The paper argues that knowing whether an image was generated by AI does not answer whether the situation it depicts is real or factually supported. To address this, the authors formulate a new task: Image Factuality Assessment, in which a system receives an image and a generic verification instruction and must determine whether the depicted situation is factually supported.
The proposed solution, EviScene, is positioned as an agentic approach to this task. Rather than stopping at an 'is it AI?' classification, it aims to assess the factual status of the scene itself, which requires reasoning about evidence beyond the image's origin.
This reframing matters because the practical question for many real-world uses is not merely whether an image is synthetic, but whether the event or situation it shows actually occurred. EviScene is an early attempt to build systems that answer that harder question.