The Tradition Behind the Technology
The breathless discourse around AI image generation often treats it as something that appeared fully-formed in the last few years — a rupture, a discontinuity, something genuinely new under the sun. The reality is more interesting. Modern AI image generation is the latest chapter in a lineage of practice stretching back to the early 1960s, when artists and engineers first started asking what images a computer could make.
Understanding this history doesn't just satisfy historical curiosity. It puts contemporary AI tools in context, clarifies what is genuinely new about them, and connects practitioners today to a tradition of thinking about computation and creativity that is rich, varied, and still evolving.
The 1960s: Plotter Drawings and the First Computer Artists
The first works that could seriously be called computer-generated art emerged independently in Germany and the United States around 1963–1965.
Georg Nees (1926–2016), a German mathematician and computer scientist at Siemens, began using a Zuse Graphomat Z64 plotter to generate abstract drawings in 1964. His "Schotter" (1968) — a grid of squares that progressively disorder into random orientations — became one of the canonical early works of generative art. It was made entirely by algorithm: a precise mathematical description that the plotter executed mechanically.
Vera Molnár (1924–2023), a Hungarian-French artist, began working with computers in 1968 after a decade of hand-simulating algorithmic processes she called "machine imaginaire." Her systematic explorations of geometric variation, disruption, and iteration — first by hand, then by computer — established many of the conceptual concerns that still animate generative art today: order and disorder, variation within constraint, the aesthetics of rule-governed process.
A. Michael Noll at Bell Laboratories produced works in 1962 that deliberately mimicked the aesthetic of Mondrian's compositions using random statistical processes — and then conducted one of the first empirical studies of computer art by asking subjects to distinguish the Mondrian from the computer imitation. (Most preferred the computer version, which Noll found delightfully problematic.)
What united these early practitioners was a focus on process: the code, the algorithm, the rule set was the work. The image on the plotter paper was a trace of the process, not the thing itself. This distinction — between the generative process and its output — remains central to how generative art is thought about.
The 1970s–80s: Harold Cohen and AARON
Harold Cohen (1928–2016) was a British painter who, in 1968, was already a successful abstract expressionist represented by major galleries when he encountered computers at the University of California San Diego. He became obsessed with a question that sounds deceptively simple: what are the minimum conditions under which a mark can be called a drawing?
The result was AARON, a program Cohen developed over the next four decades, continually extending it as his understanding of the problem deepened. AARON began by generating abstract line drawings based on simple rules. Over years it developed the ability to draw figures, then faces, then full scenes with people and objects. Later versions painted in colour using a custom robotic painting arm.
AARON was not "learning" in the machine learning sense — it was a hand-coded expert system, its knowledge of mark-making and composition encoded directly by Cohen in explicit rules. Its significance was philosophical and artistic: it was the first sustained attempt by any artist to build a system that generated novel visual work autonomously, without Cohen controlling each output. Cohen spent years debating with other artists and philosophers whether AARON was "creative," whether its outputs could be called "art," and whether his own role as its maker made him the artist even when he didn't control the specific image produced.
These debates — about authorship, agency, creativity, and machine autonomy — are exactly the debates that resurface today around AI-generated art. Cohen was having them in the 1970s.
The 1990s–2000s: Emergence and Complexity
As computing power grew through the 1990s, generative art expanded beyond geometric abstraction into complexity-based systems — simulations of natural processes like growth, flocking, weather, and evolution.
Karl Sims created "Evolved Virtual Creatures" (1994), using genetic algorithms to evolve simulated 3D creatures that learned to move through a virtual environment. The results — creatures that swam, crawled, and competed — emerged from the evolutionary process rather than being designed. Sims wasn't depicting creatures; the process was generating them.
Casey Reas and Ben Fry developed Processing in 2001 — a programming language specifically designed to make visual and interactive computation accessible to artists. Processing became the tool through which a generation of artists engaged with code-based image-making. Many contemporary AI artists came to computation through Processing.
During this period, the concept of emergence became central: complex, unpredictable visual behaviour arising from simple rules. Cellular automata, L-systems for plant growth simulation, flocking algorithms — all produced images of startling organic complexity from minimal instructions. The artist's role was to design the conditions; what emerged was genuinely surprising.
2014–2018: Neural Networks and the GAN Revolution
The first major inflection point toward modern AI image generation came in 2014 when Ian Goodfellow and colleagues at the University of Montreal published a paper introducing Generative Adversarial Networks (GANs).
The GAN architecture is elegantly adversarial: two neural networks compete. The generator produces images; the discriminator tries to distinguish generated images from real ones. Each network improves in response to the other — the generator learns to produce more convincing images to fool the discriminator; the discriminator learns to spot ever-more-convincing fakes. The result of this competitive training is a generator capable of producing highly realistic images.
Early GANs produced blurry, low-resolution faces. By 2018, NVIDIA's StyleGAN was generating photorealistic human faces that many observers couldn't distinguish from photographs. The website "This Person Does Not Exist" — which displays a new StyleGAN-generated face on each visit — demonstrated the capability to a global audience and became one of the most discussed demonstrations of AI capability of that period.
Artists quickly engaged with GAN capabilities. Refik Anadol used GANs trained on architectural archives and ocean data to produce immersive large-scale installations. Memo Akten used them for meditations on perception and consciousness. The collective Obvious used a GAN to produce "Portrait of Edmond de Belamy" (2018), which sold at Christie's for $432,500 — triggering another round of the debate about AI, authorship, and value that Cohen had been having in the 1970s.
2021–2022: CLIP, DALL-E, and the Text-to-Image Revolution
GANs required either training on a specific domain (faces, landscapes) or extremely careful manipulation to produce images responding to text descriptions. The breakthrough that changed this was CLIP (Contrastive Language-Image Pre-training), published by OpenAI in early 2021.
CLIP learned to map images and text into a shared semantic space — it understood the relationship between language and visual content at a scale and depth that hadn't been possible before. Combined with generative models, this enabled text-guided image generation: a new way of working where the artist's interface was natural language rather than code, sliders, or careful GAN manipulation.
OpenAI's DALL-E (2021) and DALL-E 2 (2022) demonstrated text-to-image generation to a mass audience. Midjourney's beta launch in 2022 made high-quality image generation accessible without a technical background. And Stable Diffusion, released as an open-source model by Stability AI in August 2022, made the technology openly available for the first time — anyone could run it, modify it, and build with it.
The scale of adoption was immediate and enormous. Within months of Stable Diffusion's release, thousands of custom models, fine-tunes, and derivative tools had been built by the community. The technology moved from research labs to mainstream creative practice in a period of months rather than years.
2023–2026: Diffusion Maturity and the Creative Community
The years since 2022 have seen rapid quality improvement and diversification. SDXL improved resolution and detail. FLUX introduced dramatically better text rendering. DALL-E 3 improved natural language understanding. Midjourney progressed through versions that each raised the quality ceiling. The gap between AI-generated and human-produced professional illustration narrowed in ways that were genuinely consequential for practitioners in affected fields.
Simultaneously, the creative community has developed practices, aesthetics, and communities around these tools. "Prompt engineering" emerged as a recognized skill. AI art communities on Discord, Reddit, and dedicated platforms share techniques, prompts, and workflows. Major museums and galleries have begun acquiring and exhibiting AI-assisted works. The question has shifted from "can AI make art?" to "what kind of art is this, and what does it mean for the people making it?"
What This History Suggests About the Present
Three themes run through this history that resonate clearly with the present moment:
The process-versus-output debate is not new. Every wave of generative art has confronted questions about authorship, creativity, and what it means for a machine to produce something that looks like art. These aren't new questions created by diffusion models — they're the questions the field has been wrestling with for 70 years. The answers haven't become clearer, but the questions have become harder to ignore.
Tool transitions always shift who can participate. Each generation of tools — from plotters to Processing to GANs to text-to-image — has expanded access to generative image-making. Each expansion has brought new practitioners who would never have engaged with the previous generation of tools. The current expansion is the largest in scale, making sophisticated image generation available to anyone with a computer and an internet connection. What emerges from that democratisation is genuinely unpredictable.
The most interesting work uses the tool's specific character. The generative art that has aged best isn't the work that imitated existing media — it's the work that did something only the specific tool and process could do. Nees's disrupted grids, Cohen's rule-governed compositions, Sims's evolved creatures — each used computation to produce images that couldn't have come from any other process. The current generation of AI tools has its own specific character: its particular aesthetic tendencies, its surprising associations, its specific failures and strengths. The artists and practitioners who engage with that specific character — rather than trying to make AI produce something that looks like something else — will produce the most interesting work.
Key Figures Who Shaped the Field
Understanding the people who built this tradition enriches its history. A brief guide to the most significant contributors:
- Georg Nees (1926–2016): One of the three pioneers of computer art (alongside Vera Molnár and Frieder Nake). His 1965 exhibition at the Studiengalerie Stuttgart is considered the first public exhibition of computer-generated art.
- Vera Molnár (1924–2023): The longest-practising computer artist, spanning from the 1960s to her late 90s. Her systematic geometric explorations remain among the most rigorous works of generative art. She was awarded the Wolf Prize in Arts in 2023.
- Harold Cohen (1928–2016): Painter who created AARON, the most sophisticated early autonomous drawing program. His philosophical writing on creativity and computation remains essential reading for anyone interested in AI and art.
- Karl Sims (1962–): Pioneered evolutionary approaches to computer art and artificial life simulation. His "Evolved Virtual Creatures" (1994) is among the most cited works in generative art history.
- Refik Anadol (1985–): Contemporary media artist who uses machine learning and large datasets to create immersive architectural installations. His work has been exhibited at MoMA, the Serpentine, and major institutions globally.
- Ian Goodfellow (1985–): Deep learning researcher who invented Generative Adversarial Networks in 2014. His 2014 paper is among the most cited in machine learning history and directly enabled the first generation of AI image creation tools.
Generative Art in India: A Growing Tradition
The history of generative art is primarily narrated through Western institutions and practitioners. The Indian contribution to this tradition is less documented but real and growing.
Indian artists have engaged with computational art since the 1980s through institutions like the National Institute of Design (NID) in Ahmedabad, where interaction between art, design, and technology was a founding concern. Contemporary Indian new media artists — Shilpa Gupta, Raqs Media Collective, Ravi Poovaiah at IIT Bombay — have all engaged with computation as an artistic medium.
The explosion of AI image generation since 2022 has catalysed a new wave of Indian practitioners. The specific strengths of Indian creative culture — an extraordinarily rich visual tradition spanning millennia, regional diversity of aesthetic practice, and a rapidly growing digital creative economy — position Indian AI artists particularly well to produce work that is both technically sophisticated and culturally distinctive.
The conversation about what AI image generation can mean for Indian creative culture is still early. The historical lineage described in this article suggests that the most significant works will come from practitioners who bring deep creative and cultural knowledge to these tools — not from those who use the tools most fluently in isolation.