Near-perfect realism with one inconsistent detail is what actually feels eerie, not imperfection itself

What it is

The uncanny valley is Mori's 1970 hypothesis: affinity for a robot rises with human likeness, plunges near-but-not-quite human, and recovers only at full realism. Current evidence narrows it to a sharper statement: the valley is not produced by anthropomorphism itself but by high realism with inconsistent detail — realism lifts the perceiver's category expectation to "this is a person," and single cues (the eyes, the skin, a stiff micro-motion) that are harmless at low anthropomorphism become threatening anomalies at high fidelity. The half of that sentence that does design work is the second one: the valley is a consistency failure, not a realism failure.

Why it happens

Realism's first work is to raise the category prior: the more human the appearance, the earlier the stimulus is sorted into the human category, and the more the perceiver expects every other cue to be coupled — person perception is not the reading of single features but of a multi-channel ensemble (expression, micro-motion, skin, voice, timing). One channel falling behind is irrelevant at low anthropomorphism — plastic skin on a geometric face was never expected to be human — while the same-grade flaw at high fidelity (an eye that skips a blink, skin too smooth, a blink half a beat late) lands as prediction error against a strong prior. The candidate explanations come in a pair: category uncertainty — the stimulus fits neither the human nor the artifact box, and the suspended failed sorting is itself aversive; and expectation violation — realism raises coupled expectations that a single lagging channel breaks. A threat-avoidance line adds that near-human anomaly cues (skin, impaired motion) share an alarm system with disease detection, which is why the uncanny reading is social — "not quite right, like someone ill" — and why descriptions of the valley's bottom reach for corpses. All three accounts share one engineering conclusion: the pathology is in the inconsistency, not the realism.

Studying it

Mori's original text is a 1970 essay in the Japanese journal Energy — a hypothesis, not a measured curve — and the English translation appeared only in 2012. The dominant paradigm is a human-likeness continuum: morphs from real faces to robots/CG, likeness steps as the independent variable, affinity, warmth, and likability ratings as the outcome. Dynamic stimuli (video, physical robots, facial movement) produce valley-shaped dips far more often than stills, and one neuroimaging line found that humanlike appearance paired with mechanical motion drives abnormal responses in motor-prediction regions — support for the expectation-violation side. Two paradigm-level caveats: most studies use static face images, on which the dip is shallow or absent anyway, so static evidence cannot adjudicate the hypothesis; and the curve's shape (position and depth of the dip) is unstable across continua — Mori's original curve should not be used as a quantitative prediction.

Where it stops holding

This is the softest-evidenced claim in its family, and the boundary deserves an honest statement. The effect's existence is contested: some labs find a monotonic rise with no dip; whether a valley appears depends on how the continuum is sampled, on stimulus dynamics, and on exposure duration; and effect sizes are far smaller than the drama of the original figure. Cross-cultural claims ("the Japanese accept robots more") have not held up well in systematic comparison. A second boundary is categorical: the empirical record is almost entirely about human faces and bodies; extending "uncanny valley" to mascots, icons, or voice personas has no direct support — character carried in text does not fall into any valley, and using the term as a general risk word for anthropomorphism is a misuse. A third is contextual: passive viewing (distance, editing) forgives far more than face-to-face interaction; high-fidelity digital humans are accepted in film and games, which does not license them in support or companionship products — interaction amplifies every inconsistent cue.

Applying it

• Treat anthropomorphism as a continuous dial, not a yes/no: choose a position between "obviously artificial" and "nearly human," defaulting to the left of the valley — stylization (cartoon, geometric, exaggerated) is the safe harbor, because it openly declares "not human" and no human expectation forms.
• If you choose high realism, budget for whole-ensemble consistency: gaze and blink timing, skin, motion dynamics, voice and lip sync, micro-expressions; one lagging channel can sink the ensemble, and when the budget cannot cover every channel, downshifting is cheaper than patching.
• Evaluate with moving material: a still-image pass is not a safety certificate; acceptance material is video and an interactive mock, with attention on motion and timing rather than single-frame likeness.
• To validate: run a likeness sweep — the same character across several anthropomorphism steps, collecting warmth ratings and "creeped-out" self-reports plus free descriptions coded for creepy vocabulary; which step and which channel produce the dip is the useful output — it names the position to retreat to.

Related

• Same group: P1.10.1 Names and pronouns are the strongest anthropomorphic switch · P1.10.2 Anthropomorphism makes users read system failures as attitude problems · P1.10.3 A sense of character, once established, is hard to withdraw
• Nearby: P1.04.3 Anthropomorphism must match actual capability · P1.01.1 The visceral layer triggers immediate reactions directly from appearance
• Search terms: uncanny valley · category uncertainty · prediction error · android science