The Autonomous Machine

A painter turns to software

By his mid-thirties Harold Cohen was near the top of the British art world. He represented the United Kingdom at Documenta 3 in Kassel in 1964 and again, two years later, at the 33rd Venice Biennale in 1966. He was an established painter long before he touched a computer.

In 1968 he joined the Visual Arts Department at the University of California, San Diego, and got interested in computing and artificial intelligence. By the early 1970s he was writing software meant to paint on its own, not to help him paint. He named it AARON. Sometime in the early-to-mid 1970s he was invited to Stanford for a two-year Stanford AI Lab (SAIL) fellowship, where he could work seriously on whether a painter's decisions could be handed to a machine.

How AARON was built

Cohen described the program in 1979:

"AARON is a knowledge-based program, in which knowledge of image-making is represented in rule form. I have been my own source of specialized knowledge, and I have served also as my own knowledge-engineer." — Harold Cohen, "What is an Image?", 1979

The program held roughly three hundred small rules, or productions, in a strict hierarchy. At the top sat a level Cohen called ARTWORK, which made whole-drawing decisions: how to start, when to stop. Below it, levels for mapping and planning allocated space and developed individual figures, keeping them from colliding. Below those, levels named LINES, SECTORS, and CURVES generated the path of the pen.

At each level the program picked, pseudo-randomly, among the options its rules judged valid, and it built its own picture of the developing drawing as it went. Cohen wrote every rule, and the program made its choices within them.

What AARON was not

Cohen was explicit that AARON took no direction from a user:

"AARON is not an 'artists' tool.' I mean that it is not interactive, it is not designed to implement key decisions made by the user, and it does not do transformations upon input data." — Harold Cohen, "What is an Image?", 1979

"AARON is a complete and functionally independent entity, capable of generating autonomously an endless succession of different drawings. The program starts each drawing with a clean sheet of paper, no data, and generates everything it needs as it goes along." — Harold Cohen, "What is an Image?", 1979

How AARON drew a line

AARON never drew a straight line from A to B. Cohen explained the mechanism:

"SECTORS produces a series of 'imagined' partial destinations, signposts, as it were, whose function is to generate a series of movements of the pen which will make it veer towards, rather than actually to reach, the current signpost." — Harold Cohen, "What is an Image?", 1979

The level below SECTORS, called CURVES, moved the pen. Because it aimed at each signpost without arriving, AARON's marks look loose and freehand, where ordinary plotter output looks ruled.

The wobble slider in the tool controls this. At zero the pen heads straight for each imagined destination and the drawing looks mechanical. Raise it and the freehand quality returns.

Cognitive primitives

Before 1980, every version of AARON worked with only a few low-level, largely universal aspects of human cognition, as Cohen described them: figure against ground, closed form against open form, similarity, division, repetition. It had no object-specific knowledge of the external world. It did not know what a bird or a rock or a person looked like. Yet people looked at its drawings and, almost without exception, saw landscapes inhabited by creatures.

Cohen called the underlying idea "standing-for-ness": the minimum condition under which a set of marks functions as an image at all, before and independent of any specific meaning being intended. In his account of this era, a handful of cognitive primitives and no knowledge of the world were enough to produce marks that viewers read as living things, though no one had intended them to mean anything.

No memory, no learning

Some later accounts picture AARON remembering what worked and getting better. Cohen's own account of this era says it did not:

"It is not a learning program, has no archival memory, is quite simple and not particularly clever." — Harold Cohen, "What is an Image?", 1979

It began each drawing as if it had never made one. The tool works the same way. The same seed with the same preferences gives the same drawing every time, and a new seed gives an unrelated one.

The Turtle at Documenta 6

In the fall of 1976 Cohen built a small two-wheeled robot he called the Turtle, steered by sonar feedback rather than dead reckoning, to draw AARON's output on real paper. It made its public debut at Documenta 6 in Kassel (24 June to 2 October 1977), then traveled to the Stedelijk Museum in Amsterdam (25 November 1977 to 8 January 1978). It drew continuously and live, on a PDP-11/34, in full view of visitors. Cohen counted the audience for both shows:

"The DOCUMENTA 6 international exhibition in Kassel, Germany, and the prestigious Stedelijk Museum in Amsterdam, the two exhibits together running for almost five months and with a total audience of almost half a million museum-goers." — Harold Cohen, "What is an Image?", 1979

Visitors read the abstract forms as animals and landscapes, and once as San Francisco. The tool models this period: irregular blob-shaped creatures with internal detail, scattered across the sheet.

Color came later, and by hand

Cohen colored the black-and-white drawings of the 1970s himself, afterward. Frieder Nake, who was in Amsterdam, recorded that Cohen "wanted to, and did, colour them in by hand according to his own will and taste." AARON did not choose color on its own until the first painting machine of 1995, which mixed its own paints and applied them with brushes.

In the tool, the optional washes stand for Cohen's hand, and the control is labeled as his. Toggle hand-color off and on and the geometry stays put, because in this period AARON did not choose the color.

Rewriting AARON for color

AARON's code passed through FORTRAN and C before Cohen rewrote it in Lisp. He judged C "too inflexible, too inexpressive... to deal with something as conceptually complex as color."

How far the autonomy went

AARON's autonomy was compositional and procedural. It never learned and never improved itself. The rules were Cohen's, written by hand over years out of his own expertise as a painter, and the program never revised them. He resisted calling AARON creative in the human sense. By his account, if it ever signed a painting on its own initiative, that alone would settle whether it was truly intelligent.

Inside those limits, AARON made thousands of distinct, unrepeated decisions about closure, placement, and proportion, using a compact set of rules and the record of its own drawing in progress.

Cohen's one directly documented tie to this library is Frieder Nake, who watched the Turtle draw in Amsterdam (see the Nees and Nake lesson). Sol LeWitt and Vera Molnár were also moving aesthetic decisions into rule systems at about the same time, though no personal contact between them and Cohen is documented.