The Painter Who Taught a Machine to Make Art

The Painter Who Taught a Machine to Make Art

In 1968, Harold Cohen stood before a room-sized computer at the University of California, San Diego, and posed a question that seemed absurd to the engineers around him. A respected British painter with works in major galleries, Cohen had not come to calculate trajectories or process data. He wanted to know whether a machine could make the same kinds of decisions he made every day in his studio: where to place a line, how to balance a composition, when a color felt right. The engineers gave him time on the machine. Cohen, who had never written a line of code, began teaching himself to program.

The Painter's Question

Cohen's fascination was not with the computer as a tool for rendering — plotters had been drawing precise lines for years — but with the computer as a potential agent. He later described being "fascinated by the question of what constitutes an image and how artists make decisions." That question had driven his own painting practice, a successful career built on representational work shown at the Venice Biennale and Documenta. But the more he painted, the more he wondered whether the intuitive choices he made — the slight shift of a figure, the weight of a contour — could be externalized, examined, and perhaps delegated.

He began his experiment in the early 1970s, naming the program AARON. The name was deliberate: it started with 'A' so it would appear first in file directories, and it carried no gender or cultural baggage. From the start, Cohen insisted AARON was not a random generator. It was, he said, "designed to understand and apply rules of drawing, composition, and even color." He would spend the next half-century refining those rules.

Teaching a Machine to See

The first version of AARON knew nothing of the world. It could only produce abstract marks — clusters of lines, enclosed shapes — guided by procedural rules Cohen encoded. He treated the program like a student, feeding it the logic of representation piece by piece. Over years, he taught it to distinguish figure from ground, to suggest volume through closure, to distribute visual weight across a page. He gave it a vocabulary of body parts: a head required certain proportions, limbs had plausible ranges of motion, figures needed to occupy space without intersecting impossibly.

Each advancement required Cohen to articulate, in algorithmic terms, what he had once done by instinct. He later recalled that the process forced him to "make explicit what had been tacit." When AARON produced a drawing that felt wrong — a figure floating unmoored, a composition collapsing to one side — Cohen did not edit the image. He rewrote the rule that had permitted the error. The program grew to contain thousands of such rules, a codified aesthetics built line by line.

A Collaboration of Unequal Partners

By the 1980s, AARON was producing drawings of startling coherence: figures in interior spaces, botanical forms, crowded scenes with multiple interacting bodies. The program chose compositions, assigned colors, and varied its own output — no two runs were identical. Cohen would watch the plotter arm move across paper, sometimes for hours, as AARON executed a drawing the artist himself had never imagined.

But the collaboration did not end at the plotter. Cohen frequently took AARON's output back into his studio and painted over it. He would add washes of color, reinforce contours, adjust tones — decisions the program could not yet make. The resulting works were hybrids: machine logic beneath human gesture. Exhibited in galleries and museums, they unsettled viewers who tried to locate the "artist." Was it the man who wrote the code? The program that chose the composition? The hand that held the brush?

The Authorship Crisis

Cohen welcomed the confusion. He had built AARON precisely to provoke "profound philosophical discussions" about artistic genius and authorship. In lectures and writings, he argued that if a program could make aesthetic judgments indistinguishable from a human's, the Romantic notion of the solitary creator — the genius touched by inspiration — needed rethinking. He pointed out that he had not "drawn" the figures in AARON's work any more than a composer "plays" every note of a symphony. The score, in this case, was algorithmic.

Critics were divided. Some dismissed AARON's output as mere simulation, a parlor trick lacking the "soul" of human creation. Others recognized that Cohen had externalized a creative process so completely that the boundary between tool and collaborator had dissolved. The debate anticipated, by decades, the controversies that would erupt when neural networks began generating images from text prompts. Cohen had already lived through the core questions: Who decides? Who deserves credit? What remains of the artist when the machine chooses?

Five Decades of Evolution

AARON never reached a final form. Cohen rewrote it repeatedly, porting it from mainframes to workstations to personal computers, each migration an opportunity to restructure its knowledge. In the 1990s, he gave it color — not as a fill tool, but as a compositional system that balanced hue, saturation, and value across a scene. In the 2000s, he taught it to represent plants with botanical plausibility, each leaf placed according to growth logic rather than decorative whim. Near the end of his life, he was still adjusting rules, still watching the plotter, still painting over its lines.

When Cohen died in 2016, AARON had been running, in one version or another, for nearly fifty years. No other generative art system has been maintained and developed by a single artist for so long. The program exists today as a historical artifact — Cohen's estate preserves the code — but its real legacy is the framework it established. Long before "prompt engineering" entered the vocabulary, Cohen demonstrated that directing an autonomous creative system is itself a creative act, one requiring deep artistic knowledge, relentless critical judgment, and a willingness to surrender control.

The Mirror AARON Held Up

Today, as AI image generators produce millions of pictures daily, Cohen's experiment reads like a controlled study conducted half a century ahead of its time. He proved that aesthetic decision-making can be decomposed, encoded, and executed without human presence in the moment of creation. He also proved that the human does not vanish — Cohen's hand is visible in every rule AARON followed, every painting he layered atop a machine drawing.

The plotter in Cohen's studio is silent now. But the question he brought to that computer lab in 1968 remains open, more urgent than ever: when the machine chooses, what is left for the artist? Cohen's answer was to keep choosing — to keep teaching, refining, painting, and asking. He showed that the relationship between human and algorithm need not be replacement. It can be a conversation, conducted in code and pigment, that lasts a lifetime.

This is one episode in a much longer story. For the full account of digital art and traditional techniques, read “Pixels and Paints” by Benjamin Ryan on MixCache.com.