Breaking the filter bubble: democracy and design

The classical liberal view of democracy attempts to uphold the values of freedom of choice, reason, and freedom from tyranny, absolutism and religious intolerance (Dunn 1979; Held 2006) Liberalism started as a way to challenge the powers of “despotic monarchies” and tried to define a political sphere independent of church and state. Once liberalism achieved victory over these old “absolute powers”, many liberal thinkers, began to express fear about the rising power of the “demos” (Madison 1787; Mill 1859; Held 2006). They were concerned by the new dangers to liberty posed by majority rule against minorities and the risk of the majority ‘tyrannizing over itself, leading to a need for people to ‘limit their power over themselves’.

Bentham (1780) argues that, since those who govern will not act the same way as the governed, the government must always be accountable to an electorate called upon frequently and this electorate should be able to decide whether their objectives have been met. Next to voting, ‘competition’ between potential political representatives, ‘separation of powers’, ‘freedom of the media, speech and public association’ should be ensured to sustain ‘the interest of the community in general’ (Bentham 1780). Individuals must be able to pursue their interests without the risk of arbitrary political interference, to participate freely in economic transactions, to exchange labor and goods on the market and to appropriate resources privately.

The liberal view of democracy is often criticized, because it construes democracy as an aggregation of individual preferences through a contest (in the form of voting), so that the preferences of the majority win the policy battle. However, this model has no way of distinguishing normatively legitimate outcomes from the preferences and the desires of the powerful, and makes no distinction between purely subjective preferences and legitimate and shared (quasi objective) judgments (Cohen 1997, 2009; Young 2002).

Filter bubbles are a problem according to the liberal view, because the non-transparent filters employed by online algorithms limit the freedom of choice. In addition, the liberal view states that citizens must be aware of different opinions and options, in order to make a reasonable decision. A filter imposed on users—unbeknownst to them—will violate their autonomy, as it will interfere with their ability to choose freely, and to be the judge of their own interests. Further, the principle of separation of powers and the freedom of the media can also be in danger, if the algorithms are designed in such a manner as to serve the interests of certain individuals or groups. Finally, filters might damage the “liberty of thought”. Liberty of thought, discussion and action are the necessary conditions for the development of independence of mind and autonomous judgment. Liberty of thought creates reason and rationality, and in turn the cultivation of reason stimulates and sustains liberty. If one is ‘coerced’ by the filters, reason will also diminish. While some thinkers such as Mill (1859) also emphasize the diversity of opinion, most liberal thinkers do not mention this as a requirement. Liberal citizens must be ‘potentially’ informed so that the elected act accountably, but deliberation according to the liberal view is not necessary. Loss of autonomy caused by filters seems to be the main issue, according to the liberal view, while diversity of opinions and perspectives is not a concern.

Elster (1997) characterizes deliberative democracy as “decision making by discussion among free and equal citizens”. Deliberative democrats propose that citizens address societal problems and matters of public concern by reasoning together about how to best solve them. This can be made possible by deliberative procedures, which help to reach a moral consensus that satisfies both rationality (defense of liberal rights) and legitimacy (as represented by popular sovereignty) (Gutmann and Thompson 2004). Individuals participating in the democratic process can change their minds and preferences as a result of reflection. According to Cohen (2009), deliberative democracy can be seen (1) as a matter of forming a public opinion through open public discussion and translating that opinion into legitimate law; (2) as a way to ensure elections are themselves infused with information and reasoning; (3) as a way to bring reasoning by citizens directly to bear on addressing regulatory issues. In all cases the goal is to use the common reason of equal citizens who are affected by decisions, policies or laws, instead of having them enter into bargaining processes or represent them by means of the aggregation of their individual preferences. Democracy, no matter how fair, no matter how informed, no matter how participatory, does not qualify as deliberative unless reasoning is central to the process of collective decision-making.

There are different versions of deliberative democracy. Rawls’ (1971, 1997) conception of deliberation is based on the idea of public reason, which is defined as “the basic moral and political values that are to determine a constitutional democratic government’s relation to its citizens and their relation to one another”. By means of public deliberation, people settle their disputes with respect and mutual recognition towards each other. Habermas (1998) provides similar conditions in his concept of the “ideal speech situation”. The Rawlsian approach aims at ‘accommodation’ of differences in a pluralistic society without criticizing people’s fundamental views of life, their so-called ‘comprehensive doctrines’ or ‘bringing them into deliberative discussion’. Habermas’ approach does the opposite, by also making moral or philosophical ideas and ideals part of the deliberative challenge. Both Rawls and Habermas advocate a ‘rational consensus’ rather than ‘mere agreement’ in political deliberation. For this purpose, Rawls uses the term ‘reasonable’, and Habermas introduces the notion of ‘communicative rationality’.

Deliberative democrats argue that deliberation (1) enlarges the pools of ideas and information (Cohen 2009), (2) helps us discover truths (Manin 1997; Talisse 2005), (3) can lead us to a better grasp of facts (Hardin 2009), (4) can lead us to discover diverse perspectives, practical stances towards the social world that are informed by experiences that agents have (Bohman 2006), (5) can help us discover the seriousness of our disagreements and discover that there is a disagreement after all (Cohen 1986), (6) can lead to a consensus on the “better or more “reasonable” solution (Landemore 2012), (7) promotes justice, as it requires full information and equal standing, (8) lead to better epistemic justification and legitimacy than simply voting (Hardin 2009). This is because political decisions based on deliberation are not simply a product of power and interest. It involves public reasons to justify decisions, policies or laws, (9) lead to better arguments, since citizens have to defend their proposals with reasons that are capable of being acknowledged as such by others (Cohen 2009), (10) allows citizens to reflect on their own arguments, that will lead to self-discovery and refined arguments (Cohen 1986), (11) promotes respect, as it requires people to consider the opinions of others, despite fundamental differences of outlook (Hardin 2009).

Critics of deliberative democracy argue that full fledged deliberation is difficult to attain because (1) there is inequality in deliberative capabilities of citizens, which gives advantages to the rhetorically gifted and those who possess cultural capital and argumentative confidence in leading the discussions (Ahlström 2012), (2) there is widespread incompetence and political ignorance among the masses (Ahlström 2012), (3) voters are not interested in the common good, but only in self-interests (Caplan 2008), (4) people are biased and may hold beliefs without investigation. Majority rule will amplify these mistakes and make democratic decisions worse (Caplan 2008), (5) while participation of citizens is possible in small nations, vast numbers of people will inevitably entail deterioration of participation (Held 2006). Past a certain threshold, deliberation turns into a chaotic mess (Landemore 2012), (6) most citizens cannot spend the time to master the issues well enough to take meaningful stands on major issues. The information processing cost and transaction cost is too high (den Hoven 2005), (7) deliberation among like-minded users can cause polarization. When people deliberate on a relatively homogenous argument pool, they consolidate fairly easily, which is bad for outsiders. Evidence from social psychology suggests that it is the viewpoints of the majority, not of the informed minorities, that can be expected to drive the relevant group judgments (Ahlström 2012). The informed minorities may refrain from disclosing what they know due to social pressure and be reluctant to dissent, thus not submitting the information to deliberation (Sunstein 2007), (8) forcing participants to deliberation with limiting their arguments due to commonly shared rational premises, public reason or common good will prevent dissenting voices to share their perspectives and identities on their own terms (Young 2002).

Filter bubbles are a problem for deliberative democrats, mainly because of the low quality of information and the diminishing of information diversity. If bubbles exist, the pool of available information and ideas will be less diverse and discovering new perspectives, ideas or facts will be more difficult. If we only get to see the things we already agree with on the Internet, discovering disagreement and the unknown will be quite difficult, considering the increasing popularity of the Internet and social media as a source of political information and news (Mitchell et al. 2014). Our arguments will not be refined, as they are not challenged by opposing viewpoints. We will not contest our own ideas and viewpoints and as a result, only receive confirming information. This will lead us not to be aware of disagreements. As a consequence, the quality of arguments and information and respect toward one other will suffer.

In contemporary political theory and philosophy, republicanism focuses on political liberty, understood as non-domination or independence from arbitrary power. The republican conception of political liberty defines freedom as a sort of structural independence—as the condition of not being subject to arbitrary or uncontrolled power. Pettit (1999) argues that people are free to the extent that no other group has “the capacity to interfere in their affairs on an arbitrary basis”. To ensure that, according to Pettit (1999), there must be an “active, concerned citizenry who invigilate the exercise of government power, challenge its abuses and seek office where necessary”. In this theory, freedom as non-domination supports a conception of democracy where contestability takes the place usually given to consent. The most important implication is not that the government does what the people want, but that people can always contest whatever decision the government has taken. While the republican tradition does not overlook the importance of democratic participation, the primary focus is clearly on avoiding the evils associated with interference and oppression.

Pettit (1999) argues that the media has a major role in forming the public opinion, ensuring non-domination and the possibility of effective contestation. However, Pettit argues, the media often fail badly in performing these roles. According to Pettit, at every site of decision-making (legislative, administrative and judicial), there must be procedures in place to identify and display the considerations relevant to the decision. The citizens should be able to contest these decisions if they find that the considerations did not actually determine the outcome. The decisions must be made under transparency, under threat of scrutiny, and under freedom of information. A group, even if they are a minority, should be able to voice contestation and must be able to speak out in a way that is liable to affect the proposed legislation. They must be able to contest in an effective manner, and they must be able to make themselves heard in decision-making quarters. To provide this, there must be reliable channels of publicity and information in place, so that the performance of the governing parties is systematically brought to attention.

If we apply these norms to the design of online platforms, we can argue that online information platforms (1) must make the right information available to the citizens and should allow them to track when something important or relevant happens. In this way, citizens can become aware of possible oppression and can become active when they feel there is a need to. This can for instance be achieved by human curation that aims at including important events that might affect the whole of society, in the information diet of everyone. It can also be achieved by means of personalization, so that, an event that is particularly important for a user can be highlighted for that user, (2) provide effective methods of contestation, so that citizens can make themselves heard with their contestations and affect the proposed legislation or policy. This means that people should not only be able to contest, but also that the contestation should reach a large public so that it can result in an effective and inclusive discussion.

While most deliberative democracy models aim for consensus concerning a ‘common interest’, agonists see politics as a realm of conflict and competition and argue that disagreement is inevitable even in a well-structured deliberative democratic setting, and even if the ideal of consensus regulates meaningful dialogues (Mouffe 2009). According to these critics, different and irreconcilable views will coexist and an overlapping final consensus can never be achieved. Having consensus as the main goal and the refusal of a vibrant clash of democratic but opposing political positions will lead to “apathy and disaffection with political participation” (Mouffe 1999; Young 2002). According to Mouffe (2009), the aim of democratic politics according to advocates of this agonistic conception of democracy should not be seen as overcoming conflict and reaching consensus, because such a consensus would actually be a consensus of the hegemony.

The aim of ‘agonistic pluralism’ then, is to construct the ‘them’ (opposing viewpoint) in such a way that it is no longer perceived as an enemy to be destroyed, but as an ‘adversary’. Thus, conflict must be in center stage in politics and it must only be contained by democratic limits.

An adversary is “somebody whose ideas we combat but whose right to defend those ideas we do not put into question” (Mouffe 2009). Agonistic pluralism requires providing channels through which collective passions will be given ways to express themselves over issues which, while allowing enough possibility for identification, will not construct the opponent as the enemy. The difference with “deliberative democracy” is that ‘agonistic pluralism’ does not eliminate passions from the sphere of the public, in order to reach a consensus, but mobilizes those passions towards democratic designs. Democracy should then be designed so that conflict is accommodated and unequal power relations and hegemony in the society is revealed.

Mouffe (1999) argues that although the advocates of deliberative democracy claim to address pluralism and the complexity of the society, their reference to reason and rationality tends to exclude certain groups from the political arena; therefore, they are essentially not pluralistic.

Similarly, Young (2002) argues that if consensus becomes the ultimate goal, some difficult issues or issues that only concern a minority might be removed from discussion for the sake of agreement and preservation of the common good (Young 2002). The idea of a generalized and impartial public interest that transcends all difference, diversity and division is problematic, because the participants in a political discussion most likely differ in social position or culture. Our democracies contain structural inequalities (e.g., wealth, social and economic power, access to knowledge, status). Some groups have greater material privilege than others, or there might be socially or economically weak minorities. Therefore in such settings “the common good” is likely to express the interests and perspectives of the dominant groups (Young 2002). The perspectives and demands of the less privileged may be asked to be put aside for the sake of a common good whose definition is biased against them.

Young (2002) argues that when there are structural conflicts of interest which generate deep conflicts of interest, processes of political communication are more about struggle than about agreement. However, according to Young, the field of struggle is not equal; some groups and sectors are often at a disadvantage. Fair, open, and inclusive democratic processes should then attend to such disadvantages and institutionalize compensatory measures for exclusion. Democratic institutions and practices must take measures explicitly to include the representation of social groups, relatively small minorities, or socially or economically disadvantaged ones. Disorderly, disruptive, annoying, or distracting means of communication are often necessary or effective elements in such efforts to engage others in debate over issues and outcomes. Christiano (2006) argues that due to cultural differences in society, deep cognitive biases make individuals fallible in understanding their own and other’s interests and compare the importance of others’ interest with their own. By default, people will fail to realize equal advancement of interests in society. Thus, special measures must be taken to make sure that equality is satisfied.

Filter bubbles are a problem for agonists and supporters of inclusive political communication, because they hide or remove channels through which opposing viewpoints can clash vibrantly. Minorities, and those who are disadvantaged due to structural inequalities need special exposure to be able to reach out with their voice to larger publics. However, filters that show us what we already agree with usually do not include such minority voices. If filters only show us what they consider “relevant” for us, then, the only way to reach a large public will be through advertisements or by gaming the filters. This will violate the inclusion norm of modern democracies, as only the wealthy who can afford such advertisements, or technologically advanced minds who can use algorithms to their own advantage will be able to express themselves.

Table 1 summarizes the democracy models we have introduced, the benchmarks they require, the points of critique they imply concerning the phenomenon of filter bubble. Liberal democrats stress the importance of self-determination, awareness, being able to make choices and respect for individuals. Filter bubbles are a problem for the liberal democrats especially due to restrictions on individual liberty, restrictions on choice and the increase in unawareness. Deliberative democracy attempts to increase information quality, discover the truth, discover facts, discover perspectives and discover disagreements. This in the end leads to better epistemic justifications, better arguments and it increases legitimacy and respect towards one other. The filter bubble, according to deliberative democrats, hurts the civic discourse, mutual understanding and sensemaking. Contestatory democracy on the other hand focuses on channels that allow citizens to be able to contest effectively, if there is a need. It does not aim for deliberation, but it requires citizens to have key information on important issues, and be aware of the oppressors. In contestatory democracy, the media should thus provide reliable channels of publicity, so that the performance of the governing parties is systematically brought to attention and can be contested. The filter bubble is a problem for contestatory democracy, because it removes the reliable channels so that key information on both topics and grounds of contestation cannot be sent and received. Agonists criticize the consensus goal of deliberative democrats and argue that other norms such as inclusion should also be the goal of democracy. They argue that special attention must be paid to the voice of minorities and other disadvantaged members of society and by making sure that dissent is continuously present. The filter bubble is a problem for agonists, because it will silence radical voices, will only reflect the viewpoints and perspectives of the mainstream and it will change agonism to antagonism.

As we have stated in “Liberal view of democracy” section, in the liberal view of democracy, filter bubbles can be seen as a form of market failure that diminishes user control and hence autonomy, hide available options and coerce people in such a way that they cannot get what they want. Users will not get the search results they were looking for, or do not receive the updates from friends they want to in a social networking platform. Designers that takes this view will develop tools that aim to promote awareness of filter bubbles and attempt to give users some sense of control. User satisfaction and awareness of options and choice seem to be the most common goals. As we will show in this subsection, this view of the filter bubble can be realized by giving users the control over the filters, increasing awareness of their own biases or increasing the awareness of the presence of filters that are implemented in common web services.

Munson et al. (2013) developed a browser tool called Balancer, that tracks users’ reading activities and shows their reading behavior and biases, in order to increase awareness (See Also Fig. 1b). Munson et al. argue that, while many people agree that reading a diverse set of news is good, many do not realize how skewed their own reading behavior is. Balancer therefore shows an approximate histogram of the user’s liberal and conservative pages, with the hope that the given feedback will nudge users to make their reading behavior more balanced. Munson et al. (2013) found that very low number of users changed their reading habits (conservatives consuming more liberal items and liberals more conservative). The majority of the users did not change their reading habits at all. While Balancer aims for users to reflect their preferences and on the long-term increase the epistemic quality of the incoming information, the primary goal is to increase user-awareness. Hence this tool belongs to the user autonomy enhancing technologies that are motivated by a liberal conception of democracy.

Scoopinion¹ is a browser add-on that tracks news sites and the type of stories one reads while using the browser. Scoopinion (See Fig. 1a) provides a visual summary of one’s reading habits by displaying user’s media fingerprint. The tool also personalizes recommended stories based upon user’s reading habits, but by displaying the media fingerprint, it assumes that the user will choose to read more diversely. It works with a white list of news sites and does not make diverse recommendations. It provides a visualization of users’ information consumption habit to increase their autonomy, but it has no clear goals such as tolerance or better information quality. Again this fits a liberal conception of democracy and prioritizes the value of choice autonomy.

Xing et al. (2014) developed a browser add-on called Bobble that allows users to compare their Google search results with other profiles worldwide. The tool (See Fig. 2) uses hundreds of nodes to distribute a user’s Google search queries worldwide each time the user performs a Google search. For example, when a user performs a Google search with keyword “Obamacare”, this search keyword is distributed to 40+ worldwide Bobble clients that perform the same Google search and return corresponding search returns. Users then can see which results are displayed on their browser, but not on others, and vice versa. It is a tool for users to get an idea of the extent of personalization taking place. The tool aims to increase user’s awareness of Google’s filters. However, it does not aim to increase deliberation or provide challenging information by its design.

Nagulendra and Vassileva (2014) developed a visualization design to display to users their filter bubbles (Fig. 3). This “control and visualization” tool helps users understand how information filtering works in an online peer-to-peer social network. The tool shows users which categories and friends are in their bubble and which ones are not. Further, it allows them to control the algorithm by manipulating the visualization to “escape” the bubble, namely adding/removing friends on a certain topic to the filters. The tool aims to maximize users’ control over their filter bubbles, increase awareness of the filter bubble, promote understandability of the filtering mechanism and ultimately increase user satisfaction. It, however, does not make an attempt to expose users into challenging information. If the user wants to remain in a bubble, the tool will allow them to do that. Also in this case, a liberal notion of democracy with an emphasis on user autonomy is at the background of the development of this tool.

As we have mentioned in “Deliberative democracy” section, filter bubbles can be seen as a problem, not because they prevent users getting what they want, but because they diminish the quality of the public discussion. Deliberative democracy assumes that users are, or should be, exposed to diverse viewpoints, so that they can discover disagreements, truths, perspectives and finally make better decisions. Polarized users or users exposed to low quality (but agreeable and relevant) information will have bad consequences. In order to increase the epistemic quality of information, a wide range of opinions and perspectives on a particular topic may be made more visible and users can compare their opinions with others, even if they are opposing their own views. In the end, respect, legitimacy and consensus can be reached. In this subsection, we will list some of the tools that allow users to discover different viewpoints by visualization, showing pro/con arguments for a controversial topic, nudging them to listen to others, or by diversifying search results by modifying them for political search queries.

Microsoft’s search engine Bing studied the effect of used language for nudging Bing Search engine users (Yom-Tov et al. 2013). In this study (which we will simply refer as “the Bing Study”), a sample of 179,195 people who used news related queries were selected and then their political behavior and their link click pattern were observed. Researchers found that, while 81 % (76 %) of Republicans (Democrats) click on items from one of the most polarized outlets of their own view, they rarely clicked on polarized outlets of the other side (4 and 6 % respectively), suggesting a filter bubble in action. The researchers then modified the Bing search engine’s results page. They matched Democratic to Republican-leaning queries on the same topic manually (e.g., Obamacare and affordable health care). They then modified the results for the queries for a subset of people who issued them (treatment group), resulting in a diversified set of results: the results contained items from both republican and liberal sources, regardless of what the user has searched for. This did not increase the number of clicks on items from the opposing political news outlets. However, when the authors chose websites that use a language similar to the user’s own language, they observed a change of 25 % toward the center. The authors thus conclude that when the language model of a document is closer to an individual’s language model, it has a higher chance of being read despite it describing an opposite viewpoint. The researchers aimed for “increasing exposure to varied political opinions with a goal of improving (and enhancing) civil discourse” (Yom-Tov et al. 2013).

ConsiderIt (Kriplean et al. 2012; Freelon et al. 2012) is a deliberation (pro/con) tool that is developed with the aims of (1) helping people learn about political topics and possible tradeoffs between different opinions, (2) nudging them toward reflective consideration of other voters’ thoughts, (3) enabling users to see how others consider tradeoffs. ConsiderIt (Fig. 4) provides an interface where users can create pro/con lists by including existing arguments others have contributed, to contribute new points themselves, and to use the results of these personal deliberations to expose salient points by summarizing their stance rather than a yes/no vote. Users can see ranked lists of items that were popular full opposers, firm opposers, slight opposers, neutrals, slight supporters, firm supporters and full supporters. In a pilot study called “The Living Voters Guide” (LVG), the system was put into testing during the 2010 Washington state elections that had certain proposals on areas of tax, sale of alcohol, candy or bottled water, state debt, bail and other political topics. In LVG, 8823 unique visitors browsed the site and 468 people submitted a position on at least one item. In a small survey of 7 users, 46.3 % of them have reported that they have actually changed their stances on at least one measure and 56 % of them saying they switched from support to oppose or vice versa. 32 % of them have reported that they moderated their stances and 12 % saying they strengthened them (Kriplean et al. 2012).

OpinionSpace (Faridani et al. 2010) plots on a two-dimensional map the individual comments in a web forum, based on the commenters’ responses to a short value-based questionnaire. By navigating this space, readers are better able to seek out a diversity of comments as well as prime themselves for engaging the perspective of someone with different values (Fig. 5). When users interrogate an individual comment, they are prompted to rate comments for how much they agree with and respect it. The size of the comment’s dot on the map then grows when people with different values than the speaker respect and/or agree with it, facilitating users in seeking out comments that resonate widely.

Reflect (Kriplean et al. 2011) modifies the comments of webpages in order to encourage listening and perspective taking. It adds a listening box next to every comment, where other users are encouraged to succinctly restate the points that the commenter is making, even if there is disagreement (Fig. 6). This is a nudge to listen to other users. Other readers can afterwards read the original comment and other listeners’ interpretations of what was said, supporting broader understanding of the discussion. In this way, users do not have to “like” or “recommend” the comment to recognize or appreciate the speaker. By nudging towards listening and reflecting, an empathetic and constructive normative environment is formed, where not only those who speak and reflect are positively affected, but those who read as well. In mid-September 2011, the popular online discussion platform Slashdot enabled Reflect on four stories. During the trial, 734 reflections were written by 247 discussants, an average of 1.0 reflection per comment. While flaming and pure replies were present (31 %), the majority of the reflections were neutral, different neutral interpretations or meta observations. The tool also allowed the community to rate reflections, making certain reflections under a threshold invisible. After users downvoted flaming or cheeky replies on those reflections, almost 80 % of all the visible reflections were neutral reflections.

Rbutr² is a community driven Chrome add-on, that informs a user when the webpage they are viewing has been disputed, rebutted or contradicted elsewhere on the Internet (Fig. 7). Users can add opposing viewpoints for an item, so that future users will see that an opposing viewpoint exists for the item they are reading. Rbutr aims to increase information quality and informed opinions by promoting fact and logic-checking.

There are other tools and studies that aim to increase epistemic quality of information. Liao and Fu (2013, 2014) studied the effect of the perceived threat, the level of topic involvement, and the effect of expertise and position indicators. Munson and Resnick (2010) studied the effect of nudging by sorting or highlighting agreeable news items and experimenting with the ratio of challenging and agreeable news items. Newscube (Park et al. 2009, 2011) is a tool that detects different aspects of a news using keyword analysis, and displays users news items with different perspectives in order to decrease media bias. Hypothes.is³ is a community peer-review tool that allows the users to highlight text and add comments and sentence-level critic. Political Blend (Doris-Down et al. 2013) is a mobile application that matches people with different political views and nudges them to have a cup of coffee face to face and discuss politics.

Table 2 summarizes our analysis of the studied tools.