CYBERCASCADES CYBERCASCADES Any discussion of social fragmentation and online behavior requires an understanding of social cascades—above all because they become more likely when information, including false information, can be spread to hundreds, thousands, or even millions by the simple press of a button. Cascades are often hard or even impossible to predict, but they are all around us, and they organize our culture and even our lives. Increasingly, cascades are a product of social media. They occur within isolated communities, which develop a commitment to certain products, films, books, or ideas. Terrorists, rebels, and revolutionaries attempt to create and use them. Frequently cascades take hold far more generally, helping to produce (for example) a right to same-sex marriage, a rebellion against an authoritarian government, a nation's exit from the European Union, a new president, or a massively popular new cell phone. It is obvious that many social groups, both large and small, move rapidly and dramatically in the direction of one or another set of beliefs or actions.1 These sorts of cascades typically involve the spread of information; in fact, they are usually driven by information. Almost all of us lack direct or entirely reliable information about many matters of importance—whether George Washington actually lived, whether the earth goes around the sun, whether matter contains molecules, whether dinosaurs existed, whether there is a risk of war in India, whether the Islamic State of Iraq and the Levant (ISIL) is dangerous, whether a lot of sugar is actually bad for you, or whether Mars is real. For the vast majority of your beliefs, you really don't have direct information. You rely on the statements or actions of trusted others. #98 TWO KINDS OF CASCADES To understand the social dynamics here, we need to distinguish between two kinds of cascades: informational and reputational. Informational cascades. In an informational cascade, people cease relying at a certain point on their private information or opinions. They decide instead on the basis of the signals conveyed by others. It follows that the behavior of the first few people, or even one, can in theory produce similar behavior from countless followers. To use a stylized example, suppose that Joan is unsure whether climate change is a serious problem. She may be moved to think that it is if her friend Mary thinks and says that climate change is .a serious problem. If Joan and Mary are both favorably alarmed /about climate change, their friend Carl may end up agreeing with them, at least if he lacks reliable independent information to the contrary. If Joan, Mary, and Carl believe that climate change is a serious problem, their friend Don will need to have a good deal of /confidence to reject their shared conclusion. And if Joan, Mary, ;£arl, and Don present a united front on the issue, their other ■friends and even acquaintances may well go along. Something like this happens online every day. It is important to emphasize a wrinkle here, which is that if one person sees that five, ten, a hundred, or a thousand people are inclined to say or do something, there is a tendency to think that each and every individual has made an independent decision to say or do it. The reality may well be that only a small fraction of the group made an independent decision. The rest are following the crowd, thus amplifying the very signal to which they were themselves subject.. That signal may be extremely loud and seem quite impressive even though it incorporates the judgments of remarkably few people. 1 Environmental issues provide examples of how information travels, and can become quite widespread and entrenched, whether or not it is right. A disturbing illustration is the widespread popular belief that abandoned hazardous waste dumps rank among the 99 CHAPTER 4 most serious environmental problems; science does not that belief, which seems to have spread via cascade.2 Anothe ' vironmental example is the widespread and false belief that f containing GMOs are hazardous to people's health; the scien; ^ consensus is that they are not. Many cascades are widespread Wk local; consider the view, which had real currency in some aSJm| American communities in the 1980s, that white doctors are respon sible for the spread of AIDS among African Americans. Or consider the notion, apparently widely held among American conservatives^ that President Obama was not born in the United States—and Ml opinion, held by many parents and apparently defended at one's point by Donald Trump, that vaccinations cause autism. One group j may end up believing something, and another the exact opposite/^ and the reason is the rapid transmission of information within one!! group but not the other. Even among specialists and indeed doctors, cascades are com-| mon. "Most doctors are not at the cutting edge of research; their} inevitable reliance upon what colleagues have done and are doing leads to numerous surgical fads and treatment-caused illnesses."3 Thus an article in the influential New England journal of Medicine' explores "bandwagon diseases" in which doctors act like "lem-.: mings, episodically and with a blind infectious enthusiasm pushing certain diseases and treatments primarily because everyone else is doing the same."4 It should be easy to see how cascades might develop among groups of citizens. And when informational cascades are operating, there is a serious social problem: people who are in the cascade do not disclose to their successors and the public the information (or reservations) that they privately hold. Reputational cascades. We can also imagine the possibility 01 reputational cascades, parallel to their informational siblings. In a reputational cascade, people think that they know, what is right, or what is likely to be right, but they nonetheless go along with the crowd in order to maintain the good opinion of others. Even the most confident people sometimes fall prey to this pressure, silencing themselves in the process. Fearing the wrath of others, #100 CYBERCASCADES •crht not publicly contest practices and values that they practice of sexual harassment long predated the legal abhor icial ■rivately ■The social F <"•■•"•---- " • 0f "sexual harassment," and the innumerable women who """"subject to harassment did not like it. But mostly they were '^lent simply because they feared the consequences of public ompla'nt- 1S mterest'n8t0 wonder how many current practices fall in the same general category—they produce harm, and are known ?° produce harm, but they persist because most of those who are harmed believe that they will suffer if they object in public. Whole governments can fall once reputational cascades start growing, as they often do when people learn that their disaffection is widely shared. To see how a reputational cascade might work, suppose that Albert suggests vaccinations can cause autism, and Barbara concurs with Albert, not because she actually thinks that Albert is right, but because she does not wish to seem, to Albert, to be ignorant or indifferent to a serious risk faced by children. If Albert and Barbara seem to agree that vaccinations can cause autism, Cynthia might not contradict them publicly and might even appear to share their judgment, not because she believes that judgment to be correct, but because she does not want to face their hostility or lose their good opinion. It.is easy to see how this process might generate a reputational cascade. Once Albert, Barbara, and Cynthia offer a united front on the issue, their friend David might be most reluctant to contradict them even if he thinks that they are wrong. The apparent views ■?f Albert, Barbara, and Cynthia carry information; that apparent •• Vlew might be right. But even if David thinks that they are wrong ar)d has information supporting that conclusion, he might be most '^luctant to take them on publicly. Reputational cascades impose '"creasing pressure as larger numbers of people join the cascade. A position that was once highly unpopular, leading people to silence .themselves, may come to seem widely held, so much so that people ■;v|}sk their reputations if they oppose it. #101 CHAPTER 4 INFORMATION AS WILDFIRE AND TIPPING POlNTs The Internet greatly increases the likelihood of diverse but' sistent cascades. Cybercascades occur every day. On Twitter^ Facebook, you can find them in an instant. They might involve pofj tics, miraculous products, deadly diseases, conspiracies, unsafe fo0|? supposed events in Moscow or Berlin, or anything else. Here is some fun and illuminating evidence of how online cas": cades can happen, from the domain of music.6 A team of expert menters, led by Matthew Salganik, Peter Dodds, and Duncan Watts, created an artificial music lab, including 14,341 partici? pants. The participants were given a list of dozens of previously un-: known songs from unknown bands; they were asked to listen to a" brief selection of any songs of interest to them, decide what songs' (if any) to download, and assign a rating to the songs they chose;; About half the participants were asked to make their decisions independently, based on the names of the bands and the songs and their own judgment about the quality of the music. About half the participants could see how many times each song had been downloaded by other participants. These participants were also randomly assigned to one or another of eight possible "worlds," or subgroups, with each evolving on its own; those in any particular world could see only the downloads in their own world. A key question was whether people would be affected by the choices of others—and whether different music would become popular in the different "worlds." Did social influences matter? Did cascades develop? There is not the slightest doubt. In all eight worlds, individuals were more likely to download songs that had been previously downloaded in significant numbers, and less likely to download those that had not been so popular. Most strikingly, the success of songs turned out to be almost entirely unpredictable! Almost all the songs could become popular or unpopular, with everything depending on the choices of the first downloaders. The identical song could be a hit or a failure—simply because other people, at the start, were seen CYBERCASCADES f: 'to download it or not. (Think for a moment about how t0 ^ rs spread or fail to spread on social media.) I ratn°Tfre sure, there is some relationship between quality and suc-| "In general, the 'best' songs never do very badly, and the j ^orst'songs never do extremely well, but almost any other result ossible."7 But even for the best and worst songs, there's a high [ degree of unpredictability in terms of ultimate market shares, de-ending on whether they benefit from early popularity—and for j vast majority of songs, everything turns on social influences. | Salganik, Dodds, and Watts acknowledge that in many ways, the real world is different from this experiment. They in fact con-[ trolled numerous variables, ensuring that their results are weaker than what happens in actual markets, where unpredictability is even greater, and where cascades are inevitable. Media attention, marketing efforts, critical reviews, and other pressures inflate the role of social influences. When experts fail to predict success, it is "because when individual decisions are subject to social influence, ■markets do not simply aggregate pre-existing individual preferences."8 Note here that marketers often try hard to create early online "buzz" by suggesting that a certain cultural product is already popular; indeed, some marketing efforts actually involve artificial ■ -efforts to overstate the demand for the product, through purchases not by ordinary people, but by those allied with the artist. Social media are full of such efforts. An acquaintance of mine, the author of an excellent book in the general domain of behavioral science, has tweeted on numerous occasions something like, My book is doing great and well above expectations! Thanks for : .the support!" Actually the book isn't doing so great, but the author knows well that if people think that other people are buying it, they'll be more likely to buy it themselves, pgiv Cpnsider in this regard the 2013 Oscar winner for best docu-; mentary, Searching for Sugar Man, a stunning film about an un-g^successful Detroit singer-songwriter named Sixto Rodriguez, who 0 released two long-forgotten albums in the early 1970s. Almost no one bought his albums, and his label dropped him. Rodriguez CHAPTER 4 stopped making records and worked as a demolition man. What Rodriguez didn't know, while working in demolition, was that he had become a spectacular success in South Africa—a giant, a leg. end, comparable to the Beatles and the Rolling Stones. Describing him as "the soundtrack to our lives," South Africans bought hundreds of thousands of copies of his albums, starting in the 1970s. Searching for Sugar Man is about the contrast between the failed career of Detroit's obscure demolition man and the renown of South Africa's mysterious rock icon. The film is easily taken as a real-world fairy tale, barely believable—a story so extraordinary that it gives new meaning to "you couldn't make it up." But as the music lab experiment shows, it is a bit less extraordinary than it seems, and it offers a profound lesson not only for music and culture markets but for business and politics as well. We like to think that intrinsic quality produces success, and that quality will ultimately prevail in free markets. To be sure, quality is usually necessary, but it's not enough. Social dynamics—who is conveying enthusiasm to whom, and how loudly, and where, and exactly when—can separate the rock icon from the demolition man, and mark the line between stunning success and crashing failure. And if this is true for online music, it is likely to be so for many other things as well, including movies, books, political candidates, and even ideas. ("Everyone is flocking to candidate X," or "idea Y is really catching on.") Candidates and ideas may enjoy stunning success (or failure) simply because social dynamics give them an early boost (or not). Here we can see a large effect from collaborative filtering, which may help move or entrench, and not merely reflect, individual preferences. POLITICAL CASCADES AND TURBULENCE These points suggest a hypothesis: political life is a lot like the music lab. In fact, it is a kind of real-world politics lab. Bill Clinton, George W. Bush, Obama, and Trump succeeded not only because #104 CYBERCASCADES f their evident talents but also because they received the equivalent of many early downloads. There are many talented politicians who never succeeded, and the reason isn't that they were not quite talented enough. It is that they failed to attract the right level of attention, either early or at some crucial time. The same is true for policy reforms. It is not simple to test this hypothesis, but Helen Margetts, Peter John, Scott Hale, and Taha Yasseri have made significant strides in ' their book, Political Turbulence? Their subtitle is How Social Media /Shape Collective Action, but their thesis is far more specific and >. striking. They argue that there is a great deal of unpredictability in .^modern political life, that the level of predictability is significantly ^increased by social media, and that social influences heighten unpredictability. Explicitly referring to the music lab experiment, they claim that in the age of social media, political movements are /likely to be highly turbulent. With respect to social influences, some of their best evidence comes from petitions. Both the United Kingdom and the United : States have created online petition platforms. Most petitions fail— and fail quickly. No one pays them the slightest attention. As it turns out, the first day that a petition becomes public is critical. Early popularity makes all the difference, because political momentum builds on itself. In the United Kingdom, five hundred signatures are required to obtain an official response, and a large : percentage of successful petitions get there within two days. It is reasonable to think that a certain (small) number of petitions spur and benefit from early cascade effects; they are a lot like Rodriguez in South Africa. But the vast majority of petitions fail to do that; I'they are just like Rodriguez in the United States. That is indeed a reasonable thought, but it is not the only reading of the data. It is possible that some petitions receive large numbers of independent signatories, and that social influences do not much matter. But Margetts and her colleagues offer strong reasons to think otherwise. For one thing, social media have a large effect. The number of signatures and the number of tweets are closely no5 CHAPTER 4 CYBERCASCADES correlated; the more tweets, the more signatures. The authors' analysis of both timing and content suggests that tweets are driving signatures, rather than the other way around. But the strongest evidence of the power of social influences comes from the fact that in April 2012, the UK Cabinet Office introduced "trending petitions" information on its web page so that everyone could see which petitions were succeeding and how many other people had signed. Margetts and her colleagues explore the effects of that information. Somewhat surprisingly, it had no effect on the overall level of petition signing. But it greatly affected the distribution of signatures. Using a method akin to that in the music lab study, the researchers find that after the trending information was introduced, signatures were much more concentrated on a small number of petitions. That is important evidence that "the information-rich get richer, and the information-poor get poorer."10 Note that we are speaking here of what kinds of petitions receive attention from high levels of government. On that question, what is observed mirrors the music lab experiment. As one might expect, Margetts and her colleagues find that small design changes can have large and unintended consequences. The United Kingdom lists the top six petitions (measured by number of signatures) in order on its website, and it also provides visitors with the option to click to see six more. Margetts and her colleagues tested whether and how the trending information affected people's signatures. The details of the test need not detain us, but the central finding is major: the first-ranked positions received more concentrated attention—and signatures—as a result of that information. The upshot is that "the addition of the trending petitions facility causes the most popular trending petitions to receive more signatures, and that these signatures come at the expense of signatures to other petitions on the site."11 We can undoubtedly reach broadly similar conclusions about how social media might promote or undermine political candidates. One result is unpredictability and turbulence, as modest differences in initial popularity map onto long-term variations. As a further test of social influences, the researchers enlisted a vebsite, WriteToThem, designed to help visitors write to public officials- The site reduces the costs of citizen engagement. In an experiment, people were randomly assigned to one of two groups. The first was the control, in which visitors to the site saw no social information. The second was the treatment, in which visitors could see how many others had written to a particular representative. Overall, 39 percent of those visitors who went to the page for their own representative ended up sending a letter. But there was a substantial difference between the two groups: 32.6 percent in the control, and 49.1 percent in the treatment. Surprisingly, it did not matter, in the treatment group, whether the social information showed low, medium, or high levels of writing from previous visitors. What mattered was the information as such. One reason may be that people did not have the comparative information right before them; without a little work, they would not see that some representatives had higher percentages ;than others. Another reason may be that the differences between ■low (around 47 percent) and high (around 53 percent) were quite • modest. With a wider range, we might expect something more like the music lab experiment and the petition data, where variations in numbers really did matter (and were visible). |Jp This expectation is strongly supported by another experiment by > the same researchers, testing people's willingness to sign petitions /.and pledge small donations on an assortment of political issues, such as climate change, protection of humpback whales, protection =fr:;of the people of Darfur, negotiation of new trade rules to combat B poverty, and protection of human rights in China. In the control % group, people saw the petitions in random order. In the treatment /. groups, people did so as well, but they were also given social infor-rnation, stating whether there were already large numbers of signatories (over one million), small numbers (less than a hundred), or II medium numbers (ranging from a hundred to one million). Overall, people in the control group signed 61.5 percent of the I petitions. As compared to the control, small and medium numbers 106 107 CHAPTER 4 had no significant effect on whether people signed. But for those who saw large numbers, there was a real impact; that treatment group signed 66.7 percent of the time. True, that is not the level of effect that is observed in the music lab experiment (or the petition study). But it makes sense to think that the participants in this experiment were already inclined to sign, as reflected by the 61.5 percent overall signature rate, so that a massive difference among treatment conditions should not be expected. What matters is that high numbers had a consistent and statistically significant impact on the likelihood of signing—consistent in the sense that it cut across every one of the tested issues, notwithstanding varying levels of initial support. Emphasizing that people with different personality traits show different propensities to engage, that it matters whether engagement is visible, and that people show different levels of susceptibility to social influences, Margetts and her colleagues conclude that "tiny acts," made possible by social media, are "a growing form of political participation, which in some countries and contexts is overtaking voting as the political act that people are most likely to undertake."11 Their most striking finding involves the nature of the underlying social dynamics. According to the researchers, "extroversion" predicts a willingness to participate at an early stage; if there is a significant number of extroverts, people with higher thresholds for participation might be moved—and once they are moved, those with lower thresholds will join, eventually encompassing large numbers of people. Because the costs of participating are so low (if only via a "like," a retweet, or a signature), millions of people can form a movement in this way. And indeed, processes of this general kind seemed to have played a role in the collapse of authoritarian nations in North Africa—and in the fullness of time, they are likely to have large effects elsewhere as well.13 RUMORS AND TIPPING On the Internet, rumors often spread rapidly, and cascades are frequently involved. Many of us have been deluged with e-mails about 108 CYBERCASCADES tjje need to contact our representatives about some bill or other— only to learn that the bill did not exist, and the whole problem was a joke or a fraud. Even more of us have been earnestly warned about the need to take precautions against viruses that do not exist. In . the 1990s, many thousands of hours of Internet time were spent on elaborating paranoid claims about alleged nefarious activities, including murder, on the part of President Clinton. Numerous sites, discussion groups, and social media posts spread rumors and conspiracy theories of various sorts. An old one: "Electrified by the In-; ternet, suspicions about the crash of TWA Flight 800 were almost instantly transmuted into convictions that it was the result of friendly i fire.... It was all linked to Whitewater.... Ideas become E-mail to be duplicated and duplicated again."14 In 2000, an e-mail rumor ' specifically targeted at African Americans alleged that "No Fear" bumper stickers bearing the logo of the sportswear company of the same name really promote a racist organization headed by former Ku Klux Klan grand wizard David Duke. Both terrorism and voting behavior have been prime areas for false rumors, fake news, and cascade effects. In 2002, a widely circulated e-mail said that a Boeing aircraft had not in fact hit the Pentagon on September 11. In 2004, many people were duly informed > that electronic voting machines had been hacked, producing mas-l-,.sive fraud. (If you're interested in more examples, you might con-. suit www.snopes.com, a website dedicated to widely disseminated ; falsehoods, many of them spread via the Internet.) During the Obama presidency, countless e-mails were widely -circulated about the alleged misconduct, incompetence, lying, disloyalty, and weirdness of President Obama and those who worked for him. The idea that Obama was born in Kenya (propagated by •'".•'Donald Trump, among many others) is just one prominent exam-|gP.le; another is that he is a Muslim. From 2009 to 2012,1 had the honor of working in the Obama administration, and I was stunned jo .to see the spread of false rumors about my own conduct and beliefs. (Some people said that I wanted to "steal people's organs"; others .-said that I was behind Wikileaks.) What is especially interesting is #109 CHAPTER 4 CYBERCASCADES that those who believe such rumors need not be irrational, Th I are simply reacting to what other people seem to believe. Most of these examples are innocuous, because no real harm; j done, and because many cascades can be corrected. But as a dis turbingly harmful illustration, consider widespread doubts | South Africa in the 1980s about the connection between HIV and AIDS. Because the AIDS virus infected a significant percentage of the adult population, any such doubts were especially troublesome. South African president Thabo Mbeki was a well-known Internet surfer, and he learned the views of the "denialists" after stumbling across one of their websites. The views of the denialists were and are not scientifically respectable—but to a nonspecialist, many of the claims on their (many) sites seemed plausible. At least for a period, President Mbeki both fell victim to a cybercascade and, through his public statements, helped to accelerate one—to the point where many South Africans at serious risk were not convinced of an association between HIV and AIDS. It is highly likely that this cascade effect produced a number of unnecessary infections and deaths. It literally killed people. Recall the existence of cascade effects among those who believe that childhood vaccinations are harmful and can in particular cause autism. If apparently reliable reports suggest that vaccinations cause autism, many parents will refuse them. That's hardly innocuous. It can result in illness and death. In fact, the Internet is a breeding ground for false information about health and risk avoidance. It also provides reams of truth and makes it available to all. But every day, damaging falsehoods spread through informational cascades; consider the problem of fake news. With respect to information in general, there is even a tipping point phenomenon, creating a potential for dramatic shifts in opinion. After being presented with new information, people typically have different "thresholds" for choosing to believe or do something new or different. As the more likely believers—that is, people with low thresholds—come to a certain belief or action, people with somewhat higher thresholds then join them, soon producing a --no ificant group in favor of the view in question. At that point, those S'ith still higher thresholds may join, possibly to a point where a crit-'cal maSS iS reacneQ,> making large groups, societies, or even nations "tip "IS/^ne resu^ °f tnis Process can De t0 produce cascade effects, as larffe groups of people end up believing something—whether or not that something is true or false—simply because other people in the relevant community seem to believe that it is true. There is a great deal of experimental evidence of informational cascades, which are easy to induce in the laboratory; real-world phenomena also have a great deal to do with cascade effects.16 Consider, for example, going to college, smoking, participating in political protests, voting for third-party candidates, striking, recycling, 1 filing lawsuits, using birth control, rioting, or even leaving bad , dinner parties.17 In all these cases, people are greatly influenced by what others do. Often a tipping point will be reached. Sometimes : we give an aura of inevitability to social developments, with the thought that deep cultural forces have led to (for instance) an increase in smoking, protesting, or a candidate's success, when in fact social influences have produced an outcome that could easily have been avoided. Social media provide an obvious breeding ground for cascades, and as a result, thousands or even millions of people jkwho consult sources of a particular kind will move in one or another direction, or even believe something that is quite false. Ihe good news is that the Internet, including social media, is easily enlisted to debunk false rumors as well as start them. Online, '.people can correct those rumors in a hurry. For this reason, most g.such rumors do no harm. But it remains true that the opportunity gfto spread apparently credible information to so many people can induce fear, error, and confusion in a way that threatens many social goals, including democratic ones. As we have seen, this danger takes on a particular form in a balkanized speech market as local cascades lead people in dramatically different directions. When this happens, correctives, even via the Internet, may work too slowly or not at all, simply because people are not listening to one another. Recall the (terrible) problem of the backfiring correction. #111 CHAPTER 4 UP AND DOWN VOTES We continue to learn more about how social influences work online. Lev Muchnik, a professor at the Hebrew University of Jerusalem, and his colleagues carried out an ingenious experiment on a particular website—one that displays a diverse array of stories and allows people to post comments, which can in turn be voted "up" or "down."18 With respect to the posted comments, the website compiles an aggregate score, which comes from subtracting the number of down votes from the number of up votes. To study the effects of social influences, the researchers explored three conditions: "up-treated," in which a comment, when it appeared, was automatically and artificially given an immediate up vote; "down-treated," in which a comment, when it appeared, was automatically and artificially given an immediate down vote; and "control," in which comments did not receive any artificial initial signal. Millions of site visitors were randomly assigned to one of the three conditions. The question was simple: What would be the ultimate effect of an initial up or down vote? You might well think that after so many visitors (and hundreds of thousands of ratings), a single initial vote could not possibly matter. Some comments are good, and some comments are bad, and in the end, quality will win out. It's a sensible idea, but if you thought it, you would be wrong. After seeing an initial up vote (and recall that it was entirely artificial), the next viewer became 32 percent more likely to give an up vote too. What's more, this effect persisted over time. After a period of five months, a single positive initial vote artificially increased the mean rating of comments by a whopping 25 percent! It also significantly increased "turnout" (the total number of ratings). With respect to negative votes, the picture was not at all symmetrical—an intriguing finding. True, the initial down vote did increase the likelihood that the first viewer would also give a down vote. But that effect was rapidly corrected. After a period of five months, the artificial down vote had zero effect on median ratings (although it did increase turnout). Muchnik and his, colleagues CYBERCASCADES conclude that "whereas positive social influence accumulates, cremating a tendency toward ratings bubbles, negative social influence is neutralized by crowd correction."19 They think that their findings have implications for product recommendations, stock market predictions, and electoral polling. Maybe an initial positive reaction, or just a few such reactions, can have major effects on ultimate outcomes—a conclusion very much in line with Salganik, Dodds, and ;Watts's evidence. But maybe negative reactions will get corrected -pretty quicldy. It's an interesting thought, but we should be careful before drawling large lessons from a single study, particularly when participants had no money on the line. It's possible that negative reactions can ; have long-term effects on products, people, movements, and ideas. But there is no question that when groups move in the direction of one or more of these, it may not be because of their intrinsic merits • but instead because of the functional equivalent of early up votes. ; • (Politicians, including Barack Obama and Donald Trump, often succeed as a result.) There are lessons here about the extraordinary unpredictability of groups—and their frequent lack of wisdom. | Of course Muchnik and his colleagues' own study involved large ... groups. But the same thing can happen in small ones, sometimes even more dramatically, because an initial up vote—in favor of .•.•.sonic plan, product, or verdict—has a large effect on others. HOW MANY MURDERS? •■vi. Here's a clean test of group wisdom and social influences. The me-% dian estimate of a large group is often amazingly accurate. But what :.\ happens if people in the group know what one another are saying? I You might think that knowledge of this kind will help, but the picture is a lot more complicated. Jan Lorenz, a researcher in Zurich, worked with several col-|f leagues to learn what happens when people are asked to estimate I certain values, such as the number of assaults, rapes, and murders in H Switzerland.20 They found that when people were informed about V112 i:113 CHAPTER 4 CYBERCASCADES the estimates of others, there was a significant reduction in the di versity of opinions, which tended to make the crowd less wise21 There's another problem with the crowd, which is that becaus people hear about other estimates, they also become more confi dent. Notably, people in the study received monetary payments for getting the right answer, so their mistakes were really mistakes-not an effort to curry favor with others. The authors conclude that for decision makers, the advice given by a group "may be thoroughly misleading, because closely related, seemingly independent advice may pretend certainty despite substantial deviations from the correct solution."22 There's a lesson there for the wisdom of crowds in online settings. Because people are interacting with one another, they might not be so wise. SEGREGATION, MIGRATION, AND INTEGRATION The Daily Me is not a lived reality, at least for most of us. Facebook, Twitter, Instagram, and Snapchat accounts can certainly spread diverse points of view, and many people use them in exactly that way. Facts and opinions on liberal sites often migrate to conservative sites, and vice versa. We have seen that even if opinions are clustering, society can benefit from the wide range of arguments that ultimately make their way to the general public. And for many of us, voluntary choices do not produce clustering. But there is also evidence of an echo chamber effect, at least for some of us. For example, a 2009 study finds modest but clear evidence of such an effect.23 Examining the behavior of 727 people over a six-week period, R. Kelly Garrett found that people are significantly more likely to click on information that reinforces their views, and somewhat less likely to expose themselves to information that contradicts those views. In her account, people seek support for their own positions, and they do so consistently. It follows that people "are more likely to be interested in reading a story that they expect to support their opinion, and they spend more time reading it. They are also marginally less likely to be interested in -'114 tories containing opinion-challenging information, but they do ot systematically avoid them."24 The fact that people spend more Ime with stories that support their views is worth underlining. The echo chamber effect here is not large: while people prefer information that supports their convictions, they do not run from information that undermines them. In Garrett's words, "People's desire for opinion reinforcement is stronger than their aversion to opinion challenges." Her conclusion is that people "do not seek to completely exclude other perspectives from their political universe, and there is little evidence that they will use the Internet to create echo chambers, devoid of other viewpoints, no matter how much control over their political information environment they are given."25 In short, her study finds an inclination to find like-minded sources, but importantly, they are hardly sealed. At the same time, Garrett offers an ominous projection: "Polarized news outlets serving niche audiences, which are more economically feasible online where production costs are lower, are another threat. Faced with a choice between a news source that is almost exclusively supportive of their opinions and another that almost exclusively challenges those same opinions, news consumers seem likely to choose the former."26 Garrett has done a great deal of i work on these issues, and it is broadly consistent with her central findings here and also signals the existence of that threat.27 One of the most systematic treatments of these issues comes > from economists Matthew Gentzkow and Jesse M. Shapiro, who compare ideological segregation online and offline.28 To measure ideological segregation, they use an "isolation index," which, in ;^ .their words, is equal to the average conservative exposure of conserva-:- tives minus the average conservative exposure of liberals. If 0i conservatives visit only foxnews.com and liberals only visit nytimes.com, for example, the isolation index will be equal M to 100 percentage points. But if both conservatives and liberals get all their news from cnn.com, the two groups will 115 CHAPTER 4 CYBERCASCADES have the same kind of exposures, and the isolation indPv be equal to 0.29 6X will That's a useful measure of segregation. Using data sources from i 2004 to 2009, Gentzkow and Shapiro find a clear difference be I tween what conservatives and liberals see online. On the Internet £ the average conservative's exposure to conservative news is 606 percent, while the average liberal's is 53.1 percent, producing an isolation index for the Internet of 7.5 percentage points. That's significant, but again, it's not huge. You could easily see it as modest. Gentzkow and Shapiro find that most people are not using the Internet to live in echo chambers. For example, a consumer who received news exclusively from foxnews.com would have a more conservative news diet than 99 percent of Internet news users, which suggests that the vast majority of people are clicking on sites that do not fit a narrow political profile. For four reasons, however, their data should be taken with some grains of salt, at least as applied to my concerns here. First, Gentzkow and Shapiro also find that isolation for the Internet is higher than for broadcast television news (1.8 percentage points), cable television news (3.3), magazines (4.7), and local newspapers (4.8)—though lower than that of national newspapers (10.4). The fact that it is higher than four standard sources of information is hardly comforting. Second, they are speaking of aggregate behavior, and the aggregate masks the extent to which significant sub-populations are creating echo chambers. Third, their finding's are now dated; it is possible that the degree of isolation on the Internet is increasing. Fourth, more recent work finds that the echo chamber effect is dramatically higher on social media. An intriguing qualification of the Gentzkow and Shapiro findings focuses directly on the question of subpopulations. Andrew Guess studied individual-level media consumption data to explore online behavior.30 Looking at both surveys and browsing history, he finds that the percentage of visits that involve news and information about politics is actually quite low—about 6.9 percent of ■-116 visits. Most of the time, people do not go online to explore 2 i'W