The Ghost in the Machine: A Short History of Early AI Pictures

Long before the era of text-to-image prompts, hyper-realistic digital avatars, and high-resolution diffusion models, the relationship between computers and canvas was an act of quiet, computational exploration. Today, anyone can generate a cinematic masterpiece in seconds. But the journey of early artificial intelligence pictures reveals a fascinating evolution—moving from cold, rule-bound lines drawn by robotic arms to the surreal, dream-like hallucinations of early neural networks.

Understanding where AI art started gives us a deeper appreciation for how machines learned to “see” and “imagine.”

1. The Pioneer: Harold Cohen and AARON (1970s)

When we think of early AI art, the story almost always begins with AARON. Developed in the late 1960s and early 1970s by British artist Harold Cohen at the University of California, San Diego, AARON was a groundbreaking computer program designed not to copy human art, but to autonomously create it.

In an era of massive, room-sized computers and punch cards, Cohen’s approach was revolutionary. AARON operated on symbolic, rule-based programming (GOFAI—Good Old-Fashioned AI). It didn’t use deep learning or neural networks because they didn’t exist yet; instead, Cohen encoded his own understanding of geometry, perspective, and human form into lines of code.

AARON drove mechanical drawing machines—robotic arms holding pens that physically glided across paper to execute lines and shapes. Cohen often watched in amazement as the machine produced compositions he hadn’t explicitly premeditated. Over decades, AARON evolved from drawing abstract shapes into rendering plants, figures, and complex scenes, proving that a machine could possess an independent artistic “hand.”

2. The Algorithmic Wild West (1980s–1990s)

Following Cohen’s early breakthroughs, the following decades saw artists and computer scientists experiment heavily with algorithmic art, fractals, and artificial evolution.

During the 1990s, projects like Karl Sims’ Galápagos used artificial intelligence concepts and genetic algorithms to let viewers “evolve” 3D animated forms. Similarly, Scott Draves created Electric Sheep in 1999, a distributed screensaver that used evolutionary algorithms and crowd-sourced computing power to generate infinite, morphing fractal animations that learned from their audience. These projects laid the groundwork for generative design, demonstrating that computers could handle aesthetic mutations and natural selection.

3. The Neural Awakening: Google DeepDream (2015)

The true paradigm shift toward modern AI image generation happened in the mid-2010s with the advent of deep learning. In 2015, Google introduced DeepDream (originally code-named Inceptionism), and the internet was instantly captivated.

DeepDream did something entirely unexpected: it inverted how a convolutional neural network worked. Instead of asking a computer to look at an image and tell you what was in it (like a cat or a car), researchers fed an image into the network and asked it to enhance what it thought it saw.

The results were wild, psychedelic, and instantly recognizable. Ordinary photographs of landscapes or clouds were transformed into hallucinatory vistas crawling with swirling patterns, dogs’ eyes, and bizarre hybrid creatures (affectionately dubbed “puppy slugs”). For millions of people, DeepDream was the first time they realized that artificial intelligence had its own unique, internal “vision” of reality.

4. The Adversarial Era: GANs (2014–2018)

Shortly before DeepDream went viral, a machine learning researcher named Ian Goodfellow invented Generative Adversarial Networks (GANs) in 2014, setting off a gold rush in AI imagery.

A GAN relies on a brilliant game-theoretic setup:

• The Generator tries to create fake images from scratch.
• The Discriminator acts as an art critic, trying to spot the difference between real human-made images and the generator’s fakes.

The two neural networks constantly pitted themselves against each other—the generator getting better at fooling the critic, and the critic getting better at catching flaws. This internal arms race produced the first truly convincing, high-resolution AI-generated faces and abstract portraits. By 2018, a GAN-generated artwork (Portrait of Edmond de Belamy) even sold at Christie’s auction house for over $400,000, signaling that early AI pictures had officially crashed the traditional art market.

Reflections on the Pixelated Past

Looking back at early AI pictures—from Harold Cohen’s ink-speckled plotter drawings to Google’s psychedelic dog-eye landscapes—we see a continuous human desire to collaborate with the unknown.

The tools have shifted dramatically from manual code and robotic arms to vast transformer models and diffusion pipelines. Yet, the core magic remains unchanged: the thrill of watching a machine take a set of parameters and surprise its creator with something entirely new.