The Uncanny Valley Has More Than One Cause

Published August 26, 2026

The uncanny valley is not one universal reflex hiding in the brain. It is a family of failures that can converge on the same feeling: an artificial face that will not settle into a category, a human-looking body that moves like a mechanism, a voice whose emotion does not match its words, or a familiar social signal arriving without the life that normally produces it. The practical lesson is sharper than “realistic robots are creepy.” Human likeness raises expectations. Unease appears when a design activates those expectations and then violates them.

That makes the valley less like a fixed geographical feature and more like a fault map. Different cracks open under different conditions. A wax figure, a game character, a humanoid robot, and an AI voice can all feel wrong, but not necessarily for the same reason.

Mori’s valley began as a design warning

Masahiro Mori’s 1970 essay, available in the first English translation authorized by him through IEEE Spectrum, proposed a curve relating human likeness to affinity. As a robot becomes more humanlike, our response may improve—until a near-human design produces a sudden drop into eeriness. Mori used prosthetic hands, corpses, puppets, and moving robots as examples. He also suggested that motion could deepen the drop.

The origin matters because Mori later described the idea as an intuition and advice to robot designers, not a completed scientific law. The famous graph was a hypothesis generator. Popular culture often treats it as if someone measured a universal emotional canyon and found the same coordinates in everyone. Research has produced a more complicated picture.

A 2015 review of empirical evidence found inconsistent support for the simplest claim that increasing human likeness automatically produces one predictable valley. It found stronger support for perceptual mismatch: eeriness becomes more likely when features imply incompatible kinds of entity or incompatible levels of realism. A later meta-analysis in ACM Transactions on Human-Robot Interaction found an overall uncanny-valley effect across studies while also showing that results depend heavily on how researchers create stimuli and measure response. The valley is real enough to study, but not simple enough to worship.

A five-part map of almost-human unease

1. Category conflict: what kind of thing is this?

Some artificial beings sit near a boundary: human or machine, alive or inanimate, person or object. The observer keeps trying to stabilize the answer. That ambiguity can be uncomfortable, especially when one cue says “human” while another says “manufactured.”

The category-conflict account sounds intuitive, but it is not a complete explanation. In a reappraisal of categorical perception, researchers separated difficulty identifying a stimulus from negative affect. The two did not line up cleanly enough to conclude that a hard-to-classify face must also be eerie. Ambiguity may load the weapon without always pulling the trigger.

2. Perceptual mismatch: the parts disagree

A stylized character can have enormous eyes and still feel coherent because every feature obeys the same visual world. A highly realistic face with slightly synthetic skin, rigid eyelids, or mismatched teeth can feel worse because the deviations are judged against a stricter standard. The problem is not imperfection by itself. It is inconsistency among cues.

This explains why “add more detail” is unreliable design advice. Increasing skin texture while leaving gaze, timing, and muscle deformation crude may widen the disagreement. Realism is a contract: each realistic cue promises that the next cue will belong to the same organism.

3. Prediction error: it looks human, so it should move human

Appearance prepares the observer for a style of motion. When the body looks biological but moves mechanically, the nervous system encounters a violated prediction. In an fMRI study of a robot, an android, and a human, the android’s humanlike appearance paired with mechanical motion produced the strongest mismatch response in parts of the action-perception system. The shell promised one kind of movement; the joints delivered another.

This mechanism extends beyond robots. Lip synchronization, blink timing, eye contact, conversational pauses, facial emotion, and voice prosody all create predictions. A synthetic agent can be visually convincing and still become uncanny when its timing treats social behavior as decoration rather than coordination.

4. Animacy error: a face without a living rhythm

Human observers are extremely sensitive to whether something seems alive, attentive, and responsive. A frozen smile is not merely a badly drawn mouth. It is a social signal that fails to change when the situation changes. Eyes that never quite acquire a target, expressions that arrive late, and heads that rotate without the small preparatory movements of a body can imply a mind and then withdraw the evidence.

The result is often described as deathlike, but “fear of death” should not be treated as the single master key. Corpses, dolls, masks, illness cues, and artificial agents share some visual properties, yet their meanings depend on context and learning. The safer claim is that disrupted animacy is one route into unease.

5. Social expectation: the interface claims personhood

The nearer an agent comes to human form, voice, and conversation, the more we judge it by human standards. We expect memory, reciprocal attention, emotional continuity, and sensitivity to context. A cartoon robot can fail charmingly. A photorealistic digital human that forgets the previous sentence feels less like a tool making an error and more like a person-shaped absence.

This is where the visual uncanny valley meets AI. A language model has no face, but a carefully human performance can still create a social mismatch. The issue is not that fluent software secretly becomes monstrous. It is that presentation can invite a richer theory of mind than the system can consistently support. The unease comes from the gap between the role being performed and the capacities revealed over time.

Why the valley moves

The same design will not unsettle everyone equally. Exposure changes expectations. Genre changes interpretation. A stiff face in a medical simulator, a horror game, a toy, and a customer-service kiosk carries four different promises. Cultural familiarity, individual experience, attention, and task also shape the reaction.

That is why the uncanny valley can appear to move as technology improves. Yesterday’s astonishing digital face becomes today’s obvious game asset. Viewers learn new visual grammars, while better rendering raises the standard applied to whatever remains imperfect. The target is not stationary because the observer is part of the system.

The phenomenon also overlaps with other kinds of estrangement. Liminal spaces become eerie when a place displays its purpose but lacks the people and events that complete it. Cute horror works by creating an expectation of safety and then violating it. The almost-human face follows a related grammar: recognition first, contradiction second.

A practical framework for designers

Instead of asking whether a character is “too realistic,” test the promises its design makes. Five checks are more useful than chasing a mythical percentage of human likeness.

• Category: Can viewers quickly understand what kind of entity this is, or is ambiguity intentional and meaningful?

• Consistency: Do skin, eyes, teeth, hair, voice, and movement share the same level and style of realism?

• Prediction: Does the agent move and respond the way its appearance teaches viewers to expect?

• Animacy: Do gaze, timing, posture, breathing, and expression create a coherent rhythm of attention?

• Context: Does the level of human likeness fit the role, environment, genre, and stakes of the interaction?

These checks do not demand cartoonish design. They demand coherence. Stylization can cross the valley by refusing to enter it; high realism can cross by aligning many channels at once. Horror can deliberately remain inside the fault zone, using mismatch as an artistic instrument. The right target depends on whether the design needs trust, clarity, intimacy, comedy, or dread.

The strongest counterargument

There may still be a general valley. The meta-analytic evidence suggests that highly humanlike artificial entities can produce a nonlinear drop in evaluation across many experiments. It would be equally careless to replace “one universal reflex” with “nothing but culture.” Human perception does bring shared sensitivities to faces, bodies, motion, disease, agency, and social intention.

But a recurring shape in the data does not prove one recurring cause. Several mechanisms can create similar rating curves. A doctor can observe the same fever produced by different infections; the shared temperature does not erase the diagnosis. “Uncanny” names the felt outcome. It does not finish the explanation.

The valley is a broken promise

The most useful synthesis is this: human likeness is not merely a visual scale. It is an expectation amplifier. Each human cue asks the viewer to predict more—more natural movement, more coherent emotion, more responsive attention, more evidence of an inner life. The valley opens when those predictions fail in a pattern the observer cannot comfortably ignore.

This is why the concept remains valuable long after 1970. It turns a vague shiver into a design question: what did this artificial being promise, and which signal betrayed the promise? The answer may be category, texture, motion, timing, animacy, or social behavior. Find the broken agreement and the valley stops being magic. It becomes something we can analyze, repair, or deliberately enter.

Which mismatch unsettles you most in an almost-human character—eyes, skin, movement, voice, timing, or emotional response—and what specific example made you notice it?

Explore the Cat-Coded collection for designs shaped by synthetic identity and digital strangeness, then join the Claw & Riot Salon to continue the discussion.