Eye Appearance and the Uncanny Valley

I've long been fascinated by the uncanny valley concept and its theoretical foundations. Theorists propose that the uncanny valley is a disgust of human-like, but not-quite-human figures that stems from survival instinct — an avoidance mechanism tied to cues of death or disease.

I explored facial perception research to identify a specific pain point in artificial faces likely to trigger this response, then designed an experiment to test my hypothesis.

The theory proposes that as artificial human renderings approach realism, they reach a point that triggers strongly negative emotional responses. The phenomenon was originally proposed by roboticist Masahiro Mori in 1970; the term "uncanny valley" was introduced through a Jasia Reichardt translation in 1978.

Research shows that humans fixate on the eye region 40% of the time when viewing a forward-facing face (Janik et al., 1978). Chimpanzee eye-tracking studies (Hirata et al., 2010) confirmed this pattern extends across species — attention centered on the eyes even when faces were rotated or inverted.

I hypothesized that if emotional connection flows primarily through the eyes, then eyes alone may be sufficient to trigger the uncanny valley response.

Participants rated images on a 5-point scale ranging from "I HATE this" to "This is human." Eleven stimuli were created by morphing real video footage with a 3D uncanny digital human model at varying degrees of digital eye opacity.

Eleven test stimuli were produced by morphing original stock footage with a 3D facial rig built in Autodesk Maya at increments from 10% to 100% digital eye opacity.

A local browser app built with p5.js displayed 11 looped 8-second videos in random order. Participants clicked one of five buttons to rate each clip — from "I HATE this; this is not human" to "This is a human." A printed scale was placed over the keyboard; only the trackpad was used to navigate.

Experiment 2 used male stock footage to account for gender bias. Two updates were made to the app: response time (time between image load and button click) was now logged, and a start screen was added so participants were not exposed to the first stimulus during instructions.

Ratings were converted to numerical values, averaged per image, and graphed against Mori's original uncanny valley curve. A notable pattern emerged: responses to earlier images in a session trended toward neutral, likely due to comparison effects from preceding stimuli.

Experiment 2 introduced a response-time graph. Individual data sheets showed higher response times and more neutral ratings on first-viewed images which were expected to balance out across averaged trials.