Why AI Art Feels Wrong (And You Can't Say Why) | The Uncanny Valley

In 1970, a Japanese robotics professor, Masahiro Mori, wrote a short essay about a handshake. A hand looks normal, but when you take it, the grip is limp, there's no bone in it, and it's cold. It's a very realistic fake. Mori said the warm feeling you had doesn't drop to neutral, it drops below zero. He gave that drop a name. In English we call it the uncanny valley. Today you get the same jolt from an AI picture: a lovely beach photo, and then you notice the little girl has six fingers. We climb one level at a time. Mori's graph. Why image generators used to mess up hands (small, hidden, a thousand shapes, and no skeleton underneath). Why a sixth finger hits so hard: Ayse Saygin's 2011 brain-scan study of a human, a robot and the android Repliee Q2, and the idea of a prediction error. Whether this feeling is older than horror movies: the 2009 Princeton monkeys who looked away from almost-right faces, and Mori's own guess about why. And finally you: a 2023 experiment where the same AI paintings were liked more when a random label said "made by a human". So next time you get that little "nope", ask: did you spot something, or did you already know? Sources Masahiro Mori, "The Uncanny Valley" (Energy, 1970), authorized English translation by Karl F. MacDorman and Norri Kageki, IEEE Spectrum, 2012 (the handshake, the graph, the corpse in the valley, movement makes it stronger, eeriness as self-protection (his guess), aim for the first peak, eyeglasses, the wooden hand carved by a maker of Buddha statues): https://spectrum.ieee.org/the-uncanny... Why AI image generators struggled with hands, incl. a Stability AI spokesperson on hands being less visible and smaller than faces in training pictures (PetaPixel, 2023): https://petapixel.com/2023/03/02/why-... Saygin et al. (published online 2011), "The thing that should not be: predictive coding and the uncanny valley in perceiving human and humanoid robot actions" (fMRI, 20 participants, human vs robot vs android Repliee Q2; prediction error): https://pmc.ncbi.nlm.nih.gov/articles... Steckenfinger & Ghazanfar (2009), "Monkey visual behavior falls into the uncanny valley", PNAS (five macaques, real vs unrealistic vs realistic synthetic faces; gaze, not emotion): https://pmc.ncbi.nlm.nih.gov/articles... Princeton news on the monkey study: https://www.princeton.edu/news/2009/1... Bellaiche et al. (2023), "Humans versus AI: whether and why we prefer human-created compared to AI-created artwork", Cognitive Research (AI paintings randomly labelled human- or AI-made; human label rated higher on liking, beauty, profundity and worth; effort and story mattered): https://doaj.org/article/c1d43cd30fa3... PsyPost summary of Bellaiche et al.: https://www.psypost.org/new-psycholog... The beach photo and the four paintings are example images we made with AI for this video (the study used its own AI paintings). The close-ups, gauges and bars are teaching drawings, not real data. The cartoon monkey and the "almost right" monkey (a blurred, washed-out version of a real photo) are our stand-ins, not the study's pictures. The magazine page is our drawing, and the Buddha hand photo is a real old statue hand, not the one Mori owned. Painted scenes are illustrations; no real person is painted. Photos (Wikimedia Commons) KUKA industrial robotic arm (Roy Nachum's "The Beach" installation): Oleg Yunakov, CC BY-SA 4.0 Tin wind-up "Mighty Robot" toy: D J Shin, CC BY-SA 3.0 X-ray of Albert von Kölliker's hand, 1896: Wilhelm Röntgen, public domain MRI scanner: Jan Ainali, CC BY 3.0 Repliee Q2 android: BradBeattie at English Wikipedia, CC BY-SA 3.0 Rhesus macaque ("ShyMonkey"): Tantj21, CC BY-SA 4.0 (also shown blurred and washed out as our "almost right" mock-up) Eyeglasses: Jacek Halicki, CC BY-SA 4.0 Hand of Buddha, Japan, late Nara period: Cleveland Museum of Art, CC0 Logo: Stability AI, used to name the company discussed. Chapters 0:00 Japan, 1970: a handshake that feels wrong 0:30 The uncanny valley 0:46 Six fingers on a beach photo 1:08 That "nope" came from you 1:20 Mori's graph 2:08 Level 1: why AI messes up hands 2:36 A thousand shapes, no skeleton 3:03 Level 2: why a sixth finger hits so hard 3:25 The android in the brain scanner 4:14 The same kind of mismatch 4:26 Level 3: monkeys and almost-right faces 5:05 Mori's guess: a warning 5:22 Level 4: the label experiment 5:48 Same pixels, different feeling 6:08 Two kinds of wrongness 6:40 Mori's advice: aim for the first hill 6:59 The wooden hand