The Economist - USA (2021-11-13)

TheEconomistNovember13th 2021
Graphic detail Social media

89

Comingclean

A

mongthemosthotlydebatedques­
tions on social media is how algorith­
mic  bias  affects  social  media.  In  America
conservatives  claim  that  Facebook  and
Twitter bury or outright censor their views.
The left retorts that right­wing conspiracy
theories like QAnon flourish on these sites.
An unlikely arbiter recently emerged in
this debate: Twitter itself. In October it re­
leased a paper showing that its algorithm,
which  picks  which  tweets  users  see  in
which order, favoured right­leaning Amer­
ican  news  sites.  In  six  of  the  seven  coun­
tries studied, the algorithm also gave a dis­
proportionate  boost  to  lawmakers  from
conservative  political  parties.  Twitter
shared  its  data  with  The Economist this
month, letting us test the authors’ claims.
The study relied on a large experiment.

Until 2016 userssawtweetsonlyfromac­

counts  they  followed,  shown  in  reverse

chronological order. After launching its al­

gorithm, Twitter kept 1% of users in the old

system. This let it measure how often its al­

gorithm  served  up  certain  tweets,  com­

pared with the “reverse­chron” method.

In  April­August  2020  the  authors  used

this approach on 3,634 accounts belonging

to  legislators  from  32  political  parties.  Al­

though they did not detect political bias in

the  treatment  of  individual  lawmakers,

they  did  find  a  slant  when  grouping  ac­

counts by party. In all countries but Germa­

ny,  the  algorithm’s  “amplification  ratio”

was  lower  for  members  of  leftist  parties

than for members of right­wing ones.

This  discrepancy  could  arise  for  rea­

sons besides ideology. To test alternatives,

we fed Twitter’s data into a model that ac­

counted for the amount of amplification in

each country, political parties’ vote shares

in  the  most  recent  elections  and  whether

they  were  in  government.  Yet  even  after

making  these  adjustments,  the  algorithm

still favoured conservative parties.

In contrast, the evidence for bias aiding

right­wing  American  media  seemed  less

robust.Thealgorithmdidgiveextraam­

plification  to  news  sources  that  indepen­

dent  groups  like  Ad  Fontes  Media  classify

as conservative. However, ideology and ac­

curacy (which Ad Fontes, among others, al­

so scores) are correlated to each other. And

among  the  sites  studied,  those  with  the

strictest  sourcing  and  fact­checking  also

tended to have left­of­centre politics.

In  2019  we  studied  how  Google  ranks

news stories, and found that accuracy, not

ideology, explained its rankings. This is al­

so true of Twitter. However, whereas Goo­

gle  gave  higher  rankings  to  more  reliable

sites,  we  found  that  Twitter  boosted  the

least  reliable  sources,  regardless  of  their

politics. Left­wing sites with poor accuracy

scores, like tmz, were amplified more than

credible,  conservative  ones  like  the  Wall

Street Journal.  ProPublica,  a  non­profit  fo­

cused  on  public­interest  investigations,

had one of the lowest amplification ratios.

Because  Twitter  did  not  share  the

tweets it studied, we could not identify the

type of content that its algorithm rewards.

But if the company wants to reducemisin­

formation on its site, making tweakstofa­

vour rigorous reporting might help.n

According to Twitter, Twitter’s
algorithm favours conservatives

→ Twitter gives its biggest boosts to right-wing legislators, and to less-reliable news sources regardless of ideology

Audience reached on algorithmic timeline compared with chronological timeline, Apr to Aug 2020

Tweetsfromlegislators

Groupedbypartyandorderedfromleft-to right-wing

*CombinationofAdFontesMedia,MediaBias/FactCheck

andNewsGuardscores(firstprincipalcomponent)

Sources:“AlgorithmicAmplificationofPoliticsonTwitter”,byFerenc

Huszáretal.;AdFontesMedia;MediaBias/FactCheck;NewsGuard

Tweetslinkingtonewssites

0.8x

0.9x

1x

1.1x

1.4x

1.3x

1.2x

02040 6080 100

Accuracyindex*

Moreaccurate→

In government

Algorithmgivestweets
↑biggeraudience

↓smalleraudience

Politicalslant
AdFontesscore

←Left Right→

Confidenceinterval

AverageAmericanusersper
daywhosefeedslinkedtosite

Japan

Germany

Canada

Britain

UnitedStates

Republican

Democratic

Conservative

LibDem

SNP

Labour

Conservative

BlocQuébécois

Liberal

NDP

AfD

CDU/CSU

FDP

SPD

Greens

TheLeft

LDP

Komeito

DPFP

CDP

Communist

1.5x 2x log scale

1.1x

1.1x

1.5x 2x 3x 4x

1.1x 1.5x 2x 3x 4x

1.1x 1.5x 2x 3x 4x

1.1x 1.5x 2x 3x 4x

AlJazeera

Daily Mail

Daily Wire

Breitbart FoxNews

Mashable

NewYorkPost

Newsweek

CNN

OneAmericaNews

ProPublica

RT

Snopes

TalkingPoints

Memo

TheFederalist

The Hill

The Verge

TMZ

Vanity Fair

BuzzFeed Mother

Jones

MSNBC

National

Review

TheGuardian

Time

Vice Forbes

Vox

WashingtonPost

AP

Daily Kos

WSJ

NYT

The Gateway Pundit

The Economist