MR-0002 · The Uncanny Valley and Synthetic Media

Question: Is the uncanny valley real, what causes it, and does it still help people notice AI-generated faces and voices?
Answer in one line: Yes, the effect is real and large when stimuli contain distorted or mismatched features (meta-analytic g = 1.01; face distortion g = 1.46). It is not a simple function of human likeness. Modern GAN faces avoid the trigger: they are judged as real as, or more real than, real faces and are trusted more. Only an implicit neural trace remains, and people cannot act on it.
Version: 0.4.0 · Updated: 2026-10-19 · Sources: 28 (8 checked against full text, 11 against the official abstract, 2 DOI-only, 7 unverified; see Verification status).

TL;DR

• The effect is large when the conditions are right. A meta-analysis of 72 studies (247 effect sizes) found an overall uncanny valley effect of Hedges' g = 1.01 [0.80, 1.22]. Face distortion produced the largest effect (g = 1.46) diel-2022-uv-meta.
• It is conditional. The simplest version, in which any near-human likeness triggers a dip, and Mori's claim that movement amplifies it, both lack consistent support katsyri-2015-uv-review.
• Best-supported cause: perceptual mismatch.
• Reducing realism consistency within a face made humans and animals (not objects) eerier and colder (N = 548) macdorman-chattopadhyay-2016.
• A review rates mismatch as having good support katsyri-2015-uv-review.
• Category ambiguity is not the cause.
• The most category-ambiguous stimuli were not the eeriest macdorman-chattopadhyay-2016.
• Ambiguous faces were easier to discriminate, and difficulty went with more positive affect, a "Happy Valley" cheetham-2014-happy-valley.
• Category confusion did not mediate the valley for robot faces mathur-reichling-2016-cartography.
• It reaches behaviour and value signals.
• Real robot faces produce a valley in likability and in trust decisions in an investment game mathur-reichling-2016-cartography.
• vmPFC encodes likability non-linearly. The behavioural dip appeared in 18 of 21 participants, and an amygdala signal predicted rejection of artificial agents rosenthal-von-der-putten-2019-vmpfc.
• Synthetic faces have crossed it on explicit judgments.
• StyleGAN2 faces are classified at 48.2% and rated 7.7% more trustworthy nightingale-farid-2022.
• White AI faces were judged human 65.9% of the time vs 51.1% for real faces miller-2023-ai-hyperrealism.
• A trace remains, but people can't use it. EEG reliably distinguishes AI from real faces while behaviour is near chance moshel-2022-decoding-ai-faces.
• The label alone changes perception. Merely labelling a real smiling face as fake made it rated less positive (73.12 vs 74.97 on a 0–100 scale) and removed early neural smile effects eiserbeck-2023-deepfake-smiles.

Key numbers

Key findings (citable)

Concept model

Text equivalent: there are three routes to eeriness.

1. Perceptual: distortion or mismatched realism produces a prediction error, which the vmPFC value signal registers as low likability and the amygdala as rejection.
2. Conceptual: attributing mind or feeling to a machine.
3. Belief: labelling something as AI dampens the positive response to it.

Modern GAN faces are internally consistent and average, so they never trigger the perceptual route. They are processed as human, and their averageness is read as "real", producing hyperrealism and higher trust. A weak implicit neural difference remains that does not reach judgment.

flowchart LR
    A[Humanlike face or agent] --> B{Distorted or mismatched<br/>realism / motion?}
    B -- Yes --> C[Prediction error<br/>aIPS action network]
    C --> D[Low likability: vmPFC<br/>Rejection: amygdala]
    B -- No --> E[Processed as human]
    E --> F{Seen as machine<br/>with feelings?}
    F -- Yes --> D
    E --> H{Labelled as AI?}
    H -- Yes --> I[Blunted positive affect<br/>effortful late processing]
    F -- No --> J[Modern GAN faces:<br/>hyperrealism + higher trust]
    J -. weak implicit residue .-> K[EEG-decodable difference<br/>not available to judgment]

Chart: effect size by stimulus technique

Text equivalent: in the 72-study meta-analysis, uncanny valley effect sizes (Hedges' g) were:

• face distortion 1.46 [0.69, 2.24]
• distinct entities 1.20 [1.02, 1.38]
• realism render 0.99 [0.62, 1.36]
• morphing 0.94 [0.64, 1.24]
• overall 1.01 [0.80, 1.22]

All are large by conventional benchmarks (0.8 = large).

xychart-beta
    title "Uncanny valley effect size by stimulus technique (Diel et al. 2022)"
    x-axis ["Face distortion", "Distinct entities", "Realism render", "Morphing", "Overall"]
    y-axis "Hedges' g" 0 --> 2.5
    bar [1.46, 1.20, 0.99, 0.94, 1.01]
    line [0.8, 0.8, 0.8, 0.8, 0.8]

The line marks the conventional "large effect" threshold (g = 0.8).

Mechanism scorecard

Theme 1 — Does the valley exist?

Yes, as a robust conditional effect:

• Meta-analysis: 72 of 468 screened studies, 247 effect sizes, 10 stimulus-creation techniques, 39 affect measures and 14 indirect measures. Overall g = 1.01 diel-2022-uv-meta.
• Measures with the largest affect effects: threatening, likable, aesthetics, familiarity, eeriness.
• Indirect measures with the largest effects: dislike frequency, categorization reaction time, like frequency, avoidance, viewing duration.
• Real robot faces: 80 faces showed a valley in explicit likability that extended to implicit trust in an investment game. Trust depended on subtle facial-expression cues that people also use to judge humans mathur-reichling-2016-cartography. The finding replicated across sites mathur-2020-uncanny-not-confusing (unverified).
• Robot gaze: across nine gaze-timing combinations of a virtual robot, likability followed a cubic valley over anthropomorphism; empathy rose steadily; extraversion, negative attitudes and neuroticism moderated (N = 61) hsieh-2026-robot-gaze-uncanny.
• Neural valuation: vmPFC encoded likability as a non-linear function of humanlikeness, with selectively low likability for highly humanlike agents. Behaviourally, artificial humans fell below the linear trend in 18 of 21 participants rosenthal-von-der-putten-2019-vmpfc.

But not as a universal law of human likeness:

• The simplest hypothesis and the movement claim lack consistent support. The valley "appears to occur only under specific conditions" katsyri-2015-uv-review.
• For digital faces varied along a single realism axis, eeriness has been reported as mostly linear, with non-linear effects only for conflicting or atypical features burleigh-2013-does-uv-exist (metadata verified; finding unverified).
• Realism alone did not produce unpleasantness; abnormal eye size on highly realistic faces did seyama-nagayama-2007 (unverified).

Measurement note: the meta-analysis also found that the field has no consensus on the effect's theoretical basis or methodology diel-2022-uv-meta. Effect sizes depend on how the stimuli were made: designs that distort faces produce the largest effects.

Theme 2 — What causes it?

Perceptual mismatch has the strongest support:

• Three rounds of experiments (N = 548) reduced realism consistency within a face, for example by mixing eyes-eyelashes-mouth with skin-nose-eyebrows at different realism levels. This made humans and animals, but not objects, eerier and colder macdorman-chattopadhyay-2016.
• The review rates perceptual mismatch as having good support katsyri-2015-uv-review.
• Face distortion is the stimulus technique with the largest meta-analytic effect (g = 1.46) diel-2022-uv-meta.

Category ambiguity is contradicted:

• The most category-ambiguous figures were neither the eeriest nor the coldest, so Jentsch's (1906) prediction failed macdorman-chattopadhyay-2016.
• On avatar–human morphs, ambiguous faces were easier to discriminate than human faces. Greater difficulty was associated with more positive affect, a "Happy Valley". Avatars appear to be processed at the category level and humans as individuals cheetham-2014-happy-valley.
• Category confusion may occur, but it does not mediate the robot-face valley mathur-reichling-2016-cartography.

Mind perception is a separate conceptual route (records unverified):

• A humanlike robot was unnerving, and this was mediated by attributing experience (feeling) to it gray-wegner-2012-mind-perception.
• In VR, identical characters were eeriest when their empathic behaviour was attributed to an autonomous AI stein-ohler-2017-uncanny-mind.
• With an AI counsellor (N = 193), a machine-like agent was rated best at low self-disclosure and worse at high disclosure, mediated by lower anthropomorphism; a human-like agent gained with disclosure jang-2026-uncanny-conversation.

Dehumanization: the hypothesis says agents are first processed as human, and that non-human cues then trigger devaluation wang-2015-uv-existence-explanations. It is consistent with the label effects in Theme 4.

Theme 3 — Neural correlates

Caveat: samples are small (19–33), there are few replications, and stimulus sets differ.

Theme 4 — Synthetic faces and deepfakes: crossing the valley

On conscious judgments, modern GAN faces are past the valley:

• Accuracy on StyleGAN2 faces was 48.2% (95% CI 47.1–49.2) nightingale-farid-2022.
• Synthetic faces were rated more trustworthy: 4.82 vs 4.48, +7.7%, d = 0.49 nightingale-farid-2022.
• White AI faces were judged human 65.9% of the time vs 51.1% for real faces miller-2023-ai-hyperrealism.

Why this fits the mechanism evidence (inference linking two literatures):

• The uncanny response is driven by distortion and mismatch (F3).
• GAN faces are internally consistent and close to the population average, so they lack the trigger.
• Their averageness is then misread as a sign of a real person. Participants misjudged proportions, familiarity and memorability miller-2023-ai-hyperrealism.
• Direct support for the averageness link: in two identity-trained neural networks, StyleGAN2 faces sat closer to the centre of face space than real faces. Super-recognisers used that centrality as a cue to AI and typical observers did not dunn-2026-super-recognizers-average.
• No single study tests the whole chain.

Residual signals exist but are weak or unusable:

• EEG reliably separates real from AI faces while group-level judgments are near chance, with frequent label swaps moshel-2022-decoding-ai-faces.
• A classifier on 14 human-rated attributes reached 94%. The cues are present but misweighted miller-2023-ai-hyperrealism.
• Disrupting face processing hurt human deepfake-video detection but mostly not the model's. Human detection relies on specialised face processing groh-2022-crowds-machines.
• AI faces were recognised (remembered) significantly better than real faces, yet recognition accuracy did not correlate with real-vs-AI classification accuracy; the authors suggest the advantage is partly covert sulimani-2026-recognition-ai-faces. This sits uneasily with the view that AI faces are simply less memorable; generator and N are not in the abstract.
• Self-resembling deepfake talking heads (2020-era) still produced uncanniness that reduced trust weisman-pena-2021-doppelganger (unverified).

Belief versus pixels. In Eiserbeck et al. (2023), all 540 photographs were real; only the on-screen label "REAL" or "FAKE" varied eiserbeck-2023-deepfake-smiles:

• Happy faces labelled fake were rated less positive (73.12 vs 74.97) and took longer to rate (≈1,366 vs ≈1,290 ms).
• Their early neural smile effects (P1, N170, EPN) disappeared, and a late LPP effect appeared that was absent for real-labelled smiles.
• Angry and neutral faces were unaffected by the label.
• Disclosure therefore changes affective response even when nothing about the image changes. This works in both directions: falsely labelling real content as AI also devalues it.

Voices: realism parity, no hyperrealism. Clones were rated as real as the speakers they copied (63.0 vs 63.8), not more; generic AI voices were rated less real (47.4) lavan-2025-voice-clones-hyperrealism. See MR-0003.

Motion and voice are open questions. The prediction-error findings (F6) suggest that motion or audio-visual mismatch could re-open the valley for video deepfakes. This has not been tested with current real-time or diffusion video generators.

Practical reading for detection

Editorial inference, not tested interventions:

• Gut-level uncanniness is no longer a usable signal for high-quality still faces (F9, F10). The absence of uncanniness is not evidence of authenticity.
• Mismatch is still the trigger (F3). Cheaper or real-time deepfakes are most likely to give themselves away through lip-sync errors, unusual blinking or gaze, inconsistent lighting, or audio-visual lag. These cues disappear as generators improve.
• Labels are powerful in both directions (F11). Provenance and disclosure systems such as C2PA shape how people respond, and false "AI" labels can devalue genuine content.

Evidence gaps

1. No preregistered, large-N implicit or neural studies with current diffusion-model faces.
2. No studies of uncanny responses to 2025–2026 video deepfakes or real-time face swaps in video calls.
3. Little work on synthetic voice; no robust voice valley has been established. Voice studies are not covered in v0.1.0.
4. Cross-race tests of hyperrealism with less biased generators are missing.
5. Whether people can be trained to use their implicit unease is untested.

Answers to common questions

Q: Is the uncanny valley real? A: Yes. Across 72 studies the effect is large (g = 1.01). It is conditional: it is driven by distortion or mismatch, not by human likeness alone (MR-0002.F1, F2).

Q: What causes the uncanny valley? A: The best-supported cause is perceptual mismatch or distortion. Category ambiguity is contradicted. Attributing feelings to machines is a separate route (F3–F5).

Q: Do deepfakes fall into the uncanny valley? A: High-quality synthetic still faces do not. They are judged at chance, trusted more, and can look more real than real faces (F9). Imperfect video deepfakes may still trigger uncanniness (F12).

Q: Can your brain detect deepfakes even if you can't? A: Partly. EEG reliably distinguished AI from real faces while people's judgments were near chance (F10). The signal does not reach conscious decisions.

Q: Does knowing a face is AI change how we feel about it? A: Yes. Labelling a real smiling face as "fake" made it rated less positive and removed early neural smile responses (F11).

Method of this metareport

• Type: narrative synthesis across uncanny-valley research (robots, CG, VR, androids) and synthetic-face research. No new statistical pooling.
• Confidence ratings are editorial, based on consistency, sample size, tier and verification status.
• Search limits for v0.1.0: compiled with restricted web access, so 2024–2026 literature on deepfake video and synthetic voice is under-covered. v0.3.0 (2026-10-18) added one 2026 face-memory study found via PubMed.

Verification status

Values from unverified records are marked "≈" or "unverified" in this report. Check them against the primary source before high-stakes use. Each source page states exactly what was checked and how.

References

• burleigh-2013-does-uv-exist · record · Burleigh, T. J., et al. (2013). Does the uncanny valley exist? An empirical test of the relationship between eeriness and the human likeness of digitally created faces. Computers in Human Behavior, 29(3), 759-771. · https://doi.org/10.1016/j.chb.2012.11.021
• cheetham-2014-happy-valley · record · Cheetham, M., et al. (2014). Perceptual discrimination difficulty and familiarity in the Uncanny Valley: more like a 'Happy Valley'. Frontiers in Psychology, 5, 1219. · https://doi.org/10.3389/fpsyg.2014.01219
• diel-2022-uv-meta · record · Diel, A., et al. (2022). A Meta-analysis of the Uncanny Valley's Independent and Dependent Variables. ACM Transactions on Human-Robot Interaction, 11(1), Article 1. · https://doi.org/10.1145/3470742
• dunn-2026-super-recognizers-average · record · Dunn, J. D., et al. (2026). Too good to be true: Synthetic AI faces are more average than real faces and super-recognizers know it. British Journal of Psychology. · https://doi.org/10.1111/bjop.70063
• eiserbeck-2023-deepfake-smiles · record · Eiserbeck, A., et al. (2023). Deepfake smiles matter less—the psychological and neural impact of presumed AI-generated faces. Scientific Reports, 13, 16111. · https://doi.org/10.1038/s41598-023-42802-x
• gray-wegner-2012-mind-perception · record · Gray, K., & Wegner, D. M. (2012). Feeling robots and human zombies: Mind perception and the uncanny valley. Cognition, 125(1), 125-130. · https://doi.org/10.1016/j.cognition.2012.06.007
• groh-2022-crowds-machines · record · Groh, M., et al. (2022). Deepfake detection by human crowds, machines, and machine-informed crowds. Proceedings of the National Academy of Sciences (PNAS), 119(1), e2110013119. · https://doi.org/10.1073/pnas.2110013119
• hsieh-2026-robot-gaze-uncanny · record · Hsieh, T.-Y., et al. (2026). Robot gaze-timing combinations and the uncanny valley. Computers in Human Behavior Reports, 24, 101378. · https://doi.org/10.1016/j.chbr.2026.101378
• jang-2026-uncanny-conversation · record · Jang, J., et al. (2026). Uncanny Valley in Conversation? Effects of Internal and External Human-Likeness on Intimacy and Trust. Behavioral Sciences, 16(10), 1794. · https://doi.org/10.3390/bs16101794
• katsyri-2015-uv-review · record · Kätsyri, J., et al. (2015). A review of empirical evidence on different uncanny valley hypotheses: support for perceptual mismatch as one road to the valley of eeriness. Frontiers in Psychology, 6, 390. · https://doi.org/10.3389/fpsyg.2015.00390
• lavan-2025-voice-clones-hyperrealism · record · Lavan, N., et al. (2025). Voice clones sound realistic but not (yet) hyperrealistic. PLOS ONE, 20(9), e0332692. · https://doi.org/10.1371/journal.pone.0332692
• macdorman-chattopadhyay-2016 · record · MacDorman, K. F., & Chattopadhyay, D. (2016). Reducing consistency in human realism increases the uncanny valley effect; increasing category uncertainty does not. Cognition, 146, 190-205. · https://doi.org/10.1016/j.cognition.2015.09.019
• mamsheli-2026-eeg-ai-faces · record · Mamsheli, S. T., et al. (2026). Processing of Real and AI-Generated Human Faces: An Exploratory EEG Study. Proceedings of the European Conference on Cognitive Ergonomics (ECCE 2026), 375-379. · https://doi.org/10.1145/3822301.3822351
• mathur-reichling-2016-cartography · record · Mathur, M. B., & Reichling, D. B. (2016). Navigating a social world with robot partners: A quantitative cartography of the Uncanny Valley. Cognition, 146, 22-32. · https://doi.org/10.1016/j.cognition.2015.09.008
• mathur-2020-uncanny-not-confusing · record · Mathur, M. B., et al. (2020). Uncanny but not confusing: Multisite study of perceptual category confusion in the Uncanny Valley. Computers in Human Behavior, 103, 21-30. · https://doi.org/10.1016/j.chb.2019.08.029
• miller-2023-ai-hyperrealism · record · Miller, E. J., et al. (2023). AI hyperrealism: Why AI faces are perceived as more real than human ones. Psychological Science, 34(12), 1390-1403. · https://doi.org/10.1177/09567976231207095
• mori-2012-uncanny-valley · record · Mori, M., et al. (2012). The uncanny valley [From the Field]. IEEE Robotics & Automation Magazine, 19(2), 98-100. · https://doi.org/10.1109/MRA.2012.2192811
• moshel-2022-decoding-ai-faces · record · Moshel, M. L., et al. (2022). Are you for real? Decoding realistic AI-generated faces from neural activity. Vision Research, 199, 108079. · https://doi.org/10.1016/j.visres.2022.108079
• nightingale-farid-2022 · record · Nightingale, S. J., & Farid, H. (2022). AI-synthesized faces are indistinguishable from real faces and more trustworthy. Proceedings of the National Academy of Sciences (PNAS), 119(8), e2120481119. · https://doi.org/10.1073/pnas.2120481119
• rosenthal-von-der-putten-2019-vmpfc · record · Rosenthal-von der Pütten, A. M., et al. (2019). Neural mechanisms for accepting and rejecting artificial social partners in the uncanny valley. Journal of Neuroscience, 39(33), 6555-6570. · https://doi.org/10.1523/JNEUROSCI.2956-18.2019
• saygin-2012-predictive-coding · record · Saygin, A. P., et al. (2012). The thing that should not be: predictive coding and the uncanny valley in perceiving human and humanoid robot actions. Social Cognitive and Affective Neuroscience, 7(4), 413-422. · https://doi.org/10.1093/scan/nsr025
• schindler-2017-erp-realism · record · Schindler, S., et al. (2017). Differential effects of face-realism and emotion on event-related brain potentials and their implications for the uncanny valley theory. Scientific Reports, 7, 45003. · https://doi.org/10.1038/srep45003
• seyama-nagayama-2007 · record · Seyama, J., & Nagayama, R. S. (2007). The uncanny valley: Effect of realism on the impression of artificial human faces. Presence: Teleoperators and Virtual Environments, 16(4), 337-351. · https://doi.org/10.1162/pres.16.4.337
• steckenfinger-ghazanfar-2009 · record · Steckenfinger, S. A., & Ghazanfar, A. A. (2009). Monkey visual behavior falls into the uncanny valley. Proceedings of the National Academy of Sciences (PNAS), 106(43), 18362-18366. · https://doi.org/10.1073/pnas.0910063106
• stein-ohler-2017-uncanny-mind · record · Stein, J.-P., & Ohler, P. (2017). Venturing into the uncanny valley of mind—The influence of mind attribution on the acceptance of human-like characters in a virtual reality setting. Cognition, 160, 43-50. · https://doi.org/10.1016/j.cognition.2016.12.010
• sulimani-2026-recognition-ai-faces · record · Sulimani, N., et al. (2026). Recognition of real and artificial intelligence-generated faces. Perception, 55(10), 924-937. · https://doi.org/10.1177/03010066261442089
• wang-2015-uv-existence-explanations · record · Wang, S., et al. (2015). The uncanny valley: Existence and explanations. Review of General Psychology, 19(4), 393-407. · https://doi.org/10.1037/gpr0000056
• weisman-pena-2021-doppelganger · record · Weisman, W. D., & Peña, J. F. (2021). Face the uncanny: The effects of doppelganger talking head avatars on affect-based trust toward artificial intelligence technology are mediated by uncanny valley perceptions. Cyberpsychology, Behavior, and Social Networking, 24(3), 182-187.