Exposure to social engagement metrics increases vulnerability to misinformation

---

Implications 

Online misinformation is a critical societal threat in the digital age, and social media platforms are a major vehicle used to spread it (Guess et al., 2019; Hameleers et al., 2020; Lazer et al., 2018). As an illustration, the International Fact-Checking Network found more than 3,500 false claims related to the coronavirus in less than 3.5 months.¹https://poynter.org/coronavirusfactsalliance Misinformation can cause serious societal harm in multiple ways: affecting public health (Sharma et al., 2020), influencing public policy (Lazer et al., 2018), instigating violence (Arif et al., 2018; Starbird et al., 2014), spreading conspiracies (Samory & Mitra, 2018), reducing trust in authorities (Gupta et al., 2014; Shin & Thorson, 2017; Vosoughi et al., 2018), and increasing polarization and conflict (Stewart et al., 2018). 

The growing societal impact of misinformation has driven research on technical solutions to detect and stop actors that generate and spread such content. The detection techniques have leveraged network analysis (Jin et al., 2013; Ratkiewicz et al., 2011), supervised models of automated behavior (Ferrara et al., 2016; Hui et al., 2019; Varol, Ferrara, Davis, et al., 2017; K.-C. Yang et al., 2019; K.-C. Yang et al., 2020), time series analysis to detect campaigns (Varol, Ferrara, Menczer, et al., 2017), and natural language processing for flagging factually incorrect content (Kumar et al., 2016; Pérez-Rosas et al., 2018). On the user interface side, researchers have explored the use of credibility indicators to flag misinformation and alert users (Clayton et al., 2019). Such credibility indicators can lead to a reduction in sharing the flagged content (Nyhan et al., 2019; Pennycook, Bear, et al., 2020; Pennycook, McPhetres, et al., 2020; Yaqub et al., 2020).

Studies have explored the role of environmental, emotional, and individual factors that impact online misinformation spread (Coviello et al., 2014; Ferrara & Yang, 2015; Grinberg et al., 2019; Kramer et al., 2014; Yaqub et al., 2020). However, there has been little empirical research on the effects of interface elements of social media feeds on the spread of misinformation (Hameleers et al., 2020; Shen et al., 2019). To address this gap, we empirically investigated how the spread of low-credibility content is affected by expo- sure to typical social engagement metrics, i.e., the numbers of Likes and Shares shown for a news article. Player behavior in the Fakey game shows near-perfect correlations between displayed social engagement metrics and player actions related to information from low-credibility sources. We interpret these results as suggesting that social engagement metrics amplify people’s vulnerability to low-credibility content by making it less likely that people will scrutinize potential misinformation while making it more likely that they like or share it. For example, consider the recent disinformation campaign video “Plandemic” related to the COVID-19 pandemic.²https://www.snopes.com/collections/plandemic/ Our results suggest that people may be more likely to endorse the video without verifying the content simply because they see that many other people liked or shared it. 

To interpret these findings, consider that the probability of sharing a piece of information grows with the number of times one is exposed to it, a phenomenon called complex contagion (Mønsted et al., 2017; Romero et al., 2011). Social engagement metrics are proxies for multiple exposures; therefore, they are intended to provide signals about the importance, relevance, and reliability of information — all of which contribute to people’s decisions to consume and share the information. In other words, users are likely to interpret high numbers for engagement metrics for an article as suggesting that it must be worthy of attention because many independent sources have validated it by liking or sharing it.

A key weakness in the cognitive processing of social engagement metrics is the assumption of independence; a malicious entity can trick people by falsely boosting engagement metrics to create the perception that many users endorsed an article. In fact, most disinformation campaigns rely on inauthentic social media accounts to tamper with engagement metrics, creating an initial appearance of virality that becomes reality once enough humans are deceived (Shao, Ciampaglia, et al., 2018). To prevent misinformation amplified by fake accounts from going viral, we need sophisticated algorithms capable of early-stage detection of coordinated behaviors that tamper with social engagement metrics (Hui et al., 2019; Pacheco et al., 2020; K.-C. Yang et al., 2020). 

Our findings hold important implications for the design of social media platforms. Further research is needed to investigate how alternative designs of social engagement metrics could reduce their impact on misinformation sharing (e.g., by hiding engagement metrics or making them less visible for certain posts), without negatively impacting the sharing of legitimate and reliable content. A good trade-off between these two conflicting needs requires a systematic investigation of news properties that can help determine differential display of social engagement metrics. Such properties may include types of sources (e.g., unknown/distrusted accounts) and topics (e.g., highly sensitive or polarizing matters with potential for significant impact on society). Further research is also needed to design media literacy campaigns, such as Fakey, that teach users to prioritize trustworthiness of sources over engagement metrics when consuming content on social media. Studies could explore the possibility of introducing deliberate pauses when consuming news through a social media feed (Fazio, 2020) and limiting automated or high-speed sharing. A comprehensive digital literacy approach to reduce the vulnerability of social media users to misinformation may require a combination of these interventions with additional ones, such as inoculation (Basol et al., 2020; Roozenbeek & van der Linden, 2019a, 2019b; Roozenbeek et al., 2020), civic online reasoning (McGrew, 2020), critical thinking (Lutzke et al., 2019), and examination of news feeds (Nygren et al., 2019).

Findings 

Finding 1: High levels of social engagement result in lower fact checking and higher liking/sharing, especially for low-credibility content.

For each article shown in the game, the player is presented with a photo, a headline, a description, and a randomly generated social engagement number. Based on this information, the player can Share, Like, or Fact Check the article(see Figure 1). The player must Share or Like articles from mainstream sources and/or Fact Check articles from low-credibility sources to earn points in the game. The Methods section provides details on source selection.

We measured the correlation between the social engagement number η displayed to players and the rates at which the corresponding articles from low-credibility sources were liked/shared or fact checked by the players. Given the realistically skewed distribution of η values, we sorted the data into logarithmic bins based on the shown social engagement numbers. For each bin ⌊log₁₀(η + 1)⌋, we calculated the liking/sharing and fact-checking rates across articles and players. We measured correlations using the non- parametric Spearman test as the data is not normally distributed. For articles from low-credibility sources, we found a significant positive correlation between social engagement level and liking/sharing (Spearman ρ = 0.97, p < 0.001) and a significant negative correlation between social engagement level and fact checking (Spearman ρ = −0.97, p < 0.001). We found similar statistically significant relationships between social engagement level and player behavior for mainstream news article as well, however the correlations are less strong: ρ = 0.66 (p < 0.001) for liking/sharing and ρ = −0.62 (p < 0.001) for fact checking.

Finding 2: People are more vulnerable to low-credibility content that shows high levels of social engagement.  

The previous finding is for the whole player population, with measures aggregated across all players. To delve further into the effect of social engagement exposure on individual players, we analyzed whether the displayed social engagement number influenced each player’s liking/sharing and fact-checking rates for articles from low-credibility sources. We treated each player as an independent entity and categorized social engagement numbers into three bins: low (0 ≤ η < 10²), medium (10²  ≤ η  <  10⁵),  and high (10⁵ ≤ η ≤ 10⁶). Within each social engagement bin, we counted the number of low-credibility articles to which each player was exposed. We then calculated the corresponding proportions of these articles that each player liked/shared or fact checked. Figure 2 plots the mean liking/sharing and fact-checking rates for low-credibility articles. Although players were more likely to fact check than like or share low-credibility content, Figure 2 shows that the trends observed at the population level held at the individual player level as well.

Since the data is not normally distributed (p < 0.05 using the Shapiro-Wilk test for normality), we used the Kruskal-Wallis test to compare differences among the three bins of social engagement. Liking/sharing (χ²(2) = 417.14, p < 0.001) and fact checking (χ²(2) = 214.26, p < 0.001) rates for low-credibility articles differed across bins. To determine which levels of social engagement impacted the rates at which low-credibility articles were liked/shared or fact checked, we conducted post-hoc Mann-Whitney tests across all pairs of social engagement bins, with Bonferroni correction for multiple testing. We found that ling/sharing as well as fact-checking rates were statistically significantly different across all bin pairings (p < 0.001). 

We employed the same approach to examine liking/sharing and fact-checking rates for mainstream articles across the three bins of social engagement. Similar to low-credibility articles, Kruskal-Wallis tests revealed a statistically significant effect of social engagement level on liking/sharing (χ²(2) = 161.80, p < 0.001) and fact checking (χ²(2) = 576.37, p < 0.001) rates for mainstream articles.

Methods

To investigate the effect of exposure to social engagement metrics on susceptibility to questionable content, we developed and deployed Fakey,³https://fakey.iuni.iu.edu/  an online news literacy game.  

Social media simulation  

Fakey simulates fact checking on a social media feed. The user interface of Fakey mimics the appearance of either Facebook or Twitter feeds for players who log into the game through the respective platforms. The game provides players with batches of ten news articles in the form of a news feed, as shown in Figure 1. Each article consists of elements typically displayed by popular social media platforms: photo, headline, description, and social engagement. 

For each article from mainstream as well as low-credibility sources, the game displays a single social engagement metric indicating the combined number of Likes and Shares. Having a single metric decreases the cognitive workload for players and simplifies the experimental design. The metric uses social engagement values drawn randomly from an approximately log-normal distribution with a maximum possible value (cutoff) of η = 10⁶. The distribution is such that roughly 69% of the articles display engagement values η > 10² and roughly 3% display values η > 10⁵. Although the engagement metric simulated in the game is not drawn from empirical data, the randomly generated metric numbers have a heavy tail similar to those typically observed on social media platforms (Vosoughi et al., 2018). 

Below each article is a set of buttons to Share, Like, Fact Check, or Skip the article or use a Hint (see Figure 1). Before playing the game, players are instructed that Share is equivalent to endorsing an article and sharing it with the world, Like is equivalent to endorsing the article, and Fact Check is a signal that the article is not trusted. After playing each round of ten articles, players have the option to play another round or check a leaderboard to compare their skill points with those of other players.

Content selection  

We followed the practice of analyzing content credibility at the domain (website) level rather than the article level (Bovet & Makse, 2019; Grinberg et al., 2019; Lazer et al., 2018; Pennycook & Rand, 2019; Shao, Ciampaglia, et al., 2018; Shao, Hui, et al., 2018). Each article in the game is selected from one of two types of news sources: mainstream and low-credibility. For mainstream news, we manually selected 32 sources with a balance of moderate liberal, centrist, and moderate conservative views: ABC News Australia, Al Jazeera English, Ars Technica, Associated Press, BBC News, Bloomberg, Business Insider, Buzzfeed, CNBC, CNN, Engadget, Financial Times, Fortune, Mashable, National Geographic, New Scientist, Newsweek, New York Magazine, Recode, Reuters, Techcrunch, The Economist, The Guardian, The Independent, The New York Times, The Next Web, The Telegraph, The Verge, The Wall Street Journal, The Washington Post, Time, and USA Today. The game obtains mainstream news articles via the News API.⁴https://newsapi.org

We selected the set of low-credibility sources based on flagging by various reputable news and fact-checking organizations (Shao, Ciampaglia, et al., 2018; Shao, Hui, et al., 2018). The selected low-credibility sources tend to publish fake news, conspiracy theories, clickbait, rumors, junk science, and other types of misinformation. The game uses the Hoaxy API6 to retrieve articles from these low-credibility sources. For each round, the game randomly selects five articles each from mainstream and low-credibility sources. For a given source, any article returned by the News or Hoaxy API⁵http://rapidapi.com/truthy/api/hoaxy is shown to the player regardless of topic, without further selection or filtering except for ensuring that the same story is not shown to the same player multiple times across rounds.

Data collection  

The game is available online through a standard web interface and as a mobile app via the Google Play Store and the Apple App Store. The mobile app is available in the following English-speaking countries: Australia, Canada, United Kingdom, and United States. People from other countries can play the game on the web. 

Our analysis uses data from a 19-month deployment of the game, between May 2018 and November 2019. During this period, we advertised the game through several channels, including social media (Twitter and Facebook), press releases, conferences, keynote presentations, and word of mouth. We recorded game sessions involving approximately 8,606 unique players⁶ We used recorded analytics to aggregate anonymous sessions by the same person. However, the aggregation approach cannot ascribe anonymous sessions to asingle person with complete certainty. Therefore, we cannot provide a precise number of unique players and 120,000 news articles, approximately half of which were from low-credibility sources. We did not collect demographic information, but we collected anonymous data from Google Analytics embedded by the game’s hosting service. Players originated from the United States (78%), Australia (8%), UK (4%), Canada (3%), Germany (3%), and Bulgaria (2%).

Limitations

Our news literacy game emulates relevant interface elements of popular social media platforms, such as Facebook and Twitter, without raising ethical concerns of real-world content manipulation (Kramer et al., 2014). Yet, conducting the study in a simulated game environment rather than an actual platform presents clear limitations as the experience and context are not identical. For example, we limited cognitive burden on players by capturing only Like and Share actions; these were the earliest ones deployed on social media platforms and, as such, are the most common across platforms and the most familiar to users. We used recorded analytics to aggregate anonymous sessions by the same person. 

The even mix of articles from mainstream and low-credibility sources is not necessarily representative of the proportion of misinformation to which social media users are exposed in the wild. Further, the fact-checking framing of the game primes players to expect misinformation, potentially making it more likely to be spotted. These factors might make players more suspicious within the game compared to the real world, correspondingly increasing fact-checking rates. However, there is no reason to believe that these factors impact our results regarding the influence of social engagement metrics. 

While this study is focused on user interaction elements, other factors related to users and content can affect the spread of questionable content. To respect privacy, we chose not to collect any player information apart from game analytics. However, knowledge about the background of the players (e.g., education, demographics, political affiliation) might provide further insight into vulnerability to misinformation. Similar refinements in insight would be provided by examining whether certain types of content are more likely to be influenced by social engagement metrics. These are important avenues for future research.