Does the uncanny valley require AI?

No. It can concern a designed character or robot regardless of how its behavior is generated.

Is realism always undesirable?

No. The hypothesis does not make that claim.

Can a still image and an animation receive different reactions?

Yes. Motion introduces additional features to evaluate.

AI Glossary / Reinforcement learning and robotics / Advanced

UNCANNY VALLEY

Definition

The uncanny valley is a proposed pattern in people's responses to humanlike artificial characters or robots: increasing human resemblance may initially increase affinity, but a nearly human appearance can sometimes produce unease before a convincingly human appearance restores affinity. The term describes a hypothesis about perception and response, not a universal law or a measure of AI intelligence.

What it means

A stylized character can be appealing without looking realistic. Adding realistic skin, eyes or movement does not necessarily make that character more comfortable to interact with. Near-human details may raise expectations that other features fail to meet.
The valley metaphor refers to a dip in an imagined graph relating human likeness to affinity. It does not specify a fixed percentage of realism at which discomfort begins. Different designs, tasks and audiences can produce different responses. A single disliked avatar is therefore insufficient evidence that an uncanny valley has occurred.

How it works

Roboticist Masahiro Mori proposed the idea in a 1970 essay. His conceptual account included both appearance and movement, suggesting that motion could intensify responses. The original curve was a hypothesis rather than a fitted equation that predicts every user's reaction.
For practical evaluation, a team can compare several designs and separately measure perceived human likeness and responses such as comfort or eeriness. The measures should be defined clearly: affinity, trust, attractiveness and willingness to use a service are related questions, but they are not interchangeable outcomes.
Comparisons should also consider what changes between designs. If the most realistic character has poor lip synchronization while the stylized one has smooth animation, realism and synchronization are confounded. That result cannot by itself establish which feature caused discomfort. Testing still images alone may also miss responses that emerge during conversation or movement. Design decisions require evidence from the intended interaction.

Three practical examples

Digital presenter: a team compares a stylized presenter with more realistic versions, checking whether facial animation fits the voice and whether viewers find the result comfortable.
Reception robot: designers assess appearance and motion together. A friendly-looking face does not establish that the robot's gaze, gestures or conversational timing will feel appropriate.
Training simulation: a project evaluates virtual characters with its intended learners. Visual realism is considered alongside comprehension, task performance and user response rather than treated as the sole quality target.

Think of it this way

Consider a cartoon face with simple dots for eyes. You accept its simplified expression because the design sets that expectation. A photographic-looking face with oddly timed blinking may attract more attention to the mismatch. This analogy illustrates a possible reaction; it does not mean everyone dislikes realistic characters or that all discomfort has the same cause.

Worked example

Consider five fictional character designs. Their average perceived human-likeness scores on a one-to-seven scale are 1, 3, 5, 6 and 7. Their average comfort scores on a separate one-to-seven scale are 3, 5, 3, 2 and 6. Higher numbers mean more humanlike or more comfortable, respectively.
As likeness rises from 1 to 3, comfort rises from 3 to 5. At likeness 5 and 6, comfort falls to 3 and 2. At likeness 7, comfort rises to 6. These stipulated averages draw a valley-shaped pattern: an initial increase, a decline and a recovery. The drop from comfort 5 to 2 is three scale points; the recovery from 2 to 6 is four points.
Now consider a second fictional set of comfort averages: 3, 4, 5, 5 and 6 at the same likeness scores. This sequence has no downward valley. Merely including highly realistic characters does not force the proposed pattern to appear.
Neither set constitutes research evidence. There are no actual participants, uncertainty estimates or controlled comparisons behind these illustrative values. Even with real participants, five averages alone would conceal how responses vary and whether the designs differ in unintended ways. Rating scales are also not physical measurements of realism or emotion.
Suppose a real study changed animation timing only for the fourth design. Its low comfort score could reflect timing, the appearance-motion combination or another difference. A follow-up could compare the same appearance under matched animation conditions. The calculation teaches how to recognize a proposed pattern and why a graph alone does not establish its cause.

Illustrative case study

Fictional company Cedar Learning is choosing a guide for an online training program. It prepares a stylized character and several more realistic versions, all delivering the same lesson. The team checks animation quality and collects separate responses about comfort, understanding and the guide's perceived role.
If a realistic version receives lower comfort ratings, the team investigates specific features before labeling the outcome an uncanny valley. It also checks whether learners understand that the guide is artificial. This imagined workflow reports no real study or measured learning improvement.

Common misconceptions and mistakes

The uncanny valley does not mean that more realism always makes a design worse. The proposed pattern includes recovery beyond the dip. Nor does every awkward robot, unnatural voice or disliked interface demonstrate the full pattern.
Another mistake is assuming that a comfortable character is competent, truthful or trustworthy. Appearance and behavior can influence impressions without establishing actual capability. It is also misleading to use the concept to judge people's appearance or disability: this entry concerns responses to designed artificial representations. Finally, the original hypothesis should not be presented as a precise numerical threshold or a universally settled explanation of human emotion.

Business use

The concept encourages teams to test realistic avatars and robots before committing substantial design resources. Relevant outcomes include user comfort, task success, comprehension and expectations about the system's capabilities. A simpler visual style may be a useful option, but its value should be assessed in context rather than assumed. Design testing should include the interaction people will actually experience, including speech and motion when applicable.

Check your understanding

Why was the first numerical example valley-shaped?

Answer: Comfort rose, fell and then recovered as perceived likeness increased.

Why was the example not evidence?

Answer: Its values were invented for teaching.

Why measure comfort separately from competence?

Answer: A favorable impression does not establish actual performance.

Related terms

Human-robot interaction
Anthropomorphism
Avatar
Android
Humanoid robot
Social robotics
User experience

Learn next

Compare controlled character variants, define the response measures and examine individual variation before interpreting an average curve.

Sources

Masahiro Mori, The Uncanny Valley, 1970, authorized English translation published in 2012: https://spectrum.ieee.org/the-uncanny-valley . IEEE Spectrum ROBOTS glossary: https://robotsguide.com/learn/robotics-glossary . The ratings and company scenario are fictional teaching illustrations.

Previous: Unbounded cost search Next: Uncertain data