10 July 2026 · 8 min read · Incidents · Brand safety

Shipped without looking: what three years of public AI-creative fails have in common

A third arm in a mayoral platform. Nineteen angry men. “KODAK ROMTOE” on a billboard. The catalogue of AI marketing fails keeps growing — and it is one story, told over and over.

There is now a genre of news story that did not exist three years ago: a brand publishes an AI-generated image, the internet finds the defect within hours, and the brand quietly deletes it. The examples below are all public, reported, and recent — and reading them together reveals something more useful than schadenfreude. The failures cluster into a handful of defect classes, every one of them was visible to a careful human eye before publishing, and in almost every case the audience was the first careful human eye the image ever met.

The catalogue is getting long

It started as a curiosity. In June 2023, readers of a Toronto mayoral candidate’s platform spotted a woman with a third arm jutting from under her armpit in an illustration — the campaign admitted the AI use and edited the image out. By 2024 the genre had A-list entrants: A24’s posters for Civil War were widely criticized for a three-door car, Chicago’s Marina Towers on the wrong sides of the river, and a giant swan that was probably meant to be a paddleboat. Amazon’s Freevee displayed a poster for 12 Angry Men with nineteen men — smudged faces, missing fingers — before pulling it, and months later Prime Video was mocked again over an AI poster for the 1922 classic Nosferatu, fingers melting into clothing under an impossible moon.

It has not slowed down. In February 2025, Marvel’s Fantastic Four crowd poster was accused of being AI-generated after fans found a hand waving a flag with four fingers and the same woman’s face twice. In April 2026, Manchester Airport said it would remove a giant billboard celebrating photography — on which the AI-generated tourists held cameras branded “KODAK ROMTOE” and “Kobak Pomtoe”. In May 2026 alone: the Arizona Cardinals were savaged for an AI schedule-announcement video, the St. Louis Cardinals posted an AI rivalry graphic that depicted a bear cub in an accidentally NSFW situation, and a Massachusetts House candidate published an AI-processed endorsement photo in which the endorsing senator’s face was unrecognizable and his hand had four fingers.

It is always the same few failures

Line the incidents up and the variety disappears. Hands and anatomy lead: the third arm, the missing fingers, the melting fingers, the four-finger flag. Then object-and-count logic: nineteen jurors instead of twelve, a car with three doors, landmarks rearranged, a moon that blocks clouds behind it. Then text: “KODAK ROMTOE”, the misspelled, googly-eyed snack ads that pushed New Jersey brand Chookie back to handmade marketing, and the Glasgow “Willy’s Chocolate Experience”, whose AI website promised “cartchy tuns” and “exarserdray lollipops” before the sparsely decorated warehouse behind it went viral. And finally judgment failures, where nothing is anatomically wrong but everything is: the accidentally lewd bear cub, the uncanny mascot video.

These are, not coincidentally, the exact failure classes AI image models are known for — the same ones Chekr scans as text, anatomy and object physics categories. Generators have improved every year of this timeline, and the incidents have kept coming anyway, because a lower defect rate multiplied by exploding volume still ships defects. The constant is not the model. It is that nobody looked.