Rethinking the uncanny valley as a moderated linear function: Perceptual specialization increases the uncanniness of facial distortions

Abstract

The relationship between artificial entities’ human likeness and aesthetic preference is thought to be best modelled by an N-shaped cubic “uncanny valley” function, which however suffers from conceptual criticisms and lack of parsimony. Here it is argued that uncanniness effects may instead be modelled by a linear function of deviation moderated by perceptual specialization. The two models are compared in an experiment with five incrementally distorted face types (cartoon, CG, drawing, real, robot). Recognition performance for upright and inverted faces were used as a specialization measure. Specialization significantly moderated the linear effect of distortion on uncanniness, and could explain the data better than a conventional uncanny valley. The uncanny valley may thus be better understood as a moderated linear function of specialization sensitizing the uncanniness of deviating stimuli. This simpler yet more accurate model is compatible with neurocognitive theories and can explain uncanniness effects beyond the conventional uncanny valley.