学术文献库

In recent years, deep learning methods have become increasingly capable of generating near photorealistic pictures and humanlike text up to the point that humans can no longer recognize what is real and what is AI-generated. Concerningly, there is evidence that some of these methods have already been adopted to produce fake social media profiles and content. We hypothesize that these advances have made detecting generated fake social media content in the feed extremely difficult, if not impossible, for the average user of social media. This paper presents the results of an experiment where 375 participants attempted to label real and generated profiles and posts in a simulated social media feed. The results support our hypothesis and suggest that even fully-generated fake profiles with posts written by an advanced text generator are difficult for humans to identify.