From Pixels to Psychedelics: The Fascinating History of Early AI Pictures

Long before the hyper-realistic images and sleek text-to-image generators we use today, artificial intelligence created a very different kind of visual art. Early AI pictures were often surreal, glitchy, and dreamlike. Rather than replacing human artists, early computational art was an experimental playground for scientists and programmers trying to teach machines how to “see.”

The Rule-Based Roots: AARON (1970s)

Long before deep learning existed, computer artists were already experimenting with code. In the 1970s, artist and programmer Harold Cohen created a robotic painting program called AARON.

Unlike modern AI that learns from millions of images on the internet, AARON followed a strict set of programmed rules. Cohen wrote code that taught the machine about geometry, composition, and human anatomy. AARON would then draw abstract shapes, plants, and human figures using a robotic arm holding a paintbrush, while Cohen occasionally added the final colors by hand. It proved that computers could assist in creative processes decades before the neural network boom.

The 2010s Revolution: GANs and DeepDream

The landscape of computer-generated imagery shifted dramatically in the mid-2010s with the rise of deep learning and neural networks. Two major milestones defined this era:

• Generative Adversarial Networks (GANs – 2014): Introduced by Ian Goodfellow, GANs introduced a clever concept: two neural networks competing against each other. The Generator tried to create fake images, while the Discriminator tried to spot the fakes. Through this continuous game of cat-and-mouse, the generator rapidly learned to produce increasingly realistic images.
• Google’s DeepDream (2015): Created by Alexander Mordvintsev, DeepDream used convolutional neural networks in reverse. Instead of recognizing what was in a picture, the program took existing images and amplified patterns it thought it saw—resulting in bizarre, psychedelic visuals packed with repeating dog faces, eyes, and swirling geometric patterns.

Milestones in Early AI Image Generation

| Year | Milestone / Project | Description |
| :— | :— | :— |
| 1973 | AARON | A rule-based computer program created by Harold Cohen that autonomously generated line art. |
| 2014 | GANs | Invented by Ian Goodfellow, pairing a generator against a discriminator to build synthetic images. |
| 2015 | DeepDream | Released by Google, famous for turning ordinary photographs into hallucinatory, pattern-heavy dreamscapes. |
| 2021 | Early Text-to-Image | Early iterations of contrastive language-image pre-training models began bridging text prompts with visual outputs. |

Looking Back

The strange, distorted faces and trippy landscapes of early AI pictures might look clumsy compared to modern tech, but they laid the crucial groundwork for today’s creative tools. They shifted AI from a rigid mathematical calculator into a medium capable of surprising interpretation, proving that machine creativity has a remarkably weird history.