The Filter Bubble: a threat for pluralinformation?

The dangers of personalizing filters

Pietro De Nicolao

Computer Ethics (2016-2017), Politecnico di Milano

December 13, 2016

Introduction

Concerns about the Filter Bubble

Case study: Facebook News Feed

Proposed remedies and counter-objections

Conclusions

Introduction

What is the filter bubble?

Figure 1: A unique universe of information for each of us (Pariser 2011,p. 10)

DefinitionsPersonalized search

Personalized search refers to search experiences that aretailored specifically to an individual’s interests by incorporatinginformation about the individual beyond specific query provided.(Wikipedia)

Filter bubble

A filter bubble is the restriction of a user’s perspective that canbe created by personalized search technologies. (Haughn 2015)

Political pluralism in the media

Political pluralism in the media refers to the fair and diverserepresentation of and expression by various political andideological groups, including minorities, in the media. (Leuvenet al. 2009, p. 12)

What I will show

In this presentation, I will show that:

I Personalization reduces information pluralism by givingusers only what they like to see

I Personalizing filters define our perception of the world andare not neutral intermediaries

I Recommender systems are relevance maximizersI Important but non-relevant stories can be left outI Different point of views are shown less

I Transparency about the use of data and about the algorithmsis needed

I Users must know when personalization is activeI Users should be able to control it

Concerns about the Filter Bubble

The dangers of personalization

The book The Filter Bubble (Pariser 2011) describes many risksassociated with it:

I Data collection and privacyI DemocracyI Information (I will focus on this)I FreedomI CreativityI CensorshipI Serendipity

Importance vs. relevance of news stories

Two metrics can be defined for news stories:

I Importance: intrinsic “value” of a story with respect to societyI Relevance: probability that a story will be “liked” by the user;

performance index of the recommender system

Recommender systems (personalizing filters) are relevancemaximizers

Example

“A squirrel dying in front of your house may be morerelevant to your interests right now than people dying inAfrica.”

– Mark Zuckerberg (Facebook CEO)

Concerns about information

Friendly world syndrome

Personalizing filters block important, but unpleasant things:

I Some topics will always be not likable: war, homelessness,poverty. . .

I Different point of views are less relevant to us

Autonomy

Autonomy of the readers is compromised, as they can’t choosewhat’s in or what’s out their “bubble”

Worst case scenarioDeliberate use of filters to shape the public opinion, bygovernments or multinational companies

Case study: Facebook News Feed

Facebook is too friendly!

Suppose that you are a Facebook user and you identify as a liberal,and you have both liberals and conservatives friends.

I News Feed recommendation algorithm: you get more postswhich reflect what you like (relevant to you)

I You may not see conservatives’ stories at all, if you interact lesswith your conservative friends

I Cross-cutting stories (those different from our viewpoint) areless likely to reach us

I . . . but how much?

89.4% of under-30 Italians uses Facebook (CENSIS 2016)

I The issue of biased content is certainly important!

Facebook: Exposure to ideologically diverse contentFacebook published a study (Bakshy, Messing, and Adamic 2015)on Science about how likely are users to view and interact withcross-cutting content.

algorithm sorts these articles and what indi-viduals choose to read (Fig. 3A). The order inwhich users see stories in the News Feed de-pends on many factors, including how oftenthe viewer visits Facebook, how much they in-teract with certain friends, and how often usershave clicked on links to certain websites inNews Feed in the past. We found that afterranking, there is on average slightly less cross-

cutting content: The risk ratio comparing theprobability of seeing cross-cutting content rel-ative to ideologically consistent content is 5% forconservatives and 8% for liberals (supplemen-tary materials, section S1.7).Individual choice futher limits exposure to

ideologically cross-cutting content. After adjust-ing for the effect of position [the click rate on alink is negatively correlated with its position in

the News Feed (fig. S5)], we estimated the riskratio comparing the likelihood that an individ-ual clicks on a cross-cutting content relative toa consistent content to be 17% for conservativesand 6% for liberals, a pattern that is consistentwith prior research (4, 17). Despite these tend-encies, there is substantial room for individualsto consume more media from the other side; onaverage, viewers clicked on 7% of hard contentavailable in their feeds.Our analysis has limitations. Although the vast

majority of U.S. social media users are on Face-book (18), our study is limited to active users whovolunteer an ideological affiliation on this so-cial media platform. Facebook’s users tend to beyounger, more educated, and more often femaleas compared with the U.S. population as a whole(18). Other forms of social media, such as blogsor Twitter, have been shown to exhibit differentpatterns of homophily among politically inter-ested users, largely because ties tend primarily toform based on common topical interests and/or specific content (16, 19), whereas Facebookties primarily reflect many different offline so-cial contexts: school, family, social activities, andwork, which have been found to be fertile groundfor fostering cross-cutting social ties (20). In ad-dition, our distinction between exposure andconsumption is imperfect; individuals may readthe summaries of articles that appear in the NewsFeed and therefore be exposed to some of thearticles’ content without clicking through.This work informs long-standing questions

about how media exposure is shaped by our so-cial networks. Although partisans tend to main-tain relationships with like-minded contacts[which is consistent with (21)], on average morethan 20% of an individual’s Facebook friendswho report an ideological affiliation are from theopposing party, leaving substantial room for ex-posure to opposing viewpoints (22, 23). Further-more, in contrast to concerns that people might“listen and speak only to the like-minded” whileonline (6), we found exposure to cross-cuttingcontent (Fig. 3B) along a hypothesized route:traditional media shared in social media (4, 24).Perhaps unsurprisingly, we show that the com-position of our friend networks is themost impor-tant factor limiting themix of content encounteredin social media. The way that sharing occurswithin these networks is not symmetric: Lib-erals tend to be connected to fewer friends whoshare conservative content than are conserva-tives (who tend to be linked to more friends whoshare liberal content).Within the population under study here, indi-

vidual choices (2, 13, 15, 17) more than algorithms(3, 9) limit exposure to attitude-challenging con-tent in the context of Facebook. Despite thedifferences in what individuals consume acrossideological lines, our work suggests that individ-uals are exposed to more cross-cutting discoursein social media than they would be under thedigital reality envisioned by some (2, 6). Ratherthan people browsing only ideologically alignednews sources or opting out of hard news alto-gether, our work shows that social media expose

SCIENCE sciencemag.org 5 JUNE 2015 • VOL 348 ISSUE 6239 1131

Potential from network Exposed Selected

1/3Proportion of content that is cross-cutting

Stage in mediaexposure process

1/2

+

0/1

-

+

.

+

.

+

-

+

.

+

.

+

-

+

.

20%

30%

40%

50%

Random Potentialfrom network

Exposed Selected

Per

cent

cro

ss−c

uttin

g co

nten

t

Viewer affiliationConservativeLiberal

Fig. 3. Cross-cutting content ateach stage in the diffusion pro-cess. (A) Illustration of howalgorithmic ranking and individualchoice affect the proportion of ideo-logically cross-cutting content thatindividuals encounter. Gray circlesillustrate the content present at eachstage in the media exposure process.Red circles indicate conservatives,and blue circles indicate liberals. (B)Average ideological diversity of con-tent (i) shared by random others(random), (ii) shared by friends(potential from network), (iii) actuallyappeared in users’ News Feeds(exposed), and (iv) users clicked on(selected).

Fig. 2. Homophily inself-reported ideologi-cal affiliation. Propor-tion of links to friends ofdifferent ideologicalaffiliations for liberal,moderate, and conserv-ative users. Points indi-cate medians, thick linesindicate interquartileranges, and thin linesrepresent 10th to 90thpercentile ranges.

Liberal friends

Moderate friends

Conservative friends

Liberal friends

Moderate friends

Conservative friends

Liberal friends

Moderate friends

Conservative friends

LiberalsM

oderatesC

onservatives

0% 25% 50% 75% 100%

Percentage of ties

RESEARCH | REPORTS

1. 2. 3.

Figure 2: Exposure stages of news stories

1. Potential from network: all the content shared by friends2. Exposed: content effectively shown in users’ News Feeds3. Selected: content clicked by the user

% cross-cutting content vs. exposure stage on Facebook

algorithm sorts these articles and what indi-viduals choose to read (Fig. 3A). The order inwhich users see stories in the News Feed de-pends on many factors, including how oftenthe viewer visits Facebook, how much they in-teract with certain friends, and how often usershave clicked on links to certain websites inNews Feed in the past. We found that afterranking, there is on average slightly less cross-

cutting content: The risk ratio comparing theprobability of seeing cross-cutting content rel-ative to ideologically consistent content is 5% forconservatives and 8% for liberals (supplemen-tary materials, section S1.7).Individual choice futher limits exposure to

ideologically cross-cutting content. After adjust-ing for the effect of position [the click rate on alink is negatively correlated with its position in

the News Feed (fig. S5)], we estimated the riskratio comparing the likelihood that an individ-ual clicks on a cross-cutting content relative toa consistent content to be 17% for conservativesand 6% for liberals, a pattern that is consistentwith prior research (4, 17). Despite these tend-encies, there is substantial room for individualsto consume more media from the other side; onaverage, viewers clicked on 7% of hard contentavailable in their feeds.Our analysis has limitations. Although the vast

majority of U.S. social media users are on Face-book (18), our study is limited to active users whovolunteer an ideological affiliation on this so-cial media platform. Facebook’s users tend to beyounger, more educated, and more often femaleas compared with the U.S. population as a whole(18). Other forms of social media, such as blogsor Twitter, have been shown to exhibit differentpatterns of homophily among politically inter-ested users, largely because ties tend primarily toform based on common topical interests and/or specific content (16, 19), whereas Facebookties primarily reflect many different offline so-cial contexts: school, family, social activities, andwork, which have been found to be fertile groundfor fostering cross-cutting social ties (20). In ad-dition, our distinction between exposure andconsumption is imperfect; individuals may readthe summaries of articles that appear in the NewsFeed and therefore be exposed to some of thearticles’ content without clicking through.This work informs long-standing questions

about how media exposure is shaped by our so-cial networks. Although partisans tend to main-tain relationships with like-minded contacts[which is consistent with (21)], on average morethan 20% of an individual’s Facebook friendswho report an ideological affiliation are from theopposing party, leaving substantial room for ex-posure to opposing viewpoints (22, 23). Further-more, in contrast to concerns that people might“listen and speak only to the like-minded” whileonline (6), we found exposure to cross-cuttingcontent (Fig. 3B) along a hypothesized route:traditional media shared in social media (4, 24).Perhaps unsurprisingly, we show that the com-position of our friend networks is themost impor-tant factor limiting themix of content encounteredin social media. The way that sharing occurswithin these networks is not symmetric: Lib-erals tend to be connected to fewer friends whoshare conservative content than are conserva-tives (who tend to be linked to more friends whoshare liberal content).Within the population under study here, indi-

vidual choices (2, 13, 15, 17) more than algorithms(3, 9) limit exposure to attitude-challenging con-tent in the context of Facebook. Despite thedifferences in what individuals consume acrossideological lines, our work suggests that individ-uals are exposed to more cross-cutting discoursein social media than they would be under thedigital reality envisioned by some (2, 6). Ratherthan people browsing only ideologically alignednews sources or opting out of hard news alto-gether, our work shows that social media expose

SCIENCE sciencemag.org 5 JUNE 2015 • VOL 348 ISSUE 6239 1131

Potential from network Exposed Selected

1/3Proportion of content that is cross-cutting

Stage in mediaexposure process

1/2

+

0/1

-

+

.

+

.

+

-

+

.

+

.

+

-

+

.

20%

30%

40%

50%

Random Potentialfrom network

Exposed Selected

Per

cent

cro

ss−c

uttin

g co

nten

tViewer affiliation

ConservativeLiberal

Fig. 3. Cross-cutting content ateach stage in the diffusion pro-cess. (A) Illustration of howalgorithmic ranking and individualchoice affect the proportion of ideo-logically cross-cutting content thatindividuals encounter. Gray circlesillustrate the content present at eachstage in the media exposure process.Red circles indicate conservatives,and blue circles indicate liberals. (B)Average ideological diversity of con-tent (i) shared by random others(random), (ii) shared by friends(potential from network), (iii) actuallyappeared in users’ News Feeds(exposed), and (iv) users clicked on(selected).

Fig. 2. Homophily inself-reported ideologi-cal affiliation. Propor-tion of links to friends ofdifferent ideologicalaffiliations for liberal,moderate, and conserv-ative users. Points indi-cate medians, thick linesindicate interquartileranges, and thin linesrepresent 10th to 90thpercentile ranges.

Liberal friends

Moderate friends

Conservative friends

Liberal friends

Moderate friends

Conservative friends

Liberal friends

Moderate friends

Conservative friends

LiberalsM

oderatesC

onservatives

0% 25% 50% 75% 100%

Percentage of ties

RESEARCH | REPORTS

Friends network News Feed User selection

Figure 3:

Facebook study: conclusions

1. The friendship network is the most important factor limitingthe mix of content encountered in social media

I if I have only friends of the same political affiliation, the filterbubble is obvious

2. The effect of News Feed ranking on cross-cutting content islimited:

I -5% for liberalsI -8% for conservatives

3. Individual choice influences the exposure to cross-cuttingcontent more than the News Feed filtering

[. . . ] we conclusively establish that on average in thecontext of Facebook, individual choices more thanalgorithms limit exposure to attitude-challengingcontent (Bakshy, Messing, and Adamic 2015)

Facebook study: criticismLimitations of the study

I Underlying (false) assumption: the building of the friendshipnetwork is independent from Facebook’s algorithms

I Friends are only partly from “offline” connectionsI Facebook suggests both pages to like and new friends

I What about sponsored content?

Methodological issues

I Sample of the study: people which declare their politicalaffiliation

I may not be representative of the entire Facebook community

I Independent researchers can’t access Facebook data andanalyze it

Ranking = visibilityThe position (rank) of a story in the News Feed is very important!

I the position in the News Feed may be used to promote somestories and not others

I money can buy rankings!I even if the algorithm is “fair” now, what about the future?

Conservative Liberal

5%

10%

15%

20%

0 10 20 30 40 0 10 20 30 40Min position

Sel

ectio

n (c

lick)

rate Content type

Ideologically congruent

Ideologically cross cutting

Viewer affiliation

Conservative

Liberal

(a)

●

●

● ●

●●

● ● ● ● ●● ● ● ● ● ● ● ● ● ● ● ●

● ● ● ● ● ● ● ● ● ● ●● ● ● ● ● ●

●

●

●●

●● ● ● ● ● ● ●

●● ● ● ● ● ●

● ● ●● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ●

20%

25%

30%

0 10 20 30 40Min position

Perc

ent c

ount

erva

iling

stor

ies

Viewer affiliation●

●

ConservativeLiberal

(b)

Figure S5: Relationship between story position and (a) click rate for ideologically congruentand cross-cutting content (b) percent of cross-cutting content shown in News Feed, for liberalsand conservatives. Note that the relationship between click-through rate and position is bothcaused by relevance (including selective exposure) and individuals’ tendencies to engage withcontent that is positioned toward the top of the News Feed.

21

Figure 4: Click rate depends on the position of the story in the News Feed.

Proposed remedies and counter-objections

Moralizing filtersProblem: the Internet is showing off what we want to see, but notwhat we need to see

I Algorithms cannot compute “what should be seen” (Morozov2011)

I What if one day Google could urge us to stop obsessing overLady Gaga’s videos and instead pay attention to Darfur?

Let’s introduce “moralizing” filters!

I Would it be a good idea to make multinational companiesmoralizing agents?

I Paternalistic, technocratic approach

I Active, educated citizens should be able to autonomouslysearch and retrieve information

I not just “ingest” whatever is thrown at them

Make the algorithms transparent

What if the algorithms and/or some of the data were public?

I The inner working of complex neural networks and machinelearning agents is not intuitively understandable

I Even if published, we may not understand those algorithms

I They are often trade secretsI Knowing at least which personal data is used to make the

recommendation may prove useful

Facebook News Feed settings

Figure 5: A rather good solution: Facebook lets users see and customizesome parameters of the News Feed algorithm

Turn off the personalization!

I What if we could turn off the personalization?I Personalization is the key feature of some services

I Facebook without personalization would be. . . Twitter?I For other services, this would be a feasible solution

I Without personalization ads would be less relevant andprofitable: no economic incentive to do so

I Users should at least know whether personalization is enabledor not

Conclusions

Conclusions

I Personalization reduces information pluralism by givingusers only what they like to see

I Recommender systems privilege relevance over importanceI These technologies and their implementations are not neutralI Transparency about the use of data and about the algorithms

is neededI Always use those services with a critical eye!

References I

Eytan Bakshy, Solomon Messing, and Lada A Adamic. “Exposure toideologically diverse news and opinion on Facebook”. In: Science348.6239 (2015).

CENSIS. 13 ◦ Rapporto Censis-Ucsi sulla comunicazione.http://www.censis.it/7?shadow_comunicato_stampa=121073. 2016.

Matthew Haughn. "filter bubble" - WhatIs.com. url:http://whatis.techtarget.com/definition/filter-bubble (visitedon 2015).

Katholieke Universitet Leuven et al. “Independent study on indicators formedia pluralism in the member states—Towards a risk-based approach”.In: Prepared for the European Commission Directorate-GeneralInformation Society and Media. 19 (2009).

Evgeny Morozov. “Your Own Facts”. In: New York Times Sunday BookReview (2011).

Eli Pariser. The filter bubble: What the Internet is hiding from you.Penguin UK, 2011.