When Did AI-Generated Images Start?

The Confusing Origins: Why "When AI Images Started" Has Two Answers

If you searched "AI image generator" in 2021, you'd find a handful of curiosities. By 2023, AI images seemed to be everywhere — in marketing campaigns, social feeds, and stock libraries — almost overnight. That suddenness is exactly why the question "when did AI-generated images start?" has two honest answers.

The first answer is technical, and it goes back further than most people expect. In the late 1960s, artist Harold Cohen began developing AARON, a symbolic rule-based art system, which was exhibited at the Los Angeles County Museum of Art in 1972 — the earliest significant AI art system on record, according to the documented history of AI art. By that measure, AI-generated images are more than half a century old.

The second answer is mainstream. The 2020s brought the text-to-image boom, when tools like Midjourney, DALL-E, and Stable Diffusion became widely available to the public, letting users generate imagery with little effort. Marketing analysts frame AI imagery as a suddenly trendy phenomenon that surged around 2023 — and for most readers, that's when AI images actually "started," in the sense of becoming visible in daily life.

Both answers are correct because the technology evolved through distinct eras rather than a single invention moment:

• Rule-based era (late 1960s–1990s): Hand-coded systems like AARON, followed by Karl Sims' Prix Ars Electronica wins in 1991–1992 and Electric Sheep in 1999.
• Deep learning and GAN era (2014–2020): Ian Goodfellow's generative adversarial networks learned aesthetics from datasets instead of following rules.
• Diffusion era (2015 onward): Diffusion models were first proposed in 2015 and surpassed GANs as state-of-the-art by early 2021.
• Consumer text-to-image boom (2022 onward): DALL-E 2, Midjourney, and Stable Diffusion put generation in anyone's hands.

Understanding both timelines matters more than trivia. The gap between a technology's invention and its mainstream arrival is where businesses get caught flat-footed — the same pattern is repeating today with AI agents and agent skills, which is why SkillzBench is building a free directory and MCP connector to help teams evaluate which skills actually deliver. The four eras below trace exactly how we got here.

Four Eras of AI Image Generation: From AARON to GANs to Diffusion

Ask someone when AI-generated images started, and they'll likely say 2022 — the year images like these flooded social media. The real answer stretches back more than five decades, through four distinct eras of technology that each redefined what machines could create.

The story begins in the late 1960s, when Harold Cohen started building AARON at UC San Diego. AARON used a symbolic, rule-based approach — artists and programmers hand-coded the logic, and the system followed it. Cohen's creation was exhibited at the Los Angeles County Museum of Art in 1972, making it the earliest significant AI art system on record. The rule-based era continued through the 1990s: Karl Sims won the Prix Ars Electronica Golden Nica in both 1991 and 1992 for his artificial life art, and in 1999 Scott Draves released Electric Sheep, an evolving fractal screensaver.

The modern era arrived in 2014, when Ian Goodfellow and colleagues at Université de Montréal invented generative adversarial networks (GANs). For the first time, a system could learn an aesthetic from example image datasets instead of following hand-coded rules. GAN milestones followed quickly — CycleGAN enabled unsupervised image conversion, StyleGAN allowed fine control over style and content, and BigGAN scaled quality through massive training runs.

But GANs had a flaw. Research comparing the two approaches found that GANs suffer from pattern collapse and unstable training, while diffusion models offer better stability and controllability. That difference decided the winner.

Diffusion models were first proposed in 2015, when Sohl-Dickstein and colleagues introduced a physics-inspired idea: "turn data into noise, so you can then turn noise into data." The research chain then built steadily:

• 2015: Sohl-Dickstein et al. publish the foundational thermodynamics-based concept.
• 2019: Yang Song and Stefano Ermon independently develop score-based generative models.
• 2020: Ho et al. publish denoising diffusion probabilistic models (DDPMs), matching GAN image quality.
• 2021: A landmark paper unifies the two mathematical approaches — and diffusion models surpass GANs as the state of the art.
• 2022: "High-Resolution Image Synthesis with Latent Diffusion Models" makes generation efficient and affordable — the research that became Stable Diffusion.

That last step mattered most for everyday users. Latent diffusion cut the computing cost dramatically, and by August 2022 Stable Diffusion was released alongside Midjourney and DALL-E 2, kicking off the consumer text-to-image boom. According to IBM's technical history, diffusion now powers the leading systems, including Stable Diffusion, DALL-E, Midjourney, and Google's Imagen.

The pace hasn't slowed. Releases continued through 2024 and 2025 — Flux, GPT Image 1, MidJourney v7, and Google Imagen 4 among them — and the same compounding pattern applies to AI agent skills, which keep improving as the ecosystem matures. SkillzBench is building a free, independent directory that verifies and benchmarks agent skills so you can put that progress to work; join the waitlist to be first in line.

The 2022 Mainstream Moment: DALL-E, Midjourney, and Stable Diffusion

Ask anyone when AI-generated images "started," and they'll usually point to the moment they first typed a prompt and watched an image appear. That moment — the consumer text-to-image boom — arrived between 2021 and 2023, and it turned decades of research into something anyone with a browser could use.

The first taste came in 2021, when OpenAI released DALL-E 1, the first widely publicized system that could turn a plain-text description into an image. Per Wikipedia's AI art timeline, the same year saw EleutherAI release VQGAN-CLIP as open source, giving early adopters a DIY path into generative imagery.

Then 2022 delivered the explosion. Midjourney launched, Google announced Imagen and Parti in May, and Stable Diffusion arrived in August — notable because it was built on the latent diffusion research that made generation dramatically cheaper and more efficient, according to IBM's explainer on diffusion models. Suddenly, text-to-image tools were "widely available to the public, allowing users to quickly generate imagery with little effort."

The momentum carried straight into 2023, with DALL-E 3 and Ideogram (which specialized in rendering legible text inside images) reaching the market. Adoption numbers tell the story: Dream by WOMBO surpassed 65 million app installs, Jasper Art reached more than 70,000 brands and content creators, and Fotor counted over 14 million users worldwide, per HubSpot's analysis of AI images.

The mainstream moment had cultural markers, too. Key milestones along the way:

• 2018: The AI portrait "Edmond de Belamy," created by Paris collective Obvious, sold at Christie's for $432,500 — roughly 45 times its high estimate.
• 2021: DALL-E 1 and open-source VQGAN-CLIP put text-to-image generation in reach of early adopters.
• 2022: Midjourney and Stable Diffusion launched, making high-quality generation a consumer product.
• 2023: DALL-E 3 and Ideogram shipped, and a US district court ruled that purely AI-generated art is ineligible for copyright.

And the wave keeps building. Black Forest Labs released Flux in 2024, followed by a 2025 run of GPT Image 1 (March), MidJourney v7 (April), Flux.1 Kontext (May), Google Imagen 4, and Flux.2 (November). For business owners, the takeaway is that image generation is now table stakes — the competitive edge comes from what you build around it, like the automated workflows SkillzBench is developing with its free agent skills directory and MCP connector (join the waitlist to get early access).

If you're evaluating tools today, the practical questions are quality, cost (plans range from free up to about $40/month), and rights — Midjourney doesn't indemnify commercial users, while Adobe Firefly, Getty, and Shutterstock offer licensed, indemnified images. Whatever you choose, expect the landscape to look different six months from now.

What This Means for Your Business: Enterprise Adoption, Legal Risks, and Practical Steps

If you're still treating AI-generated images as a novelty, your competitors have already moved on. By early 2024, brands like UnderArmour, Walmart, Coca-Cola, and Nutella were running AI imagery in production marketing campaigns — and the practical questions for business owners have shifted from "should we?" to "how do we do this without legal exposure?"

Consider the scale. In Ferrero's Nutella "Unica" campaign with Ogilvy & Mather Italia, an algorithm trained on patterns and color palettes generated 7 million unique packaging designs, each with its own authentication code. Tommy Hilfiger ran an AI design campaign for 2023's AI Fashion Week, with winning designs showcased on the virtual fashion platform Dress X. AI enables professional-grade imagery in days or even hours, shaving weeks off production schedules.

But adoption comes with real risk. In August 2023, a US district court ruled that AI art is ineligible for copyright protection because it fails to meet the human authorship requirement — meaning work you generate may not be legally yours to protect. There's also an indemnification gap between tools: stock industry players like Adobe Firefly, Getty Images, and Shutterstock trained their generators on their own portfolios and offer licensed, indemnified AI images, while Midjourney does not indemnify its users. Levis' backlash over using AI to simulate model diversity shows reputational risk matters as much as legal risk.

A practical playbook for business owners:

• Match the tool to the risk. For commercial campaigns where legal exposure matters, indemnified options like Adobe Firefly, Getty, or Shutterstock offer a safety net consumer tools don't.
• Vet every output. AI images still require human review for odd body proportions, impossible architecture, and notoriously bad hands before anything reaches production.
• Stay current. The landscape is moving fast — 2025 alone brought GPT Image 1, MidJourney v7, Google Imagen 4, and Flux.2 — and copyright law continues to evolve alongside it.
• Automate the downstream work. Generating the image is only half the workflow; resizing, captioning, publishing, and QA eat the time savings.

That last step is where agent skills come in. Instruction packages for tools like Claude Code, Cursor, and Gemini CLI can automate everything after the image is generated — but quality varies wildly, and published research found roughly 26% of skills carry at least one vulnerability. That's the problem SkillzBench is building for: a free, independent directory of agent skills with security scanning, version pinning, and sandbox benchmarks, plus a free MCP connector that lets you plug vetted skills into any agent. It's pre-launch now, with the waitlist live — worth joining if AI image workflows are becoming part of your business.