Uncanny valley
Last reviewed
Sources
24 citations
Review status
Source-backed
Revision
v1 · 4,257 words
Improve this article
Add missing citations, update stale details, or suggest a clearer explanation.
Suggest edit
CtrlK
Last reviewed
Sources
24 citations
Review status
Source-backed
Revision
v1 · 4,257 words
Add missing citations, update stale details, or suggest a clearer explanation.
The uncanny valley is a hypothesised relation between the human-likeness of an entity, typically a robot or a computer-generated character, and the emotional response it evokes in human observers. As the entity becomes more human-like, observer affinity rises until the entity is nearly but not quite human, at which point affinity drops sharply into a "valley" of revulsion, eeriness, or unease, before rising again as the entity becomes truly indistinguishable from a real person. The hypothesis was proposed by Japanese roboticist Masahiro Mori in 1970 in a short essay titled "Bukimi no tani," published in the Japanese journal Energy. It remained a niche idea for decades and was rediscovered in the 2000s as android research, performance-capture cinema, and photoreal video games made the question of how close to human is too close newly urgent.
The valley lives mostly in popular discourse, where it gets invoked freely, and partly in the empirical literature, where the picture is messier than the catchphrase suggests. Some studies find a clean U-shaped curve, others find linear discomfort, and the size of the dip depends on what stimulus you use and how you measure motion. What is not in question is that something about almost-human entities reliably bothers a lot of people, which is why the valley keeps being invoked in robotics, film, video games, and now in conversations about deepfakes and generative video.
Mori was a professor at the Tokyo Institute of Technology when he wrote "Bukimi no tani" for Energy, vol. 7, no. 4, pp. 33 to 35. The essay was speculative, not empirical. Mori offered a hand-drawn graph with human-likeness on the x-axis and shinwakan on the y-axis, a Japanese term translated as "familiarity" or, in later editions, "affinity." Mori plotted two curves: one for static entities, and a more dramatic second curve for entities in motion. Both rose from industrial robots toward human likeness, dipped sharply just before reaching full humanness, then rose again. The moving curve dipped lower and rose higher.
The examples Mori arranged along his curve show how broadly he was thinking. He placed industrial robots near the bottom, toy robots and stuffed animals higher, humanoid robots and bunraku puppets near the first peak, healthy people at the top, and corpses, prosthetic hands, and zombies inside the valley. His point was not that engineers should aim for full humanness; it was the opposite. He warned designers to stop at the first peak, the friendly toy-robot zone, rather than press through the valley.
The essay was almost completely ignored for thirty years. Two things changed in the early 2000s. First, Hiroshi Ishiguro at Osaka University began building photoreal androids, including the Geminoid HI series modelled on himself. Second, Hollywood released Final Fantasy: The Spirits Within (2001) and The Polar Express (2004), both widely accused of dropping their CGI humans straight into the valley. The phrase suddenly had a use.
The canonical English translation today is by Karl F. MacDorman and Norri Kageki, published as "The Uncanny Valley" in IEEE Robotics & Automation Magazine in 2012. An earlier rough translation existed from the mid-2000s, made between 1 and 2 a.m. in a Japanese robotics lab as MacDorman has recounted, but the 2012 version is the one most cited. The English label "uncanny valley" itself comes from Jasia Reichardt's 1978 book Robots: Fact, Fiction, and Prediction, which loosely associated Mori's idea with Ernst Jentsch's 1906 essay on the psychology of the uncanny and Sigmund Freud's 1919 essay "Das Unheimliche." That association was not Mori's.
A reasonable question is why near-human entities should provoke this reaction at all. The literature offers several candidate mechanisms, none decisively confirmed, several probably operating together.
| Proposed mechanism | Core claim | Representative work |
|---|---|---|
| Pathogen avoidance | Slight deviations in a near-human face read as cues of disease, triggering disgust and avoidance evolved to keep us away from sick conspecifics. | Moosa & Ud-Deen (2010); discussed in Wang, Lilienfeld & Rochat (2015) |
| Mate selection | Subtle facial or motor abnormalities signal poor genetic fitness, so observers reject near-human stimuli for the same reason mate-choice mechanisms reject defective faces. | Green et al. (2008) |
| Mortality salience | Near-human, dead-eyed, or partially disassembled androids prime thoughts of corpses and our own mortality, producing existential discomfort. | MacDorman (2005); MacDorman & Ishiguro (2006) |
| Categorical perception conflict | Stimuli that fit cleanly into neither the "human" nor the "non-human" category create a sustained classification problem the brain experiences as eerie. | Burleigh, Schoenherr & Lacroix (2013); Yamada et al. (2013) |
| Predictive-coding violation | The brain generates strong human-shaped predictions for near-human stimuli and registers small mismatches as costly prediction errors. | Saygin et al. (2012) fMRI study |
| Configural face processing disruption | Faces are processed holistically; small structural deviations in eye spacing, skin texture, or proportion break configural processing and produce uncanniness. | Diel & Lewis (2022) |
| Empathic concern thwarting | Observers prepare empathic responses for what looks human but cannot complete them when behaviour or affect feels wrong, producing frustration and unease. | Misselhorn (2009) |
| Dehumanisation | Realistic robot faces lead observers to actively withdraw the attribution of mind, producing the same discomfort as encountering a human one has dehumanised. | Wang, Lilienfeld & Rochat (2015) |
A useful detail from Diel and Lewis's 2022 work is that the uncanny effect can be elicited by stimuli that are not obviously "almost human" at all. Their experiments show uncanny valleys for built environments and written text, which suggests that whatever drives the effect is at least partly a general response to deviations from learned templates, consistent with the predictive-coding family of explanations and pushing against the strictly evolutionary readings.
The Wang, Lilienfeld and Rochat 2015 review in Review of General Psychology concluded that evidence on whether the valley exists at all, and on which explanation fits best, was mixed. They proposed a dehumanisation hypothesis: when a robot looks too humanlike, observers actively strip mind attribution from it, and that withdrawal of mind is what feels eerie.
Mori's essay was speculative. The first attempts to test it empirically did not arrive in serious volume until after MacDorman and Ishiguro 2006 advocated for android-based experimental psychology in Interaction Studies. Their argument, sometimes called the "uncanny advantage," was that highly humanlike androids elicit human-directed cognitive expectations, which means they can be used as experimental stimuli for studying social cognition in ways mechanical robots cannot. Falling into the valley was, on this view, a feature for researchers, not just a problem for designers.
Several studies have since tried to actually plot Mori's curve.
| Study | Method | Finding |
|---|---|---|
| Burleigh, Schoenherr & Lacroix (2013) | 164 participants rated digitally created faces varying in human-likeness and prototypicality. | Linear relationship between human-likeness and emotional response in most conditions; an uncanny dip appeared only when faces straddled categorical boundaries (e.g., morphs between goat-like and human faces). |
| Wang, Lilienfeld & Rochat (2015) | Literature review and statistical reanalysis. | Evidence for the valley is real but weaker and noisier than the popular framing suggests; magnitude is often modest. |
| Mathur & Reichling (2016) | 80 real-world robot faces plus controlled morphs, rated by hundreds of participants in Cognition. | Found a robust uncanny valley curve in explicit likability ratings and in trust behaviour in an investment game. |
| Strait et al. (2017) | Multiple measures of evaluation, gaze, and avoidance of humanlike robots. | Highly humanlike but not perfect agents were rated more eerie, less likable, and provoked reduced eye contact and shorter encounters. |
| Diel & Lewis (2022) | Configural deviation studies for faces, places, and text. | Cubic, N-shaped "deep evaluative shift" function fits the data; uncanniness is driven by deviation from learned configural patterns. |
The most often-cited counterweight is David Hanson's 2005 paper "Upending the Uncanny Valley," published at AAAI. Hanson, founder of Hanson Robotics, argued from his own face-robot prototypes that the valley is not a fixed psychological law and that careful skin sculpting, eye motion, and behavioural responsiveness can carry an android through it without an aversion dip. His reading is that the valley is an artifact of bad design, not a deep feature of cognition. Critics reply that studies of Sophia, his most famous robot, do not bear this out cleanly.
Geller's 2008 critique in IEEE Spectrum asked whether the uncanny valley is just a folk concept that scientists have over-extended. Misselhorn 2009 reframed it as a problem of thwarted empathy. Diel, Sato and MacDorman's 2022 ACM meta-analysis found the effect is best supported when the stimulus space contains near-human items and the dependent measure is eeriness or trust rather than generic likability.
Cross-cultural variation is more limited than expected. Several studies comparing Japanese, Chinese, German, and American samples have found broadly similar valley shapes. Whether children show the valley as strongly as adults is still debated, with some developmental studies finding muted responses in young children.
The valley is most often invoked in four overlapping domains: physical robotics, cinema visual effects, video games, and synthetic media. The table below sketches the typical pattern in each.
| Domain | Where the valley shows up | Typical response |
|---|---|---|
| Humanoid robots | Silicone-skinned androids with limited facial actuation, asynchronous lip-sync, or dead eyes. Examples include Hiroshi Ishiguro's Geminoid HI series and Hanson Robotics' Sophia. | Designers either lean fully into mechanical aesthetics or invest heavily in skin, eye saccades, and conversational responsiveness to climb the second slope. |
| Animatronics and theme-park figures | Disney audio-animatronics, especially earlier president figures, sometimes elicit valley reactions when motion does not match facial fidelity. | The newer figures use higher actuator counts and more careful eye motion to keep observers in the friendly zone. |
| Performance-capture cinema | The Polar Express (2004), Beowulf (2007), digital Tarkin and Princess Leia in Rogue One (2016), Henry Cavill's CGI-erased mustache in Justice League (2017), and Cats (2019). | Critics described faces as "dead-eyed," "plastic," or "creepy," and several films were box-office or critical disappointments partly for this reason. |
| Photoreal CGI doubles | The Curious Case of Benjamin Button (2008), The Lion King (2019), de-aging in The Irishman (2019). | These are widely cited as having crossed the valley by committing fully to photorealism with extensive performance capture. |
| Video games | Mass Effect: Andromeda (2017) facial animation problems; Cyberpunk 2077 (2020) launch issues; Red Dead Redemption 2 (2018) generally praised. | Internet mockery of off-model faces has become a reliable signal that a game has dropped into the valley. |
| Deepfakes and synthetic video | Face-swap clips, voice clones, and full-body avatar systems in 2024 to 2026. | Improving rapidly; voices are crossing earlier than full-face video, with implications for fraud, including a 2024 Hong Kong incident in which an employee was deceived by a deepfake CFO into authorising roughly USD 25 million in transfers. |
| Virtual avatars | Meta's Codec Avatars, Apple Vision Pro Personas, Synthesia presenter avatars, Hugging Face avatar models. | Users report "close but off," especially around eye gaze, neck rigidity, and skin shading. |
| Voice synthesis | Sesame's Conversational Speech Model, OpenAI's Voice Mode, ElevenLabs voice clones. | An auditory uncanny valley has been proposed; near-human voices with subtle prosody errors register as creepy faster than appearance does. |
| AI-generated images | Diffusion-model outputs with malformed hands, inconsistent eyes, or off-axis teeth. | The classic "too many fingers" failure mode is a textbook valley artifact. |
The Polar Express deserves a separate note because it became the canonical case study. Robert Zemeckis used a then-new performance-capture pipeline to map Tom Hanks onto digital children and adults. The body work was reasonable. The faces were not. Head-mounted cameras could not be untethered, so facial motion had to be inferred from body capture and limited markers, and the result was the "dead-eyed" look that critics like Peter Travers of Rolling Stone and Stephanie Zacharek of Salon singled out. Zemeckis kept iterating with Beowulf and A Christmas Carol, but it was Avatar (2009) that demonstrated the technique could work, mostly because James Cameron's Na'vi were stylised away from the human envelope.
The Rogue One case shows what the valley does in close-up. Industrial Light & Magic resurrected Peter Cushing's Grand Moff Tarkin by digitally replacing actor Guy Henry's face, and rebuilt a young Carrie Fisher for a brief Princess Leia cameo. Tarkin appeared mostly in low light and got mixed reactions; Leia, in harsh light and tight close-up, was widely felt to fall into the valley. Lighting turns out to be a major variable.
Mori's original advice was to aim before the valley. A second strategy is to commit fully to photorealism and try to climb the far slope, expensive and risky but occasionally successful. A third is to match form to function so the question never arises.
| Strategy | Approach | Examples |
|---|---|---|
| Aim before the valley | Stylised, abstract, or cartoon-like form that is clearly non-human. | Pixar's WALL-E (2008); Boston Dynamics Spot; SoftBank Pepper; NAO; Wreck-It Ralph's game characters. |
| Aim across the valley | Commit fully to photorealism with detailed motion fidelity. | Digital Domain's Brad Pitt in The Curious Case of Benjamin Button (2008); MPC's photoreal animals in The Lion King (2019); de-aging in The Irishman (2019). |
| Match form to function | Industrial or clearly mechanical robots that do not invite human comparison. | KUKA arms, Boston Dynamics Atlas, Roomba. |
| Iterate to fix specific cues | Identify and repair the precise off-cue (eyes, lip sync, skin specularity) rather than redesign the whole entity. | The 2019 redesign of Sonic the Hedgehog after the trailer backlash; ongoing skin and eye work on Engineered Arts' Ameca. |
| Match modalities | Keep appearance, voice, and behaviour at the same level of realism. | Sesame's Conversational Speech Model paired with deliberately abstract avatar embodiments, rather than near-human faces with robotic voices. |
The "matching modalities" strategy is especially relevant in the current generation of embodied agents. Pairing a near-human face with a flat, robotic synthesised voice is one of the most reliable ways to drop straight into the valley. So is the inverse: a warm, human-sounding voice attached to a stiff, plasticky avatar.
A rough mapping of well-known humanoid robots to their position relative to Mori's curve. Reasonable people disagree on several of these placements; the table reflects the most common reading in the popular and academic press.
| Robot | Maker | Position relative to the valley |
|---|---|---|
| Atlas | Boston Dynamics | Mechanical, no face. Comfortably below the valley; impressive without being eerie. |
| Spot | Boston Dynamics | Quadruped, deliberately not anthropomorphic. Below the valley. |
| ASIMO | Honda | White plastic shell, helmet face. Past the first peak; sometimes described as cute. |
| NAO | SoftBank Robotics (originally Aldebaran) | Small, rounded, big-eyed. Friendly first-peak zone. |
| Pepper | SoftBank Robotics | Cartoon proportions. First peak. |
| Optimus | Tesla | Smooth, featureless faceplate from Gen 2 onwards; deliberately moved away from facial features. Debated; Gen 3 hands described as approaching the valley. |
| Digit | Agility Robotics | Stylised, no face. Below the valley. |
| Sophia | Hanson Robotics | Silicone face, limited actuators, expressive but inconsistent. Widely cited as in or near the valley. |
| Ameca | Engineered Arts | Grey rubber skin and visible mechanical joints; designed to avoid the valley but still produces eerie reactions for many viewers. |
| Geminoid HI | Hiroshi Ishiguro Laboratories (ATR) | Photoreal teleoperated double of Ishiguro himself. Explicitly built to be in the valley as a research instrument. |
| NEO Beta | 1X Technologies | Mannequin-like proportions, soft fabric covering. Debated; some find it disturbing in motion. |
| Disney audio-animatronics | Walt Disney Imagineering | Range from clearly non-human to highly photoreal; older Hall of Presidents figures were a frequent valley reference. |
| HRP-4C Miim | AIST (Japan) | Female-presenting fashion-show android; discussed as a valley case in several papers. |
Position on the valley is a function of who is looking, in what light, and with what audio. The same robot rated in a quiet lab and on a viral video clip can score very differently.
Mori was thinking about physical robots and to a lesser extent prosthetics. The phrase has since expanded in three directions, all recent.
First, generative video. Models like OpenAI Sora, Google Veo, and Runway Gen-3 generate clips of synthetic people who often look photoreal at first glance and then betray the valley on closer inspection: a smile that holds a fraction of a second too long, a hand that morphs, a tongue that vanishes into a mouth without occlusion. These are exactly the configural deviations Diel and Lewis 2022 predict should produce uncanniness.
Second, voice. Sesame published a piece titled "Crossing the uncanny valley of conversational voice" in 2025, describing their Conversational Speech Model that integrates text and audio generation rather than treating speech as TTS over a separate text stream. The argument is that prosody, breath, and conversational alignment are what make voice cross the valley, not raw waveform fidelity. OpenAI's Voice Mode and ElevenLabs sit in similar territory; users report the better systems can pass the valley in short bursts but get caught by it in longer interactions when emotional context drifts.
Third, behavioural and conversational uncanny effects. As LLM-driven chatbots get embodied in physical or virtual humanoid forms, observers increasingly report a behavioural valley: the avatar looks and sounds correct but says something fluent yet hollow, or remembers nothing about a previous turn. The mismatch between human-shaped form and non-human conversational depth produces the same family of reactions as a poorly rendered face. Some researchers now talk about a "moral uncanny valley" that arises when humanlike robots appear to make moral decisions in ways that feel slightly off.
Deepfakes are the case where the valley has become a security and policy issue rather than a design issue. The 2024 Hong Kong scam in which an employee transferred roughly USD 25 million after a deepfake video call with what appeared to be the company's CFO is the high-profile example, but smaller versions of the same attack have proliferated. As synthetic video pulls out of the valley, the social mechanism that used to flag "this is not quite a real person" weakens.
A few critiques are worth taking seriously. The first is that "uncanny valley" is invoked too freely. Almost any unsettling near-human appearance gets labelled as valley territory, including stimuli where the mechanism is more likely simple ugliness, ordinary fear of unfamiliar machines, or basic disgust at a corpse-like image.
The second is that Mori's original curve is one specific shape, and the empirical literature does not always reproduce it. Burleigh et al. 2013 found mostly linear relationships outside a narrow categorical boundary. Several studies find dips that depend strongly on the dependent measure. The cubic N-shape Diel and Lewis 2022 propose fits some data better than the original U-shape.
The third is that individual variation is large. Need-for-structure and disgust-sensitivity scales correlate with valley sensitivity in several studies. So do age and prior exposure, with younger observers who grew up on photoreal CGI sometimes exhibiting more muted reactions than older ones.
The fourth is the cultural-context point. Mori was writing in a Japanese intellectual context in which Buddhism, animism, and a long tradition of bunraku puppetry and karakuri ningyō dolls shaped how he imagined the relation between objects and life. Some scholars, including Joseph Borody in a 2013 Journal of Evolution and Technology piece, have argued that translating the essay into a context dominated by the Jentsch and Freud notions of the uncanny has subtly distorted Mori's meaning. The Japanese concept bukimi, often translated as "eeriness" or "weirdness," carries a slightly different valence than the English "uncanny." Mori's own later commentary is more cautious than the popular framing suggests; he treated the curve as a heuristic, not a law, and explicitly invited designers to test it.
In Japan, the valley has functioned as both a research target, especially in the work of Ishiguro and his collaborators, and as a design constraint, especially in commercial and care robotics where companies like SoftBank have leaned into below-the-valley aesthetics for products like Pepper and NAO. The cultural acceptance of robots in Japanese popular media, from Astro Boy onwards, is often invoked to explain why Japanese robotics has been comparatively willing to risk the valley in pursuit of photorealism.
In the United States and Europe, the valley has been more often cited as a reason to avoid photoreal humanoid form, especially in care and customer-facing applications. The Boston Dynamics design language, Agility Robotics' Digit, and the recent move at Tesla toward a featureless Optimus faceplate read as conscious choices to step away from the valley. In cinema, the valley has driven a steady migration from full performance-captured digital humans toward hybrid approaches, like high-fidelity de-aging applied to live actors rather than fully synthetic faces. The 2019 Sonic the Hedgehog trailer redesign, in which Paramount delayed the film to remake the title character after a fan backlash about its human-style teeth and eyes, is a reminder that even animated characters can fall in.
The original Mori essay is now over fifty years old, and it has aged into something like a productive provocation: a sharp claim, framed in a single line drawing, that has spawned a research programme bigger than the essay itself.