What Is the Uncanny Valley? Why Almost Human Is Worse Than Obviously Not

Mori, a robotics professor at the Tokyo Institute of Technology, published a short essay in an obscure journal describing a graph with human likeness on one axis and a quality he called shinwakan, usually translated as affinity or familiarity, on the other. The curve rises through industrial robots, toy robots and humanoid robots, plunges into negative territory around prosthetic hands and corpses, and rises again to a healthy person. He added a second curve for moving figures, arguing that motion amplifies everything, so a moving figure in the valley is far more disturbing than a still one, which is why a zombie sits at the bottom. He offered practical advice to designers, which was not to aim for the far peak but to stop at the first one and make robots clearly non-human and appealing, on the grounds that the risk of falling into the valley outweighs the gain. The essay was not translated into English in full until 2012, by which time the term had been in use for decades.

Several hypotheses compete and the evidence supports parts of each:

• •Disease avoidance, which proposes that a figure with subtle abnormalities triggers the pathogen-detection responses that make us avoid sick people, and which predicts the reaction should be disgust rather than fear
• •Mortality salience, which reads an almost-human figure as reminiscent of a corpse and therefore as an unwelcome reminder of death
• •Categorical uncertainty, the claim that the discomfort comes from the figure resisting classification as either human or not, and that any boundary case is unsettling
• •Perceptual mismatch, a more specific version holding that the problem arises when features from different levels of realism are combined, such as photorealistic skin with cartoon eye movement, so that each feature sets up expectations the others violate
• •Prediction error, framed in terms of the brain continuously predicting what a face will do next and registering a strong error signal when a highly human-looking face moves wrongly, which fits the finding that motion matters more than appearance
• •Threat to human distinctiveness, the suggestion that a machine convincingly resembling a person is unsettling because it undermines a category we are invested in

The effect is real and messier than the clean curve implies. Brain imaging studies have found activity in areas involved in predicting others' actions when participants view an android whose appearance and movement mismatch, which supports the prediction error account. Studies morphing between a real face and a synthetic one reliably find a dip in ratings near the boundary. Monkeys have shown an aversion to realistic synthetic monkey faces, which argues against explanations requiring human self-concept. Infants show related effects from around twelve months. The complications are equally consistent: the dip is not a single curve but varies enormously between individuals, with some people showing almost none, it depends heavily on which features are mismatched rather than on overall realism, and a substantial number of studies fail to reproduce a clean valley at all. The eyes are implicated repeatedly, and the difficulty of rendering them convincingly has a physical cause, since real eyes are wet, translucent and constantly making small movements that are hard to reproduce and immediately noticed when wrong.

The commercial consequences have been expensive. Several animated films aiming at photorealistic humans were criticised for characters that audiences found dead-eyed and unpleasant, most famously The Polar Express in 2004, and the industry response was to retreat towards stylisation, which is why the most successful animation studios keep their human characters deliberately cartoonish while rendering fur, water and cloth with total realism. Game development hit the same wall, and the standard solution is either strong stylisation or an enormous investment in facial capture. Robotics has largely taken Mori's advice, which is why commercial social robots are small, rounded and obviously mechanical with simple expressive faces, and why the very realistic androids that exist are research projects and exhibition pieces. The dip also appears in synthetic voices, where a nearly natural voice with wrong prosody is more irritating than an obviously robotic one, which has shaped how voice assistants are designed.

Two things may flatten the curve. The first is technical, since synthetic humans generated by recent systems are frequently indistinguishable from photographs, which puts them past the valley rather than in it, and the consequences are now a problem of deception rather than discomfort. The second is habituation: there is evidence that repeated exposure reduces the effect, and a generation that grows up among synthetic faces may simply not experience the dip, in the way that early cinema audiences reportedly found moving images unsettling in ways nobody does now. Against that, the valley may be doing something useful, since a reliable signal that something is almost but not quite a person is exactly the discrimination that becomes valuable when convincing fakes are cheap. The practical question for designers has shifted accordingly, from how to climb out of the valley to whether a figure should be marked as synthetic at all.

Mori proposed in 1970 that affinity for a figure rises with human likeness, drops sharply just short of indistinguishable and recovers afterwards, with movement amplifying the effect. Competing explanations invoke disease avoidance, reminders of death, categories that resist classification, mismatched levels of realism and prediction error, and imaging evidence supports the prediction account. The dip varies greatly between individuals and depends more on which features clash than on overall realism, and it has pushed animation and robotics towards deliberate stylisation.