As generative AI becomes more advanced, some see it as machines taking over, but a recent experiment shows humans remain central to AI art in unexpected ways. The process, akin to the party game 'Eat Poop You Cat', involves translating text to images and back, revealing how AI interprets human prompts.
In summer 2023, a researcher used Stable Diffusion and Clip Interrogator to play two rounds of the game. Starting with the prompt 'Eat Poop You Cat', Stable Diffusion generated an image of a tabby cat with green eyes and distorted text. Clip Interrogator then described the image with a long, surreal list of descriptors including 'epic urban background', 'golden toilet', and 'exploitable image'.
A second round produced an image of a cat on a Frank Lloyd Wright-style toilet, described as 'a painting of a cat sitting on a toilet, PlayStation 2 gameplay still, in style of pop-art'. The prompt included references to artists like Ignacy Witkiewicz and Phil Foglio, as well as terms like 'molecular gastronomy' and 'tarot card design'.
The researcher noted that Stable Diffusion maps language to visual variables, while Clip Interrogator does the inverse, reading pixels and correlating them with terms. The seemingly random outputs reflect neural networks' opaque interpretation of images, often including cultural references like trap artist 'Destroy Lonely' or the vacuum cleaner 'Roborock'.
This interplay highlights that AI image generation is not purely mechanical but involves human-like associations and errors, keeping humans at the centre of the creative process.
