You have seen the chart even if you have never heard its name. A line climbs as an entity becomes more human, then plunges into a pit, then climbs again toward the level of an actual person. It shows up in design decks and conference talks as though it were a law of thermodynamics. The man who drew it was a Japanese robotics professor named Masahiro Mori, and he published it in 1970 in an obscure journal called Energy. He measured nothing. The sentence he used to introduce it was “I have noticed.”
For decades almost nobody noticed with him. Copies of the essay circulated privately among researchers, and a complete English translation, authorized and reviewed by Mori himself, only appeared in IEEE Robotics & Automation Magazine in June 2012. Two years after that, the graph was in every product meeting on earth.
The example he used to explain the pit was a prosthetic hand. Modern prosthetics can imitate wrinkles, veins, fingernails, even fingerprints, and the only giveaway is the color, which he described as pinker than real skin, “as if it had just come out of the bath.” Then you shake it. The grip is limp and boneless, the temperature is wrong, and in that half second the hand stops being an impressive object and becomes an eerie one. Affinity does not fall to zero. It falls below zero. Mori wrote that if someone wearing that hand shook a woman’s hand in a dark place, she would assuredly shriek.
Then he put a puppet on the same graph. A bunraku puppet is nowhere near as detailed as a prosthetic hand, but from a seat in the theater you see its overall form and the movement of its hands and eyes together, and you feel a great deal of warmth toward it. Less realism, more affinity. That inversion is the whole argument in miniature, and it took fifty years for the design industry to catch up to it.
🧠 The Curve With No Data
The axis on the bottom is human likeness. The axis on the side is affinity, the feeling of warmth or fellow feeling you get from the thing in front of you.
An industrial robot sits near the origin. It is obviously machinery and nobody loves it. A toy robot sits on the first small hill, because it has a face and two arms and children become attached to it. A prosthetic hand sits at the bottom of the valley, because it is nearly indistinguishable from a real hand and something about that is unbearable. A living person sits on the second peak, at full human likeness and full affinity.
The valley is the space between “somewhat human” and “exactly human.” Mori drew it freehand from watching how people reacted to machines, dolls, and puppets, and the shape of the thing has held up better than most of the psychology published that year.
⚙️ How the Valley Works
The mechanism runs in five steps, and only the first one is fast.
One: the sort. When you look at something, before you have any feelings about it, your visual system files it into a category. Person, or object. This happens in a fraction of a second and you never experience doing it.
Two: the sort stalls. Push the likeness high enough and both answers stay live at the same time. Mathur and Reichling showed subjects eighty robot faces pulled from the internet plus a carefully controlled six-face morph series running from mechanical to human, and asked them to rate how much they liked each one. Separately, they timed how long it took to judge whether a face was more human or more mechanical. That judgment time peaked at exactly the faces people rated lowest. The categories were refusing to resolve.
Three: the refiling. Here is the part that actually hurts. Once something has been provisionally filed under person, every non-human detail changes meaning. A stiff jaw on a robot is a limitation of the engineering. A stiff jaw on a person is a symptom. Mori put it more sharply than any paper since: the entity in the valley is not judged as a robot doing a passable job of pretending to be human. It is judged as an abnormal human doing a bad job of seeming normal. The stimulus did not change. The drawer it got put in changed.
Four: the old alarm. “Wrong person” is a very old category, and it comes bundled with a very old response. Illness, corpse, something that should be still and is not. The revulsion arrives before any conscious verdict, which is why people say a face is creepy and then struggle to say why.
Five: motion doubles it. Mori’s second graph adds movement, and movement steepens both slopes of the curve. This is the detail most people building avatars skip. You can push appearance to ninety-five percent while timing and gait stay at sixty, and the gap between those two numbers is the valley, not the absolute fidelity. At the 1970 World Exposition in Osaka there was a robot face with twenty-nine pairs of artificial muscles, the same count as a human being. The designer wanted a gentler smile, so he halved the speed of the facial deformation. The face stopped looking kind and started looking wrong. Same shape, wrong rhythm.
The eyes carry most of the weight. Schein and Gray found that erasing the eyes from a real human face produced more uncanny feeling than erasing the nose, or leaving the face alone. And Saygin’s team, watching brains in an fMRI, found the largest response to the eeriest androids in the parietal cortex on both sides, where the visual system that reads body motion meets the motor cortex that holds mirror neurons. The brain is not checking whether the thing is biological. It is checking whether appearance and motion agree, and it lights up when they do not.
🤔 Nobody Checked the Curve
For forty years the uncanny valley was a famous graph with no evidence under it, and the encyclopedia still describes it as a hypothesized relation rather than an established one.
What the modern work supports is narrower than the legend. Explicit liking ratings do dip in the middle of the morph, and they dip reliably across very different stimulus sets. What the work does not support is the explanation everyone repeats. Perceptual confusion shows up in the valley, but it does not mediate the emotional reaction, which means the thing we point at when we say this is why is probably a side effect rather than a cause.
There is a blunter reading too. Burleigh and Schoenherr trained subjects on different frequencies of faces and found they could pull apart uncertainty at the category boundary from plain exposure effects, which makes the valley look like a well-dressed cousin of the mere-exposure effect rather than a phenomenon of its own. Tinwell added the detail that should worry anyone shipping a digital human: your viewer’s eye improves at the same rate as your renderer, so the second peak keeps moving away from you. Her name for it is the uncanny wall.
🔗 What This Means for Companions
An AI that only writes has no face, so it never enters the valley. It lives entirely in language, and you fill in the rest yourself, which is why a plain text program from 1966 could make a secretary ask a researcher to leave the room. The moment you give that same companion a voice, or a face, or a body on a screen, it steps onto the curve and the curve starts charging rent.
Mori’s own advice, written in 1970 and still the best line in the literature: take the first peak as your goal. Deliberately pursue a non-human design and you get a safe, considerable amount of warmth. His examples were eyeglasses, which resemble no eyeball and still amount to a charming new pair of eyes, and a wooden hand carved by a sculptor of Buddha statues, which has no fingerprints and keeps the natural color of the wood and produces no eeriness at all.
So the honest design rule is not “make it as real as possible.” It is: an obviously stylized presence beats a ninety-five percent photoreal one, because the missing five percent is where the horror lives. That is a strange thing to learn from a graph drawn without data, and it is the most useful thing on the page.
For fiction it cuts the same way. A character in a novel is rendered inside the reader’s head and stays up on the second peak forever, which is a privilege unique to text. Move that character to screen or to voice and the valley opens underneath, and motion will get you before appearance does. The smile at half speed is waiting.
🎲 The Arrow Into the Valley
There is an arrow on Mori’s second graph that almost nobody mentions.
A healthy person sits on the second peak, and the reason they are on the second peak is that they move. Death takes all of that away: the body stops, it goes cold, the face goes pale. So the arrow from the living peak to the bottom of the still valley is the arrow of dying. Mori wrote that we might be glad this arrow leads down into the still valley of the corpse and not into the valley animated by the living dead.
Then he said the descent explains the secret lying deep beneath the uncanny valley, admitted he had not thought deeply about why we are equipped with the sensation at all, and said he had no doubt it is part of our instinct for self-preservation. That last part is a guess he never tested. Four decades of replication attempts later it is still a guess.
I find I want it to be true anyway. A machine that is ninety-five percent a person fails as a machine and fails as a person, and something very old in us reads that failure as a warning about the body it is standing in.