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

Long before prompt boxes flickered on every smartphone and generative models conjured photorealistic universes in seconds, the intersection of art and artificial intelligence looked very different.

Today, we take text-to-image engines for granted. But the journey from room-sized mainframe computers to modern neural masterpieces spans decades of quiet experimentation, mechanical plotters, digital hallucinations, and a fundamental shift in how we define artistic authorship.

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

When people think of the first “AI artist,” they usually picture a server farm. In reality, it started with a single man, a mainframe computer, and a robotic turtle-like drawing machine.

In the late 1960s, British painter Harold Cohen—who had represented his country at prestigious global exhibitions like the Venice Biennale—left the traditional art world behind to explore computer programming at the University of California, San Diego. By 1973, he had created AARON: arguably the world’s first autonomous art-generating software.

AARON was not a neural network in the modern sense; it was a rule-based expert system. Cohen codified his own understanding of geometry, perspective, closed shapes, and human anatomy into lines of code. AARON would run computations to compose drawings from scratch, sending coordinates to a physical drawing machine (a plotter) that gripped pens and brushes to execute the lines.

The Big Question: Cohen wasn’t interested in using a computer merely as a tool. He wanted to know whether a machine could possess an autonomous understanding of what makes an image evocative. For decades, Cohen and AARON evolved together, with Cohen eventually teaching the program to handle complex color choices and organic plant-like forms.

2. Hallucinogenic Pixels: The Rise of DeepDream (2015)

For decades, computer-generated art remained a niche academic pursuit. That changed overnight in the summer of 2015, when Google released DeepDream.

Originally designed to help researchers visualize what a deep convolutional neural network was “seeing” during image classification, DeepDream turned the process inside out. Instead of asking a network to identify a cat in a photo, engineers asked it to enhance whatever patterns it already thought it recognized.

The results were psychedelic, surreal, and instantly viral:

• Ordinary landscape photos morphed into sprawling architectures filled with endless dog faces, bulging eyes, and swirling optical illusions.
• It didn’t “create” images from text, but it revealed the alien way machine vision systems parse reality—not through meaning, but through statistical pattern-matching.

DeepDream birthed a distinct internet aesthetic and proved that artificial neural networks possessed a bizarre, unauthorized form of imagination.

3. The Duelists: Generative Adversarial Networks (GANs)

Shortly after DeepDream captured the public imagination, a technical breakthrough revolutionized synthetic image creation: GANs (Generative Adversarial Networks), introduced by Ian Goodfellow in 2014 and popularized among artists shortly after.

A GAN is an artistic sparring match between two neural networks:

1. The Generator tries to create fake images from random noise, hoping to fool the critic.
2. The Discriminator acts as an art critic, examining real historical artworks alongside the generator’s fakes to spot the forgery.

Through thousands of rounds of competition, both networks grow sharper. The generator learns the precise mathematical distribution of a dataset—such as 15,000 classical portraits—and begins outputting entirely original faces that look hauntingly human, yet belong to people who never existed.

This era culminated in milestones like the 2018 auction of Portrait of Edmond de Belamy at Christie’s, where a GAN-generated canvas signed with a mathematical formula sold for hundreds of thousands of dollars, shocking the traditional art market.

Looking Back to Look Forward

| Era | Core Technology | Aesthetic Signature | What It Taught Us |
| :— | :— | :— | :— |
| 1970s–1990s | Rule-based expert systems (AARON) | Abstract lines, mechanical plotter ink, hand-colored shapes | Algorithms can capture structural rules of composition. |
| 2015 | Deep convolutional networks (DeepDream) | Hallucinogenic eyes, dog-slug hybrids, psychedelic textures | Neural networks “see” statistical correlations rather than objects. |
| Late 2010s | Generative Adversarial Networks (GANs) | Ethereal portraits, morphing faces, dreamlike landscapes | Systems can autonomously mimic human artistic genres from data. |

The early history of AI pictures proves that our current technological leap didn’t happen in a vacuum. From Harold Cohen’s robotic brushes to Google’s algorithmic hallucinations and adversarial networks, artists and engineers have spent over fifty years trying to answer a single provocative question: Can a machine learn to dream?