Abstract
Although social media has become a primary news platform, the effects of social media features on users’ information processing remains under-explored. This study explores how social media design features affect use of sources. A 2 × 2 between-subjects experiment examined effects of “information context collapse” (ICC)—where different content types are presented in the same form and location—and volume of content (VoC). These features were hypothesized to predict inattentive (System 1) processing, which predicts “source blindness”—where users fail to process source cues during news use. A mock social media site was created with participants queried about posts shown on the site. Results find that while VoC has no effect, ICC significantly predicts source blindness mediated by System 1 processing. This suggests collapsed information environments lead to inattentive processing of source information, increasing potential negative outcomes of social media news use. Implications of these findings are discussed.
In July 2018, the website react365 posted an article about a cruise ship disaster in Mexico that killed at least 32 people. The article saw over 350,000 engagements on Facebook (Silverman and Pham, 2018). The misinformation was quickly debunked by Snopes.com (Evon, 2018), who noted that react365’s homepage showed clearly it was a prank website, where users could upload their own fictitious stories. Similarly, in 2019, the site wearethellod.com, claimed former basketballer Michael Jordan was running for US senate (archive.is, 2019). The site’s about page states they are part of a “network of parody, satire and tomfoolery.” The article was also debunked by Snopes.com (Snopes.com, 2019).
These and many similar stories (Silverman and Pham, 2018) have led to increasing concerns over the potential for misinformation—“the inadvertent sharing of false information”—and disinformation—“the deliberate creation and sharing of information known to be false” (Wardle, 2017; see also Institute for Public Relations, 2019 for similar definitions) to affect public opinion and political processes. In order to alleviate these concerns many organizations, companies, and academics have produced guides to help people verify information online; offering advice targeted to citizens, journalists, and educators. Many guides recommend checking the source to verify information. FactCheck.org (run by The Annenberg Public Policy Center) recommends readers “consider the source” (Kiely and Robertson, 2016); the technology news site, Digital Trends, write readers must “consider the source before believing” information (Nicol, 2018); and software developers Mozilla, makers of the Firefox browser, write that you can verify “a story’s accuracy by tracking down the original source, and make sure that what you’re sharing isn’t perpetuating a hoax” (Schreder, 2018).
Knowledge of sources is not the only issue related to avoiding misinformation. However, verifying source credentials remains common advice in media literacy campaigns (Metzger, 2007; Potter, 2004). Despite this, research finds that while citizens believe they should check source credentials, few report doing so (Flanagin and Metzger, 2001; Hargittai et al., 2010; Metzger et al., 2003; Wineburg et al., 2016). Furthermore, experiments suggest when heuristics—such as appearance—are equal, individuals show little discernment between high- and low-credibility sources (Flanagin and Metzger, 2007; Pearson and Knobloch-Westerwick, 2018). Hence, this article investigates how design features commonly seen on social media lead to users paying reduced attention to sources—a phenomenon labeled here as “source blindness.”
Using an online experiment, the study investigates the effects of two common social media features: the breaking down of distinctions between content types—labeled “information context collapse”—and high-VoC.
Sources in the changing news environment
News media has changed substantially in the 21st century, with 38% of Americans listing online as their preferred medium for news (Pew Research Center, 2018). Furthermore, social media news use grows. Pew Research Center (2019) finds 55% of Americans get at least sometimes get news from social media, with 28% often getting news on social media. Many social media sites are used, with 52% of Americans getting news from Facebook, 28% on YouTube, 17% on Twitter, and 14% on Instagram.
On social media, new users often have both a distal source—the original authors of the content—and a proximate source—which curate news from many locations (Kang et al., 2011). This is characterized by the way-finding framework (Pearson and Kosicki, 2017), whereby instead of news users routinely returning to a trusted source, they use instincts and developed routines to navigate the troves of information from numerous sources. The user is “not choosing from a collection made by a particular source, but instead selecting from a compilation of stories from a variety of sources” (Pearson and Kosicki, 2017: 1091). User tracking data show most online news is now found through these proximate sources as opposed to direct visits (Newman et al., 2017).
In addition, users have more sources to choose from. Reduced costs of publishing online mean large resources are no longer required to distribute news (Metzger, 2007; Pearson and Kosicki, 2017). This allows new sources to emerge, many of whom have no formal journalistic training (Allen, 2006). Hence, the majority of online news outlets are not household names (Jurkowitz, 2014).
These changes result in a reduced relationship between users and news sources. While scholarship has focused on many different definitions of sources (Sundar and Nass, 2001), sources here are defined as a brand, publication, or outlet that produces original content.
Despite media literacy efforts extolling the importance of checking sources (Metzger, 2007; Potter, 2004), individuals regularly fail to investigate (Metzger et al., 2003; Pearson and Knobloch-Westerwick, 2018) or forget source information (Kalogeropoulos and Newman, 2017). Although research has always found limited source retention (e.g. Hovland and Weiss, 1951), modern news environments may exacerbate these problems (Kalogeropoulos and Newman, 2017; Kang et al., 2011).
This effect is referred to here as “source blindness,” which is defined as a state whereby individuals fail to consider source information when processing news content. Due to social media design features, users fail to connect source information to related content. While users are aware content has a source, those high in source blindness, are unlikely to recall source information or use the source to make content evaluations.
Furthermore, four proposed conceptual dimensions of source blindness are investigated. First is “source investigation,” which are actions taken during news use to investigate information about the source, a key recommendation in media literacy (Kiely and Robertson, 2016; Metzger, 2007; Potter, 2004).
The second and third dimensions share similarities, both referring to information remembered. “Source information recall” is the ability to remember information about the source, such as its name, or information pertinent to evaluating its credibility. “Source recognition,” is the ability to recognize information about a source when presented with it. This also relies memory; however, it expects only recognition rather than free recall, which has been shown to be distinct (Kintsch, 1968).
Finally, “source affect,” refers to developing an attitude toward the source, irrelevant of what information is recalled. In other words, an individual may have an evaluation of a source, without recalling specific information, in line with online memory (Lodge et al., 1989).
The focus of this study is on how the information processing strategy used and the design of social media influence these four dimensions of source blindness. These relationships are explored in the following sections.
Information processing during news use
Users are presented with many cues of information quality to consider. Prominence interpretation theory (Fogg, 2003) argues information assessments are more than the interpretation of each cue, but also how prominent each cue is. Consequently, for any cue to be used, it must be prominent enough to be noticed.
Fogg (2003) argues cue prominence can be altered by cognitive factors, user motivation, and website factors. Research has supported this theory. For instance, George et al. (2016) found the prominence of images and visual design meant users were more likely to notice these features and incorporate them in credibility assessments. Howe and Teufel (2014) found younger users were more likely to notice advertising, and subsequently rate sites as less credible, suggesting demographics influenced cue prominence.
It is also the case that information processing can alter cue prominence. For instance, Ferebee (2007) found those high in issue involvement noticed different cues and were more likely to utilize more cues which took greater effort to process.
A number of theories, falling under the general umbrella of “dual processing theories” (Evans, 2003; Metzger, 2007), propose two information processing approaches: one that is efficient and inattentive, and another that is attentive but is inefficient and requires greater effort. Kahneman (2011) labels the approaches “System 1,” which “operates automatically and quickly, with little or no effort” (p. 20), and “System 2,” which is attentive but requires a greater cognitive load. While this terminology is used, other dual processing theories, including the elaboration-likelihood model (Petty and Cacioppo, 1984), heuristic-systematic model (Chen and Chaiken, 1999), and controlled versus automatic processing (Schneider and Shiffrin, 1977) also offer relevant literature.
While System 2 processing likely offers more accurate assessments, it is limited by available cognitive resources. Consequently, when users need or want to reduce the resources consumed processing information, they can switch from System 2 to System 1 processing.
Users likely switch processing strategies when they no longer have sufficient cognitive resources. System 2 processed information is expected to be housed in working memory (Evans and Stanovich, 2013); hence, when experiencing too great a cognitive strain individuals switch to a more efficient processing strategy. Individuals may also choose System 1 processing even when not cognitively depleted. Research suggests people are generally cognitive misers (Lodge et al., 1989), and therefore, seek to use as few cognitive resources as needed to still “satisfice” perceived needs. Therefore, it is likely individuals default to System 1 processing, but shift to System 2 processing when dealing with content they perceive requires it.
Individuals often aim to process news in a surveillance approach, actively processing each item (Eveland, 2001). While there are variations in how closely people follow current affairs (Gao, 2014), it is likely the majority of people desire to process current affairs information using System 2 processing due to perceived importance.
Regarding sources, it is hypothesized that when engaging in System 1 processing, source cues will be used less. Social media sites often display headlines and proximate sources in more noticeable fonts (larger, higher contrast) than distal sources. In addition, images are usually perceived as prominent (Fogg, 2003; Zillmann et al., 2001). When engaging in System 1 processing, individuals will utilize only more prominent cues.
In addition, source cues can be harder to process. While, the interestingness of a headline can be assessed easily, source evaluations require retrieving information from long-term memory or evaluating contextual information.
Finally, it may be that System 1 processing affects which cues are sought. When cognitive resources are high, individuals are more drawn toward options perceived as better for them (Shiv and Fedorikhin, 1999). Regarding media use, Panek (2016) found less cognitively depleted users selected more hard news. Building from this, individuals will generally reason they should seek high-credibility sources, to receive higher quality information. However, as resources become depleted, they are drawn toward cues indicating immediate rewards—such as entertainment value.
Consequently, System 1 processing alters the processing of cues by making individuals: only notice the most prominent cues, process cues requiring reduced cognitive effect, and drawn toward cues indicating immediate rewards. Therefore, System 1 processing should lead to increased source blindness.
H1: System 1 processing will predict reduced (a) source investigation, (b) source information recall, (c) source recognition, and (d) source affect.
Two contextual factors seen on social media are hypothesized to increase System 1 processing: “information context collapse” (ICC) and volume of content (VoC).
Information context collapse
Social media presents current affairs alongside personal stories and soft news, collapsing barriers distinguishing content types, similar to “context collapse” (boyd, 2002; boyd and Ellison, 2007; Hogan, 2010).
Online content places sources on a level field, lessening distinguishable differences between them (Burbules, 1998). Social media likely replicates this for content, blurring indicators differentiating content types. Sites like Facebook and Twitter present current affairs identically to personal updates and entertainment, all receiving the same colors, font, and link size.
“Context collapse” describes how social media collapses boundaries between distinct social groups. Individuals desire to present themselves differently to distinct groups; however, social media collapses boundaries between social contexts, leaving users with an unspecified audience (boyd, 2002; Hogan, 2010). While the “context” has traditionally been conceived as content producers in inter-personal communication, contexts can refer “to physical arrangements, [. . .] situational definitions, temporal moments and distinct locales” (Davis and Jurgenson, 2014: 477). Consequently, ICC focusses on challenges faced by information consumers in collapsed environments.
Current affairs, entertainment, personal updates are all different information contexts. Just as individuals desire to present themselves differently to different groups, but struggle with an indistinct audience, they too contend with collapsed information environments containing indistinct boundaries between content types that require different processing approaches.
Most content on people’s social media feeds is likely non-current affairs information (Wang, 2017) and can receive reduced scrutiny. Individuals are often cognitive misers (Lodge et al., 1989), only switching to System 2 processing when required. Without clear indicators of topic, users need to expand effort deducing the content’s topic. Due to ICC, individuals may lack the cognitive resources required to switch processing styles for different content types, instead using the same routine—System 1—processing used for softer content, which forms the majority of their social media feed.
H2: Information context collapse will lead to System 1 processing.
Due to the previous hypotheses, an indirect effect can be hypothesized.
H3: Information context collapse will indirectly reduce (a) source investigation, (b) source information recall, (c) source recognition, and (d) source affect mediated by increased System 1 processing.
Volume of content
The Internet has seen a great increase in available content, which has been shown to leave news users feeling overwhelmed (Panek, 2016) and fatigued (Holton and Chyi, 2012).
Dealing with large volumes of information has always been part of news consumption (Graber, 1988), however the volume has increased dramatically online. The website of the New York Times publishes 230 stories a day, The Wall Street Journal 240, and The Washington Post 500 (increasing to 1200 stories a day if wire stories are included; Meyer, 2016). The Internet transformed “‘information explosion’ to ‘information surplus,’ from ‘information abundance’ to ‘information overabundance’” (Chyi, 2009: 455). In addition, while news broadcasts and newspapers have a fixed limited length, many websites, including Facebook and Twitter, feature “infinite scroll” designs, which continuously provide new content as the bottom of the page is reached (Estes, 2017).
There is evidence individuals switch between System 2 and System 1 processing as volume increases, including in news contexts (Panek, 2016).
H4: High volumes of content will lead to System 1 processing.
Consequently, extrapolating from H1:
H5: Volume of content will indirectly reduce (a) source investigation, (b) source information recall, (c) source recognition, and (d) source affect mediated by increased System 1 processing.
Furthermore, as the ICC increases the cognitive burden of each item seen on social media, it should increase the impact of VoC.
The desire to process different content types with different strategies (Eveland, 2001; Kahneman, 2011) means users experiencing ICC must expend cognitive resources determining the topic (news, entertainment, etc.) of each item. As this increases the cognitive burden per item, it will take a reduced VoC for the cognitive burden to exceed capacity (Panek, 2016) leading to System 1 processing.
H6: The effect of the volume of content on System 1 processing will be greater when users are experiencing information context collapse.
H7: Information context collapse will increase the indirect effect of volume of content on reduced (a) source investigation, (b) source information recall, (c) source recognition, and (d) source affect mediated by System 1 processing.
Methods
A 2 × 2 between-subjects experiment was conducted, with high versus low-ICC and high versus low-VoC as the two factors. Participants saw a mock social media feed split across four pages. After the stimuli, participants completed a questionnaire measuring attitudes toward posts from the stimuli. The experiment was conducted online between the 4th and the 9th May 2019.
Participants
The experiment utilized a sample from Dynata. Dynata use an opt-in panel, and participants receive points that are used for gifts, money, or charity donations. Initially, 633 cases were collected, of which 513 completed the study.
Some cases were removed to ensure data quality. First, using the simple non-differentiation method (Kim et al., 2019), participants who straightlined at least two of four scales (all scales had reverse coded items) were removed. This removed 57 cases. Next, cases with extreme completion times were removed. Participants who completed the experiment in less than 4 minutes (it was expected to take 20 minutes); spent less than 30 seconds on the stimuli; or spent more than 1300 seconds (21.7 minutes) on the stimuli were removed. This removed 105 cases. Finally, open-ended questions revealed three responses likely created by bots. This left 370 participants (72.1% of the original sample).
To assess the representativeness of the final sample, key demographic variables were examined. Participants were on average of 48.66 years old (SD = 16.6); 154 (41.7%) were male, 211 (57.2%) female and 4 (1.1%) other. For education, 197 participants (53.3%) had a college degree. For political identification, 152 participants (41.2%) identified as Democrat, 99 (26.8%) Republican, 103 (27.9%) as independent, and 15 (4.1%) as supporting another party. This means that while the sample is diverse, the sample does over-represent females and Democrats and is more educated.
Procedure
Sampled participants were sent a link to the study. Participants could complete the study anytime during the data collection period, however they had to complete the experiment on a desktop computer (as opposed to mobile or tablet). After the consent form, participants were asked about their everyday media use and interest in politics. Then, they were given information about the social media site (stimuli), before being redirected.
The stimuli contained four pages of either two or four posts each, dependent on condition. Participants could click forwards and backwards through the pages as they wished (see “Stimuli” section for more information). The high-volume condition showed 16 posts, with 8 shown in the low-volume condition. In the low-ICC condition, all posts from one topic were on the same page. In the high-ICC condition, each set of four posts contained one from each topic (equating to one post per topic per page in the high-volume condition, and one post per topic every two pages in the low-volume condition). At the end of the stimuli, participants clicked a button that led to the post-stimuli questionnaire.
Five random posts from the stimuli were then selected. For each, participants were shown a screenshot (with source information removed). They were then asked the measures of System 1 processing, political relevance, source information recall, and source affect. Participants were then shown the same five posts again, and asked the measure of source recognition (see Measures).
Participants were then asked to evaluate source descriptions as part of a post-test, before finally completing demographic questions.
Stimuli
The stimuli used a fictional site called “Link Me” (screenshots in Appendix 1). The stimuli used custom CSS code and graphics to alter the appearance of a Qualtrics survey. To increase ecological validity, colors and fonts were modeled on Facebook. The appearance of the stimuli was crosschecked across major browsers and operating systems.
The stimuli used posts written for the study. Each post was presented as a lead that would link to a full story—similar to presentations on social media. While posts were not specifically designed to mirror the presentation of any particular site, their length and style is similar to how leads appear on Facebook. Posts consisted of a headline, lead, an image, and an icon indicating the post’s topic (screenshots in Appendix 1). Headlines were between 8 and 10 words and leads between 55 and 56 words. Each post belonged to one of the following four topics: current affairs, entertainment, lifestyle, and personal stories. All posts were based on real articles or public social media posts (taken from Reddit or Tumblr), edited for length and to meet manipulation requirements.
The manipulation required participants to perceive the current affairs posts as being more politically relevant than the other posts. For each of the five randomly selected posts (see “Procedure” section) participants were asked, “Do you think the story is politically irrelevant or politically relevant? (By politically relevant we mean relating to issues or activities that are of societal importance).” Responses were on a five-point scale from “Very politically irrelevant” to “Very politically relevant.” A Sidak post-test confirmed all four of the current affairs posts were perceived as more politically relevant than the other posts (full results are in the Online Appendices available at: https://osf.io/uv2a5/?view_only=0a54539857bd4c24b9750763dd51d6fd.)
Each post was presented as being from a particular source. Sixteen sources were used. Beneath the source name, an “about this source” button could be clicked. This displayed additional information about the source while the participant’s cursor hovered over the button. All descriptions were between 20 and 21 words in length. Fictional sources were used, manipulated to be either high or low credibility through professionalism (e.g. high-credibility sources were manipulated to be highly professional). The stimuli alternated between showing high- and low-credibility sources for each post.
Near the end of the study participants were shown three source names with descriptions and asked, “How much do you agree or disagree with the following statement?” Two statements read, “This source is credible” and “This source is run by amateurs.” Responses were given on a seven-point scale from “Strongly disagree” to “Strongly agree.” The Sidak post-test showed all high-credibility sources were seen as more credible and more professional than all low-credibility sources, and vice versa (results are available in the Online Appendices at: https://osf.io/uv2a5/?view_only=0a54539857bd4c24b9750763dd51d6fd).
Higher-level measures
Higher-level measures are measures taken at the participant level.
Source disinterest
A measure of people’s interest in using source information was included. Participants were asked to rate their agreement with six statements: “I find it hard to name news sources off the top of my head”; “There are large differences in the quality of news produced by different news sources” (reverse coded); “Most news sources produce basically the same product”; “I rarely notice the source of the news I am using”; “I care more about how interesting the story is than which organization created it”; and “I always look to see what organization created a news story” (reverse coded). Responses were on a five-point scale from “Strongly disagree” to “Strongly agree,” with the measures aggregated to create the scale (M = 2.6, SD = 0.78, Cronbach’s α = .798).
Faith in intuition
The updated Rational-Experiential Inventory Scale by Pacini and Epstein (1999) was used. The scale uses six items, and responses were on a five-point scale from “Strongly disagree” to “Strongly agree” (M = 3.45, SD = 0.66, α = .8).
Social media news use
Participants were asked, “Thinking of the news you get online, how much do you get in the following ways?—From social media (e.g. clicking links on Facebook or Twitter).” Responses were given on a five-point scale from “None or nearly none” to “All or nearly all.” However, this question was only asked to those who previously reported getting at least some news online. As those who got no news online could not get news on social media, these participants were coded as “None or nearly none” for the social media question (M = 2.20, SD = 1.26).
Lower-level measures
Lower-level measures were taken at the level of individual posts. For each measure, there are five cases per participant, one for each of the posts displayed in the post-stimuli questionnaire.
Source investigation
Source investigation described actions taken during news use to seek additional source information. This was measured using the “about this source” button (see “Stimuli” section). Participants, on averaged, clicked the “about this source” button for 10.46% (SD = 26.93%) of the posts in the stimuli. Among the five randomly selected posts (see “Stimuli” section).
Source information recall
Participants were asked, “Was there anything about the source that made you skeptical or concerned over the quality of the story?” If a participant mentioned the sources credibility or professionalism, they were coded as recalling source information. In addition, they were asked, “Can you remember the name of the source of this post? If so, please write it below. If you are unsure, feel free to guess.” If the correct name was recalled (close matches were also accepted) they were also coded as recalling source information. Participants were coded as displaying source information recall for 4.4% of posts.
Source recognition
Participants were asked, “Which of these sources do you believe was the source of this story? (If you are unsure, feel free to guess).” Five possible answers presented in a random order were given. These included the correct answer, three random sources from the stimuli, plus one additional fictional source. The additional fictional sources were “US Press Review (www.uspressreview.com)”; “News Today (www.newstoday.com)”; “The Wake-Up Blog (www.blogspot.wakeupblog.com)”; “Fool’s Garden (www.foolsgarden.info)”; and “Central News Agency (www.cna-news.com).”
Participants recognized the source 24.4% of the time, meaning the average participant correctly recognized 1.22 (SD = 1.07) sources.
Source affect
Participants were asked, “What was your feeling towards the source of this post?” on a seven-point scale from “Strongly disliked” to “Strongly liked.” A participant who formed an attitude toward the source should have a response further from the mid-point. Therefore, source affect was measured as the distance from the mid-point (M = 0.86, SD = 1.01).
System 1 processing
Participants were asked to rate their agreement or disagreement to “I actively focused on this post attentively” on a five-point scale from “Strongly disagree” to “Strongly agree.” Scores were reverse coded, so higher scores indicated increased System 1 processing (M = 2.49, SD = 1.23).
Political relevance
Participants were asked, “Do you think the story is politically irrelevant or politically relevant?” on a five-point scale from “Very politically irrelevant” to “Very politically relevant” (M = 2.49, SD = 1.23).
Data analysis
Data analysis was conducted at the level of individual posts for each of the five randomly selected posts. These were analyzed as multi-level models with cases nested within participants. Due to some missing data, this created 1837 cases (posts) within 369 participants.
Models were analyzed with the lme4 package (Bates et al., 2014) and the lmerTest package (Kuznetsova et al., 2016) designed for the R environment. All analyses included the two conditions, ICC and VoC, as independent variables. In addition, source disinterest, faith in intuition, social media news use, and political relevance were included as control variables. System 1 processing was included as an independent variable when predicting the four source blindness dimensions. Models analyzing System 1 processing and source affect used a linear model, while source investigation, source information recall, and source recognition had a binary outcome.
All continuous variables were mean centered. In addition, to assist with model convergence, the integer scalar—“the number of points per axis for evaluating the adaptive Gauss-Hermite approximation to the log-likelihood” (Bolker, n.d.)—were increased for the models with bivariate outcomes (the lmer package defaults to one). An increased integer scalar increases the accuracy of each iteration, increasing the likelihood of model convergence at the expense of processing time. Each model was set to the lowest value where convergence was achieved; these were one for source investigation, five for source information recall, and two for source recognition.
To test mediation hypotheses, the R package mvnorm (Ripley et al., 2018; Venables and Ripley, 2013) was utilized, using Monte Carlo simulations with 1,000,000 iterations to create a confidence interval for the indirect effects.
Results
Four multi-level models were created to examine the effect of System 1 processing on the four dimensions of source blindness. These results can be seen in Table 1. Results show System 1 processing predicts three of the four dimensions of source blindness, predicting source investigation (H1a), source information recall (H1b), and source affect (H1d). However, it does not predict source recognition (p = .25), offering no support for H1b.
Results of multi-level models predicting source blindness dimensions.
ICC: information context collapse; SD: standard deviation; SE: standard error.
Residual is only provided for source affect as other models have bivariate dependent variables.
Significant at p < .05.
Significant at p < .01.
Next, the effect of the two conditions on System 1 processing was examined (results can be seen in Table 2). Initially, an interaction term between ICC and VoC was included in the model, however, this was insignificant (p = .601), offering no support for H6. So main effects could be interpreted, the interaction term was removed. The second analysis shows ICC did affect System 1 processing, supporting H2. However, the effect of VoC on System 1 processing was insignificant (p = .22); offering a lack of support for H4.
Results of multi-level analysis predicting System 1 processing.
ICC: information context collapse; SD: standard deviation; SE: standard error.
significant at p < .05.
significant at p < .01.
Given significant findings for H1a, b, and d as well as H2, it is possible to assess the indirect effects from ICC on the three significantly predicted source blindness dimensions, mediated by System 1 processing. ICC significantly indirectly predicts source investigation (β = −.18, 95% CIs = −0.36, −0.04), supporting H3a; source information recall (β = −.12, 95% CIs = −0.27, −0.02), supporting H3b; and source affect (β = −.02, 95% CIs = −0.03, −0.02), supporting H3d.
Discussion
Results found mixed support for the hypotheses. Results did show System 1 processing predicted the following three source blindness dimensions: source investigation (H1a), source information recall (H1b), and source affect (H1d). However, there was no prediction of source recognition (H1c).
It is uncertain why source recognition was not predicted, especially as source recognition should share similarities with source information recall, as both involve memory. It may be the result of error being created by random, but correct, guesses. Given the multiple-choice question had five options, participants had a 20% chance of guessing correctly. The correct source was given 24.4% of the time, suggesting many “successful” recognitions were error.
In addition, source recognition questions appeared after all other questions about the five selected posts, therefore, appearing later in the study. This means there was more time for retained information to be forgotten.
For indirect effects, results showed ICC predicted three source blindness dimensions mediated by System 1 processing, supporting H3a, b, d.
However, elsewhere hypotheses were not supported, VoC did not affect System 1 processing (H4), and consequently had no indirect effect on source blindness (H5a–d). There was also no evidence of an interaction effect between ICC and VoC (H6), and therefore, no impact of the interaction on the indirect effect (H7a–d).
It is perhaps unsurprising no interaction term was found given insignificant findings for VoC. It was expected ICC would increase the cognitive burden for each post, increasing the effect of VoC. However, as VoC was insignificant, there was no effect to increase.
The issue then becomes why VoC was insignificant, especially as previous research found a clear effect of VoC on inattentive processing of news (Panek, 2016). The most likely explanation is a manipulation failure. The stimuli were split across four pages to increase the manipulation of ICC. Participants were instructed that the stimuli contained four pages of equal length. However, it may be the unseen volume of pages to come did not register with participants. In this case, even in the high-volume condition, they would only have to prepare to process 4 posts at a time, rather than all 16 posts.
This research finds contextual features of social media sites can lead to source blindness. When content was not grouped by distinct topics, it appears participants found it harder to keep track of the topic. This additional cognitive burden likely led to participants processing all content in the same inattentive (System 1) fashion, meaning they were less likely to process source content.
This blending of content is one of the primary features of social media. Reddit advertises itself as a place to “discover breaking news first, viral video clips, funny jokes, and hot memes” (Reddit, n.d.) and Twitter claims to offer “breaking news and entertainment to sports, politics, and everyday interests” (Twitter, n.d.). However, this feature may be negatively impacting the ability of its users to process content. While offering convenience, collapsed environments also make it harder for users to apply the correct processing strategy, making it more likely, potentially important content will be processed in an inattentive fashion. Furthermore, the effects of ICC likely extend beyond source blindness, potentially increasing other impacts of inattentive processing such as increased belief in misinformation, or reduced health literacy.
This argument, however, implies people desire to process current affairs content in a System 2 fashion. While this was not measured, qualitative evidence can be found. When asked if they could recall source information (see the measure of “Source Information Recall” section) participants often responded that they did not pay attention, as the content was not important. Comments included, “I don’t remember the source it was not important for this story”; “it was a post with minimal impact”; and “it’s not an important enough story to worry about.”
ICC builds upon existing context collapse literature (boyd, 2002; Hogan, 2010), offering useful extensions to the theory. Context collapse has traditionally focused on content producers—how people struggle to present themselves to different groups online. ICC shows that this context collapse also applies to information consumers, who struggle to use the desired processing approach for different types of information on social media sites that blur content distinctions.
Source blindness was significantly predicted. While previous research has shown online news can affect use of source cues (Flanagin and Metzger, 2007; Pearson and Knobloch-Westerwick, 2018), this study extends this by offering a theoretical explication of source blindness and examining it through measurable dimensions. These dimensions cover both behavior during, and memory of, news use. Indeed, this may be the first study to experimentally manipulate people’s investigation of source information. This is noteworthy, as it suggests the effects of System 1 processing on source blindness affects both people’s ability to interpret source information as well as their motivation.
If it was only users’ ability to process source information that was affected, then simplifying source verification processes may resolve the issue. Indeed, media literacy campaigns have been criticized for expecting too great an effort from users (Metzger, 2007). However, here, even when credibility information was available by clicking a button next to the source, those showing greater System 1 processing were less likely to verify source information. While overcoming the difficulty and complexity of verifying source information is important, such interventions may not be enough, as it seems that users in collapsed environments are less willing to verify source information as well.
Long-standing research has investigated source perceptions (Hovland and Weiss, 1951; Petty and Cacioppo, 1984). However, this article suggests sources are not always accounted for, and that changes to the information environment may increase or reduce use of source cues in line with prominence interpretation theory (Fogg, 2003). Consequently, scholars may benefit from considering how the design of online environments may alter use of sources, both when assessing real-world sites and designing stimuli for future research.
Much research regarding sources argues that inattentive scanning of content leads to increased, not decreased, reliance on sources (Chen and Chaiken, 1999; Wilson and Sherrell, 1993). However, the reverse was found here. This research does not seek to invalidate the claims of this long-standing literature. It does though suggest a rethinking of how such assumptions may play out in future research. As System 1 processing increases, cues that are easier to interpret are more likely to be used. Source cues are a heuristic—a cognitive shortcut that is faster than actively engaging with content—but they may require more cognitive resources than other heuristics (e.g. the interestingness of a headline, popularity indicators, site aesthetics). Therefore, our understanding of the relationships between information processing and heuristic use could be strengthened by exploring the “ranking” of heuristics, and exploring at what point of cognitive burden different heuristics are triggered.
The study also has implications for site designers wishing to aid their users to process content effectively. Designers could increase the prominence of source cues, perhaps by increasing font size or color contrast. Steps could also reduce ICC. For instance, imagine a scenario where Facebook used a different background color for current affairs posts. In addition, features such as Twitter’s lists function allow users to create separate feeds based on the different types of accounts followed. Encouraging users to utilize such systems, or automating them, allowing users to easily distinguish between content types, may reduce the effects of ICC.
These findings also offer implications for those designing media and digital literacy guides trying to stem the flow or misinformation online. Such guides often focus on the actions taken by users after viewing information (e.g. advising them to consider potential source bias, or to verify information at fact checkers). Little focus is given to how the structure of information platforms—especially social media—may be reducing positive media literacy behaviors. Such guides may find benefit in making users cognizant as to the effects of ICC. They may also encourage users to better utilize systems that allow for more contextualized information feeds (such as Twitter’s lists feature), or even use separate accounts or platforms for different content (for instance using Facebook for social interactions and Twitter for current affairs information).
There are limitations to this research. While the study utilized a general pubic audience using quotas to ensure sample diversity, the study still used a volunteer sample, not a random sample. The representativeness of the sample is also limited by over-representation of females, Democrats, and those with a degree.
Although the study sought to maintain high ecological validity through custom CSS code, in order to create a strong manipulation for ICC, the stimuli was split across multiple pages, a feature rarely seen in real-world websites. In addition, while the stimuli were labeled a social media site, many social elements were removed. The stimuli contained no comments, popularity cues, or information on who posted the information, which can all influence news use (Kang et al., 2011; Knobloch-Westerwick et al., 2005).
Finally, the measure of System 1 processing used only a single measure. As the question was asked for all five randomly selected posts, it was only possible to include one question to avoid creating too great a burden on participants. Ideally, more items measuring System 1 processing would have been included. However, using a single item reduced variance. As results were found with only one item, additional items would likely have increased, rather than decreased, effects.
This study investigated how design features of social media could influence processing of source cues during news use, finding one feature, ICC, did predict source blindness. This study introduces two new concepts: source blindness and ICC. Both variables have the potential to be useful in future research examining online news use. Online environments are changing constantly; new websites emerge and existing ones alter their designs. Examining the underlying constructs of design features that affect how people consume news helps system designers aid users, and helps scholars better predict expectations for a variety of sites.
Footnotes
Appendix
Acknowledgements
The author wishes to thank Dr Gerald Kosicki, Dr Silvia Knobloch-Westerwick and Dr Kelly Garrett for their advice and guidance on this work.
Funding
The author(s) received no financial support for the research, authorship, and/or publication of this article.
