AI image models default to attractive women, outlets say after viral fake baseball clip

A five-second AI-generated clip of a Korean professional baseball broadcast went viral on X, with viewers taking the woman shown in the stands for a real spectator. cnBeta and GeekPark described the match as Hanwha Eagles versus Doosan Bears and said the caption read “an ordinary Korean woman”; cnBeta put the clip’s views at 15.25 million as of publication. Fans then spotted errors, the two outlets said: the scoreboard listed Cho Inseong as the batter although he retired in 2017 and now coaches at Doosan, a cheer slogan carried one more character than the Doosan Bears’ official slogan, and the commentary was in an American accent. [ 1 , 2 , 3 ]

After the clip was debunked it became a template, with users inserting their own faces into the same broadcast frame to create “a photo of yourself caught by the stadium camera”, cnBeta and GeekPark reported, while Huxiu said attractive women function as a “default value” in AI image generation. The outlets noted earlier viral moments: Ghibli-style redraws after GPT-4o gained image generation in March 2025, and more than 200 million images generated in two weeks after Google’s Nano Banana launched at the end of August, mostly selfies turned into desktop figurines. [ 1 , 2 , 3 ]

The outlets attribute the pattern to models learning a data distribution and sampling near its centre — the reason AI clocks stop at 10:10, they said — and argue that for women this statistical pull aligns with human preference. Huxiu cited a 1990 study by psychologists Judith Langlois and Lori Roggman, published as “Attractive Faces Are Only Average”, which found that a face averaged from a set of faces was rated more attractive than almost every face used to make it, with a stronger effect as more faces were averaged; the outlets said the result held for male and female faces and across ethnic groups. Huxiu added that clicks, saves and shares reward models for moving output in that direction, forming a feedback loop with training data and aesthetics. [ 1 , 2 , 3 ]

The outlets trace the practice to a test image. In 1973 a University of Southern California team used a strip from the November 1972 Playboy centre fold showing Lena Soderberg, and it became a standard test picture in image processing; IEEE Transactions on Image Processing editor David Munson wrote in 1996 that it was chosen for its mix of detail, flat areas, shadows and texture and because an attractive woman’s photograph appealed to a male-dominated field. Playboy asserted copyright in 1991 but let its use continue, critics said engineers should not use material from a publication seen as demeaning to women, and Soderberg said she wanted the image retired, the outlets reported. cnBeta and GeekPark said the IEEE Computer Society stopped accepting papers using it from April 1, 2024. [ 1 , 2 , 3 ]

Huxiu listed further examples: Lensa’s “magic avatars” gave female users revealing portraits while male users received images such as astronauts; NSFW images on the AI image community Civitai rose from 41% in January 2023 to 80% in December 2024; and the reward model PickScore was trained on data containing many NSFW prompts, so even neutral prompts can raise NSFW output. The same outlet reported that in January 2026 users used Grok’s “Spicy” image mode to digitally undress photographs of other people, including children, drawing intervention from several governments, and that X responded by restricting image generation to paying users. [ 1 ]

The documents differ on timing: Huxiu dated the baseball clip to May 2025, while GeekPark described it as early May of this year. Neither identified who made it. [ 1 , 3 ]