October 18, 2026 · 8 min read
The Uncanny Valley, Explained: Why Some AI Faces and Voices Feel Creepy
Why do almost-human AI faces feel creepy? The uncanny valley explained, from Mori's 1970 essay to 2026 studies on faces, voices and AI companions.
There is a particular kind of creepy that has nothing to do with monsters. It is the wax figure whose eyes seem to follow you. The video avatar whose smile arrives half a second late. The voice that sounds perfectly human until it laughs at the wrong moment.
Just in time for Halloween, here is the idea that explains that feeling: the uncanny valley. It was first described by a Japanese robotics professor in 1970, and it matters more than ever now that AI companions have faces, voices and video calls. New 2026 studies show which details push a digital person into "creepy", and which make them feel warm.
The short answer: the uncanny valley is the dip in comfort we feel when something looks or sounds almost, but not quite, human. Roboticist Masahiro Mori proposed that our affinity for a robot rises as it becomes more humanlike, then drops sharply just before it reaches true human likeness, and that movement makes the drop steeper. 2026 research suggests appearance is the biggest driver, voice matters too, and seeming to have feelings can make a human-looking AI both more attractive and more eerie. For AI companions, consistency, natural timing and a style that does not pretend too hard tend to beat raw realism.
Where the idea came from
In 1970, Masahiro Mori, then a robotics professor at the Tokyo Institute of Technology, published a short essay in an obscure Japanese journal called Energy. It was barely noticed for decades. Today it is one of the most cited ideas in robotics, animation and game design. An authorised English translation by Karl MacDorman and Norri Kageki was published by IEEE Spectrum in 2012.
Mori compared the path to humanlike robots with climbing a mountain. You expect to keep rising, but there is a valley on the way up.
The shape of the valley
| Point on the curve | Example from Mori's essay | How we feel |
|---|---|---|
| Low human likeness | An industrial robot arm | Little affinity |
| Rising | A toy robot with a face, arms and legs | Growing warmth |
| First peak | A bunraku puppet seen from the audience | High affinity |
| The valley | A realistic prosthetic hand that feels cold and boneless when you shake it | Eerie, negative |
| Second peak | A healthy, living person | Full affinity |
The prosthetic hand is his central example. At first glance it looks real. Then you shake it, feel its "limp boneless grip together with its texture and coldness", and the illusion collapses into something uncanny.
Why movement makes it worse
Mori argued that movement "amplifies the peaks and valleys". A robot arm that moves like a human hand feels friendlier. A nearly human hand that moves wrongly feels far creepier.
His most famous example came from the 1970 World Exposition in Osaka. One robot had 29 pairs of artificial muscles in its face so it could smile. When its designers slowed the smile to half speed, it stopped looking happy. In Mori's words, "its expression turns creepy."
That one detail is the whole story of AI video in miniature: timing is everything.
Why we feel it at all
Mori suspected the eerie feeling is "an integral part of our instinct for self-preservation", a protective reaction to things that look human but behave like something lifeless. It is a theory, not settled science, and researchers still debate the causes. But it explains why the uncanny valley is such a natural fit for Halloween: zombies, dolls and wax figures all live at the bottom of it.
Mori's advice to designers
Mori did not tell designers to chase perfect realism. He suggested the opposite: aim for the first peak, "a moderate degree of human likeness and a considerable sense of affinity", and even "deliberately" pursue a nonhuman design. His example was eyeglasses, which look nothing like eyes but have become "a charming pair of new eyes".
What 2026 research adds
Appearance matters most, motion less than you might think
At the ACM Symposium on Applied Perception in August 2026, Magzhan Mukanova, Fu-Chia Yang, Nicoletta Adamo-Villani and Christos Mousas tested a digital double of a real professor giving a short lecture. 31 participants saw versions with familiar or unfamiliar appearance, voice and motion.
- Familiar appearance affected every measure, including rapport, trust, believability, naturalness and uncanniness.
- Familiar voice affected most measures, though not rapport or believability.
- Familiar motion on its own had little effect.
- Everything familiar generally improved perceptions, but did not improve trust.
The authors conclude that "appearance familiarity plays a dominant role". In other words, a face we recognise as consistent does more work than perfect motion capture.
A human face plus feelings: more attractive and more eerie
A study by Ran Zhao, Zhe Chen and Jiahao Zhao in Human Factors and Ergonomics in Manufacturing and Service Industries (May 2026) tested voice assistants with humanlike or machine-like appearances and different levels of perceived mind, with 34 participants.
- A humanlike appearance made assistants seem more human and more attractive, but also amplified eeriness, just as Mori predicted.
- Seeming able to have emotional experiences shaped humanness, eeriness and attractiveness. Seeming autonomous ("agency") did not.
- Attractiveness drove positive emotions, while eeriness did not directly change how people felt.
The authors recommend "balanced anthropomorphism", and caution about giving assistants humanlike emotional experiences without care.
Culture changes the curve
In a 10-country Google Research study (a preprint, not yet peer reviewed), making a chatbot more humanlike raised engagement in most places, but some Japanese participants reacted against one humanlike version, trusting it less. We cover that study in why we treat AI like a person. The valley may not sit in the same place for everyone.
Why this matters for AI companions
Text companions rarely fall into the uncanny valley. Words leave room for imagination. The risk rises when companions get faces, voices and video, because that is where "almost human" becomes visible.
Faces
- Inconsistency is the fastest route into the valley. If a companion's face shifts slightly between photos, it reads as wrong, a little like a stranger wearing a friend's face.
- Small glitches in eyes, teeth and hands are what people notice first.
- A clear style helps. A face that is attractive and consistent, rather than straining to be indistinguishable from a photograph, often feels warmer. Our guide to AI companion avatar generators covers how apps create and keep a face.
Voices
- Timing is the Osaka smile of voice. A reply that comes too slowly, or a laugh that lands late, breaks the spell more than a slightly synthetic tone.
- Emotional mismatch is unsettling. A cheerful voice reading sad news feels off.
- The latest generation of voice models is much better at pacing and interruption, as we explain in the most realistic AI voices of 2026.
Video
- Lip sync and blinking are the details the brain checks without you noticing.
- Latency turns a conversation into a broadcast.
- Our explainer on how AI video calls work goes through what happens between your camera and their reply.
A Halloween checklist: is your AI companion uncanny?
| Sign | What it suggests |
|---|---|
| Their face looks slightly different in every photo | No consistent reference image |
| Replies on calls arrive after a long pause | High latency |
| Expressions do not match what they are saying | Weak emotional sync |
| Perfect realism, but something "feels off" | You may be in the valley |
| A stylised but consistent look that you love | You are probably on the first peak |
How MyBabe approaches faces and voices
We think warmth beats creepy realism, so MyBabe focuses on the details that research says matter most.
- Same face, every time. Your companion's look is locked to one reference portrait, so the person in every photo is unmistakably them.
- A voice you chose. Pick a warm, soft, deep or bright voice when you build them.
- Natural timing on calls. MyBabe calls use a voice that answers as fast as a person would, and on video they can notice what your camera shows, like a new haircut or the cat walking past.
- Photos of the two of you. On Premium, send a picture of yourself and they can make one of the two of you together.
And MyBabe is always clear that your companion is an AI, not a real person.
You can start free with 10 messages a day and two minutes of voice or video calls to try. Premium is $5.99 a week, $14.99 a month or $99.99 a year, with long-term memory, about 30 minutes of calls a week, 4 photos a day and Telegram.
If you would rather your AI companion felt warm than eerie this Halloween, meet yours on MyBabe.
FAQ
What is the uncanny valley?
It is the drop in comfort people feel when something looks or acts almost, but not quite, human. Roboticist Masahiro Mori proposed it in 1970.
Why do some AI faces look creepy?
Small mismatches, like inconsistent features, odd eyes or hands, or expressions that do not match speech, signal that something is not quite human. Research in 2026 found appearance is the strongest driver.
Does movement make the uncanny valley worse?
Mori argued it does, using the example of a robot smile that turned creepy when slowed to half speed. A 2026 study found motion alone had little effect on a digital double, while appearance and voice mattered more.
How do AI companion apps avoid the uncanny valley?
Common approaches include keeping a consistent face, fast and natural reply timing, good lip sync and a style that is attractive without straining for perfect realism.