Anthropomorphism: Why a Moving Shape Looks Like It Wants Something

Mind attribution is the default rather than a special effort. People have one rich model of a mind, their own, and they reach for it whenever something behaves in a way that needs explaining. A thermostat that keeps changing the temperature is doing something, and the fastest account of doing something is wanting something. The account in terms of sensors and set points has to be learned, and it arrives second.

Nicholas Epley, Adam Waytz and John Cacioppo set out the standard theory in 2007 and named three determinants. The first is elicited agent knowledge: human-shaped cues bring human-shaped explanations, so faces, voices, names and eyes do most of the work. The second is effectance motivation, the need to explain and predict what something will do next, which is why an unpredictable machine gets a personality and a reliable one does not. The third is sociality motivation, the need for contact, which is why lonely people in their studies attributed more mind to pets and gadgets.

Motion is a cue in its own right. A thing that starts moving on its own, changes speed, turns toward another thing or seems to respond to it reads as animate before any judgment is made. That is why the 1944 film below works with no faces at all, and it is why the verb in the demo above matters: fleeing is motion plus a reason, and once the reason is on the table the rest of a mind tends to follow.

Designers use this on purpose. Adam Waytz, Joy Heafner and Nicholas Epley put 100 people in a driving simulator in 2014, in a normal car, a self-driving car, or the same self-driving car given a name (Iris), a gender and a voice. Trust in the vehicle's competence rose with the human features, though the car's actual behavior was unchanged. The same lever sells voice assistants and makes a delivery app feel like a person who is sorry.

Calling all of this an error would be too quick. Attributing goals to a dog, a chess program or a company is often the most efficient way to predict what it will do, and the attribution is wrong only about the mechanism. The failures are specific: trusting a system because it sounds human, reading a market or a storm as having intentions, or judging an animal's welfare by assuming its feelings match yours.

Fritz Heider and Marianne Simmel published "An experimental study of apparent behavior" in The American Journal of Psychology in 1944. They made a short animated film in which a large triangle, a small triangle and a disc move around a rectangle with a hinged opening. There are no faces, no bodies and no sound. The shapes approach, collide, separate, and in the end the smaller shapes leave the screen and the large triangle breaks the walls of the rectangle.

Thirty-four observers were asked simply to write down what happened. All but one described the film as the behavior of living beings. The accounts agreed closely on the plot as well as the cast: the large triangle was a bully or an aggressor, the smaller shapes were a pair who escaped, the opening was a door. The movements on the screen are geometry; the reports were about motives, and the paper's point was that the perception of behavior, including who did what to whom, is built from the pattern of movement itself.

The film is still in use eighty years later, which is one reason Adrianna Ratajska, Matt Brown and Christopher Chabris built 32 new animations in 2020 and asked whether the finding reproduces outside the original clip. Their 66 participants, writing free narratives about videos of 13 to 23 seconds, spontaneously attributed social roles and emotions to the shapes in the same way. People differed in how much they did it, and the videos differed in how much they invited it.

What the shapes are doing matters, not just that they move. Carly Rasmussen and Yuhong Jiang had adults in 2019 hold down one button while they saw social interaction in the Heider and Simmel film and another when they did not. Judgments of interaction tracked the motion of the shapes, and playing the film backwards lowered them, which places part of the effect in low-level movement cues and part in a higher-level reading of what the movement means.

The pattern appears wherever something behaves in a way that invites a reason. The rows below are illustrations, not study findings.

Anthropomorphism is not always worth countering: treating a dog, a company or a program as having goals is often the quickest route to a correct prediction. What is worth countering is letting the human-shaped cues set how much you trust something.

• Separate the cues from the record. A name, a voice and a face are design choices; competence is a measured rate of error. Ask for the second before relying on the first.
• Restate the behavior mechanically once. If a machine is doing something you can only describe as wanting, you do not yet know how it works, which is worth knowing before you depend on it.
• For animals, ask what the species does rather than what you would feel. The welfare question is answered by veterinary and behavioral evidence, not by the emotion the posture suggests.
• Watch for the two conditions the 2007 theory names: things get minds when they are unpredictable and when the person judging is short of company. Both are reasons to slow down.
• With markets, weather and institutions, name the parts. Replacing an intention with a list of actors and constraints usually changes what you do next.
• Be careful with the inverse. The same machinery runs backwards when a person is described in mechanical terms, a far more consequential error than a scolded printer.

Anthropomorphism is about granting a mind. Its neighbors differ in what is granted, and to what.

• It is not pareidolia, seeing a face or a figure in clouds, toast or tree bark. Pareidolia is a perceptual pattern match, and it ends there unless the thing is also given motives.
• It is not the illusion of external agency, the sense that an outside force arranged your circumstances. That is a claim about who caused your situation rather than about the mind of an object.
• It is not automation bias, the tendency to defer to an automated recommendation. Anthropomorphism is one of the things that raises trust in a system; automation bias is what that trust does to a decision.
• It is not the same as empathy with animals. Granting an animal feelings is often accurate; the bias is in assuming those feelings are human ones, with human reasons attached.