned to program, and spent the remaining forty-eight years of his life writing one program. AARON, begun in 1972, drew on its own: first abstract closed forms, then rocks and plants, then human figures in rooms, then colour, then abstraction again, on plotters, a turtle robot and finally a painting machine Cohen built. It was the first program to make art autonomously, and the Whitney Museum showed it in 2024 as the earliest artificial-intelligence art program. It never learned anything. Every capability it had, Cohen had coded by hand, which is why he insisted to the end that AARON was not creative, and why it remains the clearest measure of what rule-based AI could and could not do.

The Painter

Harold Cohen was born in London on May 1, 1928, and studied at the Slade School of Fine Art, where he later taught. By the mid-1960s he was one of the better-known abstract painters in Britain: he represented the country at the Venice Biennale in 1966, with his brother Bernard Cohen among others, and showed at Documenta 4 in Kassel in 1968. That year he went to the University of California, San Diego, as a visiting lecturer for a term, and stayed for the rest of his life.

At San Diego a graduate student, Jef Raskin, later the originator of the Macintosh project at Apple, taught him to program. Cohen’s account was that he had reached the limit of what he could learn about pictures by making them, and wanted to know what he knew: if he could write down the rules by which a set of marks became an image, the rules would be a theory of image-making. The program would be the experiment.

AARON

He began AARON in 1972 and spent 1973 to 1975 as a guest at the Stanford Artificial Intelligence Laboratory, where Edward Feigenbaum’s group was building the first expert systems and Cohen absorbed their method: capture the knowledge of an expert as explicit rules and let a program apply them (see Edward Feigenbaum and the Expert System). The expert was Cohen. AARON’s first decade produced abstract drawings, closed and open forms distributed across a sheet by rules about proximity and balance, drawn by a plotter or, from 1979, by a small wheeled “turtle” that drew on paper laid on the gallery floor at the San Francisco Museum of Modern Art and, in 1983, at the Tate in London, where visitors watched the machine work. The question Cohen set himself was what the minimum conditions were under which marks read as an image, and the drawings answered it: people saw things in them.

In the 1980s he gave AARON knowledge of things. Rocks and plants first, described in the program as structures that grew according to rules rather than as stored shapes, so that no two plants were alike; then human figures, which required a model of how a body is articulated and how it occludes what is behind it; then rooms with people and plants in them. The program was written in C and, from the early 1990s, in Lisp, which Cohen said was what made colour possible: AARON’s colour rules from the 1990s produced interior scenes that it painted itself, through a painting machine Cohen designed that mixed dyes and applied them with brushes to canvas. In 2001 Ray Kurzweil’s CyberArt company released an AARON screensaver that drew a new picture on any PC. In its last decade AARON returned to abstract forms in colour, and Cohen returned to the role he had started in, printing the program’s output large and finishing it by hand.

Not Creative

Cohen was firm about what AARON was. It could produce an unlimited number of distinct pictures in a style anyone could recognise, and it produced them without him in the room, but it could not draw anything he had not taught it to draw or change its style except when he rewrote it; it was, he said, autonomous in execution and not in intention. That distinction became the standard one in the debate about machine creativity. Compared with the neural image generators that arrived after his death, which learn their rules from millions of other people’s pictures and can explain none of them, AARON is the symbolic case, every decision traceable to a line Cohen wrote; the comparison is drawn in Computer Art and Generative Art.

He was professor and chair of the visual arts department at San Diego, directed the university’s Center for Research in Computing and the Arts from 1992 to 1998, and received the ACM SIGGRAPH Distinguished Artist Award for lifetime achievement in 2014. His partner was the Japanese poet Hiromi Itō. Harold Cohen died in Encinitas, California, on April 27, 2016, at 87, with AARON still running.

📚 Sources