Before the AI Prompt: The Artists Who Taught Computers to Draw

Before somebody could ask a computer for an image in a sentence, an artist had to decide what instructions might make a picture possible at all.

That earlier challenge produced some remarkable answers. A drawing could become a set of geometric conditions. A composition could emerge from programmed variations. An artist’s knowledge could be turned into a system that operated a pen or a brush.

These results came from different technologies. Interactive drawing, art made through programmed rules and AI each have their own history. Calling them all machines becoming creative hides the interesting part: people kept changing how instructions, images and artists worked together.

For a print or design business, that history is more useful than another promise that AI will transform everything. It helps explain what assistance can do, what it cannot establish and why a convincing image is not automatically a finished design.

First, not every computer-made image is AI

A computer can draw a line because a person specifies its coordinates. It can repeat a shape according to a rule. It can let a user drag an object or apply a mathematical condition. None of those activities alone establishes that a machine-learning model is involved.

Generative art is also broader than contemporary generative AI. A system can generate variations through programmed instructions without learning from a large collection of examples. The word “generative” describes something being produced; it does not identify one universal mechanism.

These distinctions matter because otherwise the history becomes a misleading countdown to today’s software. Earlier work was not simply a crude version of a modern image generator. Artists and researchers were investigating different questions.

Some wanted precise control. Others wanted to explore combinations. Others wanted to express artistic knowledge in a form a machine could execute. The computer’s role changed with the question.

Keep those differences in mind as the story moves from a light pen to a drawing machine and eventually to learned image synthesis. The shared feature is not identical intelligence. It is an image-making process in which human decisions interact with computation.

1963: a drawing could become a conversation

Ivan Sutherland’s Sketchpad dissertation, submitted to MIT in January 1963, described a system in which people could draw directly on a computer display and apply conditions to the drawing. The Cambridge-hosted edition includes the original work with a later preface.

A user could, for example, indicate that two lines should be parallel. The system could adjust parts of the drawing to meet the specified conditions. The dissertation also described repetitive drawings and changes to existing work.

This was not a text prompt asking for an attractive scene. The person was communicating through the drawing and its relationships.

Our interpretation is that the important shift was from making individual marks to managing a structure. A line was not only a visible stroke. It could have a relationship with another line, and that relationship could remain meaningful when the drawing changed.

That distinction is familiar in design today. A layout with deliberate relationships is easier to revise than a collection of elements positioned by accident. If a heading grows, the rest of the page needs to respond sensibly rather than fall apart.

Sketchpad does not prove that every later design application follows directly from one dissertation. It gives us a documented example of a computer helping a person maintain conditions within a visual problem.

A system can obey the wrong brief perfectly

The idea of constraints is powerful, but it also exposes a limitation. A machine can satisfy the conditions it is given while failing a condition nobody thought to express.

Imagine a hypothetical poster with beautifully aligned text, consistent margins and balanced spacing. If the event date is wrong, the layout remains a bad answer. If the contact details are too small for the viewing distance, mathematical order has not solved the communication problem.

These are illustrative examples, not results from Sutherland’s experiments. They show why technical correctness and useful design should not be treated as synonyms.

The person preparing the brief still needs to understand what the work is for. Which information must be noticed first? Where will the piece be seen? What should somebody do afterwards?

A rule can preserve a decision. It cannot establish that the decision was appropriate in the first place.

The computer lab became an unlikely studio

The V&A’s history of digital art describes the practical difficulty facing artists in the 1960s: computers were largely confined to university and corporate laboratories. Artists interested in using them often needed relationships with programmers and mathematicians.

The museum identifies Vera Molnár and Manfred Mohr among the pioneers working with computing facilities in France. Its account dates Mohr’s first programmed generative drawings to 1969 and describes plotters as important output devices for early computational work.

A plotter turns instructions into physical marks by moving a drawing instrument across a surface. The machine is not simply an image displayed on a screen; it has a material relationship with paper and ink.

Our reading is that this makes early computer art particularly interesting to a print business. The result has to cross a boundary: the calculated instruction must become something a physical device can produce.

That boundary can shape the work. A line-making device encourages different possibilities from a brush, a photographic process or a raster display. The available output is not a neutral afterthought.

This does not mean artists were merely accepting whatever a machine happened to do. Choosing a system, defining a procedure and deciding which result mattered were creative decisions in their own right.

Vera Molnár imagined a machine before she could use one

Molnár’s story reverses the usual account of technology creating an idea. The V&A describes her pre-computer “machine imaginaire” method: making abstract geometric images by following predetermined compositional rules. In 1968, she gained access to a research laboratory at the Sorbonne and taught herself FORTRAN.

The procedure came before access to the machine. She could explore a structured approach to making images before using a computer to carry it out.

Our interpretation is that this challenges a familiar temptation: buying the tool before knowing what question to ask it. Molnár’s example suggests that the intellectual structure can be the starting point.

For a design project, that might mean defining the variation worth exploring. Should the hierarchy change? Should a repeated element vary slightly? What happens if an arrangement becomes more regular—or deliberately less regular?

Those questions are more useful than generating a large number of unrelated images and hoping one supplies the missing brief.

The machine can make a procedure executable. The reason for investigating that procedure belongs to the work around it.

Variation is not the same as selection

Programmed image-making brings an obvious possibility: a system can produce alternatives. But an alternative is not automatically an improvement.

A set of variations may reveal something the designer would not have considered. It may also contain results that are repetitive, unsuitable or visually interesting but irrelevant to the job.

The important next step is selection. What makes one version more useful than another? Is it clearer? Does it preserve the identity? Does it work at the required size? Does it answer the actual customer need?

These questions should not be postponed until a favourite image has been chosen. Without criteria, the selection can become a contest of immediate attraction.

The same distinction matters in assisted design today. Producing options is one task; deciding which deserve more work is another. More options do not always mean better choices.

Harold Cohen asked a different question

Harold Cohen had an established practice as a painter before his long collaboration with AARON. The Whitney Museum’s account says he conceived the software in the late 1960s at the University of California, San Diego, and named it in the early 1970s.

The museum describes Cohen’s interest in translating an artist’s knowledge and process into code. Over time, AARON produced work for drawing and painting devices as well as screens and projections. Cohen also built plotters and painting machines to execute its output.

This was more ambitious than repeating a specified shape. It raised questions about what knowledge a system might need in order to construct an image, and how the artist would work with that system.

Our interpretation is that Cohen’s project relocates some creative work. Instead of deciding only what a single finished picture should contain, the artist can work on the process that produces pictures.

That is not an absence of authorship. Nor does it settle every philosophical question about a machine’s contribution. It makes the relationship more complicated than a simple contest between a human hand and an automated one.

AARON is valuable historical context precisely because it should not be casually equated with every current image tool. The surrounding aims and methods need to be understood, not replaced with a familiar label.

The pen still needed somewhere to go

The physical output deserves attention. A system able to produce an image description still needs a way to make the image visible. A plotter, painting device, screen and printed page each introduce different conditions.

That remains true when a contemporary design moves from software to a sign, leaflet or garment. The final medium has dimensions, viewing conditions and production requirements.

A visually persuasive mock-up can disguise this transition. It can show an object in flattering light without establishing the material, print area or finishing method. It can look like a completed customer job while representing only a proposed concept.

The useful discipline is to identify what the image proves. A concept demonstrates an intended appearance. A checked production file establishes something different. A photograph of a completed item, accurately described, supplies a different kind of evidence again.

Confusing those categories makes it easier to promise a result that has not been verified.

1985: the page became a desktop project

Computer-assisted design did not develop only through experimental art. The Computer History Museum’s account of desktop publishing describes Aldus PageMaker, introduced in 1985, complementing Apple’s LaserWriter printer. Its history also discusses Adobe’s PostScript and the expanding role of design applications.

This was another relationship between person and computer. The user could assemble and revise a page rather than write an art-generating program.

Our practical reading is that easier assembly changes who can participate, but it does not abolish the need to communicate well. A person can gain access to layout controls without immediately knowing which information belongs first.

The same applies to a start-up making its first promotional material. Possessing the software is not the same as having a clear offer, suitable artwork or a reliable handover to production.

The history is therefore not one continuous movement towards less human involvement. Different tools allowed people to perform different parts of the job.

Learned image synthesis is another development, not a new name for everything

More recent research introduced image-making methods based on learned models. In their 2014 paper on generative adversarial networks, Ian Goodfellow and colleagues proposed training a generative model alongside a discriminative model through an adversarial process.

Jonathan Ho, Ajay Jain and Pieter Abbeel’s 2020 paper on denoising diffusion probabilistic models presented another approach to image synthesis. These papers concern specific research methods; neither is the beginning of all generative art.

The distinction matters for the story. A contemporary model learning patterns from data is not simply the same process as an artist explicitly programming every geometric relationship in a plotter drawing.

This article does not claim that every current commercial tool uses one of these methods, or that the papers establish the capabilities of a particular product today. Individual systems, training arrangements and interfaces differ.

For a non-specialist, the useful point is that similar-looking outputs can come from different processes. Understanding the process helps identify what needs checking.

A convincing result can still be wrong

The US National Institute of Standards and Technology’s 2024 generative AI risk profile discusses confabulation: systems generating and confidently presenting erroneous or false content. It also addresses risks surrounding information integrity and intellectual property.

That guidance is not a finding that every generated image is false. It is a reason not to treat apparent confidence or visual polish as verification.

For a design project, check the parts that matter to the actual job: names, dates, contact details, proportions, logos and the relationship between the illustrated item and the product being offered.

A plausible image cannot confirm that a business owns the depicted equipment. An invented scene cannot establish that a customer received the shown product. A neatly presented source cannot be assumed to exist merely because it looks credible.

The question is not whether the output is impressive. It is what information the audience could reasonably take from it, and whether that information is true.

Copyright is not settled by the word “generated”

The Intellectual Property Office of Ireland’s copyright guidance explains protection for categories including artistic works and owners’ control over their use, subject to exceptions.

That basic principle should not be turned into a sweeping answer about every AI workflow. A project may involve supplied photographs, brand marks, templates or other protected material as well as generated elements.

Check the relevant assets, permissions and current service conditions for the proposed use. Do not assume that paying for a tool makes every possible output or input legally uncomplicated.

This article offers no legal clearance for a particular image and no guarantee that generated work is copyright-free, exclusively owned or suitable for every commercial purpose. Where the answer affects a real project, obtain appropriate advice rather than relying on a general history article.

What a small design business can take from the history

The strongest lesson is not that a business should automate everything. It is to identify which decision the tool is helping with.

Sketchpad provides an example of expressing relationships. Molnár’s approach shows the value of defining a procedure before gaining access to its execution. Cohen’s collaboration with AARON asks how artistic knowledge might become part of a system. Desktop publishing connects editable layouts to printed output.

Each suggests a useful discipline: make the brief explicit, distinguish generation from selection and inspect the handover into the final medium.

A hypothetical start-up preparing a leaflet can apply those disciplines without pretending to be a computer-art laboratory. It can establish the main customer question, explore a few purposeful alternatives, verify the words and prepare the selected design for its intended output.

That is a proposed workflow, not a measured productivity claim. The useful result depends on the job, the people, the tool and the checking required.

If you are preparing artwork for a physical project, see our business printing service. Tell us the intended item, quantity, deadline and files available. Call 062 61524 if you are unsure what should be checked before production.

The machine changed; the question remained

There is a striking continuity across this history. The interfaces became different, access changed and the methods of making images expanded. Yet the designer still needed to ask what the work should accomplish.

A precise drawing can answer the wrong brief. A rich collection of variations can lack a useful choice. A persuasive image can conceal an unverified claim. A beautiful screen preview can fail when it becomes a physical object.

Computers did not begin helping people make images when the text prompt arrived. Earlier artists and researchers were already investigating control, variation, knowledge and collaboration.

Their work gives today’s discussion something it often lacks: perspective. The interesting question is not simply whether a machine can make a picture. It is which parts of the picture-making process have changed, and who remains responsible for making the result useful.

Sources and scope

The historical account draws on Sutherland’s dissertation, museum histories from the V&A, Whitney and Computer History Museum, and the authors’ research papers on generative adversarial networks and diffusion models. Contemporary risk guidance comes from NIST; general copyright context comes from the Intellectual Property Office of Ireland. Interpretations about briefs, selection, production and customer communication are Time for Designs’ analysis. This is not a complete history of computer art, a current software comparison or a guarantee of productivity, legal ownership or production suitability.