High but imperfect realism can trigger discomfort

What it is

The uncanny valley, or valley of eeriness, is Mori's 1970 hypothesis: affinity for an artificial figure generally rises with human-likeness, but drops sharply when the figure is close to human yet detectably imperfect, before recovering once realism becomes nearly indistinguishable from an actual person — forming a "valley." It describes a possible nonmonotonic relationship, not a law every humanoid design must pass through.

Why it happens

Researchers offer a few candidate explanations, not one confirmed answer: one is category-uncertainty — the appearance sits on the human/nonhuman classification boundary, forcing the perceptual system to judge ambiguous input, and that uncertainty itself triggers a threat- or disgust-like response; another is mind-perception mismatch — a highly realistic appearance sets an expectation of matching mental and emotional capability, and once motion or detail reveals the figure cannot deliver on that, the gap between the "promise" of the appearance and the "delivery" of the behavior is what produces discomfort. It must be stated plainly that the evidence behind this hypothesis is mixed, not settled: Mori's original curve was proposed speculatively, not fitted from systematic measurement, and later replications disagree across rating scales, static-image versus video stimuli, culture, and participant age — whether a stable valley exists at all, and where its floor falls on the realism axis, remains unresolved.

Studying it

Experiments typically manipulate the same character's morphology along a continuous realism gradient and measure affinity, eeriness, categorization time, gaze, and physiological response (e.g., skin conductance), then test with a nonlinear model whether a genuine valley appears rather than a monotonic trend in either direction. Identity, attractiveness, and rendering quality must be held constant, or the effect cannot be attributed to realism itself; a core methodological difficulty is that realism resists reduction to a single dimension (pixel-level similarity? behavioral human-likeness?), and different studies operationalize it differently, which is part of why replication has been hard.

Where it stops holding

Findings depend on the stimulus set, rating scale, culture, and familiarity, and morphing continua used to generate stimuli can create artifact-like eeriness that would not occur in an actual robot. A valley found for one stimulus set does not establish a universal threshold, and not every aversion to a robot should be filed under the uncanny valley — some discomfort has other causes, such as a capability mismatch tied directly to appearance.

Applying it

When iterating toward higher realism, test appearance, voice, motion, and response timing together for coherence, and keep several style-consistent prototypes at different realism levels rather than pushing a single line toward maximum realism. Blind-test affinity and eeriness ratings across versions, plot the curve, and diagnose the specific source of any spike (the overall form, or one local detail); if the design cannot cross the valley reliably, do not paper over a structural mismatch with finer surface texture — consider a stylized route instead.

Related

• Same group: X6.02.2 Unnatural motion triggers discomfort more readily than appearance does · X6.02.3 Stylized design can sidestep this zone
• Nearby: X6.03 Emotional expression · X1.02 Robot type and expectations
• Search terms: uncanny valley · anthropomorphism · Godspeed questionnaire · human-robot interaction