We Tricked AI Into a Game of Telephone—Here’s What Happened
1 min read
Summary
AI image-generating models are unpredictable and can imperfectly interpret complex prompts, producing erroneous images, opines author Andrei Rudenko after experimenting with Microsoft’s Bing Image Creator.
Rudenko tested the model using prompts avoiding the names of the objects, instead describing their shapes and other qualities, but found results were still inaccurate.
With more complex photographs, the model failed to understand the prompts, creating promblematic results including incorrectly interpreting a Rubik’s cube and failing to generate the correct image twice for a 1920s industrial photo.
Therefore, the fallibility of current AI image generation lies with the model itself, rather than human prompts, argues Rudenko.
