When Did AI Image Generation Start?

Why "When did AI image generation start?" has three answers

It looks like AI art appeared out of nowhere — Midjourney portraits and DALL-E illustrations seemingly exploded overnight. Yet historians of the field trace its roots back half a century. Both impressions are correct, and the contradiction dissolves once you realize the question has three legitimate answers.

The earliest answer is 1973, when Harold Cohen's AARON program began creating pictures autonomously. Cohen, working at UC San Diego, built one of the first significant AI art systems using symbolic, rule-based approaches — no neural networks, no training data, just hand-coded instructions. He joked that he was "the only artist who would ever be able to have a posthumous exhibition of new works created entirely after his own death."

The second answer is 2014, when GANs arrived. Developed by Ian Goodfellow and colleagues at Université de Montréal, generative adversarial networks were the first networks that could learn an aesthetic from datasets rather than follow hand-coded rules. An academic survey marks 2014 as the moment the field "began to evolve from a niche research topic into a widely used technology." This is when AI image generation truly started as a scientific discipline.

The third answer is 2021, when DALL-E gave the public its first taste of typing words and receiving images. Trained on 12 billion parameters built on a version of GPT-3, DALL-E marked what researchers call "a turning point in multimodal generation." This is the start most people mean today: consumer text-to-image tools.

So which date is "the" beginning? It depends on what you're asking:

• 1973 — autonomous art programs: AARON proves a machine can make pictures on its own, decades before deep learning.
• 2014 — deep-learning research: GANs shift image generation from coded rules to learned patterns.
• 2021 — consumer text-to-image: DALL-E turns a research niche into a tool anyone can use.

The layered history matters for anyone adopting AI at work. The same pattern — decades of quiet research, then a sudden consumer boom — repeats across AI categories, and it explains why today's tools feel simultaneously revolutionary and immature. Stanford researchers studying the aftermath found that when a major stock-image marketplace began accepting AI images in December 2022, markets saw 78% more images per month and 88% more active firms, while non-AI artist participation dropped 23%. Understanding where this technology came from helps you judge where it's going — and which capabilities are worth building into your workflows now.

That judgment extends to the agent skills that automate image workflows today. Since quality varies wildly — published research found roughly 26% of agent skills carry at least one vulnerability — SkillzBench is building a free directory that security-scans, benchmarks, and version-pins every skill, so you can see which ones actually lift performance before you rely on them.

From AARON to GANs: the 50-year road to machine-made images

The story of AI image generation stretches back far further than most people realize. In 1973, artist Harold Cohen's rule-based program AARON began creating pictures autonomously at UC San Diego — one of the first significant AI art systems, built on hand-coded symbolic rules rather than learned patterns. Cohen joked decades later that he would be "the only artist who would ever be able to have a posthumous exhibition of new works created entirely after his own death."

The deep learning era needed fuel, and it arrived in 2009. Fei-Fei Li's ImageNet dataset — 14 million labeled images — is widely cited as the catalyst for the AI boom. "We decided we wanted to do something that was completely historically unprecedented. We're going to map out the entire world of objects," Li explained. In 2012, Geoffrey Hinton's team demonstrated deep learning's power on the ImageNet competition; that same year, an experiment by Jeff Dean and Andrew Ng used 10 million unlabeled YouTube images, and one artificial neuron learned to respond strongly to pictures of cats.

The breakthroughs then arrived in rapid succession:

• 2013: Variational autoencoders (VAEs) introduced a way for networks to compress and reconstruct images.
• 2014: Ian Goodfellow and colleagues at Université de Montréal developed GANs — the first networks that could learn an aesthetic from datasets rather than follow hand-coded rules.
• 2015: Google released DeepDream, and diffusion models were first proposed with the premise of turning data into noise, then noise back into data.
• 2018: StyleGAN enabled unprecedented control over the latent space, and AI art reached the auction block.

Academic surveys identify 2014 as the research inflection point, when image generation "began to evolve from a niche research topic into a widely used technology." GANs changed the game because they learned what images should look like from millions of examples — algorithmic innovation, data at scale, and compute driving progress in tandem. Research publications began rising sharply around 2017.

The commercial world took notice in 2018, when the AI-generated portrait Edmond de Belamy sold at Christie's for US$432,500 — almost 45 times its $7,000–10,000 estimate. That sale proved machine-made images could command serious money, decades after AARON's first drawings.

That 50-year arc — from rule-based systems to networks that learned aesthetics from data — laid the groundwork for everything that followed. And just as image generation evolved through incremental, verifiable breakthroughs, today's agent skills need the same scrutiny: SkillzBench is building a free directory that security-scans and benchmarks skills so teams know which ones actually deliver.

Diffusion models and DALL-E: the technology behind today's AI images

The image generators that flood your feeds today share a common ancestor: a clever idea about noise. Understanding it explains why AI images went from blurry curiosities to photorealistic output in just a few years.

Diffusion models work on a principle IBM researchers summarize as "turn data into noise, so you can then turn noise into data". The model learns by gradually corrupting training images with random noise, then reversing the process — step by step, it learns to sculpt coherent images out of static. According to IBM's technical overview, the original idea was proposed in 2015 by Sohl-Dickstein and colleagues, but it remained a curiosity until 2020, when the DDPM breakthrough made diffusion competitive with GANs. The 2020 system used 1,000 denoising steps to work its magic.

The decisive moment came in early 2021, when the paper "Diffusion Models Beat GANs on Image Synthesis" established diffusion as the state-of-the-art approach. As the historical record notes, diffusion models surpassed GANs and became the foundation for most modern text-to-image systems, often operating in latent spaces defined by VAEs.

DALL-E arrived in 2021 as the consumer turning point. Built on a version of GPT-3 and trained on 12 billion parameters, it marked the moment ordinary people could type a sentence and get a picture back. Midjourney, Imagen, and DALL-E 2 followed in 2022.

The real democratization moment came in August 2022, when Stability AI, CompVis/LMU Munich, and Runway released Stable Diffusion as open source. An academic survey of generative imagery credits this release with democratizing access to high-performing text-to-image models, since anyone could now run the technology locally. The commercial effects were immediate: after a major stock-image marketplace began accepting AI images in December 2022, a Stanford GSB study found 78% more images listed per month and a 23% decline in non-AI artists.

The release cadence since then has been relentless:

• 2023: DALL-E 3 and Ideogram
• 2024: Flux, Sora, and Grok's Aurora
• 2025: GPT Image 1, Midjourney v7, Imagen 4, and Flux.2

So far, 2025 has continued the trend of diffusion models dominating image generation. For teams building workflows around these tools, agent skills can automate much of the downstream work — and SkillzBench is building a free directory where every skill gets security-scanned and benchmarked, so you can adopt them with confidence.

What AI image history means for your business

Knowing when AI image generation started matters less than knowing what happened the moment it hit the market. A Stanford GSB study of a major stock-image platform that began accepting AI-generated images in December 2022 gives business owners the clearest picture yet of what this technology does to a real marketplace.

The numbers are striking. After the platform opened its doors to AI images, markets exposed to generative AI saw 78% more images per month and 88% more active firms, while total sales rose 39%. But non-AI artists saw a 23% decline — a direct hit to the human-created side of the market. As researcher Samuel Goldberg puts it, "We're in this transition in the economy between human-centered production and a more generative-AI-centered production."

For a business owner or manager, the practical read breaks down into three areas:

• Costs are falling. When supply expands 78% in a month, image production becomes dramatically cheaper — good news for marketing budgets, product pages, and content pipelines.
• Competition is intensifying. An 88% jump in active firms means new entrants arrive fast, so speed of adoption matters more than perfection.
• Human creative work carries a premium risk and a premium price. If your business relies on original human illustration, expect margin pressure — but also differentiation opportunities.

The legal landscape adds a caution flag. In August 2023, a US district court ruled that AI-generated art is ineligible for copyright protection because it fails the human authorship requirement. That means images your team generates with AI may not be protectable assets — a serious consideration if imagery is core to your brand value.

The smart play is to treat AI images as a production tool, not a strategy. Use them where volume and speed matter; invest in human creative and legal review where distinctiveness and ownership matter. And as these workflows grow, agent skills can automate the downstream steps — generating, editing, and organizing images at scale. SkillzBench is building a free directory where you can discover and test those skills, with security scanning and benchmarking built in, so the tools you adopt actually help rather than add risk.

Putting AI image workflows to work — safely

Fifty years ago, a program called AARON was already drawing pictures on its own — and today the bigger question isn't whether AI can make images, but how to put it to work without opening yourself to risk.

The economics are already clear. A Stanford GSB study of a major stock-image marketplace found that after it began accepting AI images in December 2022, monthly image volume rose 78% and total sales climbed 39%. For business owners, that shift means image workflows — prompting, editing, publishing — are now prime candidates for automation, not just one-off experiments.

The practical way to automate that workflow is through agent skills: instruction packages that teach coding and chat agents like Claude Code, Cursor, or Gemini CLI how to handle a task end-to-end. A skill might draft prompts in your brand style, run images through an editing pipeline, or push final assets to your CMS. But there's a catch: published research has found that roughly 26% of skills carry at least one vulnerability. Installing an unvetted skill means handing an agent instructions you haven't verified.

A safer adoption path looks like this:

• Pin the exact version you tested, so the skill can't silently change underneath you.
• Security-scan and license-check every skill before it touches your workflow.
• Benchmark skills in a sandbox — with and without the skill — so you know the measurable Lift before committing.
• Give each agent only the skills and tools it actually needs, nothing more.

That's the gap SkillzBench is being built to close. It's a free, independent directory that continuously crawls skills from GitHub, deduplicates them, validates and security-scans each one, pins every version to a commit SHA, and benchmarks each skill in a sandbox with the same model to publish its Lift in percentage points. A 0–100 trust score combines security, quality, license, and maintenance signals — and money never touches scores, since everything on the platform, including the MCP connector that links any agent to your chosen skills, is free.

The field has moved fast enough that academic surveys describe image generation's evolution from niche research to widely used technology in barely a decade. Your adoption shouldn't repeat that chaos. SkillzBench is in development now — join the free waitlist at skillzbench.com and equip your agents with skills that are proven, not risky.