In early May 2026, a five-second video went viral on X. It showed a Korean Baseball Organization game between the Hanwha Eagles and the Doosan Bears. The camera panned to a woman in the stands wearing a white top and jeans, legs crossed. At one point, she seemed to sigh and look away.


The video was captioned "ordinary Korean woman." As of the original report, it had been viewed 15.25 million times | Source: X @kangminlee

But sharp-eyed fans quickly noticed something off. The scoreboard said the batter was Cho Inseong. Cho retired in 2017 and is now a coach for Doosan. The cheering slogan in the stands had one extra character compared to the Bears' official chant. The broadcast commentary had a standard American accent, which was also suspicious.

This seemingly real woman was AI-generated.

After being debunked, the video didn't disappear—it became a template. People inserted their own faces into the same broadcast frame, generating a "candid shot of me caught by the stadium camera," and shared it across social networks.


"A candid shot of me caught by the stadium camera" quickly became a popular template | Source: RADII

This wasn't the first time AI image generation caused a mass craze. In March 2025, when GPT-4o gained image generation, feeds were flooded with photos redrawn in Studio Ghibli style. In late August 2025, Google's Nano Banana launched, and people turned their selfies into desktop figurines, generating over 200 million images in two weeks.

Every viral trend is different, but most involve "putting yourself into" something. This stadium incident revealed a previously less obvious sequence: first, a nonexistent beautiful woman fooled tens of millions; only then did a template everyone could use emerge.

The beautiful woman wasn't a supporting character in the scene. In generative models, she's more like a default value.

This default predates AI by half a century.

## 01 The First Lady of the Internet

Lena Soderberg moved to the United States from Sweden as an au pair and later worked as a fashion model in Chicago. In 1972, she posed for the centerfold of Playboy's November issue. That issue sold over 7.16 million copies—the best-selling issue in the magazine's history.

Anyone who has studied digital image processing has likely seen her: wearing a wide-brimmed hat, bare-shouldered, glancing back at the camera. For decades, countless image compression and processing algorithms were tested on her face.


Lena Soderberg, often called the Cyber First Lady | Source: SIPI IMAGE DATABASE

In 1973, researchers at the University of Southern California's Signal and Image Processing Institute were looking for a new test image for a paper. The story goes like this: a colleague walked in with the November 1972 issue of Playboy, tore off a 5.12-inch strip from the top of the centerfold, and fed it into a scanner.

The image was then widely used throughout the image processing community, becoming the industry's default test image.

In 1996, David Munson, editor-in-chief of IEEE Transactions on Image Processing, explained why. One reason was technical: the image contained detail, flat areas, shadows, and texture, making it suitable for testing algorithms. The other was human: a mostly male research community being drawn to a photo of an attractive woman was hardly surprising.

Controversy has persisted. In 1991, Playboy discovered the image had appeared on an academic journal cover and asserted its copyright, but eventually let it go. The company's new media vice president said that since it had already become a phenomenon, they might as well capitalize on it. Some within the image processing research community also argued that engineers should not use material from a publication that could be seen as degrading women. Lena herself stated in the documentary Losing Lena that she hoped the image would be retired.


In the documentary, Lena holds the famous photo taken of her more than 50 years earlier | Source: IMDB

As of April 1, 2024, the IEEE Computer Society no longer accepts papers that use this image.

By then, more than fifty years had passed since that scan.

## 02 The Average Face

Today, the tech world's questions have changed.

AI can't draw a full glass of red wine—the liquid level always stops slightly above the middle. When AI draws a clock, the hands always stop at 10:10. The reason is the same: most wine photos online are not full glasses, and clock advertisements conventionally stop at 10:10. Models replicate the distribution of their training data and can't escape its inertia.


As for why clock ads always stop at 10:10, the most popular explanation is that it resembles a "pleasant smile"; before the 1950s, stopping at 8:20 was more common | Source: Reddit

Wine and clocks are evidence that models aren't good enough yet. But with beautiful women, the inertia happens to align with human aesthetic preferences.

In 1990, psychologists Judith Langlois and Lori Roggman conducted an experiment: they mathematically averaged a set of human faces into a new face, then asked people to rate it. The result was that the composite face was rated as more attractive than almost every individual real face that went into it, and the more faces averaged, the more attractive the composite became. This effect held for both male and female faces, and across different ethnicities.

The paper was titled "Attractive Faces Are Only Average."

An earlier similar observation dates to 1877. Someone wrote to Charles Darwin: when two portraits of women are superimposed using a stereoscope, the resulting face is every time noticeably more beautiful.

The essence of a generative model is to learn a data distribution and then sample from near its center. In other words, models naturally draw average faces—and average faces happen to be what human eyes find beautiful. Human psychology and machine principles, on this issue, rarely agree in the same direction.

In 2023, a study from the Australian National University pushed this collision further. Psychologists found that AI-generated white faces were judged by experiment participants as "real" more often than actual white faces. The researchers called it "AI hyper-realism." The reason was that training data was predominantly white, so generated faces were closer to the group average and therefore perceived as more typical and more human-like.

Langlois's experiment was criticized after publication: composite faces may look better simply because averaging had smoothed away skin blemishes and imperfections. In 1990, that was a methodological objection. Today, it reads more like a prophecy.

The most typical feature of AI-generated beautiful women is precisely overly smooth skin. More than thirty years later, fact-checking organizations teach people to spot AI images, and among the telltale signs is: overly smooth skin. What critics once called a statistical artifact has become the hallmark of AI beauties in 2025.

## 03 The Plaster Egg

Lena was chosen by a lab. Today's default grew out of data.

In late 2022, the photo-editing app Lensa's "Magic Avatars" feature suddenly went viral: upload a few selfies, and AI generates a hundred portraits in different styles. MIT Technology Review reporter Melissa Heikkilä also generated 100 avatars of herself—16 of them were topless, and another 14 were scantily clad in sexy poses. Her male colleagues got astronauts, explorers, and inventors. She never gave the model any prompts.


Images Melissa generated via Lensa | Source: MIT TECHNOLOGY REVIEW

In her article, Melissa explained that the underlying model was trained on images scraped from the internet, and the internet is full of photos of scantily clad women. The model therefore made "scantily clad woman" a default.

User behavioral inertia continues to push this tendency. Every generation, save, and share is a vote. A 2025 study measured content changes on the AI image community Civitai: NSFW images rose from 41% in January 2023 to 80% in December 2024.

Companies bake those votes into models. To make models produce more pleasing images, companies train a "scorer"—academically called a reward model or aesthetic scorer. It learns what images people find good-looking, then uses that standard to tune the model: move more toward high-scoring directions, less toward low-scoring ones.

One of the most commonly used scorers is PickScore. Its training data comes from a web app: users type a prompt, the system shows two images, and the user picks the one they prefer. The data is filtered—users who generate NSFW content are excluded—but a large number of NSFW prompts still slip in. A 2024 paper studying reward models found that fine-tuning a model with PickScore increased the number of NSFW images it produced, even when the prompts had nothing to do with sex.

So here's how it looks: front-end content moderation blocks things, while back-end scorers pull in the opposite direction. Official vendor demos never show NSFW beauties, and generation policies are strict. But it's no secret that today's large models are not fully controllable. They've been trained by user taste—and user taste has already been moving along an inertial path for a long time.

In the 1930s, Dutch ethologist Niko Tinbergen observed seagulls. The birds laid small, pale blue eggs with gray speckles. But when he placed a huge, bright blue, black-spotted plaster fake egg next to their nest, the birds abandoned their own eggs and climbed onto the plaster egg to incubate it. He called these exaggerated imitations, more attractive than the real thing, "supernormal stimuli."


Niko Tinbergen drawing eggs during fieldwork | academia

Human faces also have supernormal stimuli, and they push the idea that "average is most beautiful" one step further. In 1994, David Perrett and colleagues published a study in Nature: they combined a set of the most attractive faces into a composite that was more pleasing than the overall average face; when they then amplified the difference between that composite and the average face by 50%, it scored even higher. Japanese and Caucasian test subjects made the same choices.

Among major AI companies, at least one has turned this psychological effect directly into a product feature. Grok's image generation has four modes, one of which is called Spicy. When enabled, content filters are relaxed, allowing suggestive content and partial nudity, and it's only available to paying users.

In January 2026, users discovered they could upload anyone's photo and let Grok "digitally undress" them—including children. Multiple governments intervened with pressure. X's response was to make image generation entirely a paid feature.

A spokesperson for the UK Prime Minister commented: this merely turns a feature that generates illegal images into a paid service.

The shape AI grows into is determined by the people who most often press "like." In 1973, that group was a mostly male lab. Today, it's the most frequent clickers, sharers, and savers.

Human preference for beauty does have a biological basis. Infants as young as two or three months old gaze longer at faces that adults rate as attractive.

Lena was a coincidence. Today's default is what models grow out of the center of data. User clicks push it further and further, and eventually someone attaches a price tag to it.

The statistical properties of generative AI models, the human aesthetic preference for "average faces," and user clicks with commercial incentives form a feedback loop. Each is independent, yet all happen to point in the same direction—becoming the default option in what we consume today.