The Ghost in the Machine: The Surprising History of Early AI Pictures

Today, we take it for granted that typing a whimsical sentence into a computer will yield a hyper-realistic painting or a glossy digital render in mere seconds. The internet is flooded with slick avatars, dreamscapes, and synthetic photography. Yet, this high-tech Renaissance did not materialize out of thin air.

Long before text-to-image models sparked global debates about art and copyright, generations of engineers, mathematicians, and visionary artists were teaching computers to draw. Looking back at “early AI pictures” reveals a fascinating history of mechanical logic, abstract lines, and the persistent human desire to see machines create something entirely new.

1. The Rule-Bound Pioneers: Harold Cohen and AARON

When people think of artificial intelligence, they usually picture deep neural networks processing massive troves of data. But the earliest ancestor of AI art used neither data nor neural networks. It used strict, handcrafted logic.

In 1973, British painter Harold Cohen—then a celebrated artist who had represented his country at the Venice Biennale—moved to California and began working on a computer program called AARON. AARON was not designed to digitize or copy Cohen’s existing style; it was built to create.

Running on mainframe computers of the era, AARON used a system of symbolic rules about geometry, posture, and composition to drive a mechanical drawing machine (a plotter) that held a physical pen.

• The Process: AARON drew abstract patterns and organic shapes on paper using rules programmed by Cohen.
• The Twist: Cohen soon discovered that AARON would produce combinations and forms that he hadn’t explicitly anticipated, effectively surprising its own creator.

AARON evolved over four decades, eventually graduating from simple abstract line work to painting organic figures and botanicals. Long before modern algorithms, Cohen had already raised the central philosophical question of AI art: Who is the actual artist—the human programmer or the machine executing the system?

2. The Awkward Teenage Years: Pixels, Noise, and Neural Nets

For decades following AARON’s inception, computer-generated imagery remained largely confined to academic laboratories. Progress was bottlenecked by processing power and a lack of flexible learning frameworks.

By the late 20th and early 21st centuries, researchers began experimenting with statistical models—like Markov Random Fields and Boltzmann Machines—to approximate textures and patterns. These systems could generate noisy textures or rudimentary checkerboards, but they struggled with coherence. A computer could tell you if a photo contained a cat, but asking it to draw a cat resulted in a blurry, unrecognizable smudge.

The paradigm shifted dramatically in 2014 with the invention of GANs (Generative Adversarial Networks).

Popularized by researcher Ian Goodfellow, a GAN pitted two neural networks against each other in a digital game of cat-and-mouse:

1. The Generator tried to create fake images from random noise.
2. The Discriminator evaluated the images, trying to spot the fakes.

Through thousands of rounds of competition, the generator grew ruthlessly good at faking reality. This era produced the first truly surreal early AI artifacts: uncanny, warped human faces, nightmarish hybrid animals, and surrealist landscapes that felt like dreams remembered halfway. Google’s DeepDream (2015) also captured the public imagination, feeding images back into neural networks to produce psychedelic, dog-eyed digital hallucinations that looked like machine acid trips.

3. The Leap to Language: When AI Learned to Read

Despite the visual flair of GANs, creating specific images still required heavy technical programming or specialized machine-learning pipelines. A regular person couldn’t just type “an astronaut riding a horse in oil paint” and watch it appear.

The breakthrough came around 2021 with models like CLIP (Contrastive Language-Image Pre-training), which bridged the gap between human language and visual concepts. CLIP taught computers how text and images relate to one another.

By 2022, the dam broke. A wave of platforms—including DALL-E 2, Midjourney, and Stable Diffusion—hit the public sphere. Utilizing advanced diffusion models (which generate images by learning to reverse a process of gradual noise reduction), these tools transformed text prompts into jaw-dropping artwork in seconds.

What once required a mainframe, a robotic plotter, or a team of computer scientists could now be summoned by anyone with an internet connection and a vivid imagination.

Echoes of the Past

Looking back at the primitive black-and-white plots of AARON or the glitchy, distorted portraits of early GANs, it is easy to view them as obsolete stepping stones. Yet, they possess a distinct raw charm. They remind us that the journey of AI art has never just been about achieving photorealism; it has been an ongoing exploration of how machines perceive form, structure, and human expression.