Overview

• Researchers from Microsoft AI for Good Lab analyzed 287,000 image assessments from more than 12,500 participants in the Real or Not quiz.
• Participants achieved an overall success rate of 62% when distinguishing AI-generated images from real photographs.
• Detection accuracy varied by content, with faces identified correctly 65% of the time compared to 59% for landscapes and only 45% for images from GANs.
• AI-based detection tools currently outperform human participants but continue to produce errors across diverse image types.
• The study team is building a new AI detector that it says will exceed 95% accuracy on both real and synthetic images.