The Uncanny Valley
A reader's summary of Masahiro Mori's 1970 hypothesis that almost-human figures provoke revulsion rather than affinity — its robotics origin, the 35-year gap before an English translation existed, and why AI-generated faces and voices made it urgent again.
The hypothesis at a glance
The uncanny valley names a proposed non-linear relationship between how human a figure looks or moves and how comfortable people feel looking at it. Comfort rises with realism at first, then drops sharply for figures that are close to human but detectably off, and only recovers once the figure becomes indistinguishable from a real person. The “valley” is that dip — a zone of near-human likeness that reads as more disturbing than either a clearly artificial figure or a real one.
Origin
Masahiro Mori, a robotics professor at the Tokyo Institute of Technology, proposed the idea in a short 1970 essay titled “Bukimi No Tani” (“the uncanny valley”) published in the Japanese journal Energy. Mori wasn't writing philosophy of aesthetics for its own sake; he was giving robot designers a practical warning, illustrated with a simple hand-drawn graph, that pursuing ever-greater human resemblance could backfire past a certain point and that a more stylized, clearly non-human design might be the safer engineering choice.
History and context
The essay circulated within Japanese robotics circles for decades before it had a reliable English-language version. Karl MacDorman and Takashi Minato produced the first widely cited translation in 2005; a further collaboration between MacDorman and Norri Kageki in 2012 refined it further, correcting earlier translation ambiguities. In the gap, the concept entered English-language robotics and human–computer-interaction literature largely through paraphrase, which is part of why popular accounts of the idea often diverge from Mori's original, narrower claim.
Main ideas
A dip, not a slope
Mori's core claim isn't that more human-likeness is always better. Plotted against perceived familiarity, the curve rises with realism, then plunges sharply for figures that are almost-but-not-quite human — a corpse, a prosthetic hand, a moving humanoid — before climbing back up only once likeness becomes indistinguishable from an actual person.
Written for robot designers, not animators
Mori published "Bukimi No Tani" (the eerie valley) in the Japanese journal Energy in 1970 as practical guidance for robotics engineers: chase familiarity, not photorealism, because a robot that looks 90% human reads as ill or wrong in a way a cartoonish one never does. The essay predates CGI, video games, and deepfakes by decades.
It sat mostly unread in English for 35 years
No authoritative English translation existed until Karl MacDorman and Takashi Minato produced one in 2005, followed by a further-refined version with Norri Kageki in 2012. The concept spread through robotics and HCI conferences on secondhand summaries well before most Western researchers could read Mori's actual argument.
Movement steepens the drop
Mori's original diagram included a second, deeper curve for moving figures rather than static ones — a still prosthetic hand unsettles less than one that twitches. This detail is routinely dropped from popular retellings, which tend to treat the valley as a single fixed curve rather than one that gets worse with motion, exactly the variable that CGI and robotics both had to solve for.
No settled explanation for the mechanism
Competing accounts include a mismatch between category cues (something categorized as "not quite alive" triggers threat-detection), a violation of expected mind-behind-the-face, and evolutionary disease-avoidance responses to corpse-like or unhealthy features. None has been established as the single cause, and it is plausible several combine.
CGI gave it a second life before AI gave it a third
The term entered mainstream criticism through animated films chasing photorealistic humans — The Polar Express and Final Fantasy: The Spirits Within were both singled out for it in the 2000s. Generative AI video, voice cloning, and hyperrealistic avatars have since made the valley a live product-design constraint again, not a historical curiosity.
Critique
- Never rigorously validated by Mori himself. The original essay presented a hand-sketched curve and a set of illustrative examples, not measured data. Decades of later studies have tested pieces of the hypothesis with mixed results, and no single experiment has confirmed the exact shape Mori proposed.
- Used as a catch-all for “this looks off.” Popular usage frequently applies the term to any mildly unsettling rendering of a face, collapsing a specific claim about a near-human dip into a general synonym for bad character design or low-budget CGI.
- Culture and familiarity likely shift the curve. Comparative studies suggest reactions to near-human figures vary with a viewer's prior exposure to robots, animation styles, and cultural context, meaning the valley may not be a fixed human universal so much as a learned, adjustable response.
Impact
The concept outlived the narrow robotics problem it was written for. It is now the default vocabulary for why photorealistic CGI humans, AI-generated faces, cloned voices, and humanoid robots can misfire even when the underlying technology is objectively more accurate than a cruder predecessor. That makes it a natural companion to AI Slop, which names a related but distinct discomfort — not almost-human appearance, but low-effort, mass-produced content — and to parasocial relationships, since AI companions and synthetic media both have to navigate the same valley Mori sketched for robots half a century earlier.