Down in the Valley: What the Evidence Really Says About Uncanny Robots

The uncanny valley is the most famous idea in robot design and one of the least tested, so it is worth asking which parts of Masahiro Mori's 1970 hunch have survived contact with data.

A short essay with a long shadow

In 1970 the Japanese roboticist Masahiro Mori published a brief essay in a little-read journal called Energy. It contained no experiments. It offered a graph and a thought. As a robot becomes more humanlike, Mori suggested, our sense of affinity for it rises, but not smoothly. Just before full human likeness, there is a sudden plunge into eeriness, before affinity climbs again for a real healthy person. He called the dip bukimi no tani. The usual English rendering, uncanny valley, was a later choice; the 2012 authorised translation by Karl MacDorman and Norri Kageki notes that something like valley of eeriness would be closer.

Mori's examples were vivid. A realistic prosthetic hand looks fine until you shake it and feel that it is cold and limp, at which point the illusion collapses into something unsettling. He argued that movement deepens the effect, making the valley steeper, and he placed the stillness of a corpse at the bottom of the valley. His practical advice was modest: designers should aim for the first peak, a moderately humanlike appearance that earns affinity without risking the fall, rather than chasing perfect realism.

From folklore to experiment

For decades the essay was obscure outside Japan. Then computer graphics and humanoid robotics brought it roaring back, and the uncanny valley became a catch-all explanation for every box office flop with dead-eyed animated characters. Rigorous tests were slower to arrive.

One of the most cited is by Maya Mathur and David Reichling, published in Cognition in 2016. They gathered photographs of 80 real robot faces, ranging from obviously mechanical to strikingly humanlike, and had participants rate how mechanical or human each looked and how likeable it was. They also built a controlled series of digitally manipulated faces that stepped gradually from machine to human. In both sets, likability followed a curve with a clear dip near the humanlike end, much as Mori had sketched.

They went further. In an investment game, participants decided how much money to entrust to each robot, a behavioural measure of trust. The valley showed up here too, suggesting the effect reaches beyond stated preference into decisions. The trust effect was weaker and more sensitive to other features, such as the emotion the robot's face displayed. The authors also tested a popular explanation, that eeriness comes from being unable to decide whether something is human or machine. They found some sign of category confusion near the valley, but it did not explain the drop in likability.

Which valley, exactly?

Not every study finds a clean dip, and much depends on what is measured and how stimuli are built. A 2015 review by Jari Kätsyri and colleagues at Aalto University in Frontiers in Psychology sorted the empirical literature by hypothesis. They found little consistent support for the simplest version, in which humanlikeness alone produces a valley: in the studies they examined, liking more often rose fairly steadily with human likeness, and no consistent evidence that movement makes things worse, despite Mori's claim. The hypothesis with the strongest backing was perceptual mismatch: eeriness arises when cues conflict, for example realistic skin paired with cartoonish eyes, or a human face with mechanical motion. Evidence on category ambiguity was too thin to judge.

Read together, these papers suggest the valley is real in some conditions, particularly for static faces arranged along a mechanical-to-human scale, but that it is better thought of as a family of effects than a single law of nature. What unsettles people is often not too much realism but realism that is uneven.

How designers respond

The most common response is Mori's own: stay on the near side of the valley. Many commercial social robots use simplified or stylised faces, big eyes, smooth shells and limited features. A 2018 survey of 157 screen-faced robots by Alisa Kalegina, Maya Cakmak and colleagues found that very minimal faces, sometimes just eyes, are widespread, and that people's preferences for realism depend on the robot's intended role.

A second response follows from the mismatch hypothesis: whatever level of realism you choose, keep it consistent. Faces, voices and movements should belong to the same kind of creature. A highly realistic voice coming from a toy-like body, or lifelike skin with jerky motion, invites exactly the conflict that the evidence links to eeriness. Projected faces such as Furhat's take a middle path, offering a humanlike face with fluid animation on a form that is visibly a machine, though how such hybrids fare against the valley has not been settled by large controlled studies.

A third response is to question the obsession. The valley is one factor among many in how people judge robots, and it may fade with familiarity or matter less than whether the robot is useful and well behaved. Mori offered a design intuition, not a theorem. Half a century later, the intuition looks partly right, partly overstated and still worth taking seriously.

References

1. Mori, M., MacDorman, K. F., & Kageki, N. (2012). The uncanny valley [From the field]. IEEE Robotics & Automation Magazine, 19(2), 98-100. (Translation of Mori, 1970.) Link
2. 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. Link
3. Kätsyri, J., Förger, K., Mäkäräinen, M., & Takala, T. (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. Link
4. Kalegina, A., Schroeder, G., Allchin, A., Berlin, K., & Cakmak, M. (2018). Characterizing the design space of rendered robot faces. Proceedings of the 2018 ACM/IEEE International Conference on Human-Robot Interaction, 96-104. Link

Feature drafted with AI from the published research listed above, for social-robotics.com.