What Was the First AI Generated Image?

Why 'The First AI Image' Has No Simple Answer

Ask ten sources what the first AI-generated image was, and you'll get ten different answers — or, more often, no answer at all. No authoritative source names a single definitive "first," and that silence is itself the key finding: AI image generation arrived gradually, through competing definitions of what "AI-generated" even means.

The answer depends entirely on where you draw the line. Three distinct categories compete for the title:

• Rule-based algorithmic images — pictures produced by hand-coded instructions, not learned from data. Harold Cohen's AARON, conceived in the late 1960s and named in the early 1970s, is the strongest candidate here; the Whitney Museum calls it "the earliest artificial intelligence (AI) program for artmaking."
• Learned generation — images produced by models trained on datasets. This era begins with Ian Goodfellow's invention of GANs at the Université de Montréal in 2014, the first approach to learn an aesthetic from images rather than follow explicit rules.
• Text-to-image — the mainstream era most people mean today, starting with OpenAI's DALL-E in 2021 and exploding with Stable Diffusion and Midjourney in 2022.

The confusion isn't just academic. Even the most famous "first AI artwork" claim collapsed under scrutiny. When the collective Obvious sold Edmond de Belamy at Christie's in October 2018 for $432,500 — roughly 43 times its $7,000–$10,000 estimate — the auction house marketed it as the first AI-created artwork. Yet the work relied on uncredited open-source code by artist Robbie Barrat, and AI artist Mario Klingemann estimated "probably 90 percent of the actual 'work' was done by [Barrat]." Obvious's own Caselles-Dupré later admitted that "lots of false information is released with our name on it."

Even AARON's dating shifts depending on the source: the Whitney says Cohen conceived it in the late 1960s, the Computer History Museum dates development to around 1971, and Artnet News points to 1973 as the year of the "first autonomous picture machine." These aren't contradictions so much as different milestones — conception, naming, public presentation — which is precisely why a single "first" is so elusive.

Why does this matter if you're using AI visuals today? Because the same definitional fuzziness runs through every modern question about AI images: who owns them (a US district court ruled in August 2023 that AI art is ineligible for copyright due to lack of human authorship), and how to evaluate them (a peer-reviewed systematic review found people generally cannot distinguish human-made from AI-made art). Understanding that "AI image" covers everything from a 1970s rule-based drawing program to a diffusion model helps you ask sharper questions about any tool you adopt. And as with AI image generation itself, evaluating what's actually behind a tool — rather than its marketing claims — is where careful verification pays off; it's the same principle behind SkillzBench's approach to independently verifying and benchmarking agent skills before you trust them.

AARON: The 1970s Machine That Drew Before It Learned

Ask most people when AI images began and they'll say 2022, when Midjourney and Stable Diffusion flooded the internet. The real answer goes back half a century earlier, to a program that painted abstract shapes on paper while Apollo astronauts were still walking on the Moon.

That program was AARON, created by artist and computer scientist Harold Cohen. The Whitney Museum of American Art calls it the earliest artificial intelligence program for artmaking, which makes it the strongest candidate for the first AI-generated images. The dating needs care: Cohen conceived the system in the late 1960s at UC San Diego, named it in the early 1970s, and a painting system was presented at the 1971 Fall Joint Computer Conference, with the "Aaron" identity established by 1973.

AARON's early output was abstract and biomorphic — forms that could differentiate figure from ground without any knowledge of external objects, according to the Computer History Museum, which also calls it one of the longest-running, continually maintained AI systems in history. Cohen's philosophy was telling: he compared himself to a Renaissance painter and AARON to a studio assistant, a trained hand executing under direction rather than an autonomous genius.

The system evolved for four decades:

• 1973 onward: rebuilt as an "art creation expert system"
• 1980s: knowledge base expanded to objects, plants and people
• 1979: a hundred-foot AARON mural exhibited at SFMOMA
• 1995: first color-capable AARON, driving a robot arm that drew and colored images in a Boston exhibition

Artnet News describes the 1973 version as the first autonomous picture machine, noting Cohen found some instructions generated forms he hadn't imagined — allowing the machine to make something like artistic decisions.

The idea of machine-made pictures reaches even further back. Artnet points to a portrait of Joseph Jacquard woven on his punch-card loom in 1836 using 24,000 punched cards, which could reasonably be viewed as the first digitized image. Precursors stretch to Maillardet's automaton around 1800, capable of drawing and writing.

What distinguishes AARON from today's tools is instructive: it followed hand-coded rules rather than learning from data. The Whitney's 2024 exhibition explicitly framed it as historical perspective now that DALL-E and Midjourney are mainstream. That distinction — rules versus learned patterns — is exactly the kind of nuance that matters when businesses evaluate AI image tools today, and it's the lens we use at SkillzBench when assessing what AI agent skills actually deliver versus what they claim.

From GANs to DALL-E: The Milestones That Changed Everything

The story of AI-generated images is really a story of two revolutions: machines that follow rules, and machines that learn. The second revolution began in 2014, when Ian Goodfellow invented generative adversarial networks (GANs) at Université de Montréal — the first approach that learned an aesthetic from image datasets rather than following hand-coded rules, according to the history of AI art. Artists broadly adopted GANs around 2017, once tech companies open-sourced tools like Google TensorFlow and Meta's Torch.

The milestones then arrived in rapid succession:

• DeepDream (2015) produced hallucinatory, dream-like imagery by amplifying patterns inside neural networks.
• Conditional GANs (2017) generated images across 1,000 ImageNet classes, giving artists more control.
• Diffusion models, proposed in 2015, surpassed GANs in early 2021; latent diffusion became the engine behind Stable Diffusion in August 2022.
• OpenAI's DALL-E 1 (2021) marked a landmark in prompt-based image generation, followed by Midjourney, Imagen and Parti in 2022.

The art world got its wake-up call in October 2018, when the Paris collective Obvious sold Edmond de Belamy at Christie's New York for $432,500 — roughly 43 times its $7,000–$10,000 estimate, per the auction record. An anonymous phone bidder won after about seven minutes of competition among five parties, beating out everything in the sale except a Warhol.

But the "first AI portrait" label was contested almost immediately. The work relied on uncredited open-source code by AI artist Robbie Barrat, who asked publicly whether the collective had "really just used my network and are selling the results." Fellow AI artist Mario Klingemann estimated that "probably 90 percent of the actual 'work' was done by" Barrat — and even Obvious's own Caselles-Dupré later admitted that "lots of false information is released with our name on it." The portrait had been trained on 15,000 portraits spanning the 14th to 19th centuries, so the machine learned from centuries of human art.

That controversy captures something important: AI image generation has always been a collaborative, messy, contested evolution — from AARON's rule-based drawings in the early 1970s to today's diffusion models that anyone can prompt. The same is true of working with AI now. Understanding which tools and techniques actually deliver results, rather than hype, is a modern version of the Belamy problem.

That's exactly the gap SkillzBench is being built to close: a free, independent directory of AI agent skills where every skill is verified, security-scanned and benchmarked before you connect it to your agent. If you want to be first in line when it launches, the waitlist is open now.

What This History Means for Your AI Image Workflow

Fifty years of AI image history leaves you with an uncomfortable conclusion: the pictures are already indistinguishable from ours, the law hasn't caught up, and the burden of judgment has quietly shifted onto whoever clicks "generate."

That burden lands on anyone using AI images at work. A systematic review of 44 studies found that people generally could not tell the difference between human-made and AI-made art — even as human-made work was valued more highly in some studies. For a business audience, that cuts both ways: your AI-generated marketing visuals will pass as "real" to most viewers, but audiences who do detect AI involvement may discount the work.

The legal side is just as consequential. In August 2023, a US district court ruled that AI-generated art is ineligible for copyright protection because it lacks human authorship, according to Wikipedia's AI art overview. If an image matters to your brand — a logo, a campaign asset, product imagery — assuming you own it outright could be a costly mistake. Human creative input isn't just an artistic question anymore; it's a legal one.

There's also a quality lesson buried in this history. The Whitney Museum's 2024 exhibition of Harold Cohen's AARON, which it calls "the earliest artificial intelligence (AI) program for artmaking," was explicitly framed as important historical perspective now that DALL-E, Midjourney and Stable Diffusion are mainstream — and it's instructive why the Whitney chose a half-century-old program for that role. Cohen spent four decades refining AARON, treating it less like a magic box and more like a studio assistant that needed constant direction. Contrast that with the Edmond de Belamy saga: a portrait sold for $432,500 at Christie's in 2018 — roughly 43 times its estimate — only for experts to estimate that "90 percent of the actual 'work'" came from uncredited open-source code by artist Robbie Barrat. Hype outpaced scrutiny, and the buyers paid for it.

When evaluating today's image tools, that history suggests a checklist:

• Ask what's actually doing the work — whose model, whose training data, whose code.
• Treat output quality claims skeptically until you've benchmarked them yourself.
• Assume copyright protection is uncertain without meaningful human authorship.
• Prefer tools and workflows you can inspect, test, and swap out.

That last point extends beyond image generators to the agent skills people now stack on top of them. Published research found roughly 26% of agent skills carry at least one vulnerability, so a skill that automates your image pipeline deserves the same scrutiny Cohen gave AARON — and the same skepticism the Belamy sale lacked. SkillzBench is building exactly that: a free directory where skills are security-scanned, version-pinned, and benchmarked with and without the skill so you can see the measured lift before you trust it, plus a free MCP connector to load chosen skills into your agents. It's pre-launch, with the waitlist live now.

Cohen spent 20 years teaching AARON to draw. You have better tools — but the same job: verify before you trust.