The Election of 2016 and the Filter Bubble Thesis in 2017

Donald Trump’s victory seemed to have blindsided many, and in the rush to understand how so few predicted it, the filter bubble thesis has taken front stage. As one reporter noted, it “simultaneously explains why Trumpism has flourished and why so many of us are insulated from it.”

And yet, when you look into the data, the story about the filter bubble and the election is strangely paradoxical.

The term, filter bubble, comes from a book by Eli Pariser of the same name, where he questioned the benefits of personalized content like Facebook’s EdgeRank algorithm, Netflix’s movie suggestions and Amazon’s book recommendations. The book presents a persuasive argument about cause and effect of algorithms. As he explained, “The Filter Bubble introduces three dynamics we’ve never dealt with before: first, you are alone in it, as it is you own personal bubble. Second, it is invisible in its actions. Finally, you don’t choose to enter into the bubble.” As Cass Sunstein said of this phenomena, “people restrict themselves to their own points of view — liberals watching and reading mostly or only liberals; moderates, moderates; conservatives, conservatives; Neo-Nazis, Neo-Nazis,” resulting in fewer of the “unplanned, unanticipated encounters central to democracy itself.” But in the aftermath of the 2016 election, it is not clear if online technology was needed to yield the same effects.

Consider this. Studies of 2016 voters found that Trump supporters largely got their news from Fox News. Clinton voters, on the other hand, didn’t coalesce around any one single source. As Pew noted, “The digital news publishers that played prominent roles in the campaign did not appear to serve as main news sources for either Trump or Clinton voters.” So if it is true that Trump voters restricted themselves to their own points of view, algorithmically based personalized filters weren’t the cause. Algorithmic technologies weren’t needed, only TV was.

Indeed, the filter bubble thesis has seen various forms throughout history. Socrates famously lamented in the Phaedrus that writing wasn’t a complete representation of knowledge and that its use would narrow wisdom. As a result, people “will therefore seem to know many things, when they are for the most part ignorant and hard to get along with, since they are not wise, but only appear wise.” Socrates technological lament was made again when the printing press came about in the 15th century, the telegraph was introduced, the telephone proliferated, and when TV first began to broadcast.

Now, this doesn’t mean changes aren’t occurring. The consumption of news is shifting, and algorithms are playing a part in that. But what is worrying about the filter bubble theory is that it papers the complex dynamics of communities and individuals, the contexts in which we communicate, and most importantly, the hard research that has accumulated in recent years.

The rest of this piece is focused in two parts. The first explores six major dynamics that have been established from research on the topic of social networks, both online and offline. The second section pulls all of these threads together to illuminate the 2016 election. My aim is to refocus the debate.

Six Dynamics of Online Interactions

Birds of a feather flock together, even online.

The filter bubble thesis suggests that technology has a certain amount of stickiness, driving the polarization of ideology. But those tendencies already exist in the meatworld.

People who are similar tend to interact with one another. Homophily, as this is called, is a well known social dynamic and is found throughout society. Homophily is one cause of racially homogenous neighborhoods. It helps to explain why married couples tend to have similar incomes and educational backgrounds. And scientific papers within a specialty tend to be written by a cohort of connected scholars, due to homophily.

These assortative patterns exist online as well. A small portion of editors account for most of the work done and the value produced for Wikipedia, for example. A small portion of Reddit users, about 1 percent of them, account for the vast majority of high quality comments, signaled by comment karma. As one textbook of social networks explained, “Homophily provides us with a first, fundamental illustration of how a network’s surrounding contexts can drive the formation of its links.” As a first approximation, social networks shouldn’t conform to random graph, but instead should show delineations.

Facebook actually provided evidence of homophily on their network in a much talked about study.

The graph below from Facebook helps to put a finer point on this. The first point on this graph, titled random, shows the percent of hard news links on the site if everyone saw a random sample. If this were the case, then self identified liberals would have about 45 percent of their feed filled with conservative content, while conservatives would see about 40 percent liberal content. The second point or the potential from the network, shows the percentage of bipartisan links that all of a person’s friends posted. In other words, this massive drop from both sides represents the homophily present in online social networks. Moving to the right once more, the third point, titled here as a exposed, details the percentage of cross content that they actually saw. As Pariser noted, it is here where the algorithm plays in. Finally, the very last point shows what individuals actually selected.

As Eli Pariser noted “The steepest reduction comes from who one’s friends are, which makes sense: If you have only liberal friends, you’re going to see a dramatic reduction in conservative news. But the algorithm and people’s choices about what to click matter a good deal, too.”

The study engendered a lot of different criticism, some of which will be explained later, but nearly everyone focused on the filtering, the dropoff from the potential in the network point to the exposure point. Sociologist Zeynep Tufekci echoed Pariser’s concerns saying, “Facebook’s algorithm is a modest suppressor of diversity of content people see on Facebook.”

Yet, that never been completely satisfying to me. As a data scientist, it is not clear to me what is garnered by limiting items in the Newsfeed. If there is engagement with an item, even if it is small, you would include it. Notice the language used in the official Science writeup of this study: “We found that after ranking, there is on average slightly less crosscutting content.” But, why would ranking reduce the total possibility space? The importance of the Newsfeed doesn’t lay in the total possibilities, which is what the exposed point would suggest, but in the ranking of those items, as Tufekci noted. Facebook has never explicitly laid out how the ranking system works, but I think we should also consider the least nefarious option available. Perhaps Facebook cannot categorize these items. As Hanlon’s Razor offers us, don’t assume bad intentions over neglect and misunderstanding.

Online networks allow for novel information to traverse groups

Because our online presence is an extension of our social selves, some homophily should be expected. However, research finds that social media sites like Facebook and Twitter allow us to maintain more extensive social networks and build social capital. No longer are cousins, high school friends, and casual acquaintances seen just around the holidays. Online social networks allow more consistent contact.

These weak relationships are important. In the early 1970s, sociology expanded into network theory when Mark Granovetter published his seminal paper, “The Strength of Weak Ties.” Granovetter studied people who found jobs through personal contacts. Of those surveyed, nearly half said they found their then-current employment from someone that was was “not a friend, an acquaintance” — people that they knew, but had minimal contact with. From this, Granovetter correctly surmised that the people we are not close to (what he called weak ties) play an extremely important role in social cohesion and information sharing.

As Granovetter helped to show, weak ties tend to have novel information, and thus when the gap between clusters of close relations are bridged, novel information flows between the two groups. However because this doesn’t often occur, any particular piece of information is less likely to flow between two groups.

In other words, social media should increase the range of experiences and news. Previous studies of Facebook data found that weak ties actually drive the majority of information on the social network, which suggests that it is driving novel information across groups. Moreover, Facebook’s own data shows that the world is getting more connected and that their own users have decreased in average degrees of separation from 3.74 in 2011 to 3.57 in 2016. In study after study, social media networks are shown to support information diffusion across disparate groups.

Exposure to diverse viewpoints might drive polarization

Decades of work under the banner of motivated reasoning shows that new and contradictory information actually entrenches beliefs. In an effort to maintain specific beliefs and stave off cognitive dissonance, individuals will find supporting reasons for their views and will discount information that is contrary to that belief. As one study found, those that occupy either extreme of the political spectrum, be it liberal or conservative, tend to be less influenced by outside information even in simple tasks. Partisans also tend to think the media is against them even if information is neutrally presented. The idea has been dubbed the hostile media effect.

So, if we accept that filter bubbles caused by social media actually do exist, will exposure actually less ideological segregation? Cass Sunstein thought so, and run some experiments to probe the idea. Indeed, Sunstein was one of the first to sound the alarm about filter bubbles. Yet, he found instead that “in a striking empirical regularity, deliberation tends to move groups, and the individuals who compose them, toward a more extreme point in the direction indicated by their own predeliberation judgments.” He has come to dub this the “Law of Group Polarization.”

Online political conversations display unique characteristics

Though filter bubbles are often discussed in the context of politics, most assume that the bifurcation that seems to exist in political discussions will be present in all types of news content. Yet, this doesn’t seem to be the case. In an analysis of Twitter data, Pew found that conversations can fall into six basic archetypes of conversation. While polarization does exist, it is contained largely in conversational political topics. Hobbies, conferences, brands, public events, and entertainment news all display different organizational structures, detailed here:

Source: Pew analysis of Twitter data

The topic changes the network topology. Partisan networks don’t spill over into non-political conversations. Using a longitudinal method of analysis, an empirical study of the 2012 presidential election, 2013 government shutdown, the 2013 Boston Marathon bombing, and the 2014 Super Bowl found that ideological preferences in the case of the first two political issues did not occur with the last two non-political events. Discussion of the Newtown shootings in 2012 reflected a dynamic process, beginning as a national conversation before transforming into a polarized exchange. The researchers concluded saying, “previous work may have overestimated the degree of ideological segregation in social-media usage.”

So, who discusses politics online anyway?

Partisans like to discuss politics

Discussing politics is heavily dependent on how strongly you identify with party politics. Those who identify as either consistently conservative or consistently liberal are far more likely to discuss politics a few times a week or nearly every day than those who hold more centrist beliefs even within their party. Even as political news consumption has shifted online, political conversations actually occur in a wide variety of casual settings often with informal acquaintances.

Sharing political news online seems to be rare as well. As Pew found, only about 10 percent of people who consume news online are actually sharing it within their networks. Facebook data suggests only 13 percent of all stories could be counted as hard news. And yet, this number may be high. As noted earlier, the Facebook cross cutting content study faced criticism on a number of fronts, including their sample selection. The entire study only focused on the 5 percent of people within their network who self-identified as having a party. Party identification is dropping, but is still maintained by those who hold more partisan views.

Actual consumption of online media seems to be fairly balanced, but it is changing.

Few have analyzed actual news consumption, because it poses difficult data collection issues, but those that have confirm that news diets tend to be diverse. One study by economists Matthew Gentzkow and Jesse M. Shapiro found that people tend to visit a relatively centrist set of websites. Isolation indices have long been used to track racial segregation in housing. Gentzkow and Shapiro adopted the concept for online news consumption, where 100 shows ideological isolation and 0 shows none. While it is true that the news consumption online (7.5) shows more segregation as compared to broadcast television (1.8) and local news (4.8), it is still lower than national newspaper consumption (10.4). Moreover, Internet news consumption was found to be significantly lower than “actual networks formed through voluntary associations (14.5), work (16.8), neighborhoods (18.7), or family (24.3).”

The information ecosystem online seems to be diverse as well. When searching for news, people tend to first head to a news organization website or app just as much as they go to social media, both of which are about 35 percent of the time. Interestingly enough, longitudinal studies on the topic from 2016 to 2017 find that those with a college education or better actually declined in their use of social media for news. That same study also found that overall news diversity is inching up. About one quarter of all U.S. adults get news from two or more of these sites, up from just 18 percent in 2016 and 15 percent in 2013.

Putting It Together: Filter Bubbles and the 2016 Election

Christopher Hitchens wasn’t a fan of Washington punditry, but not for reasons usually cited. The audience for these authors wasn’t the broader public, he lamented. Instead, the chatterati penned private letters written to other pundits, appearing in public space.

I couldn’t help but think of this quote when considering the effect of the filter bubble on the 2016 election. I am old enough to remember 2014 and 2015, and the worry then wasn’t that Facebook would allow fake news and rumors to spread, but the site would undermine a true political conversation, that in their choice of what to show show, they could tip the election. Jonathan Zittrain said that “Facebook Could Decide an Election Without Anyone Ever Finding Out.”

And now the outcries are of the opposite nature. Why didn’t they do anything?

The dramatic turn comes from a dramatic shift in expectations. In the days leading up to the vote, many on both the left and the right believed that Clinton would win. The Princeton Election Consortium predicted Clinton had a 95 percent chance of winning. The New York Times Upshot put the chances of a Clinton win at 85 percent. It wasn’t just media either, those in Silicon Valley though Clinton had all but secured the top office. One Washington Post headline indicates the tenor, “Donald Trump wins the presidency in stunning upset over Clinton.”

Trump’s victory seems to have come out of nowhere. But as Jeff Spross noted when looking at Pew data, “A lot of media coverage is much more a matter of members of the upper class talking to the other members of the upper class than it is a matter of ‘the media’ talking to voters as a whole.” If the elites on all sides missed a potential Trump presidency, perhaps echo-chambers were at work. Something was lurking in places that very few had accessed, and that something is fake news.

Beneath the filter bubble controversy lay more existential questions that journalism has been asking itself recently. Where do we fit in this new knowledge ecosystem? Where does our authority come from? Have we been getting this objectivity thing right?

American democracy is supported by the belief in a rational voter. The filter bubble thesis is a problem in so far as it impedes the rationality necessary for democracy. Only those who are informed of the issue can vote correctly, or stated another way, those that voted incorrectly are clearly uninformed. I have come to call this, the Irrational Voter Thesis. And in those uncanvassed recesses, in those places that no respectable coastal elite would dare explore, that is where fake news existed, where the bubble was stronger, those places gave rise to Trump voters. To me, the opposite side of the coin of the filter bubble thesis is what I am calling The Specter of True Knowledge.

At first you might balk at these notions, but it is a persistent current. “What’s The Matter with Kansas” is a book length treatise on the subject of why voters don’t do so in their own interest. Fox News opinion writers and journalists often paint themselves as the counter to the mainstream, a sort of vanguard for the people. And online, Breitbart has become well known for taking on the traditional center right news system, like National Review.

All together, I have to wonder, are objectivity and rationality needed to sustain democracy? In The Good Citizen, political science Michael Schudson documents the transformation in our expectations of citizenship. He takes on the idealized “informed citizen,” which, as he rightly points out, was not an expectation in eighteenth-century political circles. Rather, it took hold in the later part of the nineteenth century as education began to spread, and objectivity became a pillar of newspapers, finally becoming the yardstick it is today when the Progressives coupled both with public education and civic participation. Schudson, however, asks a tough question that needs to be more fully considered:

“If information can be more complete, more widely disseminated, more easily tapped into by citizens at large, then democracy can flourish. This is all very well if information is at the heart of mass democracy. But it isn’t.”

If he is right, then our expectations of these relationships need to be further explored.