Climbing Out of the Uncanny Valley

Dave has black eyes, like a doll’s eyes………..

Maybe you winced or did a double take before you clicked on the tile that brought you to this column. If you didn’t, click the back arrow and check the tile again. I asked Adobe Firefly — an image-generating AI — simply to “create a tile (square) icon of a human silhouette climbing out of a valley; make the image professional-looking and not cartoonish.” I didn’t ask for Picasso’s Cubism style. I didn’t ask for three arms and four hands. I definitely didn’t ask for two hands jutting out 180 degrees apart from each other on the same wrist.

If you think that was anomalous, here are two other images the AI created from the same prompt.

Yes, the first climber has an improvement of only three hands (one for each arm), and the second climber is mostly hands-appropriate. But, he is snorting an arm out of his face.

Altogether now: EEEEUUUUWWWWW!!!

That response is exactly what’s predicted by the uncanny valley hypothesis. Coined by engineer Masahiro Mori in a 1970 essay, it predicts a drop in attraction to artificially created human faces, figures, or speech patterns when they move from something plainly machinelike to something almost human. Although there are era- and culture-based differences in personal reactions, reviews and meta-analyses broadly support this prescient Big Idea^™. At the valley floor of uncanniness, our brains react with nearly automatic discomfort or revulsion. Obvious robot WALL•E is undeniably cute. Simulated human Actroid-F is unnervingly creepy. Beyond robots, there are scads of other examples from entertainment. The animated characters in “The Polar Express”, the CGI “Cats,” and the facsimile of Carrie Fisher in “Rogue One” all tried to be humanlike. They were unintentionally horrifying instead.

Studies of the uncanny valley hypothesis concentrate mostly on faces. For those of us doing research on what happens in and around business organizations, however, the uncanny valley is relevant for user reactions to AI-generated images, voice, and text.

• Generative visuals are improving rapidly, but viewers still notice nuanced anomalies. Bad hands are legion (rim shot No. 1). Odd reflections and lighting continuity are subtle but detectable. When the content looks nearly real, these defects can create the valley feeling.
• Synthetic speech can be clear and expressive, yet still stumble in micro-timing, breath patterns, or emotional cadence. Those small mismatches matter because our decades of communication experience make us extremely sensitive to speech cues.
• A narrative written by AI can feel unsettling when it is fluent but oddly frictionless. It’s likely to have no typos, no hesitations, no minor inconsistencies, and none of the idiosyncrasies in expression we each have. It still might say “delve.”

Of course, AI outputs continually get more sophisticated. Text-based AI has already passed the Turing test. As it becomes harder to reliably tell what is AI-generated, the source of creepiness may shift. Early uncanny reactions have come from obvious glitches. Soon, users may react to the opposite: content that looks, sounds, or reads as if it’s too polished — lacking the random imperfections typical of human output. For instance, on a recent trip to a church gift shop, all the for-sale illustrations I saw of Jesus looked exactly like Bradley Cooper in “A Star is Born,” but with no razor burn.

The uncanny valley won’t shimmer away. Its depth might flatten and its terrain might shift from obvious weirdness to subtler cues about authenticity, intent, and trustworthiness. In the meantime, therefore, the basic rule of AI still holds. Take care (when consuming it) and declare (when applying it).

I can’t pretend to tee up Nathan Barrymore’s column by referring to AI. But I will say it’s about getting a hand (rim shot No. 2) up in sponsorship from one’s manager. Nathan’s Big Ideas^™ provide both systematic evidence of, and a compelling solution for, de-biasing how bosses differentially advocate for their male and female employees.

Dr. David A. Harrison
Associate Dean for Research