Signatures of the uncanny valley effect in an artificial neural network

Robots and computer graphics characters that resemble humans but are not perfectly human-like tend to evoke negative feelings in human observers, which is known as the “uncanny valley effect.” In this study, we used a recent artificial neural network called Contrastive Language-Image Pre-training (CLIP) that learns visual concepts from natural language supervision as a visual sentiment model for humans to examine the semantic match between images with graded manipulation of human-likeness and words used in previous studies to describe the uncanny valley effect. Our results showed that CLIP estimated the matching of words of negative valence to be maximal at the midpoint of the transition from a human face to other objects, thereby indicating the signature of the uncanny valley effect. Our findings suggest that visual features characteristic to the conflicts of visual cues, particularly cues related to human faces, are associated with negative verbal expressions in our everyday experiences, and CLIP learned such an association from the training datasets. Our study is a step toward exploring how visual cues are related to human observers’ sentiment using a novel psychological platform, that is, an artificial neural network.