67Who Sees Human?

Adam Waytz, John Cacioppo and Nicholas Epley, "Who Sees Human? The Stability and Importance of Individual Differences in Anthropomorphism", Perspectives on Psychological Science 5(3), pp. 219–232, 2010

Linked to 18 chapters: supports 7 · contradicts 3 · mechanism 7 · same claim 2 · pairing 4

General Motors bought time in the 2007 Super Bowl for an advertisement about manufacturing standards. A factory robot drops a screw, is fired, drifts through menial jobs and throws itself off a bridge. Viewers rated it the fourth most popular advertisement of the game, and the American Foundation for Suicide Prevention demanded its withdrawal on the ground that it showed suicide as an answer to losing work. The complaint treats the robot's despair as legible, which is exactly what the advertisement was counting on. Waytz, Cacioppo and Epley take that response as their subject. Anthropomorphism, on their definition, is the attribution of a humanlike mind — intentions, consciousness, emotion — to something that is not human, and it is neither universal, as David Hume claimed, nor the small decorative matter psychology had treated it as, "more cute than critical". Some people do it far more than others; the difference is stable across months; and it can be measured with fifteen questions. The paper builds that measure and then shows what it forecasts: how much moral care a person extends to an animal, a plant or a machine, how much responsibility and trust they place in it, and how much they behave as though it were watching them.

What anthropomorphism is

Xenophanes, in the sixth century BC, noticed that gods resemble their believers, and that observation contains the two ways of making something human. One attributes human form: a face, hands, a smile in the grille of a car. The other attributes a human mind. The authors argue that the second is the essence, and that the first matters only because it implies the second, since a humanlike face or a humanlike gait is normally taken as evidence of humanlike thought behind it. Calling a fox quick is a description of behaviour. Calling a fox wily is anthropomorphism, because guile is a mental quality. Anthropomorphism also goes past animism, which grants a thing life without granting it a mind. What makes the line worth drawing where they draw it is that both philosophical and everyday accounts of personhood rest on mental states: self-reflection, conscious intention, rational thought in the philosophers; imagination and humiliation in the traits and emotions that ordinary people treat as uniquely human. So anthropomorphism can be operationalised as one species of mental state attribution, and measured as such.

The authors set aside the question that had occupied psychology for decades, which is whether the attributions are correct. Whether a dog or a hurricane or a god really has the qualities people give it is a separate matter from what makes people give them, how often, and to what effect. Hume held the tendency to be universal — "there is a universal tendency among mankind to conceive all beings like themselves" — and marketers evidently hold it to be consequential. The paper tests both claims and finds the first overstated and the second correct.

Where the differences come from

The account of the mechanism comes from the theory Epley, Waytz and Cacioppo had published three years earlier, which makes anthropomorphism the product of three factors rather than one reflex. The first is the knowledge that gets elicited. Human beings know their own kind better and earlier than anything else, so human concepts are the most accessible model available; faced with something to explain, people reach for that model first and correct it afterwards, and often incompletely. The second is effectance motivation, the need to interact with the environment competently, which makes unpredictability aversive. An agent with intentions is an agent whose next move can be anticipated, so granting a mind to a stalling car or a falling market buys a feeling of understanding. The third is sociality motivation, the need for contact with other minds, which can be satisfied by supplying minds where none exist. Lonely people, in the same group's earlier experiments, anthropomorphised gadgets and pets more than socially connected people did.

Because the two motivational factors vary between people and the accessible knowledge varies with culture, education and habits of reasoning, the theory predicts that the tendency should differ from person to person and hold reasonably steady within a person. No one had tested that. The measures in circulation either asked directly how human an agent looked, or ran to between seventy-eight and two hundred and eight items, or measured a relationship with one particular target such as God or a television character, or tracked dispositional attributions that are not anthropomorphic at all.

The questionnaire

The authors generated items by crossing four classes of commonly humanised agent — nonhuman animals, natural entities, technological devices and spiritual agents — with five mental attributes and five plain ones. The mental attributes were having a mind of its own, free will, consciousness, intentions and the capacity to experience emotions. The plain attributes were durability, usefulness, good looks, activity and lethargy, chosen because they describe observable or functional properties and so involve no mind at all. The plain items are not part of the scale. They are there to show that a high scorer is attributing minds rather than attributing dispositions generally, and that the score does not simply record a habit of using the top of a rating scale. Every item asks how much, on a scale from zero to ten: whether the wind has intentions, whether the ocean has consciousness, whether the average computer has a mind of its own, whether a tree does.

Three hundred and forty-eight people at the University of Chicago completed the forty items in the first study. Factor analysis returned three factors, and the third was the informative one: it gathered every item about spiritual agents, the mental and the plain alike. Participants were not distinguishing between saying a god has intentions and saying a god is durable; they were answering a question about whether gods exist. Since the scale was meant to measure anthropomorphism rather than religious belief, the spiritual items were dropped. The remaining thirty items resolved into two factors, one for animals and one for everything else, with the plain items loading weakly and diffusely across both.

A second study gave the revised fifteen items, with the fifteen plain items alongside, to six hundred and nine people drawn from the general population. Confirmatory analysis reproduced the two-factor structure, and the two factors correlated at .52, which is high enough to suggest that they are aspects of one tendency rather than two. A second-order model with a single superordinate factor of general anthropomorphism fitted well, loading .88 on the animal factor and .57 on the nonanimal one. Internal consistency across all the studies reported ran at .82 or above. A third study brought participants from the first back twelve to nineteen weeks later; their scores correlated .55 with their earlier ones. The tendency is not a mood. The plain items, as the authors note in a footnote, were never internally consistent, which is what they expected, since a diffuse set of unrelated attributions is not supposed to cohere.

What it predicts

A measure of a stable trait earns its place only by forecasting something. The authors work from three consequences that follow from granting an agent a mind. An agent with a mind can experience things, and so has interests that can be harmed. An agent with a mind acts on purpose, and so can be held to account and, by the same token, relied upon. An agent with a mind can observe and judge, and so can make a person behave.

The first test used secondary emotions, the ones people take to be distinctively human — shame, remorse, nostalgia, humiliation — as against primary ones like fear and pain. Forty participants watched a short video of three kittens playing and another of two snakes fighting, and rated how much each animal felt each of twenty emotions. Scores on the questionnaire predicted the attribution of secondary emotions strongly, and still predicted them when the attribution of primary emotions was held constant, so the effect was not a general willingness to see animals as feeling anything at all.

Fifty visitors to the Museum of Science and Industry in Chicago then judged three moral dilemmas: whether it would be wrong to destroy Deep Blue, whether it would be acceptable to leave a bed of rare flowers to be demolished, whether it would be acceptable to wreck a prized motorcycle to save a life. Two further vignettes, about waiting for a computer to fall in price and swapping an ingredient in a cookie recipe, carried no moral weight and served as controls. High scorers judged the destruction of all three objects more wrong, and the effect survived controlling for the non-moral judgments; on the non-moral vignettes the questionnaire predicted nothing. A further study found the same relation with concern for the environment, measured by four ordinary attitude items about forests, pollution and the well-being of nature.

Trust was tested next, on fifty-four adults asked to choose, in six scenarios, between a person and a machine: a doctor or an algorithm to estimate heart attack risk, a trained psychologist or a polygraph to judge a murder suspect's honesty, and similar choices about ranking football teams, washing fragile dishes, costing pollution control and selecting university applicants. Controlling for age, higher scorers were more willing to hand the decision to the machine. The authors read this together with existing work in which people who had been led to anthropomorphise a robot they worked with assigned it more of the credit for the joint result. Then they name three problems with their own finding. The scenarios were hypothetical and cost the participants nothing. If a humanlike machine looks more capable, people working with it may put in less effort themselves. And if a humanlike machine looks more responsible, the people who designed it, wrote its rules or were nominally in charge of it may look less so — a question the authors raise about armed robots, admissions software and cars, and about who answers for a war crime, a discriminatory decision or a death on the road.

The last study took the third consequence, that a mind can watch. Thirty-eight participants answered eight items from the Marlowe–Crowne measure of socially desirable responding — questions such as whether they had ever taken advantage of someone — put to them on screen by Kismet, a robot built to be read as expressive. Higher scorers gave more flattering answers. Being observed changes behaviour, and it appears to change it in proportion to how much of a mind the observer is credited with.

Who counts as human

The authors spend the last third of the paper arguing that this belongs to more than psychology. In human–computer interaction, anthropomorphising an agent makes it seem more understandable and more predictable, makes avatars seem more intelligent and more credible, raises engagement with an interface, and lets companion robots give something like social support to elderly people. Against that stand the animated paperclip in Microsoft Word, disliked for interrupting like a colleague with no sense of timing; Mori's uncanny valley, where a machine that looks almost human repels; the tendency to blame a humanlike interface when the software fails and to feel less credit when it succeeds; and expectations of competence that the machine cannot meet. The authors decline to say what the right amount of anthropomorphism is. Their contribution, they say, is to establish when it occurs and in whom.

In business the same tendency runs through marketing, where brands are given personalities and a watch is photographed at ten past ten so that its hands read as a smile, and through finance, where it does more damage. Describing a market in human terms rather than mechanical ones makes people more confident that a price trend will continue, and high scorers on the questionnaire made that prediction more readily, as though a run of prices expressed the settled intentions of an agent. In law, the whole question of an agent's rights turns on what mental capacities it is credited with. Spain's lower house had just backed a manifesto extending rights to great apes on the strength of their self-recognition and emotional capacity; the same reasoning governs debates about research animals, about what may be eaten, and about people whose mental capacities are contested — the fetus, the brain-damaged patient, the person diagnosed as a sociopath.

That last list is where the paper turns. Humanness, the authors argue, is a single continuum, and the questionnaire measures a position on it that can be read in either direction. To attribute a mind to something that lacks one is anthropomorphism; to withhold one from someone who has it is dehumanisation. The evidence for treating them as one dimension is that the same variables move both. Resemblance to the human prototype raises anthropomorphism, and distance from it invites dehumanisation. Loneliness raises anthropomorphism, and social connection makes people readier to deny minds to others. The moral consequences are mirror images: attributing a mind extends care and concern, while denying one licenses aggression, discrimination and the retrospective justification of harm already done. A measure of who sees human in a robot is, on this argument, also a measure of who fails to see it in a person.

Since 2010

The questionnaire became the standard instrument for the trait and has been carried into human–robot interaction, consumer research and environmental psychology, translated for other populations and cut down to shorter forms. The trust result travelled furthest. Waytz, Heafner and Epley put drivers into a simulator in 2014 and gave one group's car a name, a gender and a voice; those drivers trusted it more, relaxed more, and blamed it less for a collision that another driver caused. The sociality leg of the theory has held up least well, since experiments that induce loneliness have produced mixed results, and the samples behind the studies reported here — between thirty-eight and fifty-four people, one city, hypothetical scenarios, correlations rather than manipulations — are the size the next decade taught the field to treat with caution. The dehumanisation claim has been contested from a different direction: Harriet Over argued in 2021, in the same journal, that people who treat others with cruelty usually understand perfectly well how human their victims are, which makes a failure of mind attribution a poor account of atrocity. What has needed no defence is the technological strand. Machines that produce fluent speech now draw mind attribution from almost everyone without any design decision aimed at producing it, and the question the authors raised in a paragraph about armed robots — whether a machine that looks like an agent absorbs the responsibility of the people behind it — has moved out of the laboratory and into court.