The research gathered data from an online game called the “Real or Not Quiz,” which had participation from over 12,500 people worldwide. This diverse group analyzed around 287,000 images, which were a random mix of both real and AI-generated visuals.
The participants’ task was to identify which images were authentic and which were fake. The findings revealed that the overall success rate of the participants was approximately 62 percent, a result that is only marginally better than chance, akin to flipping a coin.
Interestingly, the study indicated that identifying fake images of faces was somewhat easier than discerning the authenticity of landscape images. However, the improvement in accuracy was minimal, amounting to only a few percentage points.
In response to these findings, Microsoft has called for clearer labeling of AI-generated images to assist in distinguishing them from real photos. Despite this advocacy, critics have raised concerns about the proposed clarity, noting that cropping the images can easily circumvent labeling efforts and mislead viewers.
This research highlights the ongoing challenges in distinguishing between real and AI-generated content, underscoring the need for better tools and approaches to tackle image authenticity in an increasingly digital world.