Uncanny Valley

012 min

Where the uncanny valley shows up

A team building a customer-service avatar, a game studio animating a lead character, or a robotics lab designing a companion robot's face all eventually face the same design choice: how human should this thing look? The intuitive answer is "as human as we can make it" — more realism should read as more trustworthy, more relatable, more finished. Past a certain point that intuition breaks. A character that is nearly photorealistic but not quite right reads as unsettling in a way a clearly cartoonish one never does, and the discomfort can undo months of work on the rest of the product.

The failure is specific enough to be testable, not just "it looks bad." Users describe the same kind of wrongness across very different products — an eerie stillness in the eyes, a texture that looks like skin but does not move like it, a voice that is expressive in the wrong places. The uncanny valley names this dip and gives teams a reason to stop chasing realism as a single, always-better dial.

The problem has only gotten more relevant since Mori's essay, not less. AI image and video generation has made photorealistic synthetic faces cheap to produce at scale — a marketing team can generate a spokesperson, a support product can animate a talking avatar, a video-call tool can offer a synthetic stand-in for a user's camera feed. Each of these puts a synthetic human directly in front of a real one, which is exactly the situation the uncanny valley predicts trouble for, and it means the design choice Mori raised for robots in 1970 now applies just as directly to software that never touches a physical body at all.

023 min

How the uncanny valley works

Mori's original argument is built around a simple graph: human likeness on one axis, a viewer's sense of affinity — comfort, warmth, familiarity — on the other. Affinity rises steadily as a robot or character becomes more human-looking, the same way it rises for an industrial robot arm once it is covered in skin-like material and given a bit of fleshy plumpness. But somewhere before full human likeness, the line does not keep climbing. It drops sharply into negative territory — the valley — before climbing again only once the design becomes indistinguishable from an actual human being.

Mori illustrated the drop with a concrete, checkable example rather than an abstract curve. Modern prosthetic hands can look, at a glance, like real hands, down to wrinkles, veins, and fingernails.

That handshake is the mechanism in miniature: a viewer's perceptual system briefly categorizes the hand as human, then receives evidence — a boneless grip, an unnatural coolness — that contradicts the category, and the contradiction itself is what reads as eerie, not the hand's appearance on its own. Movement makes the effect sharper rather than milder. An industrial robot arm switched off just looks like machinery; the same arm animated to move like a human hand starts to earn some affinity, but a prosthetic hand already near the bottom of the valley becomes more unsettling, not less, once it starts to move, because movement gives the perceptual system more chances to catch the mismatch.

Mori pointed to a real, dated case of exactly this failure mode. Robot builders working on displays for the 1970 World Exposition in Osaka gave one robot 29 pairs of artificial muscles in the face — matching a human's own count — to produce a humanlike smile. Slowing the movement down, to make the smile build more gradually, did not make it read as gentler.

Mori also offered his own explanation for why the sensation exists at all, tying it to mortality rather than to robots specifically. A living, healthy person sits at the peak of affinity while moving; death removes movement and changes color and texture, sliding the same body down into the valley. Mori reasoned that a sensitivity to that exact transition would have had real evolutionary value: "I have no doubt it is an integral part of our instinct for self-preservation." On that account, the uncanny valley is not really a quirk of robot design — it is a corpse-and-injury detector, misfiring on an artificial hand or a slow robotic smile because both trip the same perceptual alarm a genuinely unwell or dead person would.

031 min

Where the term comes from

Masahiro Mori, then a robotics professor at the Tokyo Institute of Technology, wrote the essay "Bukimi no Tani" ("The Uncanny Valley") for the Japanese journal Energy in 1970. It went almost unnoticed outside Japan for decades — there was no complete, author-approved English translation in wide circulation. That changed in 2012, when Karl MacDorman and Norri Kageki produced a translation reviewed and authorized by Mori himself, published in IEEE Robotics & Automation Magazine under the same title used ever since.

Mori wrote as a working roboticist thinking about design goals, not as an experimental psychologist reporting data — the original essay proposes the curve as a hypothesis to guide robot design, and does not report a controlled study behind it. That distinction matters for how much weight to put on the idea: it began as a design intuition from someone who built robots for a living, and only decades later did psychologists start testing whether the effect shows up reliably in controlled conditions.

041 min

The Polar Express

The clearest mainstream case is The Polar Express, a 2004 animated film that used motion-capture technology to render photorealistic human characters. Critics converged on the same complaint almost immediately.

That review, from CNN's Paul Clinton, was not an outlier — other reviewers independently reached for the word "eerie" to describe the same characters. The technology behind the film was, by the standards of 2004, an impressive leap in realism: real actors' performances mapped onto CGI human models with unprecedented fidelity. The uncanny valley explains why that leap in realism produced worse audience comfort rather than better — the characters were human-like enough to be judged by human standards, and close enough to fail that judgment in a way a more stylized film's characters never would have been.

052 min

Designing around the uncanny valley

For a game studio or animation team, the practical response to the uncanny valley is not "never get close to realism" — plenty of photorealistic characters work fine — it is to avoid landing in the narrow band of near-human-but-detectably-wrong. One route is to stay stylized: exaggerated proportions, simplified textures, and character designs that read as clearly non-human buy a team a lot of goodwill, because nothing about them invites a human-accuracy comparison in the first place. David Hanson's robotics work demonstrated the opposite route works too: adding neotenous, cartoonish features — the large eyes and soft proportions associated with young animals — to an otherwise realistic robot face measurably restored affinity that a more literal design had lost.

For a team building a voice assistant or chatbot with a synthetic voice or an animated face, the same principle applies across modalities, not just visual ones. A voice that is highly natural but has one wrong prosody pattern, or a face that renders skin realistically but has flat, unmoving eyes, can trigger the same discomfort a visual near-miss does. The fix is usually to make the mismatched dimension less realistic to match the rest, rather than trying to perfect the one detail that is currently giving the design away.

Mori himself argued for the stylized route, using an example any designer can picture: I recommend that designers instead take the first peak as their goal, which results in a moderate degree of human likeness and a considerable sense of affinity. Eyeglasses do not resemble real eyeballs, but one could say that their design has created a charming pair of new eyes. The point generalizes past hands and faces: a product does not need to hide that it is artificial to be liked, it needs a design language that is confident about being artificial, the way glasses never try to pass as eyes.

For a product manager scoping a synthetic-human feature — an AI avatar, a virtual presenter, a digital human for customer support — the uncanny valley is a reason to user-test the character at multiple realism levels before committing, rather than assuming the most technically advanced rendering is automatically the best received one.

062 min

How solid is the uncanny valley, really

The uncanny valley is best treated as a well-supported hypothesis, not a settled law with a fixed shape. A 2016 study by Maya Mathur and David Reichling tested the idea directly, using two separate sets of face images ranging from clearly mechanical to fully human.

That is real, controlled evidence the effect exists — but it is also evidence that the effect is not uniform across every measure researchers care about; likability dipped reliably, while trust dipped only under some conditions. Other researchers have pushed further and questioned whether "the uncanny valley" is really one phenomenon at all.

Two specific complications are worth knowing. First, the effect may not be fixed across audiences: younger generations, more used to computer-generated imagery (CGI), robots, and such, may be less likely to be affected by this hypothesized issue, meaning a design that reads as uncanny to one audience may read as unremarkable to another raised on more synthetic media. Second, the original curve was a hand-drawn illustration of an idea, not a measured function — treating its exact shape, or the specific point where the "valley" begins, as a precise, universal threshold goes further than either Mori's essay or the later research supports.

072 min

Uncanny valley vs. nearby concepts

Anthropomorphism is the broader tendency to attribute human traits, emotions, or intentions to non-human things — a car with a "face," a robot vacuum with a name. The uncanny valley is a specific, negative reaction that can happen only once anthropomorphism has already succeeded enough for a viewer to perceive something as almost-human. A stick figure or a boxy robot can be anthropomorphized happily, with no uncanny reaction at all, precisely because it never gets close enough to human to invite the comparison that produces the dip.

Mirroring, in a design or product context, refers to interfaces or systems that reflect a user's own behavior or input back to them. It has no inherent connection to human likeness or realism, and a system can mirror a user's actions perfectly while looking nothing like a person. The overlap is only conversational: both concepts get invoked when a product's human-like framing is under discussion, but mirroring is about behavior and feedback, while the uncanny valley is specifically about appearance and the discomfort of near-but-not-quite human form.

Deepfakes are a separate, adjacent concern that the uncanny valley does not cover on its own. A deepfake's central risk is deception — a viewer believing a synthetic video is a real recording of a real event. The uncanny valley is about a viewer correctly recognizing something as synthetic and reacting with discomfort anyway. A well-made deepfake that fools its audience produces no uncanny reaction at all, precisely because the audience never registers it as artificial; the uncanny valley only applies once the synthetic nature of a face or voice has been detected, consciously or not.

?5 questions

Questions people ask

§3 sources

Sources and further reading