The spread of fake news by social bots
Chengcheng Shao, Giovanni Luca Ciampaglia, Onur Varol,
Alessandro Flammini, and Filippo Menczer
Indiana University, Bloomington
The massive spread of fake news has been identiﬁed as a major global
risk and has been alleged to inﬂuence elections and threaten democracies.
Communication, cognitive, social, and computer scientists are engaged
in eﬀorts to study the complex causes for the viral diﬀusion of digital
misinformation and to develop solutions, while search and social media
platforms are beginning to deploy countermeasures. However, to date,
these eﬀorts have been mainly informed by anecdotal evidence rather than
systematic data. Here we analyze 14 million messages spreading 400 thou-
sand claims on Twitter during and following the 2016 U.S. presidential
campaign and election. We ﬁnd evidence that social bots play a key role in
the spread of fake news. Accounts that actively spread misinformation are
signiﬁcantly more likely to be bots. Automated accounts are particularly
active in the early spreading phases of viral claims, and tend to target
inﬂuential users. Humans are vulnerable to this manipulation, retweeting
bots who post false news. Successful sources of false and biased claims
are heavily supported by social bots. These results suggests that curbing
social bots may be an eﬀective strategy for mitigating the spread of online
If you get your news from social media, as most Americans do , you are ex-
posed to a daily dose of false or misleading content — hoaxes, rumors, conspiracy
theories, fabricated reports, click-bait headlines, and even satire. We refer to
this misinformation collectively as false or fake news. The incentives are well
understood: traﬃc to fake news sites is easily monetized through ads , but
political motives can be equally or more powerful [18, 23]. The massive spread
of false news has been identiﬁed as a major global risk . Claims that fake
news can inﬂuence elections and threaten democracies  are hard to prove.
Yet we have witnessed abundant demonstrations of real harm caused by mis-
information spreading on social media, from dangerous health decisions  to
manipulations of the stock market .
arXiv:1707.07592v1 [cs.SI] 24 Jul 2017
A complex mix of cognitive, social, and algorithmic biases contribute to our
vulnerability to manipulation by online misinformation. Even in an ideal world
where individuals tend to recognize and avoid sharing low-quality information,
information overload and ﬁnite attention limit the capacity of social media to
discriminate information on the basis of quality. As a result, online misinforma-
tion is just as likely to go viral as reliable information . Of course, we do not
live in such an ideal world. Our online social networks are strongly polarized and
segregated along political lines [3, 2]. The resulting “echo chambers” [28, 21]
provide selective exposure to news sources, biasing our view of the world .
Furthermore, social media platforms are designed to prioritize engaging rather
than trustworthy posts. Such algorithmic popularity bias may well hinder the
selection of quality content [24, 9, 19]. All of these factors play into conﬁrmation
bias and motivated reasoning [26, 14], making the truth hard to discern.
While fake news are not a new phenomenon , the online information
ecosystem is particularly fertile ground for sowing misinformation. Social me-
dia can be easily exploited to manipulate public opinion thanks to the low cost
of producing fraudulent websites and high volumes of software-controlled pro-
ﬁles or pages, known as social bots [23, 6, 27, 29]. These fake accounts can post
content and interact with each other and with legitimate users via social connec-
tions, just like real people. People tend to trust social contacts  and can be
manipulated into believing and spreading content produced in this way . To
make matters worse, echo chambers make it easy to tailor misinformation and
target those who are most likely to be believe it. Moreover, the ampliﬁcation of
fake news through social bots overloads our fact-checking capacity due to our
ﬁnite attention, as well as our tendencies to attend to what appears popular
and to trust information in a social setting .
The ﬁght against fake news requires a grounded assessment of the mecha-
nism by which misinformation spreads online. If the problem is mainly driven
by cognitive limitations, we need to invest in news literacy education; if social
media platforms are fostering the creation of echo chambers, algorithms can be
tweaked to broaden exposure to diverse views; and if malicious bots are respon-
sible for many of the falsehoods, we can focus attention on detecting this kind of
abuse. Here we focus on gauging the latter eﬀect. There is plenty of anecdotal
evidence that social bots play a role in the spread of fake news. The earliest man-
ifestations were uncovered in 2010 [18, 23]. Since then, we have seen inﬂuential
bots aﬀect online debates about vaccination policies  and participate actively
in political campaigns, both in the U.S.  and other countries . However,
a quantitative analysis of the eﬀectiveness of misinformation-spreading attacks
based on social bots is still missing.
A large-scale, systematic analysis of the spread of fake news and its manip-
ulation by social bots is now feasible thanks to two tools developed in our lab:
the Hoaxy platform to track the online spread of claims  and the Botome-
ter machine learning algorithm to detect social bots [4, 29]. Let us examine
how social bots promoted hundreds of thousands of fake news articles spreading
through millions of Twitter posts during and following the 2016 U.S. presidential
Figure 1: Weekly tweeted claim articles, tweets/article ratio and articles/site
ratio. The collection was brieﬂy interrupted in October 2016. In December 2016
we expanded the set of claim sources, from 71 to 122 websites.
We crawled the articles published by seven independent fact-checking organiza-
tions and 122 websites that, according to established media, routinely publish
false and/or misleading news. The present analysis focuses on the period from
mid-May 2016 to the end of March 2017. During this time, we collected 15,053
fact-checking articles and 389,569 unsubstantiated or debunked claims. Using
the Twitter API, Hoaxy collected 1,133,674 public posts that included links
to fact checks and 13,617,425 public posts linking to claims. See Methods for
As shown in Fig. 1, fake news websites each produced approximately 100
articles per week, on average. The virality of these claims increased to approxi-
mately 30 tweets per article per week, on average. However, success is extremely
heterogeneous across articles. Fig. 2(a) illustrates an example of viral claim.
Whether we measure success by number of people sharing an article or number
of posts containing a link, we ﬁnd a very broad distribution of popularity span-
ning several orders of magnitude: while the majority of articles go unnoticed, a
signiﬁcant fraction go viral (Fig. 2(b,c)). Unfortunately, and consistently with
prior analysis using Facebook data , we ﬁnd that the popularity proﬁles
proﬁles of false news are indistinguishable from those of fact-checking articles.
Most claims are spread through original tweets and especially retweets, while
few are shared in replies (Fig. 3).
The claim-posting patterns shown in Fig. 4(a) highlight inorganic support.
The points aligned along the diagonal lines (on the left of the plot) indicate
that for many articles, one or two accounts are responsible for the entirety
of the activity. Furthermore, some accounts share the same claim up to 100
Figure 2: Virality of fake news. (a) Diﬀusion network for the article titled
“Spirit cooking”: Clinton campaign chairman practices bizarre occult ritual,
published by the conspiracy site Infowars.com four days before the 2016 U.S.
election. Over 30 thousand tweets shared this claim; only the largest connected
component of the network is shown. Nodes and links represent Twitter accounts
and retweets of the claim, respectively. Node size indicates account inﬂuence,
measured by the number of times an account is retweeted. Node color represents
bot score, from blue (likely human) to red (likely bot); yellow nodes cannot be
evaluated because they have either been suspended or deleted all their tweets.
An interactive version of this network is available online (iunetsci.github.
io/HoaxyBots/). The two charts plot the probability distributions (density
functions) of (b) number of tweets per article and (c) number of users per
article, for claims and fact-checking articles.
Figure 3: Distribution of types of tweet spreading claims. Each article is mapped
along three axes representing the percentages of diﬀerent types of messages that
share it: original tweets, retweets, and replies. Color represents the number of
articles in each bin, on a log-scale.
times or more. The ratio or tweets per user decreases for more viral claims,
indicating more organic spreading. But Fig. 4(b) demonstrates that for the most
viral claims, much of the spreading activity originates from a small portion of
We suspect that these super-spreaders of fake news are social bots that
automatically post links to articles, retweet other accounts, or perform more
sophisticated autonomous tasks, like following and replying to other users. To
test this hypothesis, we used the Botometer service to evaluate the Twitter
accounts that posted links to claims. For each user we computed a bot score,
which can be interpreted as the likelihood that the account is controlled by
software. Details of our detection systems can be found in Methods.
Fig. 5 conﬁrms that the super-spreaders are signiﬁcantly more likely to be
bots compared to the population of users who share claims. We hypothesize that
these bots play a critical role in driving the viral spread of fake news. To test this
conjecture, we examined the accounts that post viral claims at diﬀerent phases
of their spreading cascades. As shown in Fig. 6, bots actively share links in
the ﬁrst few seconds after they are ﬁrst posted. This early intervention exposes
many users to the fake news article, eﬀectively boosting its viral diﬀusion.
Another strategy used by bots is illustrated in Fig. 7(a): inﬂuential users are
often mentioned in tweets that link to debunked claims. Bots seem to employ
this targeting strategy repetitively; for example, a single account mentioned
@realDonaldTrump in 18 tweets linking the claim shown in the ﬁgure. For a
systematic investigation, let us use the number of followers of a Twitter user as a
Figure 4: Concentration of claim-sharing activity. (a) Scatter plot of
tweets/account ratio versus number of tweets sharing a claim. The darkness
of a point represents the number of claims. (b) Source concentration for claims
with diﬀerent popularity. We consider a collection of articles shared by a mini-
mum number of tweets as a popularity group. For claims in each of these groups,
we show the distribution of Gini coeﬃcients. A high coeﬃcient indicates that a
small subset of accounts was responsible for a large portion of the posts. In this
and the following violin plots, the width of a contour represents the probability
of the corresponding value, and the median is marked by a colored line.
Figure 5: Bot score distributions for a random sample of 915 users who posted
at least one link to a claim, and for the 961 accounts that most actively share
fake news (super-spreaders). The two groups have signiﬁcantly diﬀerent scores
(p < 10−4according to a Welch’s unequal-variances t-test).
Figure 6: Temporal evolution of bot score distributions for a sample of 60,000
accounts that participate in the spread of the 1,000 most viral claims. We
focus on the ﬁrst hour since a fake news article appears, and divide this early
spreading phase into logarithmic lag intervals.
Figure 7: (a) Example of targeting for the claim Report: three million votes
in presidential election cast by illegal aliens, published by Infowars.com on
November 14, 2016 and shared over 18 thousand times on Twitter. Only a por-
tion of the diﬀusion network is shown. Nodes stand for Twitter accounts, with
size representing number of followers. Links illustrate how the claim spreads:
by retweets and quoted tweets (blue), or by replies and mentions (red). (b) Dis-
tributions of the number of followers for Twitter users who are mentioned or
replied to in posts that link to the most viral 1000 claims. The distributions are
grouped by bot score of the account that creates the mention or reply.
Figure 8: Scatter plot of bot activity vs. diﬀerence between actual and predicted
vote margin by U.S. states. For each state, we compared the vote margin
with forecasts based on the ﬁnal polls on election day. A positive percentage
indicates a larger Republican margin or smaller Democratic margin. To gauge
fake news sharing activity by bots, we considered tweets posting links to claims
by accounts with bot score above 0.6 that reported a U.S. state location in their
proﬁle. We compared the tweet frequencies by states with those expected from
a large sample of tweets about the elections in the same period. Ratios above
one indicate states with higher than expected bot activity. We also plot a linear
regression (red line). Pearson’s correlation is ρ= 0.15.
proxy for their inﬂuence. We consider tweets that mention or reply to a user and
include a link to a viral fake news story. Tweets tend to mention popular people,
of course. However, Fig. 7(b) shows that when accounts with the highest bot
scores share these links, they tend to target users with a higher median number
of followers and lower variance. In this way bots expose inﬂuential people, such
as journalists and politicians, to a claim, creating the appearance that it is
widely shared and the chance that the targets will spread it.
We examined whether bots tended to target voters in certain states by cre-
ating the appearance of users posting claims from those locations. To this end,
we considered accounts with high bot scores that shared claims in the three
months before the election, and focused on those with a state location in their
proﬁle. The location is self-reported and thus trivial to fake. As a baseline, we
extracted state locations from a large sample of tweets about the elections in
the same period (see details in Methods). A χ2test indicates that the location
patterns produced by bots are inconsistent with the geographic distribution of
political conversations on Twitter (p < 10−4). Given the widespread but un-
proven allegations that fake news may have inﬂuenced the 2016 U.S. elections,
Figure 9: Joint distribution of the bot scores of accounts that retweeted links
to claims and accounts that had originally posted the links. Color represents
the number of retweeted messages in each bin, on a log scale. Projections show
the distributions of bot scores for retweeters (top) and for accounts retweeted
by humans (left).
we explored the relationship between bot activity and voting data. The ratio of
bot frequencies with respect to state baselines provides an indication of claim-
sharing activity by state. Fig. 8 shows a weak correlation between this ratio and
the change in actual vote margin with respect to state forecasts (see Methods).
Naturally this correlation does not imply that voters were aﬀected by bots shar-
ing fake news; many other factors can explain the election outcome. However
it is remarkable that states most actively targeted by misinformation-spreading
bots tended to have more surprising election results.
Having found that bots are employed to drive the viral spread of fake news,
let us explore how humans interact with the content shared by bots, which may
provide insight into whether and how bots are able to aﬀect public opinion.
Fig. 9 shows that human do most of the retweeting, and they retweet claims
posted by bots as much as by other humans. This suggests that humans can be
successfully manipulated through social bots.
Finally, we compared the extent to which social bots successfully manipulate
the information ecosystem in support of diﬀerent sources of online misinforma-
Figure 10: Popularity and bot support for the top 20 fake news websites. Popu-
larity is measured by total tweet volume (horizontal axis) and median number of
tweets per claim (circle area). Bot support is gauged by the median bot score of
the 100 most active accounts posting links to articles from each source (vertical
tion. We considered the most popular sources in terms of median and aggregate
article posts, and measured the bot scores of the accounts that most actively
spread their claims. As shown in Fig. 10, one site (beforeitsnews.com) stands
out in terms of manipulation, but other well-known sources also have many bots
among their promoters. At the bottom we ﬁnd satire sites like The Onion.
Our analysis provides quantitative empirical evidence of the key role played by
social bots in the viral spread of fake news online. Relatively few accounts are
responsible for a large share of the traﬃc that carries misinformation. These
accounts are likely bots, and we uncovered several manipulation strategies they
use. First, bots are particularly active in amplifying fake news in the very early
spreading moments, before a claim goes viral. Second, bots target inﬂuential
users through replies and mentions. Finally, bots may disguise their geographic
locations. People are vulnerable to these kinds of manipulation, retweeting bots
who post false news just as much as other humans. Successful sources of fake
news in the U.S., including those on both ends of the political spectrum, are
heavily supported by social bots. As a result, the virality proﬁles of false news
are indistinguishable from those of fact-checking articles. Social media platforms
are beginning to acknowledge these problems and deploy countermeasures, al-
though their eﬀectiveness is hard to evaluate [31, 17].
Our ﬁndings demonstrate that social bots are an eﬀective tool to manipulate
social media and deceive their users. Although our spreading data is collected
from Twitter, there is no reason to believe that the same kind of abuse is not
taking place on other digital platforms as well. In fact, viral conspiracy theories
spread on Facebook  among the followers of pages that, like social bots, can
easily be managed automatically and anonymously. Furthermore, just like on
Twitter, false claims on Facebook are as likely to go viral as reliable news .
While the diﬃculty to access spreading data on platforms like Facebook is a
concern, the growing popularity of ephemeral social media like Snapchat may
make future studies of this abuse all but impossible.
The results presented here suggest that curbing social bots may be an eﬀec-
tive strategy for mitigating the spread of online misinformation. Progress in this
direction may be accelerated through partnerships between social media plat-
forms and academic research. For example, our lab and others are developing
machine learning algorithms to detect social bots [6, 27, 29]. The deployment
of such tools is fraught with peril, however. While platforms have the right
to enforce their terms of service, which forbid impersonation and deception,
algorithms do make mistakes. Even a single false-positive error leading to the
suspension of a legitimate account may foster valid concerns about censorship.
This justiﬁes current human-in-the-loop solutions, which unfortunately do not
scale with the volume of abuse that is enabled by software. It is therefore
imperative to support research on improved abuse detection technology.
An alternative strategy would be to employ CAPTCHAs , challenge-
response tests to determine whether a user is human. CAPTCHAs have been
deployed widely and successfully to combat email spam and other types of online
abuse. Their use to limit automatic posting or resharing of news links could
stem bot abuse, but also add undesirable friction to benign applications of
automation by legitimate entities, such as news media and emergency response
coordinators. These are hard trade-oﬀs that must be studied carefully as we
contemplate ways to address the fake news epidemics.
The online article-sharing data was collected through Hoaxy, an open plat-
form developed at Indiana University to track the spread of fake news and
fact checking on Twitter . A search engine, interactive visualizations, and
open-source software are freely available (hoaxy.iuni.iu.edu). The data is
accessible through a public API.
The links to the stories considered here were crawled from websites that rou-
tinely publish unsubstantiated or debunked claims, according to lists compiled
by reputable third-party news and fact-checking organizations. We started the
collection in mid-May 2016 with 71 sites and added 51 more in mid-December
2016. The full list of sources is available on the Hoaxy website. The collection
period for the present analysis extends until the end of March 2017. During this
time, we collected 389,569 claims. We also tracked 15,053 stories published by
independent fact-checking organizations, such as snopes.com,politifact.com,
Using Twitter’s public streaming API, we collected 13,617,425 public posts
that included links to claims and 1,133,674 public posts linking to fact checks.
We extracted metadata about the source of each link, the account that shared it,
the original poster in case of retweet or quoted tweet, and any users mentioned
or replied to in the tweet.
We transformed URLs to their canonical forms to merge diﬀerent links re-
ferring to the same article. This happens mainly due to shortening services
(44% links are redirected) and extra parameters (34% of URLs contain analyt-
ics tracking parameters), but we also found websites that use duplicate domains
and snapshot services. Canonical URLs were obtained by resolving redirection
and removing analytics parameters.
We apply no editorial judgment about the truthfulness of individual claims;
some may be accurate (false positives) and some fake news may be missed (false
negatives). The great ma jority of claims are misleading, including fabricated
news, hoaxes, rumors, conspiracy theories, click bait, and politically biased
content. We did not exclude satire because many fake-news sources label their
content as satirical, making the distinction problematic. Furthermore, viral
satire is often mistaken for real news. The Onion is the satirical source with the
highest total volume of shares. We repeated our analyses of most viral claims
(e.g., Fig. 6) with articles from theonion.com excluded and the results were not
The bot score of Twitter accounts was computed using the Botometer ser-
vice, developed at Indiana University and available through a public API
(botometer.iuni.iu.edu). Botometer evaluates the extent to which an ac-
count exhibits similarity to the characteristics of social bots . We use the
Twitter Search API to collect up to 200 of an account’s most recent tweets and
up to 100 of the most recent tweets mentioning the account. From this data we
extract features capturing various dimensions of information diﬀusion as well
as user metadata, friend statistics, temporal patterns, part-of-speech and senti-
ment analysis. These features are fed to a machine learning algorithm trained
on thousands of examples of human and bot accounts. The system has high
accuracy  and is widely adopted, serving over 100 thousand requests daily.
The location analysis in Fig. 8 is based on 3,971 tweets that meet four
conditions: they were shared in the period between August and October 2016,
included a link to a claim, originated from an account with high bot score (above
0.6), and included one of the 51 U.S. state names or abbreviations (including
District of Columbia) in the location metadata. The baseline frequencies were
obtained from a 10% sample of public posts from the Twitter streaming API.
We considered 164,276 tweets in the same period that included hashtags with
the preﬁx #election and a U.S. state location. 2016 election forecast data was
obtained from from FiveThirtyEight (projects.fivethirtyeight.com/2016-
election-forecast/) and vote margins data from the Cook Political Report
Acknowledgments. We are grateful to Ben Serrette and Valentin Pentchev
of the Indiana University Network Science Institute (iuni.iu.edu), as well as
Lei Wang for supporting the development of the Hoaxy platform. Clayton A.
Davis developed the Botometer API. C.S. thanks the Center for Complex Net-
works and Systems Research (cnets.indiana.edu) for the hospitality during
his visit at the Indiana University School of Informatics and Computing. He
was supported by the China Scholarship Council. G.L.C. was supported by
IUNI. The development of the Botometer platform was supported in part by
DARPA (grant W911NF-12-1-0037). A.F. and F.M. were supported in part
by the James S. McDonnell Foundation (grant 220020274) and the National
Science Foundation (award CCF-1101743). The funders had no role in study
design, data collection and analysis, decision to publish or preparation of the
 A. Bessi and E. Ferrara. Social bots distort the 2016 us presidential election
online discussion. First Monday, 21(11), 2016.
 M. Conover, J. Ratkiewicz, M. Francisco, B. Gon¸calves, A. Flammini, and
F. Menczer. Political polarization on twitter. In Proc. 5th International
AAAI Conference on Weblogs and Social Media (ICWSM), 2011.
 M. D. Conover, B. Gon¸calves, A. Flammini, and F. Menczer. Partisan
asymmetries in online political activity. EPJ Data Science, 1:6, 2012.
 C. A. Davis, O. Varol, E. Ferrara, A. Flammini, and F. Menczer. Botornot:
A system to evaluate social bots. In Proceedings of the 25th International
Conference Companion on World Wide Web, pages 273–274, 2016.
 M. Del Vicario, A. Bessi, F. Zollo, F. Petroni, A. Scala, G. Caldarelli, H. E.
Stanley, and W. Quattrociocchi. The spreading of misinformation online.
Proc. National Academy of Sciences, 113(3):554–559, 2016.
 E. Ferrara, O. Varol, C. Davis, F. Menczer, and A. Flammini. The rise of
social bots. Comm. ACM, 59(7):96–104, 2016.
 J. Gottfried and E. Shearer. News use across social media platforms 2016.
White paper, Pew Research Center, May 2016.
 L. Gu, V. Kropotov, and F. Yarochkin. The fake news machine: How
propagandists abuse the internet and manipulate the public. Trendlabs
research paper, Trend Macro, 2017.
 N. O. Hodas and K. Lerman. How limited visibility and divided attention
constrain social contagion. In Proc. ASE/IEEE International Conference
on Social Computing, 2012.
 P. J. Hotez. Texas and its measles epidemics. PLOS Medicine, 13(10):1–5,
 L. Howell et al. Digital wildﬁres in a hyperconnected world. In Global
Risks. World Economic Forum, 2013.
 T. Jagatic, N. Johnson, M. Jakobsson, and F. Menczer. Social phishing.
Communications of the ACM, 50(10):94–100, October 2007.
 Y. Jun, R. Meng, and G. V. Johar. Perceived social presence reduces fact-
checking. Proceedings of the National Academy of Sciences, 114(23):5976–
 M. S. Levendusky. Why do partisan media polarize viewers? American
Journal of Political Science, 57(3):611—-623, 2013.
 W. Lippmann. Public Opinion. Harcourt, Brace and Company, 1922.
 B. Markines, C. Cattuto, and F. Menczer. Social spam detection. In Proc.
5th International Workshop on Adversarial Information Retrieval on the
Web (AIRWeb), 2009.
 A. Mosseri. News feed fyi: Showing more informative links in news feed.
press release, Facebook, June 2017.
 E. Mustafaraj and P. T. Metaxas. From obscurity to prominence in minutes:
Political speech and real-time search. In Proc. Web Science Conference:
Extending the Frontiers of Society On-Line, 2010.
 A. Nematzadeh, G. L. Ciampaglia, F. Menczer, and A. Flammini. How al-
gorithmic popularity bias hinders or promotes quality. Preprint 1707.00574,
 D. Nikolov, D. F. M. Oliveira, A. Flammini, and F. Menczer. Measuring
online social bubbles. PeerJ Computer Science, 1(e38), 2015.
 E. Pariser. The ﬁlter bubble: How the new personalized Web is changing
what we read and how we think. Penguin, 2011.
 X. Qiu, D. F. M. Oliveira, A. S. Shirazi, A. Flammini, and F. Menczer.
Limited individual attention and online virality of low-quality information.
Nature Human Behavior, 1:0132, 2017.
 J. Ratkiewicz, M. Conover, M. Meiss, B. Gon¸calves, A. Flammini, and
F. Menczer. Detecting and tracking political abuse in social media. In
Proc. 5th International AAAI Conference on Weblogs and Social Media
 M. J. Salganik, P. S. Dodds, and D. J. Watts. Experimental study of
inequality and unpredictability in an artiﬁcial cultural market. Science,
 C. Shao, G. L. Ciampaglia, A. Flammini, and F. Menczer. Hoaxy: A
platform for tracking online misinformation. In Proceedings of the 25th
International Conference Companion on World Wide Web, pages 745–750,
 N. Stroud. Niche News: The Politics of News Choice. Oxford University
 V. Subrahmanian, A. Azaria, S. Durst, V. Kagan, A. Galstyan, K. Lerman,
L. Zhu, E. Ferrara, A. Flammini, F. Menczer, A. Stevens, A. Dekhtyar,
S. Gao, T. Hogg, F. Kooti, Y. Liu, O. Varol, P. Shiralkar, V. Vydiswaran,
Q. Mei, and T. Hwang. The darpa twitter bot challenge. IEEE Computer,
 C. R. Sunstein. Going to Extremes: How Like Minds Unite and Divide.
Oxford University Press, 2009.
 O. Varol, E. Ferrara, C. A. Davis, F. Menczer, and A. Flammini. Online
human-bot interactions: Detection, estimation, and characterization. In
Proc. Intl. AAAI Conf. on Web and Social Media (ICWSM), 2017.
 L. von Ahn, M. Blum, N. J. Hopper, and J. Langford. Captcha: Using
hard ai problems for security. In E. Biham, editor, Advances in Cryptol-
ogy — Proceedings of EUROCRYPT 2003: International Conference on
the Theory and Applications of Cryptographic Techniques, pages 294–311.
 J. Weedon, W. Nuland, and A. Stamos. Information operations and face-
book. white paper, Facebook, April 2017.
 S. C. Woolley and P. N. Howard. Computational propaganda worldwide:
Executive summary. Working Paper 2017.11, Oxford Internet Institute,