The uncanny valley is the region on a chart where a figure looks almost human but not quite human enough — and the small residual wrongness evokes unease, revulsion, or dread rather than recognition or warmth. It is the specific creepiness of a wax museum, of a too-smooth CGI face, of a doll whose eyes are the right size but whose gaze is slightly off.
The phrase comes from the Japanese roboticist Masahiro Mori, who in a short 1970 essay coined the term bukimi no tani genshō (不気味の谷現象), literally "the phenomenon of the eerie valley". His famous sketch plots the human-likeness of a figure on the x-axis against the emotional response it elicits on the y-axis. As realism grows, affinity rises — a toy robot is cute, a humanoid robot is endearing — until the figure approaches true human appearance, at which point affinity plunges into negative territory. The bottom of that dip is the valley. On the far side, a being fully indistinguishable from a living human climbs back up out of it.
What we are reacting to
There is no single agreed explanation for why the valley exists. Competing theories include:
- Threat detection. The small deviations from normal human appearance look like symptoms of disease, death, or mental illness — exactly the cues our ancestors needed to avoid.
- Categorical ambiguity. The figure is neither clearly alive nor clearly inanimate, and the mind's inability to file it cleanly registers as discomfort.
- Failed empathy. We begin to extend the full social machinery — reading expressions, inferring intent — and then something fails to come back the way it should, breaking the feedback loop.
All three are probably partially right. What unites them is that the reaction is triggered by realism itself. A cheap toy robot does not give us the creeps; a beautifully crafted mannequin with almost-right eyes does.
Everywhere, now
For decades the uncanny valley lived mostly as a concern of animators and roboticists. It explained why certain films — The Polar Express, Final Fantasy: The Spirits Within — left audiences cold, while stylised characters like Pixar's Wall-E did not. It was a cautionary graph on the wall of every CG studio.
That era is over. The valley is now a general feature of everyday life, because generative models have crossed its threshold in images, voice, and text. A face subtly produced by a diffusion model, a voice cloned from one minute of training audio, a chatbot that almost remembers who it is talking to — each occupies some point along Mori's curve, and we are collectively developing a finer-grained response to the exact type of wrongness each one produces. Six fingers. A shadow with no source. A cadence that is too smooth. A reply that is fluent and empty. These are the new uncanny valley tells.
The interesting thing is that the valley is not a fixed landscape. As we grow used to a class of synthetic artefact, the dip softens, and the creepiness moves elsewhere — to whatever is almost convincing now. The valley is less a place than a moving frontier, tracking the exact distance between what a machine can currently fake and what a human still notices.