AI-generated imagery : what makes the content convincing or not for human observers?

Sammanfattning

We conducted an experiment to study how and why participants classified 24 imagesas either AI-generated or man-made, and which image types were hardest to classify correctly. The four image types were People, Animals, Landscapes and Abstract. 12 imageswere man-made, and 12 were AI-generated, and each image type category consisted ofthree images each. We used eye-tracking to see how the participants observed the images and gain insight on subconscious processes, and verbal motivations to study theirconscious decision-making processes. We found that the verbal motivations and the eyetracking data often gave different results, and that the different image types had differentaspects and parts that affected the participants’ decisions most. The hardest image typesof AI-generated images to classify correctly were Landscapes, followed by Abstract, thenAnimals, and lastly People. The hardest man-made image types to classify were People,then Landscapes, then Abstract, and lastly Animals.