66The Media Equation

Byron Reeves and Clifford Nass, The Media Equation: How People Treat Computers, Television, and New Media Like Real People and Places, CSLI / Cambridge University Press, 1996

Linked to 20 chapters: supports 12 · contradicts 3 · mechanism 10 · same claim 2 · pairing 5

Byron Reeves and Clifford Nass arrived at Stanford's communication department in 1986 with offices two doors apart and, for several years, no reason to talk: one studied what television did to viewers, the other what computers did to users. What they eventually found the two had in common became a research programme and then this book. Its claim is a single equation, "media equal real life", and its method is a piece of borrowing so plain it sounds like a joke. Take a published finding about how people treat one another — that they are polite to a man who asks after his own performance, that they like strangers whose personalities resemble their own. Cross out the human being and write in a medium. Then run the same experiment, with the same measures and the same statistics, and see whether the result survives the substitution. Across thirty-five studies it did, and it did on plain text on a black-and-white screen as reliably as on a ninety-inch home theatre. People were polite to a computer that asked about itself, were moved by flattery they had been told was random, preferred the machine whose conversational style matched their own, took praise more seriously from a male voice, rated identical news more highly on a television set with a sign on it reading News Television, and changed their answers to agree with a machine they had been told shared their grade. The participants were daily computer users, many of them programmers, and every one of them, on being told afterwards what had been measured, denied doing it. Reeves and Nass argue that the denial is honest and the behaviour is real, and that what explains both is not ignorance but evolution: the brain took shape in a world where anything that acted socially was a person, and there is no switch in it that can be thrown when the thing on the other side of the conversation is made of wire.

The equation, and how it was tested

The book opens with four scenes and an easy dismissal of each. Preschoolers say that the popcorn on a television screen will spill if the set is turned over; children take television literally and will grow out of it. A novice says hello back to an animated dog that greets him; novices like special treatment and experts do not need it. A candidate in a televised debate walks toward the questioner and so toward every viewer at home, and viewers warm to him; they were distracted and would not have been fooled had they stopped to think. A robotics team describes a one-legged hopping machine that has memory, can learn, can accept punishment and reward, and has a personality; the engineers were using a convenient metaphor. Each dismissal turns on age, inexperience, distraction or figure of speech, and each concludes that the confusion is rare, inconsequential and curable by thought. Reeves and Nass argue that it is none of those things: it is common, easy to produce, cheap to provoke, and thinking will not make it go away.

The procedure is set out as seven steps, offered with what the authors call simplifying overstatement. Pick a social-science finding about how people respond to each other or to the physical world. Find the sentence where the rule is stated — people like to be praised even when the praise is undeserved; the onset of motion draws intense focus. Cross out person and write in medium, which yields a prediction with its explanation already attached. Find the passage describing how the rule was tested, cross out one of the people there too, and you have a method. Run it, and draw the implications. The authors insist on the sixth step: they assumed nothing about whether the results would carry over, and an experiment counted as successful only when its numbers matched the original study's.

Two readings are ruled out early. The first is anthropomorphism, the belief that an inanimate object is really human; that, the authors say, is rare, and it is not what is happening, because the social responses appear in people who know the responses are inappropriate and say so afterwards. The second is that people willingly suspend disbelief, pretending for the pleasure of it. Nobody in any experiment reported pretending. What is automatic, Reeves and Nass argue, is belief; it is the correction that takes work. A mental reminder that it is only a movie does release a person from a frightening film, but it costs enough attention to make the plot hard to follow, and it cannot be sustained.

Their own explanation is evolutionary and stated without hedging. For nearly all of the two hundred thousand years in which Homo sapiens has existed, anything that behaved socially really was a person and anything that appeared to move toward you really was moving toward you. Those were absolute truths for long enough that acceptance of what seems real became automatic, and modern media engage brains built for that world. Because treating a person as an object costs far more than treating an object as a person, people are built to make the conservative mistake: "When in doubt, treat it as human." A medium therefore does not need to be convincing, only close enough. A puppet is close enough, which is why a senator at a hearing on television violence could ask Shari Lewis whether she agreed with Lamb Chop, and why nobody in the room laughed for a moment.

Media and manners

The prototype experiment is about politeness, and it starts from Ed Koch, three-term mayor of New York, and his habit of ending speeches by asking the crowd how he was doing. Ask a New Yorker that in person and you get a warm answer; have a pollster from the Times ask him the same thing that evening and you get a truer one. Reeves and Nass extract two rules the psychological literature supports. People who ask about themselves receive more positive answers than a third party asking the same question; and because politeness compresses the range of what may be said, those answers are also more alike.

Twenty-two people were told they would work with a computer to learn about various topics and would afterwards evaluate it. Each sat in front of a black NeXT machine with a twenty-one-inch black-and-white monitor showing nothing but text and buttons — no pictures, no voice, not even an icon. The computer presented twenty facts, one of them that according to a Harris Poll thirty per cent of American teenagers kiss on the first date, and after each it asked how much the user already knew and said it would tailor what followed; in fact everyone got the same material in the same order. It then tested them, marked the test, and announced that it had done a great job of teaching. Half the participants evaluated the session on that same computer, rating how well each of twenty-two adjectives — accurate, analytical, competent, fair, friendly, helpful and the rest — described it; the other half wheeled their chairs across the room and answered the identical questions on a second, identical machine.

The people who answered on the tutoring computer rated it more positively on twenty of the twenty-two adjectives, and the spread of their answers was narrower. Two replications closed off the obvious objections. When the other party was a paper questionnaire rather than a second machine, the same gap appeared, so the effect was not about where the boxes stood. When the whole procedure was repeated with recorded human speech, the gap appeared again and was no larger — the more telling of the two results, since voices, for all the cost of storing them, added nothing to the sense that someone was present. Afterwards, no participant guessed that the number of computers mattered or that the study concerned politeness, and all said they would never soften an evaluation to spare a machine's feelings. Those who had used two computers all assumed a single programmer had written both, and were right, which disposes of the idea that the politeness was aimed at him.

Reeves and Nass turn the finding into a design argument with a hinge: the reason machines must be polite is that people are polite to machines, and politeness is reciprocal or it is nothing. A machine that breaks a social norm is not read as under-engineered; it is read as rude. The prescription is borrowed from Grice's maxims. Quality is the one computers keep, though accuracy without cooperation is its own offence, as when a driver who asks where he is is told that he is in a car. Quantity is broken by one-word menus and by error messages that offer either a code or forty pages; relevance, by interfaces that assume a user has only the task in mind, as early television news assumed viewers watched only to be informed; clarity, by precision bought at the price of being unreadable, for which one of the authors offers a sentence of his own. Smaller rules follow: say hello and good-bye rather than evaporating, do not turn your back without announcing it, answer in the modality you were addressed in, and never let the same machine demonstrate a product and then ask what you thought of it.

The next study moves to something more primitive than etiquette. Thirty-two people watched faces in every combination of three distance cues that professionals treat as unrelated — viewing distance at four or ten feet, screen size at forty-one or fifteen inches, framing from chin-to-crown or head-to-toe — and faces that appeared close were judged more extremely in both directions, drew about ten per cent slower reactions to a tone the viewer had to answer while watching, and were recognised about twelve per cent more often.

Flattery is tested with a guessing game. Sixty people played an animal-guessing program twelve times each; whenever it failed, it asked the player for a better question. One group received no comment, one was praised, one was criticised in the same register. Crossing this, half of each group were told the evaluations came from algorithms built on hundreds of previous players, and half were told plainly not to believe a word, because the evaluation system did not yet exist and the comments were random samples. Flattered participants thought they had chosen better questions than the unevaluated group did, liked the computer more and thought it had played better — and it made no difference whatever whether they had been told the praise was random, which everyone remembered correctly afterwards. Criticism ran the other way, as the literature predicts: people told the criticism was baseless discounted it, rating themselves and the machine higher than those who believed it.

The last of the manners studies makes the participant an onlooker. Forty-four people were tutored by one computer, tested by a second and evaluated by a third, each with a different recorded male voice; for half of them the evaluation came from the tutoring machine itself. Praise of the tutor raised every rating of it, and raised participants' estimate of their own score by about two answers. But praise from an independent machine counted for more than self-praise and was better liked, a machine that criticised another was liked less than one that criticised itself, and the machine that criticised was rated more intelligent. Two on-screen financial experts who recite their own credentials, the authors note, would have been more credible and better liked had each introduced the other, at no cost in code.

Media and personality

Personality psychologists reduce the adjectives people use about each other to five dimensions, of which the first two, dominance and friendliness, carry most of the work. Two hundred children in the third, fifth and seventh grades rated the similarity of every pair of fourteen popular television characters, a method chosen because it never names a trait and so cannot plant one; dominance came out as the first dimension, with Steve Austin at one end and Laura from Little House on the Prairie at the other, and friendliness as the second. Repeated with forty-two adults and with original line drawings of animals and objects made by five artists so that nothing could be familiar, the same two traits reappeared. Give anything eyes, the authors conclude, and personality follows.

The interface study has the cleanest manipulation in the book. Two hundred people took a standard dominance questionnaire, and the twenty-four most dominant and twenty-four most submissive were invited back weeks later to work with a computer on the Desert Survival Problem, ranking twelve items salvaged from a crash. The machine differed in four respects and no others: it phrased its contributions as assertions and commands or as questions and suggestions; it displayed a confidence number averaging eight out of ten or three out of ten; it spoke before the user or after; and it was called Max or Linus. The information it gave was identical either way, and it made no attempt to model the user. Participants read the intended personality off the machine accurately, recognised their own style in the matching one, and — when the two matched — rated the same advice more intelligent, insightful and useful, enjoyed themselves more, and were more satisfied with their own rankings.

Eighty-eight people then took the same problem in two rounds, with the machine's style either constant across both or reversed between them. Similarity that had to be earned beat similarity that was simply there: a computer that began unlike the user and converged on his style was preferred, and rated more competent, over one that had matched all along; a computer that changed away from the user still did better than one that had never matched, because the attempt earned credit. This, the authors point out, is one of the few places where they put something back rather than borrowing, since gain theory had never been tested on personality in humans either.

The design consequences are deliberately cheap. Personality is not an artificial-intelligence problem; it took the two of them a couple of days to write enough of one into a desktop machine to change what users thought of the machine, of their work and of themselves. Every interface that shows words has a personality already, error messages included, and writing by committee produces a hybrid that users dislike. The strength of a personality means its identifiability, not its dominance: identify it and roughly half the audience will like it, leave it ambiguous and almost nobody will. Flaws help; inconsistency does not.

Media and emotion

Emotion is treated on the two dimensions psychologists use, valence and arousal, and tested first at the level of cortical activity. Sixteen right-handed women watched five one-minute positive scenes — weddings, births, smiles, soft focus — and five negative ones, with frontal electrodes recording an electroencephalogram. Positive material produced more left-hemisphere activity and negative material more right, the lateralisation found when the cause is a drug, a mood disorder or the taste of a lemon. There is no separate neural mode for material that arrives on a screen.

Negativity then gets three studies, on the strength of what psychologists call hedonic asymmetry: pleasure is adapted to, pain is dwelt on. Adults turning a dial to register liking as they watched political advertisements rated the negative ones lower, and did so within the first ten seconds; pictures and sounds from those advertisements were later recognised more accurately and faster, positive material taking twenty-three per cent longer to place. Ten news stories were then built in matched pairs, each with the same anchor introduction, voice-over and wrap-up, differing only in whether the middle carried graphic footage. Two months later, viewers of the graphic versions recalled more of them and remembered the wrap-up better and the introduction worse — proactive enhancement and retroactive interference, the pair of effects the literature reports for unpleasant experience generally. The gain is memory for the wrong thing: gory pictures did not teach anyone more about the event.

Arousal is then shown to be independent of valence. Eighty-three viewers rated 258 six-second clips on pictorial scales, and the map places arousing positives, such as a rocket launch at night, as far from calm positives such as a leaf on a river as from arousing negatives. Sixty of those clips, shown to 125 people, were recalled in proportion to how arousing they were, five minutes later and again after two weeks; and since arousal accumulates across events regardless of sign, the advice to an advertiser is to buy the first slot in the break after an exciting scene.

Media and social roles

Roles are where the medium being treated socially is least plausible, and the results are correspondingly blunt. A television set was given a sign. One group watched four news stories and four situation-comedy segments on a single twenty-inch set labelled News and Entertainment Television; the other watched the news on a set labelled News Television and the comedies on an identical set beside it labelled Entertainment Television. Nobody remarked on the signs. Of twenty-eight assessments of the news, twenty-six favoured the specialist set, as did all sixteen assessments of the comedies. A second study moved the label from the box to the channel, presenting the same twelve local news stories as coming either from CNN, HNN, SNC and RNN or from ABC, CBS, Fox and NBC. The specialist networks' stories were rated more important, interesting, informative and serious, and — a result the authors had not imagined, and went back to check their tapes over — clearer and better in colour.

Teams were made with a wristband. Half the participants wore blue, used a computer with a blue border and a sign reading Blue Team, and were told they would be graded on their work and the computer's together; the others wore blue, used a machine labelled the Green Computer, and were graded alone. Teamed users thought the machine more like them, thought its identical suggestions more relevant and helpful, tried harder to reach agreement — and changed more of their survival rankings to match it. Two further pairs of groups separated the ingredients that social psychology bundles together: interdependence alone produced every team effect, identity alone produced a weaker version of some, and identity added nothing once interdependence was present. The relationship, the authors conclude, should be neither one-up, the machine as wizard, nor one-down, the machine as tool, but one-across — and the language required is trivial, since it is enough for the machine to say the task will go better if they work together, which is in any case true.

Gender was given to a computer by giving it a voice. Three male and three female recorded voices tutored participants on three subjects: love and relationships, mass media, and computers. A second, text-only machine tested them; a third reported their results and praised the tutor's teaching, in a male voice for half of them and a female voice for the rest. Men and women alike were more swayed by praise from the male voice, to the point of rating the tutoring machine across the room as having taught better when a male voice had commended it, and they found the male voice's identical sentences friendlier. The female-voiced tutor was rated the better teacher of love and relationships and the worse teacher of computers. Everyone denied holding gender stereotypes, and everyone assumed the programmer was male, which was true and, the authors note, ironic. A second study electronically raised or lowered the frequencies and volume of six women's voices while leaving their faces on screen; the masculine-sounding versions of the same women were credited with more drive, willpower, reasoning and persuasiveness. A male voice, the design advice runs, is not a neutral one, and a synthesised voice does not escape the problem, because every listener assigns it a gender, is certain of the assignment, and behaves accordingly.

That raises the question of how many social actors a machine contains, and the answer is one per voice. A computer that tutored, bragged and then asked for its evaluation in the same voice got the polite, compressed ratings of the original politeness study; the same machine asking in a second male voice got honest, varied ones. The finding was replicated on a prediction that reverses by gender — a male speaker praised by another male is credited more than one who praises himself, a female speaker praised by another female is credited less — and both halves held. Moving one voice onto a second machine changed nothing. The source-orientation study then asks where responsibility settles. Half the participants were told they were working with programmers who had different tutoring styles and would evaluate them; the other half heard the same sentences with the word computer substituted. Those never asked to think about a programmer rated the machine friendlier, more effective, more playful and more like themselves, and were more positive on twenty-eight of thirty adjectives. People can attend to a distant source, but only on instruction and only at a cost — the cost, the authors argue, that makes watching a play while thinking about the playwright exhausting. Which inverts the usual account: it is not belief in the characters that takes the effort.

Media and form

The last group of studies drops social rules for the physical properties of a display, and it is here that the equation earns its second half: media are natural as well as social. Size was tested at the two extremes a household might contain, a twenty-two-inch diagonal and a ninety-inch one, more than sixteen times the area. The large screen raised the arousal a hundred and twenty-five adults reported after each of sixty clips, and a week later those viewers recalled more of what they had seen. The explanation is peripheral vision: on a big screen the boundary of the picture leaves the corners of the eye, so action away from the point of focus is more of a surprise.

Fidelity split in two, and the authors predicted the split. Sixteen action scenes were shown either as computer-enhanced video projected bright on a six-foot screen or as fifth-generation dubs on bad tape through a dimmer projector, with visible scan lines. Attention, recognition memory and evaluation showed no differences at all, and viewers asked directly to rate the picture quality rated the two versions the same: poor video resembles the imperfect seeing people do all day, and is not intrusive. Poor audio has no counterpart in the world — tape hiss, a truncated frequency range, sound arriving from one direction when the picture says it should come from everywhere — and in the same experiment audio fidelity moved every measure. The advice to designers follows: spend on the audio.

Timing gave the sharpest version of the argument. Audio was advanced ahead of video by five frames, one-sixth of a second, a direction chosen because it never happens in life. Asked three successively more pointed questions about whether anything had seemed strange, only half the affected participants ever said yes; the speakers in those clips were nonetheless rated less interesting, less pleasant, less influential and more agitated, and the ratings held among viewers who had noticed nothing. They did not ask what was wrong with the technology; they asked what was wrong with the speaker. Motion was tracked with electroencephalograms recorded every half second through half an hour of television, and attention rose within a second of movement beginning and fell over every still shot.

The conclusions the book draws are stated plainly. Everybody does it, across the children, adolescents, undergraduates, homemakers, executives and computer scientists they tested. Media are more alike than different, and psychologically the personal computer is not far from the television. The responses are automatic, need only minimal cues, and reach far past yelling at a quarterback, into how attention is allocated, what is stored, and what happens in the body. What seems true matters more than what is true, which is the authors' answer to the Turing test and to Searle's Chinese Room alike: people do not judge intelligence in each other by its objective merits either, so the question about a machine is not whether it is the same as its counterpart but whether it is perceived as identical. And none of it could have been learned by asking, since participants told them the opposite of what the data said, every time.

Since 1996

The research programme already had a name in the literature — computers are social actors, after a 1994 conference paper by Nass, Jonathan Steuer and Ellen Tauber — and CASA became the standard label for the paradigm and for the recipe the book describes: give a machine a minimal social cue, withhold any suggestion that a person is on the other end, and rerun a study of people. The chapters became papers, among them Fogg and Nass on silicon sycophants, Nass, Fogg and Moon on computers as teammates, and Nass, Moon and Green on gender stereotyping. B. J. Fogg's work on persuasive technology grew out of the same laboratory, and Nass carried the voice results into Wired for Speech with Scott Brave in 2005 and into a book of management advice, The Man Who Lied to His Laptop, with Corina Yen in 2010. He died in 2013.

The explanation moved before the findings did. In Machines and Mindlessness, published in the Journal of Social Issues in 2000, Nass and Moon set out a mindlessness account: people apply overlearned social scripts to cues that trigger them, without ever concluding that the computer is a person, and the marker of the phenomenon is precisely the mismatch between what participants do and what they will endorse. That is a shift of emphasis away from the old-brain story told here, which locates the cause in evolutionary hardware, and toward attention and habit, which are easier to test and easier to disrupt. Later theorists moved further. Epley, Waytz and Cacioppo's three-factor account of anthropomorphism treats attributing mind to non-humans as varying with what a person already knows, how badly they need to explain the thing's behaviour, and how much they want company, rather than as a constant; Sundar's model of how interface cues become heuristics does something similar for media specifically; and in 2020 Gambino, Fox and Ratan proposed an extended version of CASA in which people, after decades of daily use, develop scripts specific to media rather than borrowing scripts wholesale from human interaction. That revision is a direct challenge to the book's strongest claim, which is that the borrowed scripts are applied without modification.

The empirical record is thinner than the book's confident prose suggests. The samples are small — twenty-two, thirty-two, forty-two, forty-four — the designs are one-shot, effect sizes are rarely reported, and by the authors' own account in the preface the aim was a bold statement to set beside the guarded conclusions of the journals. Attempts to reproduce the politeness effect with embodied on-screen agents have returned mixed results, and the question of which CASA effects survive replication has been pressed repeatedly since psychology's reckoning with reproducibility. What has survived best is not any single experiment but the paradigm. That social cues on a screen produce social responses in people who deny giving them is now the working assumption in the study of voice assistants and conversational systems, and the gendering of those assistants — the subject of UNESCO's 2019 report on the default female voice — is argued in terms Nass's experiments supplied. The book is cited most often now by people asking whether its equation still holds when the thing on the other side of the conversation produces language well enough that treating it as a social actor is no longer a mistake about what is there. Reeves and Nass would have said the question was settled in 1996, on a black-and-white screen showing nothing but text, and that the machine's actual capacities were never what the effect depended on.