PreprintPDF Available
Preprints and early-stage research may not have been peer reviewed yet.

Abstract and Figures

Do echo chambers actually exist on social media? By focusing on how both Italian and US Facebook users relate to two distinct narratives (involving conspiracy theories and science), we offer quantitative evidence that they do. The explanation involves users' tendency to promote their favored narratives and hence to form polarized groups. Confirmation bias helps to account for users' decisions about whether to spread content, thus creating informational cascades within identifiable communities. At the same time, aggregation of favored information within those communities reinforces selective exposure and group polarization. We provide empirical evidence that because they focus on their preferred narratives, users tend to assimilate only confirming claims and to ignore apparent refutations.
Content may be subject to copyright.
Electronic copy available at: http://ssrn.com/abstract=2795110
1
Very preliminary draft 6/13/2016
Not yet for publication; for discussion purposes
All rights reserved
Echo Chambers on Facebook
Walter Quattrociocchi1,2, Antonio Scala1,2, Cass R. Sunstein3
1 Laboratory of Computational Social Science, IMT Lucca, 55100, Lucca Italy
2 Institute of Complex Systems, CNR, 00100, Rome Italy
3 Harvard Law School, Cambridge, MA, US
Abstract
Do echo chambers actually exist on social media? By focusing on how both Italian and
US Facebook users relate to two distinct narratives (involving conspiracy theories and
science), we offer quantitative evidence that they do. The explanation involves users’
tendency to promote their favored narratives and hence to form polarized groups.
Confirmation bias helps to account for users’ decisions about whether to spread content,
thus creating informational cascades within identifiable communities. At the same time,
aggregation of favored information within those communities reinforces selective
exposure and group polarization. We provide empirical evidence that because they focus
on their preferred narratives, users tend to assimilate only confirming claims and to
ignore apparent refutations.
Introduction
Do echo chambers exist on social media? To answer this question, we compiled a
massive data set to explore the treatment of two distinct narratives on Facebook,
involving the spread of conspiracy theories and scientific information.
It is well-established that many people seek information that supports their current
convictions1,2 - the phenomenon of confirmation bias,” That phenomenon significantly
affects decisions about whether to spread content, potentially creating informational
1 Mocanu, Delia et al. "Collective attention in the age of (mis) information." Computers in Human
Behavior 51 (2015): 1198-1204.
2 Bessi, Alessandro et al. "Science vs conspiracy: Collective narratives in the age of
misinformation." PloS one 10.2 (2015): e0118093.
Electronic copy available at: http://ssrn.com/abstract=2795110
2
cascades within identifiable communities3,4. In these circumstances, online behavior can
promote group polarization 5,6,7.
To explore the role of confirmation bias in the selection of content, we test how
users who are interested in information involving conspiracy theories respond to a)
intentionally false claims that deliberately mock conspiracy stories, even though they
apparently confirm their narratives and to b) debunking information – i.e. attempts to
correct unverified rumors 8,9. We find that intentionally false claims are accepted and
shared, while debunking information is mainly ignored. As a result, exposure to
debunking information may even increase the commitments of users who favor
conspiracy theories. We also compare the reception of scientific information to the
reception of conspiracy theories, showing how Facebook users create communities of
like-minded types.
The paper is structured as follows. In section 1 we describe the datasets. In section
2 we discuss the evidence of echo chambers on Facebook in both Italy and the United
States. In section 3 we show the power of confirmation bias by measuring the
susceptibility of conspiracy users to both confirming and debunking information.
Setting Up the (Data) Experiment
Conspiracy theories often simplify causality and reduce the complexity of reality.
Such theories may or may not be formulated in a way that allows individuals to tolerate a
certain level of uncertainty. Of course some conspiracy theories turn out to be true.
Scientific information disseminates advances and exposes the larger public to how
scientists think. Of course some scientific information turn out to be false.
The domain of conspiracy theories is exceptionally wide, and sometimes the
arguments on their behalf invoke explanations designed to replace scientific evidence.
The conspiracy theories traced here involve the allegedly secret plots of “Big Pharma”;
the power and plans of the “New World Order”; the absence of a link between HIV and
AIDS (and the conspiracy to make people think that there is such a link); and cancer
3 Anagnostopoulos, Aris et al. "Viral misinformation: The role of homophily and polarization."
arXiv preprint arXiv:1411.2893 (2014).
4 Del Vicario, Michela et al. "The spreading of misinformation online." Proceedings of the National
Academy of Sciences 113.3 (2016): 554-559.
5 Zollo, Fabiana et al. "Emotional dynamics in the age of misinformation." PloS one 10.9 (2015):
e0138740.
6 Bessi, Alessandro et al. "Trend of Narratives in the Age of Misinformation." PloS one 10.8
(2015): e0134641.
7 Bessi, Alessandro et al. "The economy of attention in the age of (mis) information." Journal of
Trust Management 1.1 (2014): 1-13.
8 Zollo, Fabiana et al. "Debunking in a World of Tribes." arXiv preprint arXiv:1510.04267 (2015).
9 Bessi, Alessandro et al. "Social determinants of content selection in the age of (mis)
information." Social Informatics (2014): 259-268.
3
cures. By contrast, the scientific news reports on the most recent research findings, such
as the discovery of gravitational waves and the Higgs boson.
To produce the data set, we built a large atlas of Facebook pages, with the
assistance of various groups (Skepti Forum, Skeptical spectacles, Butac, Protesi di
Complotto), which helped in labeling and sorting both conspiracy and scientific sources.
(We emphasize that other kinds of data sets may not show the particular patterns that we
observe here.) To validate the list, all pages have been manually checked looking at their
self-description and the type of promoted content. We analyzed users’ interaction through
Facebook posts with respect to these two kinds of information over a time span of five
years (2010-2014) in the Italian and US contexts (see Table 1 for a breakdown of the
dataset). Note that the list refers to public Facebook pages dedicated to the diffusion of
claims from the two kinds of narratives. Some examples of science pages are
https://www.facebook.com/ScienceNOW (2 million likes);
https://www.facebook.com/livescience (1.2 million likes); and
https://www.facebook.com/sciencenews (2.5 million of likes). Some examples of
conspiracy theory pages are https://www.facebook.com/TheConspiracyArchives (200k
likes) and https://www.facebook.com/CancerTruth-348939748204/ (250k likes).
Numbers reported in Table 1 refer to the posts total number of likes and comments to
each page’s post on the overall time window.
We measured the reaction of Facebook users who are exposed to different posts,
in particular:
a) For Italy, troll posts, sarcastic, and paradoxical messages mocking conspiracy
thinking (e.g., chem-trails containing Viagra)
b) For the United States, debunking posts, involving information attempting to
correct false conspiracy theories circulating online.
4
Table 1. Breakdown of the Italian and US Facebook datasets grouped by page category.
Polarized Communities
On Facebook, actions like “share,” “comment,” or “like” have distinctive
meanings. In most cases, a “like” stands for a positive feedback to the post; a “share”
expresses the desire to increase the visibility of a given information; and a “comment”
reflects a contribution to an online debate, which may contain negative or positive
feedback to the post.
Our analysis shows that in these domains, users are highly polarized and tend to
focus their attention exclusively on one of the two types of information. We also find that
users belonging to different communities tend not to interact and that they tend to be
connected only with like-minded people.
More precisely, we analyzed users’ engagement with respect to content as the
percentage of a user’s “likes” on each content category. We considered a user to be
polarized in science or conspiracy narratives when 95% of his “likes” is on either
conspiracy or science posts. With this stringent test, we find that the most users are
highly polarized, with especially high percentages on conspiracy posts: there are 255,225
polarized users on scientific pages (76.79% of all users who interacted on scientific
pages), and there are 790,899 polarized users on conspiracy pages (91.53% of all users
who interacted with conspiracy posts).
5
Figure 1 shows the probability density function (PDF) of users’ polarization. We
found that there are distinct communities that correspond to the two sharp peaks near ρ =
-1 (science) and ρ = 1 (conspiracy).
Figure 1: Users are polarized. The probability density function (PDF) of the frequency that a user has
polarization ρ is remarkably concentrated in two peaks near the values ρ = -1 (science) and ρ = 1
(conspiracy), indicating that users are split into two distinct communities.
In short, the Facebook users we studied mainly focus on a single type of narrative,
at least in the contexts studied here. As a further step, we tested whether different
narratives present different information consumption patterns. Figure 2 shows the
statistics CCDF (Complementary Cumulative Distribution Function) for likes, comments,
shares and post lifetime for both types of information.
The shape of the CCDFs is remarkably similar, indicating that conspiracy and
scientific information on Facebook is consumed in essentially the same way. The same
pattern holds if we look at the liking and commenting activity of polarized users (Figure
3). The bottom right panel of Figure 3 shows the few posts - 7,751 (1,991 from science
news and 5,760 from conspiracy news) that were commented on by polarized users of
the two communities.
6
Figure 2: Scientific and conspiracy narratives experience of similar user interactions regarding the statistics
of likes, comments, shares and post lifetimes.
7
Figure 3: The user polarization of scientific and conspiracy narratives shows statistically similar
interactions respect to the number of “likes” and “comments” to a post. The number of users debating with
the other community is a very small fraction of the polarized users.
Figure 4 shows that the more active a polarized user is on a specific content, the
higher the number of friends who display the same behavior For each polarized user, we
consider the fraction of y friends who share the same polarization and compare it with
that user’s engagement θ (number of likes) on the specific narrative. Social interaction is
“homophily driven” i.e., users with similar polarization tend to aggregate together. It
follows that the two groups of polarized users (science and conspiracy) share not only
similar information consumption patterns but also a similar social network structure.
Figure 4: Homophily and activism: the more a polarized user is active (larger θ), the more the user has
friends with similar profiles (larger y).
To check whether the observed effects might be limited to the Italian Facebook,
we perform a similar analysis on the conspiracy and science page of Facebook US. As
expected, we find the same patterns. Contents of the two narratives aggregate users into
different communities. The consumption patterns of users’ communities are very similar.
Figure 5 provides a summary of the polarization and consumption patterns in terms of
likes, comments, shares. and lifetime for US Facebook users.
In the top panel, we can observe the users’ polarization histograms sorted on the
basis of both likes (on the left) and comments (on the right). As in the Italian case (Figure
1), polarization is sharply bimodal, with most of the users concentrated around the
extreme values ρ(u) = 1, 1 (respectively science and conspiracy). The bottom left panels
of Figure 5 show that the CCDFs of the number of likes, comments, and shares are heavy
8
tailed and similar for both groups, thus indicating similar activity and consumption
patterns for both types of users. Finally, in the bottom right panel of we plot the lifetime
of posts belonging to conspiracy and scientific news, and even here they are hardly
distinguishable.
Figure 5: Analysis of US Facebook: as in the Italian case, contents related to distinct narratives aggregate
users into different communities and users’ attention patterns are similar in both communities in terms of
interaction and attention to posts. Top panels: histograms of users’ polarization calculated compared to the
number of likes (left) and the number of comments (right). Left bottom panels: the statistics of likes and
comments are similar both for science and for conspiracy users. Right bottom panel: the Kaplan-Meier
estimates of survival functions of posts in science and conspiracy (measuring the fraction of posts which
are still active after a given time from their publication) are hardly distinguishable.
Information Spreading and Emotions
Cascades
Thus far, we have considered users’ interaction with information. We now focus
on the spread of information among users. We show how homophily produces
informational cascades and how these are mostly confined inside the echo chambers. We
start the analysis by looking at the statistical signatures of cascades related to science and
conspiracy news.
9
Measuring the distance in time between the first and last user sharing a post can
approximate the lifetime of cascade effects. In Figure 6 we show the PDF of the cascade
lifetime (using hours as time units) for science and conspiracy. In both categories we find
a first peak at 1–2 hours and a second one at 20 hours. Temporal patterns are similar. We
also find that a significant portion of the information diffuses rapidly (24% for science
news and 21% for conspiracy rumors diffused in less than 2 hours, and 39% of science
news and 41% of conspiracy theories in less than 5 hours). Only 27% of the diffusion of
science news and 18% of conspiracy lasts more than one day.
Figure 6: cascade lifetimes for science and conspiracy are very similar.
In Figure 7, we show that the majority of shares pass from users with similar
polarization, i.e. users belonging to the same echo chamber. In particular, the average
edge homogeneity (measuring the users’ similarity) of all cascades shows that it is highly
unlikely that a path might include users from different groups. Contents tend to be
confined only within echo chambers, and the cascade size is well approximated by the
dimension of the echo chamber.
10
Figure 7: Confinement of cascades within echo chambers: a positive edge homogeneity indicates that
information propagates among users with similar beliefs. We do not observe cascades with a negative mean
edge homogeneity and that the values are most likely to be concentrated around the maximum edge
homogeneity value of ~1, indicating a confinement of the cascades within echo chambers.
In Figure 8, we show the lifetime of a cascade as a function of the cascade size,
i.e. the number of users sharing a post. Thus far we have seen similar signatures for both
the science and the conspiracy echo chambers, but we now observe, for the first time,
some differences between the two. In short: For conspiracy-related content, the lifetime
of a post shows a monotonic growth respect to the cascade size, but for science news, we
observe instead a peak in the lifetime corresponding to a cascade size of 100÷200 users.
News is assimilated very differently. Science news reaches a higher level of diffusion
more quickly, and a longer lifetime does not correspond to a higher level of interest but
most likely to a prolonged discussion within a specialized group of experts.
By contrast, conspiracy rumors diffuse more slowly and show a positive relation
between lifetime and size. Long-lived posts tend to be discussed by a higher number of
users.
11
Figure 8: Lifetime of a post in comparison to cascading size (number of users discussing the post). The
grey area represents the variability of the lifetime for a given size.
Extremity and Emotions
Consistent with a large body of research, it might be hypothesized that users,
discussing issues online, will become increasingly extreme in their beliefs after those
discussions. As a result, their views will be reinforced and polarized. We provide some
support for this hypothesis by showing the results of the sentiment analysis10 applied to
user discussion (comments) within the Italian Facebook echo chambers.
On Facebook, comments are, of course, the medium of online debate by which
users express their views about the post or about the discussion itself. Sentiment analysis
is a computational tool to approximate the emotional attitude of users’ comments. It is
based on a supervised machine learning approach, where we first manually annotate a
substantial sample of comments, and then build a classification model. The model
associates each comment with one sentiment value: negative, neutral, or positive. The
value expresses the emotional attitude of Facebook users when posting comments.
We find that for both kinds of content, the longer the discussion, the more
negative is the sentiment. Comments on conspiracy posts tend to be more negative than
those on science posts. Moreover, the higher the engagement of a user in the echo
chamber, the stronger the probability of expressing negative emotion when commenting -
- both on science and conspiracy. In Figure 9 we show the average sentiment of polarized
10 Sentiment analysis has been performed with supervised training of Support Vector Machines
12
users as a function of their number of comments. The more active a polarized user is, the
more the user tends towards negative values both on science and conspiracy posts.
Figure 9: Sentiment (positive, negative, neutral) of polarized users comments as a function of their
engagement.
Response to confirmatory and debunking
opinions
We have seen that users tend to aggregate around preferred contents and form
polarized groups. We now attempt to sharpen this claim by testing how users respond to
information that either confirms or debunks their beliefs.
We first test the attitudes of conspiracy and science users in interacting with false
information. To avoid biases in the determination of the truth of a post, we have used
information that is false intentionally. For the Italian data set, we collected a set of Troll
posts that were satirical imitations of conspiracy information sources. All these posts
contain clearly unrealistic and satirical claims. Examples include posts declaring that a
new lamp made of actinides (e.g. plutonium and uranium) might solve problems of
energy gathering with a lower impact on the environment, or that chemical analysis
reveals that chem-trails contains sildenafil citratum (the active ingredient of ViagraTM).
We find that polarized users on conspiracy pages are consistently more active in liking
and commenting on intentionally false claims (80% of the pool) when compared to
science users (Fig 10).
13
Users usually exposed to conspiracy claims are more likely to jump the credulity
barrier. Even when information is deliberately false and framed with a satirical purpose,
its conformity with the conspiracy narrative transforms it into suitable (and welcome)
content for the conspiracy echo chamber. Confirmation bias evidently plays a pivotal role
in the selection of content.
Figure 10: interaction of polarized users with deliberately false information (Trolls’ posts). We observe that
both in terms of comments and likes conspiracy users represent approximately 80% of the pool.
Debunking information aims to correct falsehoods. For the U.S. data set, we use
debunking posts to test their efficacy and, more generally, to characterize the effect of
such information on conspiracy users. The first interesting result is that out of 9,790,906
polarized conspiracy users, just 117,736 interacted with debunking posts that is,
commented on a debunking post at least once. Among these conspiracy users, those with
persistence in the conspiracy echo chamber greater than one day were only 5,831 -- 5,403
who registered likes, and only 2,851 offered comments. (The latter numbers exceed 5,831
because some people did both.) Hence, the impact of debunking appears marginal; fewer
than 1.3% of conspiracy users interact with it.
A good approximation of a user’s commitment to the echo chamber is given by
his or her daily number of likes and comments (i.e., liking and commenting rates). We
use this measure to compare the activity of the user before and after the interaction with a
debunking post (dissenting information if the user belongs to the conspiracy echo
chamber). Fig. 11 shows that the exposure to debunking actually induces a shift towards
higher liking and commenting rates.
14
Figure 11: Reinforcement of conspiracy beliefs after exposure to debunking. We have observed that the
small fraction of conspiracy users that interact with debunking posts tend to increase their activity within
the conspiracy echo chamber
Debunking is ignored (by ~99.98% of conspiracy users) or produces the unwanted
effect of reinforcing the very beliefs that it was supposed to correct. It follows that in our
data set, attempts to convince conspiracists that their beliefs are false generally seem to
fail.
Conclusions
At least in the areas studied here, Facebook users are highly polarized. Their
polarization creates largely closed, mostly non-interacting communities centered on
different narratives - i.e. echo chambers. The echo chambers are statistically similar in
terms of how communities interact with posts. For both scientific information and
conspiracy theories, the more active a user is within an echo chamber, the more that user
will interact with others with similar beliefs. The spreading of information tends to be
confined to communities of like-minded people. We have also found at least indirect
evidence of group polarization within those communities.
In the discussions here, users show a tendency to seek out and receive information
that strengthens their preferred narrative (see the reaction to trolling posts in conspiracy
echo chambers) and to reject information that undermines it (see the failure of
debunking) The absorption of trolls’ intentionally false conspiracy theories into echo
chambers shows how confirmation bias operates to create a kind of cognitive inoculation.
We emphasize that our data sets are limited, of course, to particular data sets, and
so we cannot venture any general claims about echo chambers on social media. Contexts
differ, and far more research would be necessary to support any such general claims. But
15
at least in the domains studied here, people are using Facebook to create enclaves of like-
minded people, spreading information in strikingly similar ways.
Acknowledgments
For the development of this work we thanks Alessandro Bessi, Michela Del Vicario,
Fabiana Zollo, Delia Mocanu, Qian Zhang, Laura Tommaso, Guido Caldarelli and Gene
Stanley. A special thanks to Giulia Borrione for her valuable thoughts and insights that
greatly contributed to the development of this work.
Furthermore, we want to thank for valuable suggestions and discussion Geoff Hall,
“Protesi di Protesi di Complotto,” “Che vuol dire reale,” “La menzogna diventa verita e
passa alla storia,” “Simply Humans,” “Semplicemente me,” Salvatore Previti, Elio
Gabalo, Sandro Forgione, Francesco Pertini, and Franco Lechner. Thanks to the Program
on Behavioral Economics and Public Policy at Harvard Law School.
MULTIPLEX (Foundational Research on MULTIlevel comPLEX networks and system) grant
number 317532, SIMPOL (Financial Systems SIMulation and POLicy Modelling) grant number
610704, DOLFINS (Distributed Global Financial Systems for Society) grant number 640772, and
SoBigData (Social Mining & Big Data Ecosystem) grant number 654024.
ResearchGate has not been able to resolve any citations for this publication.
ResearchGate has not been able to resolve any references for this publication.