Overview

• Quiz participants scored 65 percent accuracy on AI-generated human portraits versus 59 percent on natural scenes
• The analysis drew on nearly 287,000 image evaluations from over 12,500 global users of Microsoft’s “Real or Not” quiz
• Researchers warn that as generative models become more realistic, human performance is likely to decline further
• Preliminary tests of the lab’s own AI detector demonstrate over 95 percent success in distinguishing real from synthetic images
• The study clarifies that output realism depends more on training data than on specific model architectures