What is the uncanny valley? Why almost-human feels wrong

The uncanny valley is the idea that people warm to things the more human they look, until those things become almost human, at which point warmth turns sharply into unease. The Japanese roboticist Masahiro Mori proposed it in 1970. Plotted as a graph, comfort climbs, plunges into a "valley" of eeriness for near-human robots and faces, then climbs again for real people. What causes the dip, and how general it is, is still argued.

In short

• Mori's curve plots how much affinity people feel against how human something looks, with a dip just short of fully human.
• Experiments find the dip under some conditions, especially when parts of a face don't match each other, but not as a single universal curve.
• AI-generated still faces now fool most viewers, so the valley has moved to motion, timing and voice.

What did Mori actually claim?

Mori, an engineer at the Tokyo Institute of Technology, set out the idea in a short essay, "Bukimi no tani", in the Japanese journal Energy in 1970. The first English translation he authorized, by Karl MacDorman and Norri Kageki, appeared in IEEE Spectrum in 2012. The English name "uncanny valley" was in use long before then.

His graph has two axes. Across the bottom is human likeness, from nothing like a person to a healthy human. Up the side is what he called shinwakan, usually translated as affinity: how warm or at ease a person feels toward the thing. An industrial robot arm sits near the start, with little likeness and little feeling either way. A toy robot is halfway up the first hill. Then the line drops. Mori placed a realistic prosthetic hand near the bottom: it looks real until you shake it and feel a limp, boneless, cold grip, and the feeling of affinity turns negative. On his graph a corpse sits at the bottom of the valley: dying, he wrote, is a fall into it.

He made two further claims. Movement, he said, steepens the curve, deepening the valley as well as raising the peaks: a moving prosthetic hand is eerier than a still one. And he gave designers advice: aim for the first peak, with a deliberately nonhuman design, rather than risk falling into the valley while chasing the second. He wrote that he had "no doubt" the eeriness was "an integral part of our instinct for self-preservation," though he admitted he had not yet considered the question deeply, and the essay offered no data for it.

A walk along the curve

Mori's own examples were robots, prosthetics and puppets. Walking newer examples along his curve shows how the idea is used today, and where it strains.

Example	How human it looks	Typical reaction
Industrial robot arm	Barely	Neutral: it is a machine
Cartoon character or toy robot	Simplified, with big eyes and few details	Warm: the first peak
Humanoid robot with rubbery skin	Close, but stiff, with eyes that don't track	Uneasy: the slope into the valley
Computer-animated human in a film	Very close, with lifeless eyes or skin	Often uneasy: critics have complained of "dead eyes" in films that aimed for realism
AI-generated still photo of a face	Indistinguishable from a photo	Usually taken as a real person, so no dip
AI-generated talking face or voice	Close, with small slips in timing, blinking or breath	Uneasy once the slips are noticed

The last two rows matter most now. The valley isn't fixed to a level of realism; it sits wherever the cues stop matching. A still face that passes for real gives the viewer nothing to catch. Put the same face in motion, and a lip movement half a beat off from the voice can bring the unease back.

What have experiments found?

More than for many famous ideas, the tests have been mixed. In 2015, Jari Kätsyri and colleagues at Aalto University reviewed the experimental evidence and found little support for the simple version, a valley that appears whenever anything is made more human: only one of the eight studies they examined found it. Most found instead that liking rose fairly steadily with human likeness. Mori's claim that movement deepens the valley wasn't supported either. Their conclusion was that "the uncanny valley exists only under specific conditions."

Some studies do find a clear dip. In 2016 the researchers Maya Mathur and David Reichling had people rate 80 real-world robot faces, plus a controlled set of edited faces, and found a valley in how likable they were. In an "investment game", people also trusted near-human robots less with money, which suggests the effect reaches decisions as well as feelings.

Why would almost-human feel wrong?

There is no single agreed explanation. The main ones, and how well each has held up:

• Mismatched cues. Realistic skin with cartoonish eyes, or a human face with oddly proportioned features, sends signals that conflict. Kätsyri's review found this the best supported: mismatched realism produced the effect in four of four studies.
• Category confusion. People are unsettled by not knowing whether something is human or machine. Studies show people do sort faces into categories sharply, but the review found little evidence that this doubt is what makes them dislike the face.
• A sense of mind. Kurt Gray and Daniel Wegner found in 2012 that machines become unnerving when people think they can feel, not when they think they can act. A machine described as able to feel was unnerving even without a human appearance.
• Avoiding disease and death. Some researchers argue that a face with subtle flaws resembles a sick or dead person, and that disgust evolved to keep people away from both. This fits Mori's corpse example, but it is hard to test directly.

Still argued

Evolutionary explanations, disease avoidance and reminders of death, are often presented as settled; they aren't. They are plausible stories with limited direct evidence. The perceptual-mismatch account has the strongest experimental support, but researchers still disagree about whether the uncanny valley is one effect or several, and whether it fades as people get used to near-human faces.

Why do AI faces and voices sometimes feel off?

Because still images crossed the valley before anything else did. In a 2022 study in PNAS, Sophie Nightingale and Hany Farid showed 315 people real photos and faces made by an AI image generator. Their accuracy at telling them apart was 48.2%, close to the 50% expected from guessing; after training on what to look for, it rose only to 59.0%. Another group rated the AI faces 7.7% more trustworthy than the real ones. The authors wrote that synthesis engines had "passed through the uncanny valley."

That is a problem as well as a feat. A face that triggers no unease also triggers no suspicion, which is what makes deepfakes effective. Where unease does remain, it is usually in motion and sound: generative AI video can still slip in the timing of a blink or a breath, and a cloned voice can be right in every word and wrong in rhythm. Mori's warning still applies to design. Many assistant robots and chatbot avatars are made deliberately cute or abstract, a bet on the first peak. Looking human and seeming human are separate tests: the Turing test asks whether a machine's conversation can pass, while the uncanny valley asks about its face and body. Both sit inside the larger story of artificial intelligence.

Check yourself

• Draw Mori's curve from memory: what goes on each axis, and where do a toy robot, a prosthetic hand and a healthy person sit?
• Why did the Kätsyri review conclude the valley exists "only under specific conditions"?
• Explain why an AI-generated still face can pass when an AI-generated talking face may not.

Sources

• The Uncanny Valley: The Original Essay by Masahiro Mori — Masahiro Mori, translated by Karl F. MacDorman and Norri Kageki, IEEE Spectrum, 2012
• A review of empirical evidence on different uncanny valley hypotheses — Jari Kätsyri and others, Frontiers in Psychology, 2015
• Navigating a social world with robot partners: A quantitative cartography of the Uncanny Valley — Maya B. Mathur and David B. Reichling, Cognition, 2016
• Feeling robots and human zombies: Mind perception and the uncanny valley — Kurt Gray and Daniel M. Wegner, Cognition, 2012
• AI-synthesized faces are indistinguishable from real faces and more trustworthy — Sophie J. Nightingale and Hany Farid, PNAS, 2022
• Uncanny valley — Wikipedia