Ghosts in the Machine: The Strange, Beautiful Evolution of Early AI Pictures

Long before prompt engineering became a modern tech buzzword and hyper-realistic synthetic portraits flooded our social feeds, artificial intelligence was already trying to draw.

Today’s text-to-image powerhouses can render photorealistic landscapes, intricate fantasy characters, and flawless typography in mere seconds. Yet, looking back at the ancestral roots of AI-generated imagery reveals a fascinating history. It is a story not of instant perfection, but of clumsy algorithms, robotic arms, psychedelic software glitches, and a gradual, uncanny journey toward machine imagination.

The Pioneer: Harold Cohen and AARON (1970s)

When most people think of early AI art, they picture pixelated computer graphics from the late 20th century. But the foundational roots of artificial intelligence creating imagery go back even further—to a British painter named Harold Cohen.

In 1973, while working at the University of California, San Diego, Cohen began developing a computer program that would eventually be known as AARON. AARON was not a neural network in the modern sense; it was a rule-based expert system. Cohen codified his own understanding of drawing—rules about perspective, closed versus open shapes, shading, and composition—into software code.

The output was striking. AARON would output code that drove a custom-built, mechanical turtle-like drawing machine (or an x-y plotter) equipped with a pen or brush across a physical sheet of paper to create abstract, organic compositions.

Cohen didn’t treat the machine as a mere tool; he viewed AARON as a collaborative entity. For decades, AARON evolved. By the 1990s and 2000s, it could autonomously choose colors, paint complex botanical and jungle-like scenes, and project them onto gallery walls. AARON proved a profound concept early on: a machine could internalize human rules of visual logic and execute them to create original art.

The Digital Awakening: Pixels, Digits, and Early Neural Nets

As computing power grew through the 1990s and 2000s, researchers began experimenting with early machine learning models designed to recognize patterns. Before these systems could generate pictures, they had to learn how to see them.

Early databases like MNIST (introduced in 1994) trained basic algorithms to recognize handwritten digits. Naturally, researchers soon flipped the process: If a network can learn what a handwritten “5” looks like, can it draw one?

These early generation attempts resulted in blurry, ghostly gray smudges on dark backgrounds. They weren’t art; they were mathematical proofs of concept. The images possessed zero aesthetic polish—they were low-resolution grids of pixels that looked like digital static attempting to organize itself into a shape.

The Acid Trip Era: Google DeepDream (2015)

If AARON was the thoughtful elder statesman of AI art, DeepDream was its wild, psychedelic teenager.

Released by Google in mid-2015 and engineered by Alexander Mordvintsev, DeepDream originally began as a tool to help computer scientists visualize what a convolutional neural network was “seeing” when it analyzed an image for object recognition (such as spotting a dog or a car in a photo).

When researchers ran the algorithm backward—asking the network to amplify what it thought it saw—the results went viral overnight. DeepDream took ordinary photographs and saturated them with layers upon layers of swirling eyes, dog snouts, paisley patterns, and monstrous architectural structures.

[ Normal Photo ] ──► [ Convolutional Neural Network ] ──► [ "Over-processed" Hallucination ]

The resulting pictures looked like a digital fever dream. While DeepDream couldn’t conjure an image out of nothing from a text prompt, it marked a massive cultural milestone: it was the first time millions of ordinary internet users looked at machine-processed images and felt that the algorithm possessed a bizarre, alien form of imagination.

The Adversarial Arms Race: GANs Take Over (2014–2018)

The true technological leap toward modern AI pictures came with the introduction of Generative Adversarial Networks (GANs), introduced by Ian Goodfellow and his colleagues in 2014.

A GAN relies on a clever game-theory architecture featuring two neural networks competing against one another:

1. The Generator: Tries to create fake images from random noise.
2. The Discriminator: Examines images and tries to guess whether they are real (from a training dataset) or fake (produced by the generator).

As they compete, both networks get ruthlessly efficient. By the late 2010s, models like StyleGAN began producing hyper-realistic faces of people who did not exist (famously captured in websites like ThisPersonDoesNotExist.com).

Early GAN pictures, however, came with a distinct uncanny valley. Viewers frequently encountered telltale artifacts: mismatched earrings, asymmetrical glasses, background walls warping into impossible geometry, or teeth that bled together into a terrifying white blur. These “glitches” became the aesthetic signature of early deep-learning art.

From Artifacts to Masterpieces

Looking back at early AI pictures—from Cohen’s ink-plotting mechanical arms and blurry pixelated digits to Google’s psychedelic dog-slugs and warped GAN portraits—highlights a remarkable trajectory.

Early AI images were defined by their limitations: low resolutions, abstract distortions, and structural chaos. Yet those very limitations gave them a distinct, haunting poetry. They reminded us that machines do not perceive the world through human eyes, memories, or emotions. Instead, they see the world as data, probability, and math—translating human visual culture back to us through a delightfully strange digital mirror.