Sources:Twitter;TechnologyReview;Chenetal.(workingpaper);TheEconomist
*Onalgorithmicfeed,wordsgreaterthantwocharacters,excludingstopwords
-0.6 -0.4 -0.2 0 0.2 0.4 0.6
DailyCaller
DailyWire
FreeBeacon
Gateway
Pundit
0 0.5 1.0 1.5 2.0 2.5
Hannity
Disgust
Sadness
Joy
Anger
Surprise
Fear
Tr u s t
Anticipation
20 30 40 50 60
Chronological
Bynewsfeedtype
Algorithmic
Chronological Algorithmic
←Negative Positive→ Chronological Algorithmic
Averagesentimentoftweetscontainingthesewords
Trump’stweets
Sentimentoftweetsservedtoa cloneofDonaldTrump’saccount
Averagefortweetscontaining 40 mostfrequentwords*,Sep-Dec 2019
Trumpclone’snewsfeed,hyper-partisannews
Tweetscontaininglinkfromselectedsites,%
Trumpclone’snewsfeed,emotions
Tweetscontainingemotion,%
→Comparedwitha chronologicalnewsfeed,Twitter’salgorithmtendstoshowtweetsthataremoreemotive
↓Thesewordswereaccompanied
bymoreemotivelanguageona
chronologicalfeed
→Tweetsonthealgorithmicfeed
usedmoreemotivelanguageto
describethesewordsthantweets
ona chronologicalfeeddid
Hannity
Ukraine
Schiff
Adam
Dems
political
people
Joe
Democratic
Democrats
impeachment
whistleblower
Pelosi
media
time
American
Donald
President
House
news
White
day
call
committee
world
election
Rep
Biden
Americans
Trump
Potus
vote
live
country
national
Greta
watch
week
United
support
TheEconomistAugust 1st 2020 73
S
ince launchinga policy on “mislead-
ing information” in May, Twitter has
clashed with President Donald Trump.
When he described mail-in ballots as “sub-
stantially fraudulent”, the platform told us-
ers to “get the facts” and linked to articles
that proved otherwise. After Mr Trump
threatened looters with death—“when the
looting starts, the shooting starts”—Twit-
ter said his tweet broke its rules against
“glorifying violence”. On July 28th the site
took down a tweet by Donald Trump junior
promoting a malaria drug for covid-19 that
plenty of studies discredit.
The president says that “social media
platforms totally silence conservatives’
voices.” However, a study by The Economist
finds the opposite. Twitter’s feed used to
show people the latest posts from accounts
they followed, but in 2016 it launched an al-
gorithm to serve “relevant” tweets to users,
even if they were days old and from unfa-
miliar accounts. We compared the two sys-
tems, and found that the recommendation
engine appears to reward inflammatory
language and outlandish claims.
Our experiment began in June 2019,
when we created a clone of Mr Trump’s pro-
file. This bot used his picture, biography
and location, and followed the same peo-
ple as he did. We used it to re-post some of
the president’s old tweets over several
weeks, so that the algorithm could learn
what our Trump clone cared about.
Then from September to December we
checked every ten minutes if Mr Trump had
tweeted something. If so, three things hap-
pened. First, our clone repeated the tweet.
Second, we checked its Twitter feed and re-
corded the first 24 posts served by the algo-
rithm. Finally, we simulated what a chro-
nological feed might have looked like,
using the 24 most recent tweets by ac-
counts that Mr Trump follows.
Our algorithmic and chronological
feeds differed starkly. Nearly half the rec-
ommended tweets were from users whom
Mr Trump does not follow. Using senti-
ment-analysis tools to extract feelings
from text, we found the average curated
tweet was more emotive, on every scale,
than its chronological equivalent—and
more so than Mr Trump’s own posts, too.
Sentiment analysis can be confusing.
The emotional scores assigned to tweets
by, say, Sean Hannity, a right-wing pundit,
might be highly negative—not because
they reflect poorly on him, but because he
stridently criticises others, such as Demo-
crats. Nonetheless, in a sample of 120,000
tweets, the posts recommended by the al-
gorithm were more likely to sit near either
end of a positive-to-negative spectrum.
Twitter might also boost extreme views.
Researchers at Indiana University have
classified a list of left- and right-wing web-
sites as untrustworthy or hyper-partisan.
We found 1,647 links to such domains on
our clone’s algorithmic feed, but only 895
on the chronological one. (Almost all cases
on both feeds were right-wing.)
Our experiment ended when a change
in Twitter’s interface broke our bot. The
platform also suspended another Trump
clone that copied his looting threat. How-
ever, if an algorithmic penchant for sensa-
tionalism has remained, then Twitter may
be amplifying and profiting from mislead-
ing tweets, rather than removing them. Its
business is serving ads to 330m users, even
if that means grabbing their attention by
showing them exactly what they want to
believe. Flagging a presidential whopper
every now and then will not change that. 7
The firm may serve Donald Trump
crankier tweets than he posts himself
Relevant content
Graphic detailTwitter’s algorithm