Abstract
Viewing artificial objects and images that are designed to appear human can elicit a sense of unease, referred to as the 'uncanny valley' effect. Here we investigate neural correlates of the uncanny valley, using still images of androids (robots designed to look human), and humans wearing hyper-realistic silicone masks, as well as still images of real humans, in two experiments. In both experiments, human-like stimuli were harder to distinguish from real human faces than stimuli that were clearly not designed to mimic humans but contain facial features (mechanical robots and Halloween masks). Stimulus evoked potentials (electromagnetic brain responses) did not show convincing differences between faces and either androids or realistic masks when using traditional univariate statistical tests. However, a more sensitive multivariate analysis identified two regions of above-chance decoding, indicating neural differences in the response between human faces and androids/realistic masks. The first time window was around 100-200 ms post stimulus onset, and most likely corresponds to low-level image differences between conditions. The second time window was around 600 ms post stimulus onset, and may reflect top-down processing, and may correspond to the subjective sense of unease characteristic of the uncanny valley effect. Objective neural components might be used in future to rapidly train generative artificial intelligence systems to produce more realistic images that are perceived as natural by human observers.