Early AI-Generated Images: Evolution and History of AI Art

The Strange and Surreal World of Early AI-Generated Images

Long before modern tools could render photorealistic portraits or crisp digital landscapes from a single sentence, the world of computer-generated art looked entirely different. Far from the polished pixels of today, early artificial intelligence art was characterized by dreamlike hallucinations, geometric experiments, and a fascinating glimpse into how machines first learned to “see.”

Understanding where this technology started reveals how far it has traveled—and why those early, glitchy creations still capture our imagination.

Rule-Based Roots and Robotic Painters

The journey of algorithmic imagery didn’t start with deep learning; it began with code and raw logic. In the late 1960s, artist and programmer Harold Cohen created AARON, a computer program designed to generate original drawings. AARON wasn’t using neural networks; instead, it followed complex sets of symbolic rules written by Cohen to paint and draw using physical robotic arms.

As computing power grew through the decades, artists and scientists began experimenting with evolutionary algorithms—systems that could mutate digital shapes, fractal flames, and virtual creatures over generations. These systems laid the conceptual groundwork: humans provided the system and the rules, and the computer produced variations beyond manual repetition.

The Era of Hallucinations: Google DeepDream

A massive cultural turning point arrived in 2015 with the introduction of DeepDream by Google. Originally built to help researchers understand how computer vision networks classify images, DeepDream used a convolutional neural network to find and enhance patterns in photos.

The results were unforgettable. When fed ordinary photographs, DeepDream would aggressively over-process them, filling skies and landscapes with swirling organic patterns, repeating dog faces, and bulging animal eyes.

• The Aesthetic: DeepDream imagery looked genuinely psychedelic and surreal.
• The Significance: It gave the public a direct look inside the “mind” of an early neural network, proving that machines didn’t see the world like humans did—they saw patterns, edges, and features that they amplified to an extreme degree.

The Cat-and-Mouse Game: GANs Enter the Scene

Shortly after DeepDream, a breakthrough framework called Generative Adversarial Networks (GANs) arrived in 2014, fundamentally changing how machines create. A GAN relies on two distinct neural networks playing a high-stakes game against each other:

1. The Generator: Tries to create fake images from random noise to fool the other network.
2. The Discriminator: Inspects the images and tries to determine whether they are real (from a training dataset) or fake (created by the generator).

Through thousands of rounds of competition, both networks improve. Early GAN outputs were often ghostly, abstract, or distorted—faces with melting features, bizarre limbs, or surreal smudges that hovered unsettlingly between recognizable objects and pure abstraction.

By the late 2010s, GANs had advanced enough to generate convincing historic-style portraits, such as the infamous Portrait of Edmond de Belamy, which famously sold at Christie’s auction house in 2018 and sparked global conversations about authorship and machine creativity.

Why Early AI Art Still Matters

Looking back at early AI imagery highlights a distinct charm that modern, hyper-realistic generators often lack. Early algorithms were constrained by data limitations and architectural quirks, resulting in an unpolished, avant-garde aesthetic full of happy accidents. They forced creators to embrace unpredictability, treating the computer not just as a tool, but as an unpredictable digital collaborator.