Science Exemplars in the Eye of the Beholder

Exemplification, Selective Exposure, and Attitude Impacts

The often abstract nature of scientific information may deter many lay recipients, who are generally thought to find personalization in media coverage appealing (e.g., Bennett, 2009). The present study draws on exemplification theory (Zillmann, 1999), which conceptualizes the use of vivid, concrete exemplars (case illustrations) in contrast to the use of perceptually pallid statistics (statistical, baserate information in numeric formats) in media messages. Exemplars “describe causes, importance, and consequences of the problem under consideration from the unique perspective of an individual” (Brosius & Bathelt, 1994, p. 48), which may be a more accessible, intuitive format to present science information to lay audiences. Empirical research has yielded significant effects of inclusion of exemplars in news—such news affects opinions on controversial topics more strongly (Perry & Gonzenbach, 1997), yields more narrative engagement (Kim, Bigman, Leader, Lerman, & Cappella, 2012), and affects issue perceptions as well as behavior beyond short-term effects, for up to 2 weeks (Gibson & Zillmann, 1994; Knobloch-Westerwick & Sarge, 2015; Zillmann, Gibson, Sundar, & Perkins, 1996).

The outlined research implies that exemplars in science information could make such messages more attractive and foster more exposure to science information. At least in the context of health news, exemplars have yielded a positive effect on exposure when compared with statistics-based news (Hastall & Knobloch-Westerwick, 2013; Knobloch-Westerwick & Sarge, 2015). On the other hand, however, science information with statistics may convey more “scientificness” (Thomm & Bromme, 2012) and thus attract more exposure, despite less intuitive appeal. Based on exemplification theory (Zillmann, 1999) and related empirical work (Hastall & Knobloch-Westerwick, 2013; Knobloch-Westerwick & Sarge, 2015), the following hypothesis will be tested:

As it is possible that different recipients approach science information very specifically, and because science news often features much numeric information (Griffin, 1999) as opposed to the otherwise often case-illustrated news and media messages (Bennett, 2009; Zillmann, 1999), it is proposed that information consumers’ characteristics will affect the impact put forward in Hypothesis 1. Drawing further on exemplification research (Gibson, Callison, & Zillmann, 2011) as well as scholarship on numeracy (Peters, 2012), the relative attraction to exemplars or statistics in science information should depend on recipients’ numeracy and trait empathy. Numeracy involves greater attentiveness to quantitative versus qualitative information (Zillmann, Callison, & Gibson, 2009), whereas trait empathy (Davis, 1980) should facilitate perceived emotional connection with portrayed individuals that serve as exemplars.

Given earlier indications of effects of exemplification on perceptions, attitudes, and behavior (Gibson & Zillmann, 1994; Knobloch-Westerwick & Sarge, 2015; Perry & Gonzenbach, 1997; Zillmann et al., 1996), a research question on the role of exemplification for persuasive effects will be examined.

Online Science Information

As Internet users commonly access science messages online (e.g., Harrigan, 2006), the present study will examine science information use in the online context. Furthermore, scholars have called for research to address web-based seeking of science information (e.g., M. C. Nisbet & Goidel, 2007; see also Becker, Dalrymple, Brossard, Scheufele, & Gunther, 2010; Weeks, Friedenberg, Southwell, & Slater, 2012). The online information search context differs from exposure to science information in traditional media, in that users will typically encounter several messages on a topic lined up together—on the same online search results list, contradictory scientific information may appear. This circumstance affects the information selection and processing in important ways.

Experimental evidence has yielded that readers of science information are more uncertain regarding the issue if a single presented science message presents conflicting information on that matter, as opposed to a single message featuring only consensus (Dixon & Clarke, 2013). Yet other similar work did not find consistent effects on certainty across different science topics (Jensen & Hurley, 2012). Experimental work on how lay recipients respond to conflicting science information from several web messages has yielded that those messages that are deemed more plausible are more likely to be recalled and integrated into a broader understanding of the topic at hand (Maier & Richter, 2013), similar to a confirmation bias of selective exposure, processing, and recall (e.g., DeFleur & Ball-Rokeach, 1989). Several studies (e.g., Bråten & Strømsø, 2006; Stadtler, Scharrer, Brummernhenrich, & Bromme, 2013) demonstrated that lay recipients of science information notice and integrate conflicting evidence to a greater extent if they read several messages, as opposed to reading the same text presented in one message. Hence, the online presentation of conflicting science messages might instigate more elaboration and subsequent attitude impacts than messages in traditional formats. But will online users indeed attend to conflicting and counterattitudinal messages if they are free to select?

The Role of Values and Attitudes in Science Information Use

Following the perspective of the science literacy model (as explained by Miller, 1998), a higher level of science knowledge in the general population is thought to foster more favorable attitudes. However, the suggested relationship between knowledge and attitudes is relatively weak in empirical studies, as a large cross-national meta-analysis found (Allum, Sturgis, Tabourazi, & Brunton-Smith, 2008). Hence, more recent work emphasizes general values and specific attitudes held by lay recipients of science information, as those may greatly affect the information intake and processing and then, in turn, knowledge and support for science (Ho, Brossard, & Scheufele, 2008; E. C. Nisbet, Hart, Myers, & Ellithorpe, 2013; M. C. Nisbet & Goidel, 2007). Citizens’ values and attitudes appear to channel their involvement with science information—“Faced with a daily torrent of news, citizens use their value predispositions (such as political or religious beliefs) as perceptual screens, selecting news outlets and Web sites whose outlooks match their own” (M. C. Nisbet & Mooney, 2007, p. 56). Such bias from preexisting attitudes and values on what science information is selected for actual consumption has important implications because it fosters polarization in society regarding risks and scientific innovation (e.g., Kahan, 2012).

Unfortunately, the existing research on biased processing of science information cannot shed light on some of the most crucial processes involved. Specifically, with much of the related research relying on cross-sectional data (e.g., Besley & Shanahan, 2005; Brewer & Ley, 2011; Feldman, Maibach, Roser-Renouf, & Leiserowitz, 2012; Zhao, 2009), it is difficult to disentangle whether the attitudes were shaped by media exposure or whether the attitudes led individuals to select certain science messages from the media. In the realm of exposure to political messages, numerous studies have recently corroborated what Festinger’s (1957) theory of cognitive dissonance already suggested nearly six decades ago—individuals favor messages that align with preexisting views (e.g., Garrett, 2009; Knobloch-Westerwick & Meng, 2009; Taber & Lodge, 2006). With regard to science information, only one study has, to our knowledge, examined selective exposure. Jang (2014) largely adopted the research design and procedure that Knobloch-Westerwick and Meng (2009) had applied to study exposure to political messages in the context of an online magazine but employed messages that pertained to controversial science topics (stem cell research, genetically modified foods, global warming, and evolution). Jang (2014) found that, for two of the four topics (stem cell and GM foods), participants were more likely to click on attitude-discrepant science messages and also spent more time on them, compared with attitude-consistent messages (however, global warming and evolution did not yield any significant choice patterns). This finding appears to contradict the majority of the above-mentioned findings from political communication research, which calls for further tests of the matter for the science communication domain. Jang’s findings also highlight the necessity of accounting for relevant individual differences and their effects on exposure, because individuals high on religiosity and high on political knowledge did exhibit an attitude-consistent bias in exposure (Jang, 2014). These results were interpreted as demonstrating that individuals high on certainty regarding scientific issues were most likely to engage in confirmation bias, while those who were less certain (either less knowledgeable or less religious) sought out novel information from attitude-discrepant science articles. Finally, mixed patterns of results by topic may relate to beliefs about science and technological risk, which are influenced by cultural values such as individualism and egalitarianism (cf. Kahan et al., 2011). Thus, it remains important to continue to examine and account for differences and similarities between topics in their patterns of message exposure and effects.

Given that research on political messages has consistently found a confirmation bias in exposure and only one selective exposure study on science messages found an effect in the opposite direction, the present study tests for the following hypothesis on this matter:

Additionally, the previous research by Jang (2014) did not consider differences in how news articles were framed and presented, and it is unclear if issue stances may have been confounded with information presentation. Specifically, the present investigation examines exemplars and numerics and relevant ways of presenting science information. As exemplars are thought to attract greater exposure (see Hypothesis 1), it is worth exploring in a research question whether this exemplification effect might even override the confirmation bias suggested in Hypothesis 4.

Numeracy has further been suggested to affect the extent to which individuals engage in motivated reasoning regarding science information to bolster existing views; yet theoretical predictions and empirical evidence have been inconsistent (Kahan et al., 2012). Thus, a research question will examine whether numeracy moderates the confirmation bias in exposure suggested in Hypothesis 4.

Processing and Attitude Impacts of Science Information

Recent work has examined how science information exposure shapes opinions regarding science topics and what role factual evidence plays in that process: For instance, Druckman and Bolsen’s (2011) findings revealed limited influence of factual information on initial opinions, which did not exceed the influence of preexisting values and science credibility perceptions. Furthermore, supplementing message frames with factual information did not exert more impact on opinion formation than frames without facts. Moreover, Druckman and Bolsen (2011) concluded that recipients process additional factual information in a biased, opinion-consistent fashion once they have formed an initial opinion; evidence is viewed as more compelling if it aligns with preexisting opinions, and impartial facts are likely perceived as corroborating preexisting opinions. Similarly, an experiment found that preexisting beliefs (but not topic familiarity) predicted reinforcing effects of information exposure regarding nanotechnology (Kahan, Braman, Slovic, Gastil, & Cohen, 2009).

However, the existing work on attitudinal impact of science information did not enable recipients to select and sample from science messages, as they would in a regular online use setting. Thus, the present study examines how exposure to science information influences attitudes regarding the science topics. Although much work in the political communication context (e.g., Taber & Lodge, 2006) suggests that media users not only prefer attitude-consistent messages over attitude-discrepant messages (e.g., Knobloch-Westerwick & Meng, 2009) but also process information in attitude-bolstering fashion, evidence on how this exposure to science information affects attitudes is lacking. Specifically, Jang’s (2014) findings imply that recipients of science information might be more open to selecting attitude-challenging content but did not capture impacts on attitudes. Based on the perspective that information recipients engage in motivated reasoning and process information such that existing attitudes are reinforced (e.g., Taber & Lodge, 2006), we hypothesize only an attitude-reinforcing effect of exposure to attitude-congruent messages:

Method

Stimuli and Stimuli Pretest

Online Information Portal

An experimental site with the masthead “Wired” was employed to present alleged search results, one topic at a time (see example of search results overview page in Figure 1). To ensure ecological validity, the look of an actual site was mimicked by the experimental site. One of four topical keywords appeared in the search box, such as “fracking,” followed by four headlines with corresponding lead paragraphs relating to the search topic. Participants picked articles they wanted to read by clicking. During this time, their message exposure was recorded by logging each hyperlink click and the seconds spent on each page. Then participants were able to return to the search results overview page at any time by selecting the “back to search results” button to select other articles.

OPEN IN VIEWER

Figure 1.

Example of online search results overview page for fracking topic.

Topics

The four target topics of fracking, GM food, biofuels, and nanotechnology were selected based on high and roughly equivalent levels of press coverage during the preceding 2 years (as indicated by LexisNexis search results). Furthermore, this selection of topics featured two energy issues and two issues pertaining to technology at the molecular level. Moreover, for both the energy domain and the molecular-level technology domain, one topic was anticipated to be overall viewed negatively (fracking and GM foods) and one positively (biofuels and nanotechnology), as reported below under “Attitudinal Measures.”

Headlines, Leads, and Articles

Sixteen articles (four per topic) were compiled from news and advocacy sources and edited for length, stance, and clarity of content. Articles had a mean length of 703.5 words (SD = 7.7). For each issue, the researchers created headlines (4-5 words each) and lead paragraphs (25 words each; see Table 1 for a complete list of headlines and leads), two in support of the particular technology or scientific innovation and two in opposition. Article headlines and leads featured either exemplar or numeric information as evidence. For each topic, two of four articles featured exemplars—one supporting and one opposing the science innovation or technology—and two featured numeric information—one in support and one in opposition. The individuals mentioned in the exemplar articles had first names that are commonly used for both genders (e.g., Morgan) so that participants of both sexes would be equally likely to relate to these characters.

OPEN IN VIEWER

Table 1.

Stimuli Headlines and Leads.

Fracking	Support	Oppose
Exemplar	People save money through fracking—Chris Jenkins of Texas saved enough on energy since switching to natural gas to establish retirement funds. “I support fracking—it is energy independence.”	Fracking cripples Pennsylvania town’s resident—Fracking near Dimock, Pennsylvania, contaminated local drinking water. For resident Terry Walsh, fracking “hospitalized my daughter Lisa, killed my cattle, and makes my tap flammable.”
Numeric	Fracking cuts gas costs 61%—Households saved $32.5 billion in 2012 with natural gas. Over 52% of Americans rely on fracking for providing energy, saving an average home around $430.	Fracking leaves 600 toxins behind—Over 40,000 gallons of chemicals go in the average frack. Flammable, toxic methane leaks into drinking water, generating 1,000 cases of contamination across 31 states.
Genetically modified foods	Support	Oppose
Exemplar	Genetically enhanced cows aid Minnesota family—Minnesota farmer Pat Harkins: “Genetic enhancement makes my cows produce more milk. Since I can produce more with fewer cows, I am making food affordable.”	Genetically altered beef harms consumers—St. Louis resident Sam Watson: “GMO beef has antibiotics, hormones, more fat and less essential vitamins. I am worried it is making my family sick.”
Numeric	Genetically enhanced dairy 18% more efficient—Bovine growth hormone allows 843,000 cows to produce the milk of 1,000,000, saving 2.3 million tons of feed and 540,000 acres of farmland for crops.	Genetically modified hormones increase udder infections 79%—Bovine growth hormone causes serious udder infections in a 6 million dairy cattle, resulting in 8 illegal antibiotics being found in 20% of U.S. milk.
Biofuels	Support	Oppose
Exemplar	Biofuels let citizen breathe easy—Chicago’s Morgan Ellis is enthusiastic about biofuels. “Ethanol means cleaner air, and that matters because I care about the environment and my child has asthma.”	Biofuels leave vulnerable workers hungry—“We can’t afford these prices anymore. I can’t feed my family,” says Alex Douglas, Nebraska. Biofuel demand disrupts American corn markets, threatening those like Douglas.
Numeric	Biofuels 30% cleaner, boost economy—Forty percent of U.S. corn is now for ethanol production. Nearly 15% of fuel demand is met by clean, sustainable biofuels, growing agricultural markets sixfold.	Biofuel raises food costs 7%—Nine million acres of U.S. corn produce fuel, only reducing gasoline consumption 0.5%. Food prices are subsequently soaring. Over 260 million people face biofuel-related famine.
Nanotechnology	Support	Oppose
Exemplar	Nanotechnology detects American’s heart disease—Tracy Reynolds, Kentucky: “I worry because heart disease runs in my family. As nanotechnology reduces my risk of heart attack, I am all for it.”	Nanotechnology threatens privacy, worries student—Iowa’s Jamie Nichols fears undetectable, microscopic nanobots already threaten privacy. “I don’t want to worry that the fly on the wall is listening to me.”
Numeric	Nanotechnology: Saving 600,000 lives annually—One in 4 American deaths result from heart disease. Nanotechnology can prevent roughly 715,000 heart attacks a year, saving 108.9 billion dollars on medical bills.	Nanosurveillance: Millions under invisible scrutiny—Nanotubes 100,000 times smaller than a human hair and nanoparticles 1,000,000 times smaller than an ant would allow untraceable NSA spying to increase on everyone.

The manipulations of stimuli were tested with an online survey, offered as an extra credit assignment in an undergraduate communication course at a large Midwestern university. A total of 12 men and 24 women participated in the pretest (M_age = 21.85, SD = 3.53). Participants were asked to examine the headlines and leads and to indicate the presence of either “numbers and statistics” or “examples and case descriptions.” Specifically, the prompts “I expect the article to feature many numbers and statistics” and “I expect the article to feature examples and case descriptions” were used with a 7-point scale with strongly disagree and strongly agree as anchor labels. Participants also indicated the extent to which they perceived each article would support or oppose the issue in question based on the prompt “I expect the article to support . . .” on a 7-point scale with strongly disagree to strongly agree (see Table 2). As reported in Table 2, all utilized leads were perceived as desired: For each topic, two of four articles featured exemplars—one in support and one opposing the science innovation or technology—and two featured numeric information—one in support and one opposing the science innovation or technology.

OPEN IN VIEWER

Table 2.

Stimuli Pretest Results for Article Leads Shown as Online Search Results.

	Support/oppose (7-point scale)		Exemplar/numeric (7-point scale)
Articles	M	SD	M	SD
Fracking
Pro-Exemplar	6.03 a	1.11	0.58 a	1.54
Pro-Numeric	5.75 a	1.23	−0.56 b	1.16
Anti-Exemplar	1.86 b	1.51	1.19 c	1.82
Anti-Numeric	2.36 c	1.88	−0.39 b	1.10
Genetically modified foods
Pro-Exemplar	5.89 a	1.24	1.03 a	1.72
Pro-Numeric	5.92 a	1.16	−0.72 b	1.43
Anti-Exemplar	2.58 b	2.09	1.58 c	1.83
Anti-Numeric	2.39 b	1.90	−0.47 b	1.23
Biofuels
Pro-Exemplar	6.11 a	1.14	1.14 a	1.59
Pro-Numeric	6.08 a	1.08	−0.67 b	1.24
Anti-Exemplar	2.50 b	1.84	0.75 c	1.52
Anti-Numeric	2.92 c	2.06	−0.47 b	1.44
Nanotechnology
Pro-Exemplar	6.08 a	1.23	1.36 a	1.62
Pro-Numeric	6.19 a	1.06	−0.42 b	1.13
Anti-Exemplar	2.06 b	1.57	0.78 c	1.61
Anti-Numeric	3.42 c	1.68	−0.33 b	1.69

Note. Means pertaining to the same topic with different superscripts differ at p < .05 in multiple comparisons with Sidak correction. Support/Oppose was measured with a single item for each article lead, 1 = strongly oppose to 7 = strongly support. Exemplar/Numeric score is the difference between an item rating the presence of examples and case descriptions and an item rating the presence of numbers and statistics (1 = strongly disagree to 7 = strongly agree).

Sources

To rule out that the sources (URLs) influenced the patterns of interest through varying levels of credibility or other impacts, all used URLs were established to have equally high perceived credibility. To this end, 12 women and 6 men (M_age = 21.44, SD = 1.20) from the same study population completed an online questionnaire on online source perceptions to earn extra credit. For various sites, they were asked to indicate “based on the website name and URL above, I expect the website to be credible” (1 = strongly disagree to 7 = strongly agree) in order to identify homogenous sets of high-credibility sources for use in the main study. Sixteen sources and their URLs (e.g., National Academy of Sciences; http://www.nasonline.org/) were selected, with a grand mean of 5.71 (SD = 0.19) and no significant pairwise differences among them.

Stimuli Rotations and Randomization

On the overview page for each issue, headlines and leads were randomized in a 2 (attitude-consistent or -discrepant) × 2 (exemplar or numeric) factorial design. Each participant saw each of the four issues (fracking, GM foods, biofuels, and nanotechnology) in the same sequence. For each issue, the placement of headlines and leads on the search results page was randomized, with one attitude-consistent exemplar, one attitude-consistent numerics article, one attitude-discrepant exemplar, and one attitude-discrepant numeric, all available for reading. Furthermore, each news headline was paired with a high credibility source as described above, in a rotated fashion.

Attitudinal Measures

Attitudes (Dichotomous)

Attitudes were measured as a dichotomous variable (oppose or support). Participants were first given a task to familiarize themselves with the protocol (see Knobloch-Westerwick & Meng, 2011, for further details of the procedure). Participants were asked to place one finger on the “z” key and one finger on the “/” key. Each symbol corresponded to either a negative or positive attitude. Participants then evaluated eight distracter topics as well as the four science and technology issues. The specific cues used for the four target issues were “Genetically modified foods,” “Biofuels,” “Nanotechnology,” and “Fracking.” Whether they selected “z” or “/” reflected either oppose or support as response option labels. Participants were instructed to complete the procedure as quickly as possible without compromising accuracy (see Table 3 for proportions of support and opposition).

OPEN IN VIEWER

Table 3.

Descriptive Statistics for Main Experiment Sample.

Variables	Fracking	Genetically modified foods	Biofuels	Nanotechnology
Pre-Exposure Attitude (dichotomous, % support)	24 a	31 a	81 b	76 b
Pre-Exposure Attitude (Likert, 1-7)	2.96 a (1.67)	2.75 a (1.82)	4.86 b (1.55)	4.62 c (1.56)
Pre-Exposure Attitude Certainty (Likert, 1-7)	3.88 a (2.12)	4.79 b (1.88)	4.22 c (1.80)	3.94 a (1.81)
Post-Exposure Attitude (dichotomous, % support)	33 a	32 a	74 b	67 b
Post-Exposure Attitude (Likert, 1-7)	3.21 a (1.76)	2.88 b (1.74)	4.66 c (1.52)	4.34d (1.66)
Attitude Shift (computed range −6 to 6)	−0.40 a (1.46)	−0.39 a (1.72)	−0.39 a (1.61)	−0.34 a (1.53)

Note. Values in a row with different superscripts differ significantly (p < .05) in paired t tests.

Attitude Shift

Explicit attitudes toward each of the four target technologies as well as six distracter topics were measured with the prompt “Please indicate how STRONGLY you support or oppose the following issues” on 7-point anchored scales (1 = strongly oppose to 7 = strongly support) both before and after exposure to the articles (see Table 3). Again, the specific cues used for the four target issues were “Genetically modified foods,” “Biofuels,” “Nanotechnology,” and “Fracking.” Attitude shift is the difference in pre- versus postexposure attitudes, which is then reverse scored for those who oppose the issue, so that the personal stance (pro or against) does not affect scores (see Table 3).

Attitude Certainty

The first set of attitude measures, before the message exposure task, included a rating task with the prompt “how certain are you of your opinion toward the following issues,” with a 7-point scale ranging from not at all certain to extremely certain. Descriptive statistics for attitude certainty are included in Table 3. Across the four measures, the mean attitude certainty was at M = 4.21 (SD = 1.33).

Results

Discussion

When seeking information about science topics, lay individuals commonly use online search engines to access information about scientific knowledge and innovation. This experiment presented participants with the opportunity to freely browse a series of alleged online search results on four relevant issues in science and technology. The leads of the online science information were manipulated to make use of either exemplar or numeric evidence. Furthermore, online article leads were also manipulated to indicate supportive or oppositional stances regarding the scientific innovations in question: fracking, GM foods, biofuels, and nanotechnology.

Regarding Hypothesis 1, no main effect of the message characteristic (exemplar vs. numeric) on message exposure emerged. However, exposure to exemplar versus numeric messages depended on individual differences: Users with high trait empathy dedicated more time to reading exemplar information, while those high on numeracy dedicated more time to reading numeric information. These findings provide support for Hypotheses 2 and 3. These effects persisted across topics, and after controlling for religiosity, science knowledge, and attitude toward science. Additionally, analyses found that higher trait empathy was associated with greater exposure to attitude-consistent articles, and higher science knowledge was associated with greater exposure to articles featuring numbers.

Regarding Hypothesis 4, results corroborated the suggested confirmation bias in participants’ exposure behavior, as online information users favored attitude-consistent messages. This confirmation bias emerged across all four topics. Inclusion of exemplars did not affect this pattern (Research Question 2). Even though some scholarly work suggests that recipients’ numeracy might affect the extent of confirmation bias (Kahan et al., 2012), the present data did not find such a moderating effect (Research Question 3) on the confirmation bias in exposure to science information. Shifts in attitudes were also examined as consequences of exposure to the different types of online science information. Exposure to attitude-consistent messages was found to reinforce existing attitudes. This finding supports Hypothesis 5 but did not emerge uniformly across all topics, only for fracking and biofuels. Also, while numeric attitude-consistent exposure emerged as specifically influential across all four topics, examination for specific topics yielded that both exemplar and numeric attitude-consistent reinforced attitudes. Thus, no clear evidence regarding differential persuasive effects of numeric versus exemplar messages emerged in response to Research Question 1.

Limitations of the present investigation include the finite number of science topics used, which restricts the generalizability of the results. Although the issues were relatively diverse in their baseline support (Table 3) and relative appeal to conservatives or liberals, all four represent technological innovations that pose environmental risks, which are typically seen as more beneficial and less risky by hierarchical-individualists, compared to communitarian-egalitarians (Kahan et al., 2009; Kahan et al., 2011). This is in contrast to topics such as stem-cell research or evolution (Jang, 2014) that relate differently to cultural beliefs. Future work should examine more topics.

Another limitation is the use of a sample comprised of college students, which may have more knowledge of, and more positive attitudes in general toward, science than the overall population. For example, the average participant answered 61% of items correctly on the science knowledge measure, compared with 55% for the average Western adult (Miller, 1998). In contrast to the cultural cognition model, the scientific literacy perspective (Miller, 1998) emphasizes the role of knowledge in informing attitudes toward science. This perspective can help reconcile the findings with those of Jang (2014). On the surface, the present finding of a general confirmation bias (Hypothesis 4), regardless of topic, contrasts with the previous study of selective exposure to online science information (Jang, 2014), which found a preference for attitude-discrepant exposure for at least for two of the four topics. However, in Jang’s sample, those with high perceived science knowledge and high religiosity showed attitude-consistent exposure patterns instead, which was attributed to attitude certainty. While the studies’ sample sizes were comparable, Jang’s study featured 238 participants from a stratified quota sampling method, and the present study built on a convenience sample of 229 students. The use of a representative sample by Jang may account for that study’s main effect of selective exposure to attitude-discrepant messages, as many respondents may have held relatively minimal science knowledge, and subsequently uncertain attitudes. This corresponds with the confirmation bias seen in the present study’s sample of university students. Their relatively high scores on science literacy, and moderately certain attitudes, may account for their confirmation bias. In addition to a relationship of literacy and certainty with selective exposure (as illustrated by Jang), the present results also show a link in which greater attitude uncertainty is associated with more attitude reinforcement from attitude-consistent messages. This pattern may reflect an attitude formation process, and future research should consider the interplay between scientific literacy and cultural cognition, rather than treating them as mutually exclusive propositions (e.g., Kahan et al., 2009). It appears both must be considered.

An additional limitation is that, given that exposure was tested as a self-selection phenomenon and was not experimentally assigned, there is the potential for spurious correlations between selectivity in message exposure and subsequent attitude shift. However, the inclusion of control variables helps in suggesting that it is indeed selective exposure to attitude-consistent numeric information that has a positive effect on attitude shift after exposure.

In addition to the differences in topics, samples, and moderating trait variables used in both studies, a number of methodological differences between Jang’s (2014) setup and the present study exist, that could also contribute to the differences in findings. First of all, both studies manipulated the stance of the messages to include supporting and opposing views, but Jang also included a neutral view. The presence of moderate views might dampen perceptions of ideological conflict that could otherwise activate confirmation bias (Dixon & Clarke, 2013; Jensen & Hurley, 2012; Maier & Richter, 2013). Message length in the present study was considerably longer, even after accounting for time available for browsing; accordingly, participants in the present study had 704 words of attitude-consistent material available per browsing minute, whereas Jang’s participants had only 300 words of attitude-consistent material available per browsing minute and thus may have run out of material that they preferred. This difference in total browsing time window could help account for the time spent with attitude-discrepant articles in Jang’s (2014) study, as readers often view attitude-consistent articles first and then move toward reading attitude-discrepant articles (Westerwick, Kleinman, & Knobloch-Westerwick, 2013). Finally, the present study manipulated the messages to emphasize either exemplar or numeric information, message characteristics which influenced participants’ exposure. Accounting for differences in information presentation in the present study removes any possible confounds of presentation style and allows for a more precise assessment of whether attitude-consistent or -discrepant articles are chosen at differential rates.

The present study demonstrated that confirmation bias in message exposure was evident across multiple topics, providing support in the science communication setting for the classic pattern of selective exposure seen in political and other contexts. However, although a consistent effect, the size of that effect was modest (4.6% of variance explained). This is in line with Garrett’s (2009) emphasis on the distinction between selective exposure and selective avoidance. Although individuals may select attitude-consistent messages at a higher rate, it does not follow that they completely cocoon themselves off from attitude-discrepant messages and reside in echo chambers. In the present study, participants did spend a substantial (albeit lesser) amount of time with science articles that challenged their attitudes. Even when accounting for moderating traits, selectivity remained modest, and the absolute amount of time spent with attitude-discrepant information did not approach zero. This lack of selective avoidance is consistent with Jang’s (2014) findings, where some participants (depending on individual differences) even spent more time reading attitude-discrepant information. Furthermore, the lack of selective avoidance has important implications, as demonstrated by the results, because exposure to attitude-discrepant information does not have the same polarizing effect on attitude shift that exposure to attitude-consistent information does.

In addition to testing for selective exposure in the science communication context, the present study built on previous work (Jang, 2014) by also considering how exemplification (Zillmann, 1999) might affect message exposure. Using vivid case studies and personal stories is a promising technique for conveying science information to the public. However, exemplification had a surprisingly limited influence. Individual differences played a key role, as empathetic people spent more time viewing articles with exemplars, while people high on numeracy spent more time viewing articles with numeric information. Additional findings indicated that empathy heightened the extent of confirmation bias, but attitude certainty did not. However, attitude certainty appears to have played a role in the effects of message exposure on subsequent attitudes. The selection of attitude-consistent messages produced reinforcing effects on attitudes, but these effects were inconsistent across science topics. Attitude-consistent numeric messages were most likely to consistently produce reinforcing attitude shifts. Future research is needed to not only examine how exposure to different types of science communication messages have the potential to contribute to public awareness of science but also to polarize public opinion regarding controversial technological advances.

Scientific innovation can benefit societal change, but change also depends on public opinion and individual attitudes. It is pivotal to understand how individuals select information about science and how the subsequent exposure shapes their attitudes. The present work may aid the tailoring of science information to different population segments—for instance, for information users with high numeracy, the present evidence suggests that the use of exemplars might deter them from closer reading of online content. On the other hand, population segments with high trait empathy should be more drawn to science information that features exemplars. An important challenge for general outreach appears to reside in preexisting attitudes—once an attitude on a topic has been formed, it can be difficult to attract information consumers to messages that do not align with said attitudes. Along these lines, the present data showed a bias toward attitude-confirming content. Given that scientific discovery often yields contradictory evidence, and expert consensus evolves only over time, lay individuals may often stick with a view formed earlier that may no longer converge with state-of-the-art scientific knowledge. With these dynamics in the development of scientific knowledge, confirmation bias may be particularly harmful toward spreading scientific advancements to the general population.