ArticlePDF Available

Same Candidates, Different Faces: Uncovering Media Bias in Visual Portrayals of Presidential Candidates with Computer Vision


Abstract and Figures

How do today’s partisan media outlets produce ideological bias in their visual coverage of political candidates? Applying computer vision techniques, this study examined 13,026 images from 15 news websites about the two candidates in the 2016 U.S. presidential election. The analysis unveils a set of visual attributes (e.g., facial expressions, face size, skin condition) that were adopted by media outlets of varying ideologies to differentially portray these two candidates. In addition, this study recruited 596 crowdsourced workers to rate a subset of 1,200 images and demonstrated that some visual features also effectively shape viewers’ perceptions of media slant and impressions of the candidates. For example, Clinton was portrayed with more expressions of happiness, which rendered her as more favorable, whereas Trump was associated with more expressions of anger, which made him look less positive but more dominant. These differences in facial expressions varied in line with media outlets’ political leanings.
Content may be subject to copyright.
Journal of Communication ISSN 0021-9916
Same Candidates, Different Faces:
Uncovering Media Bias in Visual Portrayals of
Presidential Candidates with Computer
Yilang Peng
Annenberg School for Communication, University of Pennsylvania, Philadelphia, PA 19104, USA
How do todays partisan media outlets produce ideological bias in their visual coverage
of political candidates? Applying computer vision techniques, this study examined
13,026 images from 15 news websites about the two candidates in the 2016 U.S. presi-
dential election. The analysis unveils a set of visual attributes (e.g., facial expressions,
face size, skin condition) that were adopted by media outlets of varying ideologies to dif-
ferentially portray these two candidates. In addition, this study recruited 596 crowd-
sourced workers to rate a subset of 1,200 images and demonstrated that some visual
features also eectively shape viewersperceptions of media slant and impressions of the
candidates. For example, Clinton was portrayed with more expressions of happiness,
which rendered her as more favorable, whereas Trump was associated with more
expressions of anger, which made him look less positive but more dominant. These dif-
ferences in facial expressions varied in line with media outletspolitical leanings.
Keywords: Media Bias, Visual Bias, Face Perception, Trait Perception, Non-Verbal
Communication, Computer Vision, Computational Social Science, Crowdsourcing.
Increasingly, media outlets are explicitly labeling their political aliations to market
themselves (Groeling, 2013). In the hope of exposing media consumers to more bal-
anced and diverse viewpoints, scholars have made various attempts to automatically
label political slants in media content (Park, Kang, Chung, & Song, 2009). Given the
diculty in analyzing images on a large scale, prior research examining content bias
across media outlets has been mostly limited to textual data (Groeling, 2013). Yet,
visual content proliferates in the digital media environment and is widely used in
political communication (Grabe & Bucy, 2009;Verser & Wicks, 2006). Empirical
studies also demonstrate that visual attributes such as facial expressions and face
Corresponding author: Yilang Peng; e-mail:
1Journal of Communication 00 (2018) 122 © The Author(s) 2018. Published by Oxford University Press on behalf of
International Communication Association. All rights reserved. For permissions, please e-mail:
Downloaded from by University of Pennsylvania Library user on 01 October 2018
size can eectively shape our impressions and voting preferences towards politicians
(Mutz, 2007;Tiedens, 2001). This research gap regarding visual bias should require
special attention today.
This study tackled the challenge of analyzing large-scale visual media by bridg-
ing communication research and computer vision, a eld that trains computers to
understand digital imagery (Szeliski, 2010). Applying computer vision techniques
such as facial detection and emotional analysis, this study examined 13,026 images
from 15 news websites covering the 2016 U.S. presidential candidatesHillary
Clinton and Donald Trumpregarding various visual features. Beyond this meth-
odological contribution, this study also advances our understanding of visual bias in
the following ways. First, while extensive research has been devoted to quantifying
the direction and magnitude of partisan bias, relatively limited studies have exam-
ined the forms of bias in visual content. Based on the ideological positions of media
outlets established by prior research (Budak, Goel, & Rao, 2016;Flaxman, Goel, &
Rao, 2016;Mitchell, Gottfried, Kiley, & Matsa, 2014), this study reveals the specic
portrayals that media outlets use to convey their partisan views about the two candi-
dates. In addition, this research also recruited crowdsourced workers to rate a sub-
sample of images on their perceptions of media slant and impressions of the two
candidates. By doing so, this study further illuminates how dierent visual represen-
tations adopted by partisan media potentially aect audiences of varying ideologies.
Measuring partisan media bias in visual content
This study denes partisan bias as systematic patterns in media content that favor
one political party, candidate, or ideology over another (Groeling, 2013;Waldman
& Devitt, 1998). First, bias should be distinguished from slant in a specic piece of
media content, as bias should be systematic, rather than anecdotal, episodic, or
eeting(Groeling, 2013, p. 133). In addition, as Entman (2007, p. 166) argued, to
establish media bias, researchers need to show patterns of slant that regularly
prime audiences, consciously or unconsciously, to support the interests of particular
holders or seekers of political power.Our operationalization of bias thus retires the
notion that it is a deviation from the truth but instead incorporates its potential
impact on audiences. If a piece of content is determined as favoring certain political
actors, it is reasonable to expect that viewers exposed to it should form favorable
impressions about the targets or at least perceive the intended slant. While prior
research has proposed other types of partisan bias, such as coverage bias (which
focuses on the volume of coverage), this study focused on presentation bias, which
specically deals with the favorability of media coverage toward one party or ideol-
ogy over the other (DAlessio & Allen, 2000;Groeling, 2013).
Previous attempts at measuring visual slant can be broadly categorized into two
approaches. Some have used more objective measures by coding visual content on a
checklist of features that researchers predene as (un)favorable treatment of politi-
cians. The criteria used to determine visual slant usually include a politicians
2Journal of Communication 00 (2018) 122
Media Bias in Visual Portrayals of Presidential Candidates Y. Peng
Downloaded from by University of Pennsylvania Library user on 01 October 2018
nonverbal behaviors (e.g., facial expressions, hand gestures, activity), contextual fea-
tures (e.g., photographic settings, other objects, and people in the same picture),
and structural features (e.g., camera angle, color; Grabe & Bucy, 2009;Moriarty &
Popovich, 1991;Verser & Wicks, 2006;Waldman & Devitt, 1998). This checklist
approach oers nuanced understandings of how bias is embodied in specic visual
portrayals, although, as scholars have noted, dierent visual features do not contrib-
ute equally to the favorability of images (Barrett & Barrington, 2005). It also
remains unknown whether these attributes, selected by scholars, indeed inuence
audience interpretations (Lobinger & Brantner, 2015).
Others have used more subjective measures by instructing coders to rate the
favorability of media content. For example, in Barrett and Barrington (2005), three
coders rated photos of politicians on a highly unfavorableto highly favorable
scale. In Hehman, Graber, Homan, and Gaertner (2012), six coders evaluated can-
didates in photos on warmth and competence. With crowdsourcing platforms like
Amazon Mechanical Turk, scholars can recruit more coders from diverse demo-
graphic and ideological backgrounds to assess bias on a large scale (Budak et al.,
2016). Nevertheless, this approach still relies on coderssubjective interpretations
and answers only how much, but not how, media content is biased. As Grabe and
Bucy (2009, p. 101) argued, visual analyses should move beyond the positive ver-
sus negativeindex measures and investigate more specic and nuanced character
framebuilding dimensions.
These two approaches can be regarded as not only dierent methods for quanti-
fying bias, but also two routes of conceptualizing bias that complement each other.
Partisan bias should rst be established as systematic patterns of dierential treat-
ment of political actors in media content. However, dierences alone do not guaran-
tee favorability; it requires additional eorts to demonstrate that these patterns
indeed (dis)advantage certain actors among audiences. This study thus integrated a
content analysis that investigated whether the visual coverage of candidates does
vary by media outlets with a survey that asked crowdsourced workers to rate the
favorability of these pictures. By doing so, we hoped this research would reveal
visual cues that (a) reveal media outletsideological positions and (b) inuence
audience perceptions of favorability.
Selection of visual features
The rst step in our approach was to select visual features that could be captured by
computer vision applications on a large scale and that should, theoretically, reect
media bias. Based on previous scholarship on visual bias, social cognition, and polit-
ical psychology, this study focused on the following features:
Facial orientation
Current facial detection algorithms often show facial orientation in three angles:
pitch refers to the extent of a head bowing down or raising up; roll reects the
3Journal of Communication 00 (2018) 122
Y. Peng Media Bias in Visual Portrayals of Presidential Candidates
Downloaded from by University of Pennsylvania Library user on 01 October 2018
extent of a head tilting to the left or to the right; and yaw shows the extent of a face
turning to the left or to the right (Figure 1). The pitch angle can be seen as a proxy
of camera angle, a criterion frequently used in prior research to indicate visual
favorability. A low camera angle (face pitching upward) is regarded as a better por-
trayal, as it could convey a sense of dominance and power compared with a high
angle (face pitching downward; Grabe & Bucy, 2009;Waldman & Devitt, 1998). It
is unclear, however, if the other two anglesroll and yaware related to visual
Face size and location
As a cue for judging interpersonal distance, face size might pose mixed impacts on
perceptions of other people. On one hand, compared with a long shot, a close-up
portrait makes a person seem closer to viewers, thus appearing more intimate or
dominant (in the sense of face-to-face confrontation; Archer, Iritani, Kimes, &
Barrios, 1983;Grabe & Bucy, 2009). On the other hand, an extreme close-up might
be a negative portrayal, as it resembles an extremely close physical distance that vio-
lates the notion of personal space. It also brings a persons face under detailed scru-
tiny, revealing skin aws or awkward expressions (Grabe & Bucy, 2009). In
addition, prior research has also claimed that featuring a candidate dominating the
photo or as the center of attention positively portrays the candidate (Verser &
Wicks, 2006). A photo that locates a politicians face closer to the center should cast
the person in a better light.
Facial expressions
Facial expressions of emotionmotions or positions of facial muscles that convey
the emotional state (Ekman & Friesen, 2003)have also been used to evaluate
visual bias. Looking happy or condent is usually coded as a positive representation
of politicians, while frowning or looking sad, worried, or tired is seen as negative
(Moriarty & Popovich, 1991;Waldman & Devitt, 1998). Happy faces are also per-
ceived as more trustworthy, attractive, and dominant (Knutson, 1996;Oosterhof &
Todorov, 2008;Sutherland et al., 2013). Besides coding facial expressions on a posi-
tivenegative spectrum, scholars have also argued for distinctions among discrete
emotions. Ekman and Friesen (2003) identied six basic facial expressions: happi-
ness, sadness, fear, anger, surprise, and disgust. Grabe and Bucy (2009) also distin-
guished among anger/threat, fear/evasion, and happiness/reassurance. Dierent
negative emotions may produce distinct impressions: for example, individuals
showing anger or disgust are often perceived as more dominant and powerful than
those showing sadness or fear (Knutson, 1996).
Eye and mouth status
Eye status is another criterion frequently used to evaluate the favorability of photo-
graphs. Looking directly at the camera or at someone in pictures is coded as a posi-
tive portrayal of the candidate, whereas closed eyes portray a politician negatively
(Moriarty & Popovich, 1991;Verser & Wicks, 2006). Research has shown that eye
4Journal of Communication 00 (2018) 122
Media Bias in Visual Portrayals of Presidential Candidates Y. Peng
Downloaded from by University of Pennsylvania Library user on 01 October 2018
openness could make people look more intelligent and attractive (Talamas, Mavor,
Axelsson, Sundelin, & Perrett, 2016). Mouth openness should also be an indicator
of favorability. Prior research has regarded exhibiting dynamic behaviors such as
speaking as a favorable depiction of politicians (Verser & Wicks, 2006). Mouth
openness also reects the intensity of smiling, although it can also indicate yelling
and shouting that convey aggressiveness.
Skin condition
Skin condition has rarely been examined in visual bias literature, but its eects are
frequently documented in face perception research. First, darkening a faces skin
tone often leads to more negative reactions to it (Alter, Stern, Granot, & Balcetis,
2016;Ronquillo et al., 2007). An analysis of news articles showed that how posi-
tively an article portrayed a person correlated with the persons skin lightness in the
articles visuals (Alter et al., 2016). Facial skin coloration is another important cue
in impression formation. Skin redness and yellowness can be used as cues for infer-
ring health status (Stephen, Smith, Stirrat, & Perrett, 2009). Signaling increased
blood ow, facial redness is also linked to perceptions of attractiveness, dominance,
and aggressiveness (Stephen, Oldham, Perrett, & Barton, 2012). Lastly, faces with
healthier skin are often rated as more attractive, whereas skin imperfections such as
wrinkles and uneven pigmentation make a person look older, less healthy, and less
attractive (Fink, Grammer, & Matts, 2006;Jones, Little, Burt, & Perrett, 2004).
Other people
In prior research, presenting a cheering crowd or attentive colleagues together with
a politician has often been coded as a positive representation, whereas featuring the
politician alone or with inattentive crowds or colleagues has been seen as negative
(Moriarty & Popovich, 1991;Verser & Wicks, 2006). Being accompanied by other
Figure 1 Facial orientation in pitch, roll, and yaw angles. Credit: Kyle Cassidy.
5Journal of Communication 00 (2018) 122
Y. Peng Media Bias in Visual Portrayals of Presidential Candidates
Downloaded from by University of Pennsylvania Library user on 01 October 2018
people in pictures would also make a person appear more attractive than being
photographed alone (Walker & Vul, 2014). Therefore, the presence of other people
in pictures, as well as their facial expressions and eye openness, can also be proxies
of media bias.
Visual bias across liberal and conservative media outlets
Having proposed a list of computer vision features that should theoretically reect
media bias, this study then asked how bias would be embodied in dierential por-
trayals of politicians. Past research on partisan bias often focused on whether the
media deviates from the norm of balance and exhibits an overall liberal or conserva-
tive bias, but no consistent patterns have emerged. For example, regarding television
news, DAlessio and Allens (2000) meta-analysis revealed a detectable but small
pro-Democrat bias, whereas Grabe and Bucy (2009) found a persistent pro-
Republican one. Given the rise of partisan media outlets that explicitly favor one
side instead of sticking to the norms of objectivity and balance, scholarship has
gradually shifted to quantifying bias at the individual media outlet level (Budak
et al., 2016). Regarding visual bias, research has shown that media outlets like news-
papers (Barrett & Barrington, 2005) and websites (Hehman et al., 2012) indeed pub-
lish photos that portray candidates they endorse more favorably than the
candidatesopponents. Therefore, we should expect that liberal media portrayed
Clinton better than Trump and conservative media acted reversely prior to the elec-
tion. However, it remains unknown what visual cues are adopted by media outlets
of various political aliations in their (un)favorable treatment of politicians, which
is one focus of this study.
The analysis of partisan bias was further complicated in this case because the
past election had both a female and a male candidate. Although this research can
uncover dierences in visual representations of Trump and Clinton, it is dicult to
attribute these dierences solely to partisan bias or gender bias, as some attributes
related to favorability are also linked to gender. For example, faces are often shown
as more prominent in visual depictions of men than those of women, a phenome-
non termed as face-ismin prior research (Archer et al., 1983). Gender is also ste-
reotypically linked to dierent emotions: anger is often seen as more masculine,
signifying aggressiveness and dominance, whereas happiness, sadness, and fear are
seen as more feminine, showing friendliness or weakness (Plant, Hyde, Keltner, &
Devine, 2000). Therefore, instead of focusing on the overall contrast between
Clinton and Trump, this study asked whether the dierential treatment between the
two candidates varied by media outletsideological positions, thus uncovering visual
cues that signalled their political orientation.
RQ1: Which visual featuresfacial orientation (pitch, roll, yaw angles), face size,
face location, facial expressions (e.g., happiness, anger, sadness), eye openness, eye
gaze direction, mouth openness, facial skin condition (lightness, redness,
6Journal of Communication 00 (2018) 122
Media Bias in Visual Portrayals of Presidential Candidates Y. Peng
Downloaded from by University of Pennsylvania Library user on 01 October 2018
yellowness, health), presence of other people, other peoples facial expressions, and
eye opennesswere used by liberal and conservative media to dierently portray
Clinton and Trump?
Favorability and its dimensions
As noted earlier, the notion of favorability in media bias implies that biased content
should indeed advantage certain political actors among media consumers. One
important function of visual media in political communication is to convey cues
that help us judge politicianstraits and characters (e.g., warmth, competence;
Lobinger & Brantner, 2016), which in turn could inuence voting preferences
(Caprara & Zimbardo, 2004). Nevertheless, the majority of media bias studies often
see bias as a single unfavorable-versus-favorable spectrum, which might not capture
the diverse eects of visual portrayals on viewersimpressions. For example, show-
ing the negative emotion of anger might make a person look unfriendly, but simul-
taneously dominant (Knutson, 1996;Tiedens, 2001). In addition, dierent traits
also correspond to gender stereotypes: women are expected to be more friendly and
kind, whereas men are perceived to be more assertive and aggressive (Prentice &
Carranza, 2002). This study thus asked whether dierent visual representations of
politicians actually aect viewersperceptions of media slant, as well as evaluations
of politicians, on separate trait dimensions.
Prior research is still divided on what specic dimensions govern our judgment
of people from visual portrayals. Research in person perception claims that two fun-
damental dimensions underlie our judgments of people: one dimension (commu-
nion) captures traits related to perceived intent and could be further divided into
two sub-dimensions, warmth (e.g., sociable, friendly) and morality (e.g., trustwor-
thy, sincere); and the other dimension (agency) captures traits about perceived abil-
ity and incorporates two sub-dimensions as well, competence (e.g., intelligent,
competent) and dominance (e.g., dominant, assertive; Abele et al., 2016). Evaluating
politicians could be seen as a specic case of person perception. Caprara and
Zimbardo (2004) also found a two-factor structure in judging personalities of politi-
ciansenergy and agreeablenesswhich largely overlap agency and communion.
Research in face perception also proposes that we use multiple dimensions to
infer traits from human faces. The rst dimension, labeled as valence, incorporates
traits related to warmth, morality, and competence, indicating an overall favorabil-
ity in impressions. The second dimension reects only dominance (Oosterhof &
Todorov, 2008). Attractiveness has also been proposed as a third factor in perceiving
human faces (Sutherland et al., 2013). In this study, news photographs of politicians
include not only their faces, but also social information such as their interactions
with other people. Therefore, this study rst examined the structure underlying view-
ersperceptions of candidates in images and then investigated whether visual features
inuence these dimensions dierently. Here, this study proposes two research ques-
tions regarding the potential eects of various visual features.
7Journal of Communication 00 (2018) 122
Y. Peng Media Bias in Visual Portrayals of Presidential Candidates
Downloaded from by University of Pennsylvania Library user on 01 October 2018
RQ2: Among the visual features proposed in RQ1, what features could best predict
viewersjudgment of media slant in images of the two candidates?
RQ3: (a) Among warmth, morality, competence, dominance, and attractiveness,
what dimensions underlie peoples perceptions of candidates in news photographs?
(b) And what visual features proposed in RQ1 could best predict audience
perceptions of these dimensions?
Data preparation
Prior research has already placed a list of popular news websites on the liberalcon-
servative spectrum. This research combined insights from several recent studies:
one that averaged crowdsourced workersperceived slant of each media outlets
news articles (Budak et al., 2016) and two using aggregated political orientation of
each outlets audience as a proxy of its ideological position (Flaxman et al., 2016;
Mitchell et al., 2014). The sample included eight liberal sites (Daily Kos, Slate,The
New York Times,The Hungton Post,The Washington Post, MSNBC, BBC, CNN),
four relatively neutral sites (USA Today, Reuters, NBC News, The Wall Street
Journal), and three conservative sites (Fox News, Breitbart, and TheBlaze). With
Google search, the study searched for images of the two candidates limited to a spe-
cic news site (e.g., Hillary Clinton A total of 20,702 still images
were retrieved in the last week of November 2016.
All images were transformed to
JPEG format and large images were resized so both width and height did not exceed
1000 pixels.
With a computer vision service, Face++, this study then identied images with
visible faces of the two candidates. The analysis rst prepared a face set that
included the two candidates. Next, for each image, the analysis detected whether
this image contained faces (facial detection). For each detected face, Face++ com-
pared this face to faces in the face set and returned the most similar-looking face,
along with a condence score of the two faces belonging to the same person (facial
Pictures without faces of Clinton or Trump were excluded. Based on
the URLs of images returned by Google, this study then determined the date of each
photo. The majority of URLs embedded the date when an article or a photo was
published. For remaining images, a combination of web scraping and manual
checking was used. Images from 20152016 (N=13,026; 6,543 for Clinton) were
kept in further analysis (Figure 2).
Participant ratings
From Amazon Mechanical Turk, 596 U.S. crowdsourced workers who had com-
pleted at least 100 tasks with an approval rate above 98% were recruited to rate a
subset of 1,200 images randomly selected from the sample (40 for each candidate
and each outlet). Each participant rated a random set of 20 images for each candi-
date. On average, each image received 19.9 ratings (SD =1.66, range =1832).
8Journal of Communication 00 (2018) 122
Media Bias in Visual Portrayals of Presidential Candidates Y. Peng
Downloaded from by University of Pennsylvania Library user on 01 October 2018
Ratings of each image were averaged across participants and used in the later analy-
sis (Oosterhof & Todorov, 2008).
Each image was rated on the following questions.
1. Perceived slant: on ve-point scales, participants rated how an image negatively
or positively (1 =extremely negative, 5 =extremely positive) and unfavorably or
favorably (1 =extremely unfavorable, 5 =extremely favorable) portrayed a can-
didate (α=.99; M
=3.32, SD =.47; M
=2.84, SD =.58).
2. Traits perceptions: on ve-point scales (1 =strongly disagree; 5 =strongly
agree), participants rated in each image whether the candidate looked friendly,
warm, honest, trustworthy, dominant, assertive, competent, intelligent, attractive,
youthful, and healthy. These adjectives correspond to various constructs pro-
posed in prior research (Table 3;Abele et al., 2016;Oosterhof & Todorov, 2008;
Sutherland et al., 2013).
Figure 2 Data preparation procedures.
9Journal of Communication 00 (2018) 122
Y. Peng Media Bias in Visual Portrayals of Presidential Candidates
Downloaded from by University of Pennsylvania Library user on 01 October 2018
3. Perceived facial expression: given the subjectivity in human perceptions of facial
expressions, this study also asked 396 participants to rate a subsample of 800
images to determine the extent to which the candidatesfaces displayed anger,
disgust, fear, happiness, sadness, surprise, and emotional neutrality (1 =not at
all,5=to the full extent). These ratings were used to validate emotional analysis
results from computer vision services.
Computer vision analysis
For each image, Face++ provided the yaw, roll, and pitch angles that represented
facial orientation. Face++ also provided each faces location in the picture as a facial
rectangle that bounded the face region (Figure 2). Face size was calculated as the
ratio between the size of the face rectangle and the size of the image. Face location
was calculated as 12×d
, in which d
was the distance between the face rectan-
gles center and the pictures center, and d
was the length of the pictures diagonal.
A higher value indicated that the face was closer to the center. Face++ also provided
the extent of eye openness and mouth openness, as well as eye gaze direction, as a
three-dimensional vector. The analysis calculated the angle between the gaze direc-
tion and the image, so a higher value approaching 90 degrees would indicate the eye
was looking towards the camera rather than looking elsewhere. Values for eye open-
ness and eye gaze direction were averaged across both eyes.
Regarding facial expressions, prior research has shown that computer vision
tools can accurately detect happiness, but might not identify other facial expressions
well (Dehghan, Ortiz, Shu, & Masood, 2017). The analysis thus compared emotions
Table 3 Principal Component Analysis of Perceived Traits
(N=1200) Clinton (600) Trump (600)
Constructs Items 121212
Communion (warmth) Friendly (.89) .87 .38 .85 .44 .82 .43
Warm (.88) .89 .35 .87 .41 .86 .40
Communion (morality) Honest (.65) .95 .01 .94 .01 .94 .05
Trustworthy (.69) .96 .01 .95 .02 .94 .03
Agency (dominance) Dominant (.64) .29 .92 .29 .91 .47 .84
Assertive (.67) .35 .89 .35 .89 .45 .84
Agency (competence) Competent (.73) .92 .24 .88 .35 .92 .20
Intelligent (.74) .90 .22 .85 .38 .91 .19
Attractiveness Attractive (.78) .94 .12 .93 .15 .91 .13
Youthful (.78) .85 .24 .85 .28 .79 .28
Healthy (.73) .93 .02 .92 .07 .90 .02
Variance explained 66.5% 22.4% 62.2% 26.6% 61.1% 25.1%
Note: Inter-rater reliability for each item is provided in brackets. Factor loadings larger than
.5 are shown in bold.
10 Journal of Communication 00 (2018) 122
Media Bias in Visual Portrayals of Presidential Candidates Y. Peng
Downloaded from by University of Pennsylvania Library user on 01 October 2018
detected by four popular emotion analysis servicesMicrosoft, Face++, Sighthound,
and Google Visionwith participantsperceived emotions (Table 1). All services
detected happiness well, with Microsoft performing best (r=.85). Microsoft also pre-
dicted anger (r=.54) and emotional neutrality (r=.63) relatively accurately. Yet, the
detection of other emotions, such as fear and sadness, was not satisfactory. Therefore,
scores of anger, happiness, and emotional neutrality from Microsoft were used and
otheremotionswereexcludedfromtheanalysis. Where Microsoftfailedtodetecta
candidates face in a picture, Face++sresultswereused.
Face++ provided facial landmarks of each face, which were locations of impor-
tant face components, such as face contour, eyes, and mouth (Figure 2). Based on
facial landmarks, the analysis then identied the facial skin region as the facial
region excluding eyebrows, eyes, nose, and mouth. The image was transformed into
the CIELab color space. Lightness, redness, and yellowness of facial skin were calcu-
lated as the average L, a, and b values, respectively, of pixels inside the facial skin
region (Stephen et al., 2012). In addition, Face++ also returned likelihoods about a
faces skin condition. Skin health was calculated as the dierence between the likeli-
hood that the skin was healthy and average likelihood that the skin showed dierent
types of ill conditions, such as dark circles and stains.
The number of other peoples faces detected by Face++ was used to indicate
whether a candidate was presented with other people or not. If the candidate was
accompanied by multiple individuals in an image, the means of other faceshappi-
ness and eye openness were used as predictors. For a small portion of images with
multiple faces of the same candidate (e.g., Trump standing in front of a screen
showing his face; .9%), face-related visual attributes were averaged across the candi-
dates faces. The number of repetitions of the candidates face was included as a
control variable. A few computationally-calculated aesthetical features were also
included as control variables, including brightness and contrast, measured as the
mean and the standard deviation of all pixelsperceived luminance values; colorfulness,
Peng &
Jemmott, 2018); and image size and aspect ratio, measured as the product and the quo-
tient of an images width and height.
The study rst examined whether the treatment between Clinton and Trump dif-
fered between liberal and conservative media regarding various visual features
(RQ1). The analysis constructed a candidate variable (0 =Clinton, 1 =Trump) and
a media outlet variable (1 =liberal media, 2 =relatively neutral, 3 =conservative
media) based on the ideological positions of media outlets, as quantied by prior
research (see Method section). A series of moderated multiple regressions were con-
ducted (Jaccard & Turrisi, 2003), with each regression using the candidate variable,
the media outlet variable, and their interaction to predict one visual feature
(Table 2). For each visual feature, a statistically-signicant coecient of the
11Journal of Communication 00 (2018) 122
Y. Peng Media Bias in Visual Portrayals of Presidential Candidates
Downloaded from by University of Pennsylvania Library user on 01 October 2018
Table 1 Correlations Between Human-Perceived and Computer Vision ServicesDetected
Facial Expressions
Face++ Microsoft Sighthound Google Vision
Anger (.88) .29*** .54*** .31*** .23***
Disgust (.81) .25*** .36*** .09* NA
Fear (.50) .08* .19*** .11** NA
Happiness (.96) .79*** .85*** .71*** .80***
Sadness (.64) .04 .28*** .18*** .17***
Surprise (.77) .24*** .39*** .23*** .36***
Neutral (.74) .46*** .67*** .36*** NA
N791 738 783 772
Note: Analyses were performed on images containing only one face of the candidate (N=
791). The sample size varied, as some faces detected by Face++ were not detected by other
services. NA =not available. The inter-rater reliability for each item is provided in brackets.
*p<.05, **p<.01, ***p<.001.
Table 2 Distribution of Visual Features Across Candidates and Media Outlets
Candidate Media Outlet Candidate ×Media Outlet
Face orientation: Pitch .14*** .01 .02
Roll .11*** .02* .02
Yaw .02 .01 .01
Face size .14*** .06*** .12***
Face location .09*** .18*** .02
Facial expressions: Happiness .46*** .06*** .08***
Anger .42*** .02
Neutral .09*** .01 .02
Eye openness .30*** .01 .04
Eye gaze .22*** .03* .01
Mouth openness .11*** .00 .03
Skin condition: Skin lightness .19*** .03** .04
Skin redness .40*** .01 .02
Skin yellowness .08*** .04** .02
Skin health .09*** .05*** .07**
Number of other peoples faces .12*** .03** .07**
Other faceshappiness .04 .01 .02
Other faceseye openness .13*** .02 .00
Note:N=13026 (N=4939 for other faceshappiness and eye openness). Candidate: 0 =
Clinton, 1 =Trump. Media outlet: 1 =liberal, 2 =neutral, 3 =conservative. Each line repre-
sents one regression model. Standardized coecients are shown.
p<.10, *p<.05,
**p<.01, ***p<.001.
12 Journal of Communication 00 (2018) 122
Media Bias in Visual Portrayals of Presidential Candidates Y. Peng
Downloaded from by University of Pennsylvania Library user on 01 October 2018
candidate variable would imply that media overall portrayed the two candidates dif-
ferently regarding that feature. As noted earlier, this dierence could be a mix of
partisan and gender biases, or characteristics of these two candidates (e.g., Trump
and Clinton dier in their skin tones). A signicant interaction would imply that
the size of this dierence was moderated by media outletspolitical orientations,
suggesting the presence of partisan bias.
In overall media coverage, compared with Clinton, Trump images had larger
faces (β=.14), showed less happiness (β=.46) but more anger (β=.42), and por-
trayed less healthy facial skin (β=.09) and fewer other peoples faces (β=.12,
all ps<.001). As indicated by signicant interactions, these gaps regarding face size
(β=.12, p<.001), happiness (β=.08, p<.001), anger (β=.06, p=.002), skin
health (β=.07, p=.002), and number of other faces (β=.07, p=.001) narrowed
or reversed as the media outletspolitical orientations moved from liberal to conser-
vative (Figure 3), implying that these attributes were adopted by outlets to dieren-
tially portray the two candidates. For example, regarding happiness, the gap between
Clinton and Trump was 32 overall (on a 1100 scale), and was wider in liberal media
(34) than in conservative media (27). The gap was most pronounced in Daily Kos (46)
and least in Breitbart (23), two sites situated at two extreme ends of the ideological
spectrum. Face size served as another example. While liberal and relatively neutral
outlets almost universally portrayed Trump images with larger faces than Clinton, the
gap diminished or reversed in conservative sites such as Fox News, TheBlaze, and
Breitbart. A similar pattern occurred for skin health. With two outliers (The New York
Times and The Hungton Post), liberal media, especially Daily Kos and Slate, por-
trayed Clinton with healthier facial skin, but this gap narrowed in conservative media.
Having shown what visual features were adopted by media outlets of varying
positions as signals of their political leanings, the analysis then investigated whether
these features indeed impacted viewersperceptions (RQ2). The inter-rater reliabil-
ity (IRR) of participantsratings was calculated based on intraclass correlation coef-
cients (see Kim, 2014, p. 167). The IRR for perceived slant was quite high (.87),
suggesting a high degree of agreement among participants. An ordinary least
squares regression used computer vision features to predict averaged perceived
slant, controlled for which candidate an image featured (Figure 4a). Given that
raters were randomly assigned to a large pool of images, it was unlikely that the
raterscharacteristics would be exactly the same across all the images. The model
thus controlled for the number of raters assigned to each image and the aggregated
characteristics of each images raters, including percentages of raters who were women
and White and means of ratersages, education levels, and political orientations.
Detected emotional neutrality highly correlated with happiness (r=.71, p<.001),
and was therefore removed from the model. All variance ination factors were below 3.
Among all the attributes, expressions of happiness had the largest eect size
(β=.48). Images with large face sizes (β=.19) and expressions of anger (β=
.15) were rated as negative portrayals of candidates, whereas skin health (β=.09;
all ps<.001) positively contributed to favorability. These results showed a large
13Journal of Communication 00 (2018) 122
Y. Peng Media Bias in Visual Portrayals of Presidential Candidates
Downloaded from by University of Pennsylvania Library user on 01 October 2018
overlap between visual features that dierentiated liberal and conservative media and
features that inuenced audience perceptions of favorability. Based on unstandardized
coecients, a completely happy face, a completely angry face, and a face with per-
fectly healthy skin would impact a pictures favorability by +.8, .7, and +.4, respec-
tively, on a 5-point scale. In addition, mouth openness (β=.10, p<.001) and other
faceshappiness (β=.07, p=.002) also slightly enhanced perceived favorability.
Next, this study looked at the dimensionality in the evaluation of traits from
photographs (RQ3a). The IRR in perceiving dierent traits ranged from .64 to .89,
indicating reasonably high agreement among participants, particularly for traits
related to warmth, competence, and attractiveness, but less for dominance and
morality (Table 3). A two-factor structure emerged from exploratory factor analy-
ses. Two terms related to dominance (dominant and assertive) loaded on one factor
(IRR =.68; α=.95; M
=3.25, SD =.45; M
=3.25, SD =.46). Warmth,
morality, competence, and attractiveness did not form distinct concepts, as some
prior research had implied; instead, these traits converged on another factor,
referred to as valence (IRR =.77; α=.98; M
=3.09, SD =.51; M
SD =.41; Table 3).
The analysis then looked at what visual features impacted audience perceptions
on these two factors (RQ3b). Interestingly, the valence dimension highly correlated
with perceived slant (r=.93, p<.001). Most features predicting perceived slant
Figure 3 Distribution of visual features across media outlets and candidates. Note:BB=
Breitbart; DK =Daily Kos; HP =The Hungton Post; NYT =The New York Times;UT=
USA Today; WP =The Washington Post;WSJ=The Wall Street Journal. Scales: happiness/
anger, 0100; skin health, 100100; face size, 0100%.
14 Journal of Communication 00 (2018) 122
Media Bias in Visual Portrayals of Presidential Candidates Y. Peng
Downloaded from by University of Pennsylvania Library user on 01 October 2018
also inuenced judgment on the valance dimension, including face size (β=.16),
happiness (β=.43), anger (β=.14), mouth openness (β=.10), skin health
(β=.08), and other faceshappiness (β=.07, all ps<.001). The number of other
faces also slightly increased valence (β=.04, p=.03; Figure 4a).
In contrast, the eects of visual attributes on dominance, which correlated with
perceived slant only to some degree (r=.23, p<.001), showed a dierent pattern.
A few attributes inuenced dominance in the same direction of predicting valence,
such as face size (β=.14, p<.001) and mouth openness (β=.09, p=.01). Yet
anger (β=.25) increased dominance, while happiness (β=.13, both ps<.001)
acted negatively: these measures were in the opposite directions of predicting
valence. Completely happy and angry faces would impact dominance by .1 and
Figure 4 Eects of visual features on perceived slant, valence, and dominance. Note:
Regarding other faceshappiness and eye openness, mean substitution was applied to photos
with only the candidatefaces. Candidate: 1 =Trump, 0 =Clinton. The three adjusted R
values are in the order of predicting perceived slant, valence, and dominance.
15Journal of Communication 00 (2018) 122
Y. Peng Media Bias in Visual Portrayals of Presidential Candidates
Downloaded from by University of Pennsylvania Library user on 01 October 2018
+.8 on a 5-point scale, respectively. Skin redness (β=.09, p=.01) and eye gaze
(β=.08, p=.02) also had small eects on dominance (Figure 4a).
In summary, this research advances our understanding of visual bias in the follow-
ing ways. First, by integrating the objective checklist and subjective rating
approaches developed in prior research, this study extends conventional analysis,
which often focuses on the magnitude and direction of bias, to understanding how
partisan media bias is constructed in visual portrayals of politicians and how these
portrayals then inuence audience interpretations. Regarding audience perceptions,
this research also extends visual favorability from a single positivenegative spec-
trum to a two-factor space that includes valence and dominance. Dierent visual
cues can exert both similar and reversed impacts on perceptions of these two
dimensions. And last, this research also demonstrates that the use of computer
vision tools greatly expands the scope of visual analysis and could better equip com-
munication scholars to study visual content, which is becoming increasingly ubiqui-
tous in our media environment.
Partisan bias in visual content
With a few exceptions, visual features that dierentiate liberal from conservative
media largely overlap features that impact viewersperceptions of slant, including
facial expressions, face size, and skin condition. Facial expressions, and particularly
happiness, play essential roles in shaping participantsperceptions of media slant
and impressions of politicians. Indeed, in face perception literature, perceived hap-
piness in faces and valance-related traits such as trustworthiness, intelligence, and
attractiveness often load on the same dimension in factor analysis (Oosterhof &
Todorov, 2008;Sutherland et al., 2013). In interviews, Lobinger and Brantner
(2015) also found that viewers heavily relied on politiciansfacial expressions to
judge the slant in news images. Therefore, computationally-detected happiness in
politiciansfaces could be a simplied but ecient proxy of visual slant. This mea-
sure might be particularly useful for information platforms that primarily circulate
visual data, such as Instagram and YouTube.
One genre of negative portrayals in media coverage might require further atten-
tion: images featuring candidates with a large face, occupying almost the entire
image, which highlights their skin aws and negative emotional expressions.
research in face-ism, which often compares full-body with half-body images, has
shown a positive inuence of facial prominence on impression formation (Archer
et al., 1983). The negative impact of face size found in this study somewhat contra-
dicts face-ism, but echoes what Mutz (2007) has termed as in-your-face politicsin
television discourse and Grabe and Bucys (2009) interpretation of extreme close-
ups. Close-up shots of politicians make them seem to be in the face of viewers and
create a sense of discomfort and uneasiness among viewers, which could intensify
16 Journal of Communication 00 (2018) 122
Media Bias in Visual Portrayals of Presidential Candidates Y. Peng
Downloaded from by University of Pennsylvania Library user on 01 October 2018
viewerspreexisting negative feelings towards them. Furthermore, in fashion and
advertising, media professionals can manipulate skin condition (e.g., remove wrin-
kles) to make models look more attractive in photos. In political coverage, media
outlets might also intentionally make a politician look unfavorable by highlighting
facial skin aws. We should also note that both candidates examined here are
White. The eects of skin condition might be more complicated if politicians of
other races are considered, which could be an arena for future research.
Audience perceptions of favorability
Echoing prior research, this study revealed the multiple dimensions people use to
evaluate politicians and highlights the need to study the eects of visual portrayals
on dierent trait perceptions (Grabe & Bucy, 2009). Conrming previous research
in face perception, a two-factor structure emerged from the data (Oosterhof &
Todorov, 2008). While perceived slant converges with the valence dimension, domi-
nance forms a separate dimension that should require further attention. These two
factors do not completely converge. For example, some pictures intended to make
candidates look bad by emphasizing their angry expressions and aggressive beha-
viors, like yelling and shouting, might simultaneously render them as more power-
ful and dominant.
Prior research has frequently shown the eects of trait perceptions on voting
preferences (Caprara & Zimbardo, 2004), but where do people get their impressions
of politicians? This study suggests that visual portrayals of politicians might be one
source of trait perceptions. In the results, across dierent media outlets, Trump
images expressed more anger and less happiness than Clinton, which should make
him look less favorable and friendly but more dominant and aggressive. Indeed, a
study conducted before the election showed that people considered Clinton as more
caring and competent and Trump as more dominant (Kakkar & Sivanathan, 2017).
Furthermore, trait perceptions on dierent dimensions might have distinct impacts
on candidate preferences. Voters might prefer more dominant, aggressive leaders to
more competent or caring ones when feeling threatened by outsiders or in uncertain
situations (Kakkar & Sivanathan, 2017). Future work could experimentally examine
how dominance in visual portrayals inuences viewersjudgment of candidates and
subsequent voting behaviors.
With the increasing visibility of female politicians in contemporary politics, it is
also important to further bridge partisan bias and gender bias research. Gender bias
might operate on multiple levels. First, media content might associate dierent
norms with female and male politicians. In the results, Trump was portrayed using
images with larger faces, less happiness, more anger, and fewer people around than
Clinton. This could result from an unfavorable treatment of Trump across media
outlets, as well as gender bias in face-ism and gender stereotypes that expect women
to be more friendly and sociable and less aggressive (Archer et al., 1983;Prentice &
Carranza, 2002). Moreover, the eects of visual cues on audience perceptions might
dier between female and male politicians. Additional ordinary least squares
17Journal of Communication 00 (2018) 122
Y. Peng Media Bias in Visual Portrayals of Presidential Candidates
Downloaded from by University of Pennsylvania Library user on 01 October 2018
regressions using photos of Clinton and Trump partially armed this possibility
(Figure 4b and c). For example, the positive impact of happiness on perceived slant
was more salient regarding Clintons images than Trumps, potentially reecting the
gender norm in facial expressions that it was more proper or rewarding for women
than men to smile (Plant et al., 2000). The negative eect of face size on dominance
was also more pronounced for Clinton than Trump. Nevertheless, neither Clinton
nor Trump represents all female or male politicians. Future research could include
multiple female and male politicians to better study the interplay between political
and gender biases.
The diversity in audience interpretations of visual content might also require
our attention (Lobinger & Brantner, 2016). As Figure 4b and c shows, photos of
candidates that matched (mismatched) viewerspolitical orientation were rated as
more positive (negative), independently from visual attributes. This pattern was
especially salient for Trump, a controversial gure who invited polarized responses
from participants. As one crowdsourced worker commented at the end of the rating
survey: It was very dicult to rank Trump anything but the lowest in honesty no
matter what the picture was.We also see that, compared with slant and valence,
viewersagreement on dominance was relatively low. This shows that crowdsourced
workers can reach an agreement regarding certain evaluations, but there might be
some inherent variability in peoples interpretations of visual dominance, which
requires further study. Future research could also study how other individual char-
acteristics, such as political knowledge and visual literacy, might aect how viewers
attend to and process visual portrayals of politicians.
Computer vision, limitations, and future research
This research demonstrates the potential of applying computer vision tools in ana-
lyzing large-scale visual media. However, several limitations exist. Current face
detection services only work well with near-frontal faces. If algorithms fail to detect
faces in a photo, this might already indicate an unfavorable portrayal of politicians
(e.g., blocking a politicians face); therefore, this studysndings should only apply
to images featuring identiable faces of politicians. In addition, the computer vision
services examined could not accurately identify facial expressions such as fear and
sadness. This is unfortunate, as dierent negative emotions often produce diering
eects regarding the evaluation of valence and dominance (Knutson, 1996).
Nevertheless, this low accuracy could partially be due to the consideration that
some expressions were not prevalent in the dataset and some were highly ambigu-
ous: raters themselves couldnt agree upon which face should count as fearful
(IRR =.50). Also, visual features such as politiciansgestures and activities might
also reect media bias, but could not be captured by the currently-available com-
puter vision services. As the eld of computer vision grows, future research could
also incorporate these attributes.
One limitation of this study is that the images were resized and decontextualized
from their original content. News stories associated with images may further
18 Journal of Communication 00 (2018) 122
Media Bias in Visual Portrayals of Presidential Candidates Y. Peng
Downloaded from by University of Pennsylvania Library user on 01 October 2018
inuence how photos of politicians are picked up by media practitioners and inter-
preted by audiences. The design components in news websites might also reect
media bias. For example, the size of a photo might reect the visibility a media out-
let intends to give to a politician and inuence readersallocation of attention (Barrett
& Barrington, 2005). Bringing tools of already widely-applied computational textual
analysis, the next step in this line of research might be multi-modal analyses that
investigate the interplay between textual and visual biases in media outlets. This
study also only looked at still images, which are only one component of visual polit-
ical communication. Future work might also apply computer vision techniques to
analyze moving images, such as online videos.
This work was supported by the Dissertation Research Fellowship at the Annenberg
School for Communication. The author thanks Sandra González-Bailón, Paul
Messaris, Jessa Lingel, Michael X. Delli Carpini, Sharrona Pearl, members of the
DiMeNet research group, and four anonymous reviewers for their feedback on ear-
lier versions of this article. The author also thanks Hyun Suk Kim for his help in
data analysis.
1 Although BBC is a British outlet, it attracts a sizable U.S. readership and has been
classied as left-leaning by prior research (Mitchell et al., 2014), so it was included in our
sample. Two strategies were used to address the concern about the potential bias in search
engine results: (a) this study predened a list of media sources and limited each search to
one source, so whether Google prioritized certain sources should have had a limited
impact on the results; and (b) all the images returned by Google were retrieved, so
whether Google ranked certain types of entries ahead of others should not have generated
substantial biases in the data.
2 The face set included the two candidates and their family members (e.g., Ivanka Trump),
running mates (e.g., Tim Kaine), competitors (e.g., Bernie Sanders), third-party
candidates (e.g., Gary Johnson), and other gures who frequently appeared in the
campaign (e.g., Barack Obama). For each person, the face set included about 10 images
that covered a diversity of facial expressions and luminance conditions. To validate results
from Face++, the study manually coded 400 images selected from all the retrieved images
on whether the image had visible faces of Clinton and Trump. The analysis tried a series
of thresholds (.60, .65, .70, , .95) as cut-opoints to accept whether the face should be
determined as Clinton or Trump. The overall accuracy rate was maximized when the
threshold of .75 was used both for identifying Clinton (accuracy =97.75%, false negative =
2%) and Trump (accuracy =97.25%, false negative =1.5%).
3 Participants reported their gender (female =50.0%), age (M=39.3, SD =12.6), race/
ethnicity (multiple answers allowed: Hispanic =7.6%; African American =8.6%; Asian =
8.1%; White =83.2%), education (high school or less =7.9%; some college =21.0%;
Associates degree =13.7%; Bachelors degree =37.4%; Masters degree or higher =
19Journal of Communication 00 (2018) 122
Y. Peng Media Bias in Visual Portrayals of Presidential Candidates
Downloaded from by University of Pennsylvania Library user on 01 October 2018
19.8%), political ideology (M=3.73, SD =1.78; 1 =extremely liberal,7=extremely
conservative), and party aliation (M=3.86, SD =2.13; 1 =strong Democrat,7=strong
Republican). The last two (r=.82) were combined into one political orientation scale. A
screening survey sampled a roughly equal number of women and men and of Democrats
(34.3%) and Republicans (33.8%).
4 To assess the accuracy in other computer vision features, on three-point scales (2 =in the
middle/cant decide), this study also manually coded a subset of 150 images on pitch (1 =
head bowing, 3 =raising; r=.60), roll (1 =head tilting right, 3 =left; r=.72), and yaw
(1 =facing right, 3 =left; r=.78) angles of the face; eye (1 =closed, 3 =open; r=.71)
and mouth openness (1 =closed, 3 =open; r=.67); eye gaze direction (1 =looking
elsewhere; 3 =nearly vertical to the image; r=.50); and skin health (1 =visible skin
aws, 3 =smooth and healthy; r=.48), which well correlated with computationally
calculated features.
5 Face size negatively correlated with skin health (r=.21) and happiness (r=.09),
implying that media combined these features to portray a candidate positively/negatively.
Using regressions, tests of interactions suggested that the impact of happiness (but not
skin health) on slant (β=.04, p=.048) and valence (β=.05, p=.012) was larger for
images with smaller faces. Given the potential mixed eect of face size, the quadratic term
of face size was also tested and was signicant in predicting slant (β=.09, p=.007) and
valence (β=.09, p=.009), implying that images with extremely large faces did not
further reduce favorability compared with only fairly large ones.
Abele, A. E., Hauke, N., Peters, K., Louvet, E., Szymkow, A., & Duan, Y. (2016). Facets of the
fundamental content dimensions: Agency with competence and assertiveness
communion with warmth and morality. Frontiers in Psychology, 7, 1810. doi:10.3389/
Alter, A. L., Stern, C., Granot, Y., & Balcetis, E. (2016). The bad is Blackeect: Why
people believe evildoers have darker skin than do-gooders. Personality and Social
Psychology Bulletin, 42(12), 16531665. doi:10.1177/0146167216669123
Archer, D., Iritani, B., Kimes, D. D., & Barrios, M. (1983). Face-ism: Five studies of sex
dierences in facial prominence. Journal of Personality and Social Psychology, 45(4),
725735. doi:10.1037//0022-3514.45.4.725
Barrett, A. W., & Barrington, L. W. (2005). Bias in newspaper photograph selection. Political
Research Quarterly, 58(4), 609618. doi:10.2307/3595646
Budak, C., Goel, S., & Rao, J. M. (2016). Fair and balanced? Quantifying media bias through
crowdsourced content analysis. Public Opinion Quarterly, 80(S1), 250271. doi:10.1093/
Caprara, G. V., & Zimbardo, P. G. (2004). Personalizing politics: A congruency model of
political preference. American Psychologist, 59(7), 581594. doi:10.1037/0003-
DAlessio, D., & Allen, M. (2000). Media bias in presidential elections: a metaanalysis.
Journal of Communication, 50(4), 133156. doi:10.1093/joc/50.4.133
20 Journal of Communication 00 (2018) 122
Media Bias in Visual Portrayals of Presidential Candidates Y. Peng
Downloaded from by University of Pennsylvania Library user on 01 October 2018
Dehghan, A., Ortiz, E. G., Shu, G., & Masood, S. Z. (2017). DAGER: Deep age, gender and
emotion recognition using convolutional neural network. Retrieved from https://arxiv.
org/abs/1702.04280 (date last accessed September 1, 2018)
Ekman, P., & Friesen, W. V. (2003). Unmasking the face: A guide to recognizing emotions
from facial clues. Los Altos, CA: Malor Books.
Entman, R. M. (2007). Framing bias: Media in the distribution of power. Journal of
Communication, 57(1), 163173 doi:10.1111/j.1460-2466.2006.00336.x
Fink, B., Grammer, K., & Matts, P. J. (2006). Visible skin color distribution plays a role in
the perception of age, attractiveness, and health in female faces. Evolution and Human
Behavior, 27(6), 433442. doi:10.1016/j.evolhumbehav.2006.08.007
Flaxman, S., Goel, S., & Rao, J. M. (2016). Filter bubbles, echo chambers, and online news
consumption. Public Opinion Quarterly, 80(S1), 298320. doi:10.1093/poq/nfw006
Grabe, M. E., & Bucy, E. P. (2009). Image bite politics: News and the visual framing of
elections. Oxford, England: Oxford University Press.
Groeling, T. (2013). Media bias by the numbers: Challenges and opportunities in the
empirical study of partisan news. Annual Review of Political Science, 16(1), 129151. doi:
Hehman, E., Graber, E. C., Homan, L. H., & Gaertner, S. L. (2012). Warmth and
competence: A content analysis of photographs depicting American presidents.
Psychology of Popular Media Culture, 1(1), 4652. doi:10.1037/a0026513
Jaccard, J., & Turrisi, R. (2003). Interaction eects in multiple regression. Thousand Oaks,
CA: SAGE Publications.
Jones, B. C., Little, A. C., Burt, D. M., & Perrett, D. I. (2004). When facial attractiveness is
only skin deep. Perception, 33(5), 569576. doi:10.1068/p3463
Kakkar, H., & Sivanathan, N. (2017). When the appeal of a dominant leader is greater than a
prestige leader. Proceedings of the National Academy of Sciences, 114(26), 67346739.
Kim, H. S. (2014). Attractability and virality: The role of message features and social
inuence in health news diusion. Retrieved from Publicly Accessible Penn Dissertations.
Knutson, B. (1996). Facial expressions of emotion inuence interpersonal trait inferences.
Journal of Nonverbal Behavior, 20(3), 165182. doi:10.1007/bf02281954
Lobinger, K., & Brantner, C. (2015). Likable, funny or ridiculous? A Q-sort study on
audience perceptions of visual portrayals of politicians. Visual Communication, 14(1),
1540. doi:10.1177/1470357214554888
Lobinger, K., & Brantner, C. (2016). Dierent ways of seeing political depictions: A
qualitativequantitative analysis using Q methodology. Communications, 41(1), 4769.
Mitchell, A., Gottfried, J., Kiley, J., & Matsa, K. E. (2014). Political polarization & media
habits. Retrieved from
media-habits/ (date last accessed September 1, 2018)
Moriarty, S. E., & Popovich, M. N. (1991). Newsmagazine visuals and the 1988 presidential
election. Journalism & Mass Communication Quarterly, 68(3), 371380. doi:10.1177/
21Journal of Communication 00 (2018) 122
Y. Peng Media Bias in Visual Portrayals of Presidential Candidates
Downloaded from by University of Pennsylvania Library user on 01 October 2018
Mutz, D. C. (2007). Eects of in-your-facetelevision discourse on perceptions of a
legitimate opposition. American Political Science Review, 101(4), 621635. doi:10.1017/
Oosterhof, N. N., & Todorov, A. (2008). The functional basis of face evaluation. Proceedings
of the National Academy of Sciences, 105(32), 1108711092. doi:10.1073/
Park, S., Kang, S., Chung, S., & Song, J. (2009). NewsCube: delivering multiple aspects of
news to mitigate media bias. In Proceedings of the SIGCHI Conference on Human Factors
in Computing Systems (pp. 443452). New York, NY: ACM. doi:10.1145/
Peng, Y., & Jemmott, J. B., III (2018). Feast for the eyes: Eects of food perceptions and
computer vision features on food photo popularity. International Journal of
Communication, 12, 313336.
Plant, E. A., Hyde, J. S., Keltner, D., & Devine, P. G. (2000). The gender stereotyping of
emotions. Psychology of Women Quarterly, 24(1), 8192. doi:10.1111/j.1471-6402.2000.
Prentice, D. A., & Carranza, E. (2002). What women and men should be, shouldnt be, are
allowed to be, and dont have to be: The contents of prescriptive gender stereotypes.
Psychology of Women Quarterly, 26(4), 269281. doi:10.1111/1471-6402.t01-1-00066
Ronquillo, J., Denson, T. F., Lickel, B., Lu, Z. L., Nandy, A., & Maddox, K. B. (2007). The
eects of skin tone on race-related amygdala activity: An fMRI investigation. Social
Cognitive and Aective Neuroscience, 2(1), 3944. doi:10.1093/scan/nsl043
Stephen, I. D., Oldham, F. H., Perrett, D. I., & Barton, R. A. (2012). Redness enhances
perceived aggression, dominance and attractiveness in mens faces. Evolutionary
Psychology, 10(3), 562572. doi:10.1177/147470491201000312
Stephen, I. D., Smith, M. J. L., Stirrat, M. R., & Perrett, D. I. (2009). Facial skin coloration
aects perceived health of human faces. International Journal of Primatology, 30(6),
845857. doi:10.1007/s10764-009-9380-z
Sutherland, C. A., Oldmeadow, J. A., Santos, I. M., Towler, J., Burt, D. M., & Young, A. W.
(2013). Social inferences from faces: Ambient images generate a three-dimensional
model. Cognition, 127(1), 105118. doi:10.1016/j.cognition.2012.12.001
Szeliski, R. (2010). Computer vision: Algorithms and applications. Berlin, Germany: Springer.
Talamas, S. N., Mavor, K. I., Axelsson, J., Sundelin, T., & Perrett, D. I. (2016). Eyelid-
openness and mouth curvature inuence perceived intelligence beyond attractiveness.
Journal of Experimental Psychology: General, 145(5), 603620. doi:10.1037/xge0000152
Tiedens, L. Z. (2001). Anger and advancement versus sadness and subjugation: the eect of
negative emotion expressions on social status conferral. Journal of Personality and Social
Psychology, 80(1), 8694. doi:10.1037/0022-3514.80.1.86
Verser, R., & Wicks, R. H. (2006). Managing voter impressions: The use of images on
presidential candidate web sites during the 2000 campaign. Journal of Communication,
56(1), 178197. doi:10.1111/j.1460-2466.2006.00009.x
Waldman, P., & Devitt, J. (1998). Newspaper photographs and the 1996 presidential
election: The question of bias. Journalism & Mass Communication Quarterly, 75(2),
302311. doi:10.1177/107769909807500206
Walker, D., & Vul, E. (2014). Hierarchical encoding makes individuals in a group seem
more attractive. Psychological Science, 25(1), 230235. doi:10.1177/0956797613497969
22 Journal of Communication 00 (2018) 122
Media Bias in Visual Portrayals of Presidential Candidates Y. Peng
Downloaded from by University of Pennsylvania Library user on 01 October 2018
... On the contrary, showing sadness, anger, disgust, or fear is usually connected to negative emotions (Peng, 2018); however, anger or disgust may also give an impression of dominance and powerfulness. ...
... Following Kress and van Leeuwen (1996), the smile and direct look generated a more positive perception because they activated the interpersonal function and reduced the social distance between the represented actor and the potential reader-voter. Furthermore, positive character traits, such as good leadership or trust, may be inferred from a smiling face (Kilgo et al., 2018;Peng, 2018). However, the difference obtained in the chi-square test was not significant. ...
... In the second question, participants were asked to rank candidates' credibility on a scale from 1 to 5. Table 4 shows the results of the answers grouped in three categories: low credibility (ratings of 1 and 2), neutrality, and high credibility (ratings of 4 and 5). (Kilgo et al., 2018;Knutson, 1996;Oosterhof & Todorov, 2008;Peng, 2018;Sutherland et al., 2013), the smiling expression or the direct look does not seem to lower the value of a positive character trait such as credibility. ...
Campaign posters are semiotic-discursive resources that form multimodal units of meaning. In political communication, voters' decision making is affected not only by the verbal message, but also by nonverbal indications or physical features (visual attributes) of the candidates. This study analyzed the relationship between visual communication and citizens' voting decisions in a political campaign in Colombia. An analysis of campaign posters in the 2019 Bogotá mayoral elections was designed using a multi-technique methodology. First, two eye-tracking experiments were conducted to assess attention patterns. Then, a series of surveys measured emotions in slogans. It was concluded that, in the Colombian scene, visual elements related to candidates' physical attributes have a small influence over voters' decision making. This finding contradicts the results of studies carried out in different contexts, namely in Europe and the United States.
... An emerging line of social science scholarship uses automated image analysis to answer questions relevant to social scientific research. For example, image analysis has been used for detecting partisan media bias in news photographs (Peng 2018), predicting poverty from satellite images (Jean et al. 2016), and detecting protests and estimating their size from social media images (Sobolev et al. 2020;Zhang and Pan 2019). The community of computational social science has also provided reviews and tutorials that explain the task of image classification (Joo and Steinert-Threlkeld 2018;Torres and Cantu 2021;Williams, Casas, and Wilkerson 2020). ...
Automated image analysis has received increasing attention in social scientific research, yet existing scholarship has mostly covered the application of supervised learning to classify images into predefined categories. This study focuses on the task of unsupervised image clustering, which aims to automatically discover categories from unlabelled image data. We first review the steps to perform image clustering and then focus on one key challenge in this task—finding intermediate representations of images. We present several methods of extracting intermediate image representations, including the bag-of-visual-words model, self-supervised learning, and transfer learning (in particular, feature extraction with pretrained models). We compare these methods using various visual datasets, including images related to protests in China from Weibo, images about climate change on Instagram, and profile images of the Russian Internet Research Agency on Twitter. In addition, we propose a systematic way to interpret and validate clustering solutions. Results show that transfer learning significantly outperforms the other methods. The dataset used in the pretrained model critically determines what categories the algorithms can discover.
... Although facial recognition is now regularly employed in commercial settings, few studies have used these tools for visual-based content analyses. Furthermore, most existing studies have been published outside of communication (e.g., Zhu et al., 2013), and only very few have examined video content specifically (e.g., Joo et al., 2018;Guha et al., 2015), focusing instead on face recognition from photographs (e.g., Peng, 2018). This seems particularly notable since a handful of research advancing facial recognition and other computer vision tools have used the same type of video content that mass communication scholars are interested in studying (i.e., movies and television) as training sets for large-scale machine-learning models (e.g., Nagrani & Zisserman, 2017;Patron-Perez et al., 2010). ...
... In these and in numerous later studies in the same vein, researchers suggested that those expressions were latent attempts of hidden persuasion. More recent studies analyzed the appearance and nonverbal behavior of presidential candidates (Banning & Coleman, 2009;Peng, 2018) and of other politicians in real-life and laboratory situations (Haumer & Donsbach, 2009;Kilgo et al., 2018). Olivola and Todorov (2010) reported that facial appearance-based trait inferences following rapid exposure can predict electoral success. ...
Full-text available
The Media Bias Effect (MBE) represents the biasing influence of the nonverbal behavior of a TV interviewer on viewers’ impressions of the interviewee. In the MBE experiment, participants view a 4-min made-up political interview in which they are exposed only to the nonverbal behavior of the actors. The interviewer is friendly toward the politician in one experimental condition and hostile in the other. The interviewee was a confederate filmed in the same studio, and his clips are identical in the two conditions. This experiment was used successfully in a series of studies in several countries (Babad and Peer in J Nonverbal Behav 34(1):57–78, 2010. and was administered in the present research. The present investigation focused on the interviewer's source credibility and its persuasive influence. The viewers filled out questionnaires about their impressions of both the interviewer and the interviewee. A component of "interviewer's authority" was derived in PCA, with substantial variance in viewers' impressions of the interviewer. In our design, we preferred the conception of Epistemic Authority (Kruglanski et al. in Adv Exp Soc Psychol 37:345–392, 2005)—based on viewers' subjective perceptions for deriving authority status—to the conventional design of source credibility studies, where dimensions of authority are pre-determined as independent variables. The results demonstrated that viewers who perceived the interviewer as an effective leader demonstrated a clear MBE and were susceptible to his influencing bias, but no bias effect was found for viewers who did not perceive the interviewer as an effective leader. Thus, Epistemic Authority (source credibility) moderated the Media Bias Effect.
... of user-related variables on the perceptions of bias [18], [19], [33], [34], [35], or the perception of bias in particular topics [20]. Other main interests in the existing research are topicdependent text perception [36], user comments [21], [37], and visual features [38]. ...
Full-text available
Media coverage has a substantial effect on the public perception of events. Nevertheless, media outlets are often biased. One way to bias news articles is by altering the word choice. The automatic identification of bias by word choice is challenging, primarily due to the lack of gold-standard data sets and high context dependencies. In this research project, I aim to devise data sets and methods to identify media bias. To achieve this, I plan to research methods using natural language processing and deep learning while employing models and using analysis concepts from psychology and linguistics. The first results indicate the effectiveness of an interdisciplinary research approach. My vision is to devise a system that helps news readers become aware of media coverage differences caused by bias. So far, my best performing BERT-based model is pre-trained on a larger corpus consisting of distant labels, indicating that distant supervision has the potential to become a solution for the difficult task of bias detection.
... of user-related variables on the perceptions of bias [18], [19], [33], [34], [35], or the perception of bias in particular topics [20]. Other main interests in the existing research are topicdependent text perception [36], user comments [21], [37], and visual features [38]. ...
Conference Paper
Full-text available
Media coverage has a substantial effect on the public perception of events. Nevertheless, media outlets are often biased. One way to bias news articles is by altering the word choice. The automatic identification of bias by word choice is challenging, primarily due to the lack of gold-standard data sets and high context dependencies. In this research project, I aim to devise data sets and methods to identify media bias. To achieve this, I plan to research methods using natural language processing and deep learning while employing models and using analysis concepts from psychology and linguistics. The first results indicate the effectiveness of an interdisciplinary research approach. My vision is to devise a system that helps news readers become aware of media coverage differences caused by bias. So far, my best performing BERT-based model is pre-trained on a larger corpus consisting of distant labels, indicating that distant supervision has the potential to become a solution for the difficult task of bias detection.
... It must be "volitional", "putatively influential", "be reasonably, plausibly threatening to conventional values or institutions" and "sustained", according to the classical definition by Williams [2]. Substantial resources have been invested recently to refine the theoretical definition of media bias and devise algorithms to detect it in practice [3,4,5,6,7] (see [1] for a review). Media bias can manipulate and skew the opinions of consumers of media products and trigger a change in voting behavior [8,3]. ...
Perceptions of political bias in the media are formed directly, through the independent consumption of the published outputs of a media organization, and indirectly, through observing the collective responses of political allies and opponents to the same published outputs. A network of Bayesian learners is constructed to model this system, in which the bias perceived by each agent obeys a probability density function, which is updated according to Bayes's theorem given data about the published outputs and the beliefs of the agent's political allies and opponents. The Bayesian framework allows for uncertain beliefs, multimodal probability distribution functions, and antagonistic interactions with opponents, not just cooperation with allies. Numerical simulations are performed to test the idealized example of inferring the bias of a coin. It is found that some agents converge on the wrong conclusion faster than others converge on the right conclusion under a surprisingly broad range of conditions, when antagonistic interactions are present which "lock out" some agents from the truth, e.g. in Barab\'asi-Albert networks. It is also found that structurally unbalanced networks routinely experience turbulent nonconvergence, where some agents fail to achieve a steady-state belief, e.g. when they are allies of two agents who are opponents themselves. The subtle phenomenon of long-term intermittency is also explored.
On July 17, 2021, the CDU's chancellor candidate Armin Laschet was photographed laughing during a speech by the German Federal President in the flood-stricken city of Erftstadt. The photographic images caused an uproar and contributed to the CDU's defeat in the September 23 election. The paper analyzes why these images resonated with such damaging effects. Theoretically, it sets the analysis on the background of the moralization and personalization of politics and argues that photography, with its ability to capture behavior at a distance, plays a prominent role in these processes. While this condition explains why an image of a laughing politician can generate such indignation in the first place, the paper discusses how this effect was amplified in the case of Laschet by a range of contextual features: (a) the timing of the images in the middle of an election period where politicians come under intense scrutiny; (b) their appearance in a crisis situation (the German flooding disaster) where politicians are surrounded by other role expectations than in routine periods; (3) Laschet's new, insecure position as leader of the CDU; (d) his history of scandals and poor political judgment; and (e) the frivolous and boisterous manner of his laughter. At a general theoretical level, the paper's insights caution us to avoid prima facie conclusions about the autonomous power of photographs. Instead, they encourage analytical sensitivity to the importance of timing and context as explanatory elements in our understanding of photographic exposés.
The objective of this study was to explore the potential of the opinion mining methodology in sociopolitical research, its techniques, and applications from the field of reflexivity. It is a non-experimental and exploratory study. It concludes that advances in the field of artificial intelligence have provided the sociopolitical sciences with tools that make it possible to approach the dominant trends of opinion in society during specific junctures where decision-making and/or the positioning of an idea, public policy, or social project can be measured in real-time, with broad demographic scopes that can be segmented, bringing the researcher closer to the subject of study with minimum levels of bias. Opinion mining research continues to be dominated by electoral and marketing topics; however, there are potentialities in the research of public policies, social programs, democracy, and governance that are still waiting for the application of opinion mining as a sociopolitical research methodology.
Full-text available
The widely circulated food photos online have become an important part of our visual culture. Combining human ratings of food characteristics and computational analysis of visual aesthetics, we examined what contributed to the aesthetic appeal of a diversity of food photographs (N = 300) and likes and comments they received in an artificial newsfeed from participants (N = 399). The results revealed that people tended to like and share images containing tasty foods. Both healthy and unhealthy foods were able to gain likes. Aesthetic appeal and specific visual features, such as the use of arousing colors and different components of visual complexity, also influenced the popularity of food images. This work demonstrates the potential of applying computer vision methods in visual analysis, offers insights into image virality, and provides practical guidelines for communicating healthy eating.
Full-text available
This paper describes the details of Sighthound's fully automated age, gender and emotion recognition system. The backbone of our system consists of several deep convolutional neural networks that are not only computationally inexpensive, but also provide state-of-the-art results on several competitive benchmarks. To power our novel deep networks, we collected large labeled datasets through a semi-supervised pipeline to reduce the annotation effort/time. We tested our system on several public benchmarks and report outstanding results. Our age, gender and emotion recognition models are available to developers through the Sighthound Cloud API at
Full-text available
Agency (A) and communion (C) are fundamental content dimensions. We propose a facet-model that differentiates A into assertiveness (AA) and competence (AC) and C into warmth (CW) and morality (CM). We tested the model in a cross-cultural study by comparing data from Asia, Australia, Europe, and the USA (overall N = 1.808). Exploratory and confirmatory factor analyses supported our model. Both the two-factor model and the four-factor model showed good fit indices across countries. Participants answered additional measures intended to demonstrate the fruitfulness of distinguishing the facets. The findings support the model's construct validity by positioning the fundamental dimensions and their facets within a network of self-construal, values, impression management, and the Big Five personality factors: In all countries, A was related to independent self-construal and to agentic values, C was related to interdependent self-construal and to communal values. Regarding the facets, AA was always related to A values, but the association of AC with A values fell below our effect size criterion in four of the five countries. A (both AA and AC) was related to agentic impression management. However, C (both CW and CM) was neither related to communal nor to agentic impression management. Regarding the Big Five personality factors, A was related to emotional stability, to extraversion, and to conscientiousness. C was related to agreeableness and to extraversion. AA was more strongly related to emotional stability and extraversion than AC. CW was more strongly related to extraversion and agreeableness than CM. We could also show that self-esteem was more related to AA than AC; and that it was related to CM, but not to CW. Our research shows that (a) the fundamental dimensions of A and C are stable across cultures; and (b) that the here proposed distinction of facets of A and C is fruitful in analyzing self-perception. The here proposed measure, the AC-IN, may be a useful tool in this research area. Applications of the facet model in social perception research are discussed.
Across the globe we witness the rise of populist authoritarian leaders who are overbearing in their narrative, aggressive in behavior, and often exhibit questionable moral character. Drawing on evolutionary theory of leadership emergence, in which dominance and prestige are seen as dual routes to leadership, we provide a situational and psychological account for when and why dominant leaders are preferred over other respected and admired candidates. We test our hypothesis using three studies, encompassing more than 140,000 participants, across 69 countries and spanning the past two decades. We find robust support for our hypothesis that under a situational threat of economic uncertainty (as exemplified by the poverty rate, the housing vacancy rate, and the unemployment rate) people escalate their support for dominant leaders. Further, we find that this phenomenon is mediated by participants’ psychological sense of a lack of personal control. Together, these results provide large-scale, globally representative evidence for the structural and psychological antecedents that increase the preference for dominant leaders over their prestigious counterparts.
Previous research has shown that visual images of political candidates can influence voter perceptions. This study examines newspaper photographs of candidates to determine whether the favorableness of these pictures is related to the “political atmosphere” of individual newspapers. In particular, we examine 435 candidate photographs from several races covered by seven newspapers during the 1998 and 2002 general election seasons. Based on our analysis, we conclude that candidates endorsed by a particular newspaper—or whose political leanings match the political atmosphere of a given paper—generally have more favorable photographs of them published than their opponents.
Across six studies, people used a “bad is black” heuristic in social judgment and assumed that immoral acts were committed by people with darker skin tones, regardless of the racial background of those immoral actors. In archival studies of news articles written about Black and White celebrities in popular culture magazines (Study 1a) and American politicians (Study 1b), the more critical rather than complimentary the stories, the darker the skin tone of the photographs printed with the article. In the remaining four studies, participants associated immoral acts with darker skinned people when examining surveillance footage (Studies 2 and 4), and when matching headshots to good and bad actions (Studies 3 and 5). We additionally found that both race-based (Studies 2, 3, and 5) and shade-based (Studies 4 and 5) associations between badness and darkness determine whether people demonstrate the “bad is black” effect. We discuss implications for social perception and eyewitness identification.
What makes health news articles attractable and viral? Why do some articles diffuse widely by prompting audience selections (attractability) and subsequent social retransmissions (virality), while others do not? Identifying what drives social epidemics of health news coverage is crucial to our understanding of its impact on the public, especially in the emerging media environment where news consumption has become increasingly selective and social. This dissertation examines how message features and social influence affect the volume and persistence of attractability and virality within the context of the online diffusion of New York Times (NYT) health news articles. The dissertation analyzes (1) behavioral data of audience selections and retransmissions of the NYT articles and (2) associated article content and context data that are collected using computational social science approaches (automated data mining; computer-assisted content analysis) along with more traditional methods (manual content analysis; message evaluation survey). Analyses of message effects on the total volume of attractability and virality show that articles with high informational utility and positive sentiment invite more frequent selections and retransmissions, and that articles are also more attractable when presenting controversial, emotionally evocative, and familiar content. Furthermore, these analyses reveal that informational utility and novelty have stronger positive associations with email-specific virality, while emotion-related message features, content familiarity, and exemplification play a larger role in triggering social media-based retransmissions. Temporal dynamics analyses demonstrate social influence-driven cumulative advantage effects, such that articles which stay on popular-news lists longer invite more frequent subsequent selections and retransmissions. These analyses further show that the social influence effects are stronger for articles containing message features found to enhance the total volume of attractability and virality. This suggests that those synergistic interactions might underlie the observed message effects on total selections and retransmissions. Exploratory analyses reveal that the effects of social influence and message features tend to be similar for both (1) the volume of audience news selections and retransmissions and (2) the persistence of those behaviors. However, some message features, such as expressed emotionality, are relatively unique predictors of persistence outcomes. Results are discussed in light of their implications for communication research and practice.
It is widely thought that news organizations exhibit ideological bias, but rigorously quantifying such slant has proven methodologically challenging. Through a combination of machine-learning and crowdsourcing techniques, we investigate the selection and framing of political issues in fifteen major US news outlets. Starting with 803,146 news stories published over twelve months, we first used supervised learning algorithms to identify the 14 percent of articles pertaining to political events. We then recruited 749 online human judges to classify a random subset of 10,502 of these political articles according to topic and ideological position. Our analysis yields an ideological ordering of outlets consistent with prior work. However, news outlets are considerably more similar than generally believed. Specifically, with the exception of political scandals, major news organizations present topics in a largely nonpartisan manner, casting neither Democrats nor Republicans in a particularly favorable or unfavorable light. Moreover, again with the exception of political scandals, little evidence exists of systematic differences in story selection, with all major news outlets covering a wide variety of topics with frequency largely unrelated to the outlet’s ideological position. Finally, news organizations express their ideological bias not by directly advocating for a preferred political party, but rather by disproportionately criticizing one side, a convention that further moderates overall differences.
Online publishing, social networks, and web search have dramatically lowered the costs of producing, distributing, and discovering news articles. Some scholars argue that such technological changes increase exposure to diverse perspectives, while others worry that they increase ideological segregation. We address the issue by examining web-browsing histories for 50,000 US-located users who regularly read online news. We find that social networks and search engines are associated with an increase in the mean ideological distance between individuals. However, somewhat counterintuitively, these same channels also are associated with an increase in an individual’s exposure to material from his or her less preferred side of the political spectrum. Finally, the vast majority of online news consumption is accounted for by individuals simply visiting the home pages of their favorite, typically mainstream, news outlets, tempering the consequences—both positive and negative—of recent technological changes. We thus uncover evidence for both sides of the debate, while also finding that the magnitude of the effects is relatively modest.