Can You Spot the AI-Generated Images? Distinguishing Fake Images Using Signal Detection Theory

This study explores individuals’ ability to differentiate between AI-generated and genuine human images, with a specific emphasis on different emotional states (lack of emotion, positive emotion, and negative emotion) and human behaviors (postures and activities). An experiment involving 18 participants was conducted to discern various AI-generated human images, and the results were analyzed using signal detection theory. The analysis revealed a significant variation in sensitivity (d’) based on the emotional content, with images displaying positive emotions exhibiting notably higher sensitivity compared to emotionless images. No significant sensitivity difference was observed concerning different types of behaviors. Furthermore, there were no significant variations in bias (β) in relation to emotional states and behaviors. This study holds promise for informing various user experience investigations associated with AI image generators.