First AI Generated Image: Harold Cohen to Image Generation
Quick answer
What was the first AI generated image ever made?
There is no single agreed answer, because it depends on what counts as AI. Harold Cohen's AARON drew from hand-written rules in the early 1970s. Learned, photo-style faces came with the GAN in 2014. Edmond de Belamy, a GAN portrait, drew wide attention in 2018, and by 2023 many AI pictures looked real.
The first AI generated image was not a photo of anything. It was a line drawing, and a person had typed in every rule the computer followed. That was decades before anyone said "deepfake." Today the same idea can make a picture of your teenager that looks real enough to ruin a night, or a life.
First AI generated image has no single answer: it started as rule-based drawing in the 1970s, became learned faces in 2014, and by 2023 looked like real photos, so your eyes alone can no longer tell what is fake.
First AI Generated Image: Where the Story Really Starts
Ask a search engine for the first AI generated image and you get a pile of different answers. That is not because experts are sloppy. It is because the question hides another one: what counts as "AI"? Let's walk through the candidates, because the history is more fun than a trivia answer.
Early images from a drawing program called AARON
In the early 1970s, an artist named Harold Cohen began building Harold Cohen's rule-based software program AARON. Search results often label him as the American computer scientist Harold Cohen, though he was a painter first. AARON did not learn anything from pictures. Cohen wrote the rules by hand, things like "close this shape" or "keep these two forms apart." The program then followed them and drew. Many people call it the first computer-generated art from a system that made its own choices within the rules, and historians file it under computational art.
Is that artificial intelligence? Barely. It is more like a very polite robot following a recipe. But it matters, because every early ai image worked this way. A human did the thinking. The machine did the drawing.
The 2014 shift: machine learning takes over
Here is the turn. In 2014, a researcher named Ian Goodfellow introduced the GAN, which stands for generative adversarial network (two programs trained against each other, more on that in a minute). Before that, people hand-coded rules. After that, a computer could study thousands of examples and work out the rules itself. That is machine learning: the machine learns patterns from examples instead of being told what to do. A Medium history of AI images traces this same break between the hand-coded era and the learning era.
The faces from that first year were small, grey, and blurry. Picture a grid of faces laid out by year. Top left: a smudge with two dark holes where eyes should be. You would not mistake it for a person. Keep scanning right and down, though, and the faces sharpen at a speed that should make you a little dizzy. A timeline of AI-generated faces on Wikimedia Commons lays that progress out in one picture.
From First AI Generated Image to Your Face: What Changed in 2014 and 2023
The first AI generated image to make headlines as ai art came in 2018. A French collective called Obvious created a smudgy, dark portrait titled Edmond de Belamy with a GAN. It sold at a famous auction house for $432,500. The "signature" in the corner was not a name. It was a line of math from the program. People argued for months about whether it was generative art or just a clever trick. Nobody argued about whether it looked like a real photograph, because it didn't.
How a GAN model works like two forgers
Here is the best way I know to picture a GAN. Imagine two people in a workshop. One is a forger who paints fake pictures. The other is an expert who tries to spot them. Each time the expert catches a flaw, the forger learns to hide it. Each time the forger improves, the expert has to look harder. They keep pushing each other, round after round, with nobody telling either one what a "good" fake looks like. This article is part of a series, start with Deepfake Voice Scam News Today 77 Of Targets Lost Money.
In technical terms, the forger is the generator and the expert is the discriminator. Both are neural networks, and GANs pair them into one network (a stack of math layers loosely inspired by the brain). The model never understands what a face is. It only learns that faces tend to have two eyes, that shadows fall a certain way, that skin has certain textures. It produces new pictures that match those patterns. That is why early fakes had melted earrings and teeth that did not quite line up. The patterns were right, and the understanding was missing.
How text prompts and image generators changed the game
Next came a big upgrade. Researchers built a system called CLIP that learned to match pictures with the words people use to describe them. Pair that with newer generative models and you get today's image generators: you type a sentence, called a prompt, and you get a picture back. This is image synthesis (building a picture from scratch instead of editing an existing one), and it is the heart of modern image generation.
In January 2021, OpenAI DALL-E showed that neural networks could turn a written caption into a picture. Its successor, DALL-E 2, arrived in April 2022 and used diffusion, a method that starts with random noise and removes it step by step until a picture appears. Stable Diffusion, released in August 2022, brought that diffusion approach to ordinary home computers. Tools like these turned AI image generation into a public product, and images created by such art generators no longer needed a research lab.
The famous test prompt was "astronaut riding a horse." Why that one? Because it is something nobody has ever photographed. If an image generator can draw it, the tool is not copying a stock photo. It is combining ideas. Early image generated results showed a horse with an astronaut's legs growing out of its back. Within a few years, the same prompt gave you a sharp, well-lit scene that looked like a movie still. And video generation is now following the same path, one step behind.
The researchers I read put the turning point at 2023. That is when an ai-generated image began to look photo-realistic, and when many people could no longer tell AI pictures from real ones. Nine years from grey smudge to something your own eyes cannot judge. That is the speed we are dealing with.
Can Detection Tools Spot AI Generated Content? Why Lab Scores Mislead
Often, yes, but not reliably, and not on the platforms where most of us actually see pictures. Detection tools can score well on clean test sets and then slip badly on compressed social media posts or on pictures from a model they have never met. That gap is the whole story of this section.
The misconception: "AI can catch AI"
It is a very natural thing to believe. If artificial intelligence built the fake, surely artificial intelligence can unmask it. Lab reports even back that up, with accuracy scores of 95 to 98 percent. I would believe it too. But those numbers come from tidy test sets, and real life is not tidy.
Deepfake detection is not foolproof, and performance can drop when media is compressed, noisy, multilingual, or generated by newer models.
From a 2026 guide to deepfake detection technology
Here is why. A detector does not really spot "fake." It spots the fingerprints a specific generator leaves behind: odd blinking, lighting that disagrees with itself, tiny pixel patterns. Then you post the picture online. The platform squeezes the file to save space, and that squeezing wipes out many of the pixel-level clues. Meanwhile a brand-new generator leaves fingerprints the detector has never seen. The detector is always studying yesterday's attack with yesterday's data. Previously in this series: Deepfake Brother In Laws Fake Photos Spark New Uk Crime.
| What the test sees | Lab setting | Your phone at 11pm |
|---|---|---|
| Picture quality | Clean, uncompressed ai image files | Squeezed, resized, screenshotted copies |
| The generator | Known generators the detector trained on | Newer generators it has never met |
| Reported accuracy | 95 to 98 percent | Down to about 65 percent |
| Best clue | Pixel fingerprints in generated content | Who sent it, and the story around it |
So where does facial recognition fit? At CaraComp, we teach how face comparison works, which means measuring how closely two faces match once each is mapped as a set of points. The lesson carries over: any tool that judges a face is only as good as the examples it has seen. Anyone selling certainty about a single photo is promising more than the science gives.
Why a Fake Image Can Be Deadly: The Harm Behind the Trivia
Now the part that stopped me cold while researching. In May 2025, a Texas teenager named Elijah Heacock died by suicide. According to CBS News, scammers had sent him an AI-generated nude photo of himself and demanded $3,000. That scheme is called sextortion (blackmail using a real or fake private picture). The FBI estimates that at least 20 young people have taken their own lives because of sextortion scams since 2021, and generative tools make the threat easier to carry out.
Notice something. The fake did not have to be good. A teenager in a panic does not zoom in and check the earlobes. The picture only had to be believable enough to frighten, and the fear did the rest. That is the real lesson of the first AI generated image story: the technology stopped needing to fool experts. It only needs to fool a scared 16-year-old for a few hours.
Other cases show how wide this is. In April 2024, a child psychiatrist named David Tatum was sentenced to 40 years in prison, according to the FBI, for using AI to create abusive images from photos of real children. In February 2026, a retired engineer in Surrey, UK, lost £20,000 of his savings, according to the Surrey Police and Crime Commissioner, after a deepfake video of a TV star promoted a fake investment. Different victims, same trick: fake generated content that borrows a real face.
What to do if a fake image is aimed at your family
Know these steps before you ever need them. Do not pay. Paying rarely ends it, and it often brings more demands. Stop replying, but keep everything: usernames, messages, screenshots. Tell a trusted adult or your spouse right away, and make sure a young person knows they will not be punished for what happened to them. Then report it to the police and the FBI, and use the report tools in your own account at the platform where it happened. Shame is the blackmailer's best tool, so take it away.
What You Just Learned About the First AI Generated Image
- The first one was hand-ruled: AARON drew from rules a person typed in, not from learning.
- 2014 was the real turning point: the GAN let machines learn patterns by competing against themselves.
- Detection looks for fingerprints: it finds a known generator's marks, not "fakeness" in general.
- Lab scores mislead: about 65 percent in the real world, so the source matters more than the pixels.
The Aha Moment: Look at the Sender, Not the Pixels
If detectors lose accuracy once a picture has been squeezed through a social feed, what hope does a worried parent have? More than you would think, because you are looking in the wrong place. Pixels are the generator's home turf. The sender, the timing, and the demand are the scammer's weak spot. Real photos do not arrive with a deadline and a payment request.
The first AI generated image began as rule-based drawing like AARON and ended up as a fake you cannot judge by eye, so when an ai image shows up with a demand attached, check who sent it and why, and report it, instead of squinting at the picture.
Recap: The first AI generated image grew from 1970s hand-ruled drawings into photo-real fakes by 2023. Since detection tools slip outside the lab, trust the source and the situation, not your eyes. Up next: First Ai Generated Image 9 Years From Blur To Perfect Fake.
So here is the question to take to bed tonight. If a fake picture of someone you love landed on a screen right now, where would you look first? Not at the earlobes. At the sender. That habit, more than any detector, is what the story of the first AI generated image has taught us to keep.
First AI Generated Image: Frequently Asked Questions
What was the first AI photo?
There is no single first AI photo everyone agrees on. Early computer images were drawings, not photographs. Blurry, photo-style faces made by learning systems first appeared in 2014 with the GAN, and by 2023 many AI pictures looked like real photographs. So the honest answer depends on whether you mean drawn art, learned faces, or photo-realistic results, and each of those has a different date and a different creator in the history of the field.
Who was Harold Cohen and what did AARON do?
Harold Cohen was an artist who began building a drawing program called AARON in the early 1970s. Some sources call him an American computer scientist, though he started as a painter. AARON followed rules he wrote by hand and produced original drawings. It is often named as the earliest example of computer-generated art made by a system that decided details on its own, even though it did not learn from examples the way later models did.
Why is astronaut riding a horse the classic test prompt?
Nobody has ever photographed an astronaut riding a horse, so a picture of it cannot be copied from a stock image. If an image generator draws it convincingly, the model has combined ideas it learned separately. Early results looked strange, with bodies merging into the animal. Newer generators, including diffusion models, draw it cleanly, which is a handy way to see how far image generation has moved in just a few years.
Is video generation as convincing as a still image?
Video generation follows the same path as pictures, but it is harder, because every frame has to agree with the next. Unnatural blinking, shifting light, and odd voice sounds have been common clues. Those clues fade as models improve. Compression on social platforms also hides small flaws, so treat any surprising video, especially one asking for money, as unproven until you confirm it through a separate channel.
Can I tell an ai-generated image by looking at it?
Not reliably anymore. In the early days you could spot melted hands or mismatched earrings. Since about 2023, many people cannot separate AI pictures from real photos by eye. Detection software helps, but its accuracy can fall to around 65 percent on real social media posts. A better habit is to check where the picture came from, who sent it, and whether anyone is pressuring you to act fast.
Is an AI portrait the same as a deepfake?
Not exactly. An AI portrait like Edmond de Belamy, a form of AI art, is a made-up picture that does not claim to be a real person. A deepfake uses a real person's face or voice to make something they never did. Both rely on the same kind of generative models and the same kind of image synthesis. The difference is intent: art invites you to look, while a deepfake tries to deceive you about who is in the picture.
Ready for forensic-grade facial comparison?
Full forensic reports with detailed similarity scoring. Results in seconds.
Run My First SearchMore Education
Google Face Recognition: 1 Switch Deletes Every Face Code
Your photos app has been sorting faces into piles for years. Learn what google face recognition actually does, where it gets people wrong, and the one setting that deletes it.
biometricsImmigration Biometrics: 30,000 Officials Can Search Prints
One fingerprint appointment can put your prints into a chain of linked government systems. Learn how the chain works, what the research shows, and where to look first.
biometricsRoblox Face Scan: Predators Rarely Need to Fool the Camera
A face scan now gates Roblox chat, but the real danger was never at the door. Learn what the scan does, the mistakes parents make about it, and what to check tonight.