Political Bot Bias in the Perception of Online Discourse

Abstract

Four nationally representative studies (N = 1,986; three preregistered) find evidence for a bias in how people perceive opposing viewpoints expressed through online discourse. These studies elucidate a political bot bias, where political partisans (vs. their out-party) are more likely to view counter-ideological (vs. ideologically consistent) tweets to be social media bots (vs. humans). Study 1 demonstrates that American Democrats and Republicans are more likely to attribute tweets to bots when those tweets express counter-ideological views. Study 2 demonstrated this bias with actual bot tweets generated by the Russian government and comparable human tweets. Study 3 demonstrated this bias manifests in the context of real recent elections and is associated with markers of political animosity. Study 4 experimentally demonstrates the consequences of bot attribution for perceptions of online political discourse. Our findings document a consistent bias that has implications for political discussion online and political polarization more broadly.

Keywords

social media bots political polarization social processes cognitive bias

Political polarization among Democrats and Republicans has risen in recent years (Pew Research Center, 2017), causing partisan animosity (Iyengar et al., 2019). A common explanation for polarization’s rise is that social media has sorted people into ideological echo chambers (Pariser, 2011; Sunstein, 2018), primarily exposing people to likeminded ideas and excluding opposing viewpoints (Cinelli et al., 2021; Quattrociocchi et al., 2016; Sunstein, 2002). If this homogeneity facilitates polarization, then exposing people to cross-ideological information to free them from echo chambers is an intuitive solution. Unfortunately, research shows exposure to counter-attitudinal political views can have negligible or backfiring effects on political extremity and participation (Bail et al., 2020; Matthes et al., 2019).

Why does exposure to opposing views fail to pacify political conflict? One possibility is that novel technologies that can produce online discourse (e.g., social media bots) enable people to dismiss opposing viewpoints as nonhuman. Namely, our studies show: (1) Political partisans (vs. their out-party) are more likely to view counter-ideological (vs. ideologically consistent) tweets to be social media bots (vs. humans). (2) The strength of this bias is related to markers of political polarization and dismissal of outgroup content perceived to be bot-generated. Studies 1 to 3 demonstrate this political bot bias, whereby political partisans (vs. their out-party) are more likely to view counter-ideological (vs. ideologically consistent) tweets to be social media bots (vs. humans) and Study 4 focuses on consequences of bot attribution toward outgroup tweets such as distrust and dementalization.

Social media bots have the capacity to spread misinformation, sew political discord, and increase polarization (Barberá, 2020; Bessi & Ferrara, 2016; Stella et al., 2018; Woolley, 2016). Malevolent actors can manipulate a social bot, defined as “a computer algorithm that automatically produces content and interacts with humans on social media, trying to emulate and possibly alter their behavior” (Ferrara et al., 2016), to amplify extreme or misleading political viewpoints and sway opinion. The present work suggests that bots indeed contribute to political conflict, but not for reasons commonly suggested and demonstrated (i.e., spreading misinformation to bolster partisan discord). Rather, the mere presence of bots enables people to discount opposing viewpoints easily as mindless noise rather than thoughtful or legitimate critique. Studying online behavior is particularly important because its consequences are potentially vast: There are currently 4.62 billion users on social media globally (equal to 58.4% of the global population; Kemp, 2022), and expressing moralized and aggressive views online can proliferate through social networks (Brady et al., 2017), contributing to severe offline behavior, including violence (Blake et al., 2021; Mooijman et al., 2018; Williams et al., 2020). Here, we test how bot attribution influences the perception of and engagement with online discourse that expresses an opposing viewpoint.

This politically biased reasoning is rampant (Van Bavel & Pereira, 2018) and shapes how people perceive the world (Gerber et al., 2010; Kahan, 2013; Zmigrod et al., 2019). It is also exacerbated by ambiguity: Under conditions of uncertainty, people require less proof to convince themselves of their preconceived conclusion (Dunning et al., 1989; Kunda, 1990). This may make bot attribution especially prone to biased processing. Because any account’s status as a bot is ambiguous, it is easier for people to justify attributing bots to their political enemies. As such, we predict people are more likely to attribute tweets to bots when the content of the tweets is ideologically disagreeable.

Of course, social media bots have historically been deployed asymmetrically, predominantly targeting politically conservative networks and causes (Badawy et al., 2019; Luceri et al., 2019). Therefore, for some cases, attributing discourse to bots may represent a relatively accurate view of online discourse rather than a cognitive bias. We examined this possibility through ancillary analyses of how politically independents evaluated the tweets used in this research (see Supplemental Materials). These analyses showed naturally occurring differences in some, but not all, the tweets, suggesting that politically conservative tweets are not inherently more bot-like by default, but also that a valid focal comparison for illustrating the political bot bias is to compare Republicans’ and Democrats’ evaluations of the same tweet stimulus.

Importantly, attributing discourse to bots discounts the credibility of the viewpoint expressed, and biased bot attribution may reinforce people’s ingrained negative beliefs about their political enemies. Thus, we also situate bot attribution bias within the larger U.S. political climate of extreme polarization (Finkel et al., 2020). Polarization takes many forms and can manifest in othering, or viewing the out-party as fundamentally different from oneself; this could include dementalization of the out-party, compared with one’s own group (i.e., lacking higher order cognitive and emotional abilities; Haslam & Loughnan, 2014), or believing that the out-party is more politically extreme than one’s own party (Ahler, 2014). Here, we test whether this political bot bias is related to these various indicators of polarization.

Theoretically, it is possible for the political bot bias to correspond to lower rather than higher political animosity, demonstrating the critical need for a test of this relationship. Some work shows that algorithms that produce unwanted outcomes elicit less outrage and perceptions of bias than humans that produce these outcomes (Bigman et al., 2023; Bonezzi & Ostinelli, 2021). Furthermore, attributing a counter-ideological tweet to a bot could shift hostile perceptions to the bots themselves rather than to the out-party, in line with work showing that when robots become a salient threat, intergroup hostility diminishes (Jackson et al., 2020). Attributing malevolence to bots themselves (or to a hostile third party responsible for the bots, e.g., a foreign government) could even activate a sense of common ingroup identity (Dovidio et al., 2004) with the out-party, lessening political animosity. On the contrary, further work demonstrates that failure to engage with ideologically disagreeable information contributes to conflict (Minson et al., 2020; Yeomans et al., 2020) whereas receptiveness to such information can help bridge ideological divides (Reschke et al., 2020). Another reason why the political bot bias might correspond to political animosity is that—given that bots are often used to sew political discord as noted above—people could perceive the very use of bots as an act of hostility, and this perceived hostility can contribute to political conflict (Warner & Villamil, 2017; Waytz et al., 2014). Importantly, the political bot bias is conceptually distinct from partisan hostility generally because bots represent political interference by technology, rather than by humans, and people respond to these two sources of threat differently (Gamez-Djokic & Waytz, 2020; Granulo et al., 2019). Whereas attributing an opposing political view to a human triggers people to engage with the view as human (potentially counterarguing, discounting it, or disregarding it altogether), attributing the view to a bot invalidates the view by default, lessening any obligation to engage with the view as human altogether.

We also base this prediction on a pilot study that analyzed a nationally representative Pew Research Center survey (n = 4,478) and found awareness of bots on social media was linked to perceptions of the opposing political party as more ideologically extreme (see Supplemental Materials). This finding was not contingent on frequency of social media usage, suggesting that this relationship is not exclusive to people excessively online (or not online). Given this pilot study, and work cited above on motivated political reasoning and polarization (Van Bavel & Pereira, 2018), we predicted that the political bot bias would be related to including out-party hate and perceived out-party extremity (Study 3) as well as discounting the source of the outgroup point of view (Study 4). We also examined links to dehumanization in Studies 3 and 4.

We tested our predictions in four studies (three preregistered) with nationally representative samples (see details below). We consistently find evidence for the political bot in evaluations of verified political social media bots and equivalent human accounts, as well as in the context of real political events. We also show associations between this bias and indicators of political polarization and perceptions of online political discourse. For data and supplemental material, see https://osf.io/muvfr/?view_only=d59a967af86847fcb42f2d9ab87233d1.

Study 1

Study 1 tested whether the political bot bias would emerge such that Democrats/Republicans (vs. their out-party) would be more likely to view counter-ideological (vs. ideologically consistent) tweets to be social media bots (vs. humans).

Participants

We surveyed 491 Americans via Prolific Academic in May 2019 (see Supplemental Materials for further information on sample). The samples for all studies presented here were stratified to be nationally representative to 2015 U.S. census demographics for ethnicity, sex, and age. A power analysis indicated that this sample size was sufficient to obtain a power of 80% for a small effect size (f) of .11. All studies presented here collected this sample size or greater.

Procedure

Democrats and Republicans learned that some tweets are generated by social media bots and then evaluated four liberal-leaning tweets (e.g., denigrating Donald Trump) and four conservative-leaning tweets (e.g., denigrating Barack Obama) on the extent to which they believed each tweet was generated by a bot versus a human.

Participants first read a description of what a “Twitter bot” is and were asked if they knew about Twitter bots before the study and then entered demographic information including which party (Democrat or Republican) they felt closer to in a forced, binary choice measure (that served as our measure of political affiliation).

Participants read four focal tweets (presented in randomized order, one per page; see Figure 1). Response tweets ostensibly replied to these focal tweets and were presented below each of the four focal tweets. One of the response tweets was a stereotypically conservative response while the other was a stereotypically liberal response. Participants indicated the extent to which they believed each response was a human or a bot (1 = Definitely a human to 7 = Definitely a bot).¹

Figure 1

Study 1 Tweet Stimuli and Bot Attribution Measures

We presented the tweets in isolation, rather than on a Twitter feed (that presents a scrollable webpage with tweets on numerous topics from numerous users), to avoid various potential confounding factors (e.g., number of likes and presence of additional replies) that could artificially sway bot attribution. By presenting tweets presented in isolation, we sought to incorporate ecologically valid stimuli in a more internally valid experimental setting.

Results

A three-way mixed design comparing tweet political leaning (within-subjects: liberal-leaning vs. conservative-leaning) by tweet number (within-subjects: tweet response #1 vs. #2 vs. #3 vs. #4) by political affiliation (between-subjects: Democrat vs. Republican) revealed an interaction between political affiliation and Tweet political leaning, F(1, 486) = 52.27, p < .001, ηp² = .097 (see Figure 2). That is, Republicans (M = 3.69, SD = 1.22) attributed liberal-leaning tweets to bots significantly more than Democrats (M = 3.17, SD = 1.01, t(486) = 5.07, p < .001, d = 0.46, mean difference 95% confidence interval [CI] = [0.32, 0.73]). Conversely, Democrats (M = 4.39, SD = 1.18) attributed conservative-leaning tweets to bots significantly more than Republicans (M = 3.82, SD = 1.13, t(486) = 5.12, p < .001, d = 0.49, CI = [0.35, 0.79]). For Democrats, liberal-leaning tweets were attributed significantly less to bots (M = 3.16, SD = 1.01) than conservative-leaning tweets (M = 4.40, SD = 1.18, t(326) = 14.34, p < .001, d = 0.79, CI = [1.06, 1.40]). For Republicans, liberal-leaning tweets were not attributed significantly more to bots (M = 3.69, SD = 1.22) than conservative-leaning tweets (M = 3.81, SD = 1.13, t(166) = 0.94, p = .350, d = 0.07, CI = [−0.13, 0.37]).

Figure 2

Attribution of Liberal- Versus Conservative-Leaning Tweets to Bots or Humans, Moderated by Political Affiliation (Democrat vs, Republican)

There was a significant main effect of the tweets’ political leaning, as conservative-leaning tweets were attributed significantly more to bots (M = 4.20, SD = 1.20) than liberal-leaning tweets (M = 3.35, SD = 1.11, t(490) = 11.43, p < .001, d_z = 0.52, CI = [0.71, 1.01]) when collapsing across political affiliation. No main effect of political affiliation emerged, as Democrats attributed tweets to bots (M = 3.78, SD = 0.78) to a statistically indistinguishable degree compared with Republicans (M = 3.76, SD = 0.84, t(486) = .295, p = .768). There was a three-way interaction between tweets’ political leaning, participants’ political affiliation, and tweet number, F(3, 1,458) = 3.03, p = .028, ηp² = .006, suggesting that the extent of the interaction between tweet leaning and affiliation differed for individual tweets, but the critical interaction held across tweets.

Study 1 provides the first demonstration of the political bot bias, such that people attributed tweets to bots versus humans more when those tweets presented ideologically disagreeable versus agreeable content.

Study 2

Study 2 was a preregistered (https://aspredicted.org/CSH_TQZ) test of the political bot bias using tweets from verified bots and humans. People might perceive verified as especially bot-like and verified humans as especially human-like regardless of ideology, thereby mitigating the political bot bias. Directly comparing perceptions of bot and human tweets, therefore, provides a stringent test of our predictions (see Open Science Framework (OSF) Supplemental Materials for further rationale).

Participants

We surveyed 498 Americans via Prolific Academic in August 2021 (see Supplemental Materials for further information on sample).

Tweet Selection

To sample verified political social media bots, we used The Russia Tweets (russiatweets.com), which compiled nearly 3 million tweets from the Internet Research Agency, the Russian troll factory employed to disrupt U.S. political discourse in 2016. Two research team members searched four keywords in the database relevant to the time period (“Trump,”“Clinton,”“Black Lives Matter,” and “Make America Great Again”). For each keyword, the team members independently selected the first explicitly conservative-leaning and liberal-leaning tweets that were not simply retweets of other tweets or articles. They then collectively resolved any discrepancies. Because the bot tweets were time-stamped, the team could then find human tweets from the same time period. The team members searched for the same four keywords on Twitter while restricting the date to the same date as the equivalent bot tweet. Again, they independently selected the first explicitly conservative-leaning and liberal-leaning tweets. To verify that these tweets were human-generated, each Twitter user was input to Botometer (Davis et al., 2016), which provides a score from 0.0 (most likely to be a human) to 5.0 (most likely to be a bot). To provide a high degree of certainty that a given user was human, we selected only profiles with Botometer scores below 1. All selected users scored below 0.6 (M = 0.34). The tweet selection process left us with 16 total tweets: eight bots, eight humans, eight liberal-leaning tweets, and eight conservative-leaning tweets. See OSF page for full sample of tweets.

Procedure

As in Study 1, participants first provided demographic information and read a description of social media bots, indicated whether they had heard of social media bots, and indicated their political affiliation. They then saw the tweets and grouped by topic. We measured bot attribution for each tweet as in Study 1.

Results

A three-way mixed design comparing the political leaning of the tweet (within-subjects: liberal-leaning vs. conservative-leaning) by generator of the tweet (within-subjects: bot vs. human) by participant political affiliation (between-subjects: Democrat vs. Republican) revealed an interaction between political affiliation and political leaning of the tweets, F(1, 494) = 39.57, p < .001, ηp² = 0.074. That is, Republicans (M = 3.70, SD = 0.78) attributed liberal-leaning tweets to bots significantly more than Democrats (M = 3.39, SD = 0.86, t(494) = 3.62, p < .001, d = −0.38, CI = [0.14, 0.48]). Conversely, Republicans (M = 3.85, SD = 0.82) attributed conservative-leaning tweets to bots significantly less than Democrats (M = 4.17, SD = 0.88, t(494) = 3.66, p < .001, d = 0.38, CI = [0.15, 0.50]). For Democrats, liberal-leaning tweets were attributed significantly less to bots (M = 3.39, SD = 0.86) than conservative-leaning tweets (M = 4.17, SD = 0.88, t(366) = 15.01, p < .001, d = 0.79, CI = [0.68, 0.88]). For Republicans, only a marginal effect emerged, and liberal-leaning tweets were not attributed significantly more to bots (M = 3.70, SD = 0.78) than conservative-leaning tweets (M = 3.85, SD = 0.82, t(128) = 1.70, p = .092, d = 0.15, CI = [−0.02, 0.32]). There was a significant main effect of bot versus human, F(1, 494) = 399.02, p < .001, ηp² = 0.447, wherein bot tweets were attributed to bots (M = 4.26, SD = 0.81) more than human tweets (M = 3.29, SD = 0.95, t(497) = 19.80, p < .001, d_z = 0.89, CI = [0.87, 1.06]). Critically, no significant three-way interaction emerged between generator of the tweet (i.e., bot vs. human), political leaning of the tweet, and political affiliation of the participant F(1, 494) = 0.004, p = .947.

Study 2 replicated the political bot bias. While bot tweets were perceived as generally more bot-like than human tweets, critically, the political bot bias emerged equally for both bot tweets and human tweets.

Study 3

Study 3 was a preregistered study (https://aspredicted.org/HCN_QGP) that tested the political bot bias in a different context, and also examined markers of political conflict. In this study, participants considered tweets disputing the two most recent U.S. presidential elections, to examine the bias in the context of major recent real-world political events.

Participants

We surveyed 498 Americans via Prolific Academic in October 2021 (see Supplemental Materials for further information on sample).

Procedure

Participants first provided demographic information and then considered the two most recent U.S. presidential elections in which large numbers of political tweets were generated: when Donald Trump won the 2016 presidential election and when Joe Biden won the 2020 presidential election. Both elections were contested by Democrats and Republicans, respectively (see Cheney, 2017 and Gerhart, 2021). In our study, participants judged whether, in each of these cases, bots or humans generated tweets disputing the actual results of the election. As in Study 1, participants first read a description of social media bots, indicated whether they had heard of social media bots, and indicated their political affiliation. We then measured bot attribution across two items, presented in randomized order. The first item considered tweets in support of Hillary Clinton after her loss of the 2016 U.S. presidential election:

“Just after the 2016 presidential election was declared official, immediately thousands of tweets in support of Hillary Clinton emerged online stating the election was fraudulent. Below, we would like you to evaluate these pro-Hillary Clinton tweets.”

The second item considered tweets in support of Donald Trump after his loss of the 2020 U.S. presidential election:

“Just after the 2020 presidential election was declared official, immediately thousands of tweets in support of Donald Trump emerged online stating the election was fraudulent. Below, we would like you to evaluate these pro-Donald Trump tweets.”

For each of these items, participants indicated on a 0 (definitely humans) to 100 (definitely bots) scale, how likely it was that the tweets originated from humans versus bots.

Political Conflict Measures

Following the bot attribution measure, participants completed three different measures, each in reference to the Democratic and Republican parties separately. First, as a measure of in-party love/out-party hate, participants evaluated Democrats and Republicans on feeling thermometer scales (0 = very cold/unfavorable, 50 = neutral, and 100 = very warm/favorable). Next, as a measure of mind attribution, participants evaluated both parties on the extent to which they believed members of the party (1) “can engage in a great deal of thought,” (2) “are capable of doing things on purpose,” and (3) “have complex feelings” (1 = strongly disagree to 7 = strongly agree). Finally, as a measure of perceived extremity, participants indicated how liberal members of the Democratic party are (1 = very conservative to 10 = very liberal) and how conservative members of the Republican party are (1 = very liberal to 10 = very conservative).

Results

We compared bot attribution between Democrats and Republicans using multilevel regression analyses and treating the political leaning of the tweets as a within-subjects variable alongside the participants’ political party. The interaction between participants’ party (Republican = 0.5, Democrat = −0.5) and the party being rated (Republican = 0.5, Democrat = −0.5) was significant for bot attribution (b = −19.09, t(495) = −6.39, p < .001), mind attribution (b = 1.96, t(495) = 15.02, p < .001), warmth (b = 87.81, t(495) = 27.77, p < .001), and perceived extremity (b = −1.93, t(495)= 7.05, p < .001).²

Republicans (M = 54.47, SD = 24.57) did not attribute liberal-leaning tweets to bots significantly more than Democrats (M = 54.47, SD = 24.19, t(496) = 0.99, p = .325, d = 0.10, CI = [−2.41, 7.26]). Conversely, Republicans (M = 42.58, SD = 26.18) attributed conservative-leaning tweets to bots significantly less than Democrats (M = 59.27, SD = 26.24, t(496) = 6.29, p < .001, d = 0.64, CI = [11.48, 21.91]). For Democrats, liberal-leaning tweets were attributed significantly less to bots (M = 54.47, SD = 24.19) than conservative-leaning tweets (M = 59.27, SD = 26.24, t(364) = 3.21, p = .001, d = 0.17, CI = [1.50, 1.86]). For Republicans, liberal-leaning tweets were attributed significantly more to bots (M = 56.89, SD = 24.57) than conservative-leaning tweets (M = 42.58, SD = 26.18, t(132) = −5.70, p < .001, d = −0.45, CI = [−19.79, −8.84]).

We then tested for relationships between biased bot attribution toward the outgroup and biased evaluations toward the outgroup. To create bot attribution scores (e.g., how much people attribute outgroup tweets to bots vs. ingroup tweets), we regressed ingroup ratings onto outgroup ratings for all four measures and retained the residuals of the outgroup ratings. The residual score provides an index of variability in outgroup ratings after removing the variance predicted by ingroup ratings and has been used in prior research on cognitive bias (e.g., Anderson et al., 2012).

Bot attribution scores were significantly negatively correlated with in-party love/out-party hate (r(498) = −.12, p = .006) and mind attribution (r(498) = −.12, p = .010), and marginally positively correlated with perceived extremity (r(498) = .08, p = .066). The more people attributed bots to outgroup tweets relative to ingroup tweets, the colder they felt toward the outgroup relative to the ingroup, the less they perceived the outgroup as capable of mental states relative to the ingroup, and the (marginally) more politically extreme they believed the outgroup to be relative to the ingroup.

Study 3 replicates the political bot bias in a consequential political setting and shows evidence of its relationship to political conflict. Notably, unlike Studies 1 and 2, Democrats and Republicans do not significantly differ in their judgments of liberal tweets (although this finding is directionally consistent with the political bot bias such that Republicans judge these tweets to be more bot-like). This finding likely results from the reality that, as noted in the introduction, the preponderance of bots during the 2016 election largely supported the conservative (not liberal) candidate, and so less natural variability emerges in the judgments of pro-Clinton tweets.

Given that Study 3 shows a relationship between the political bot bias and political conflict, it is important to understand the downstream consequences of bot attribution. By experimentally manipulating bot attribution, Study 4 examines how people perceive the same content differently when they believe it was generated by a social media bot versus a human, and how this affects their engagement with and trust in the content.

Study 4

Study 4 was a preregistered study (https://aspredicted.org/GBN_N87) that tested for downstream consequences of bot attribution. In this study, we randomly assigning people to perceive a tweet as generated by a bot or generated by a human and then measured mind attribution toward the tweet generator, engagement with the tweet, and trust in the tweet. We also controlled for relevant factors to test whether the influence of bot perception was robust to outgroup animosity, partisan strength, participants’ self-reported political ideology, and self-reported Twitter usage.

Participants

We surveyed a sample of 500 Americans via Prolific Academic in August 2022 (see Supplemental Materials for further information on sample).

Procedure

Study 4 was a two-cell (bot vs. human condition) between-subjects experiment. Before being introduced to the experimental manipulation, participants reported their gender and ethnicity and completed several measures related to their political beliefs. First, they indicated their feelings toward Democrats and Republicans via the thermometer scales used in Study 3. Then, they indicated their political party, using a binary scale (Democrat or Republican), followed by a measure of partisan strength where they indicated how strongly they identified as a member of their selected party (“1 = leaning,”“2 = not so strong,” and “3 = strongly”); this measure was adapted from previous research studying partisan strength (e.g., Martherus et al., 2021). Participants also indicated their political ideology using scale from 1 = “extremely liberal” to 7 = “extremely conservative.”

Next, we randomly assigned participants to experimental condition. In the human condition, we described discourse on Twitter and told participants they would view a tweet posted by a human. Democrats then viewed a conservative-leaning tweet (praising Republican Senator Ted Cruz), and Republicans viewed a liberal-leaning tweet (praising Democratic President Joe Biden). In the bot condition, we described discourse on Twitter including how social media bots operate and told participants they would view a tweet posted by a bot. To enhance the external validity of this condition, we described the practice of “Tweetstorms,” whereby social media bots generate several identical tweets to post simultaneously. We also showed bot condition participants a Tweetstorm version of the same tweet used the human condition (i.e., either praising Cruz or Biden depending on participant political identification). Critically, we then showed bot condition participants the tweet in isolation for them to evaluate, therefore ensuring participants in both conditions evaluated the same exact stimuli for the primary measures. The instructions also described social media bots neutrally, did not include any information about trustworthiness, and otherwise mirrored the human condition in all ways.

See Figure 3 for a sample comparison of Tweet and Tweetstorm (full materials are available in supplemental material).

Figure 3

An Example of a Tweet (Left) and Tweetstorm (Right) Used in Study 4

Next, participants completed several evaluative measures. For our primary dependent variables, we measured mind attribution toward the tweet generator, engagement with the tweet, and trust in the tweet. For mind attribution, we adapted scales from prior research on mind perception (Bigman & Gray, 2018; Kozak et al., 2006), measuring specifically how much participants believed generator of the tweet was capable of “intelligence,”“a great deal of thought,”“complex emotion,”“foresight,” and “empathy and compassion” (1 = not at all, 7 = very much; α = .94). For engagement with the tweet, participants completed three items of how seriously they considered the tweet, specifically, whether the message of the tweet “deserves to be taken seriously,”“deserves to be disregarded” (reverse scored), and “deserves careful consideration” (1 = strongly disagree, 7 = strongly agree; α = .86). Participants then completed a one-item measure of trust in the tweet, specifically, “to what extent can the message of this tweet be trusted” (1 = not at all, 7 = very much). Finally, participants indicated their Twitter usage (0 = I do not use Twitter, 1 = very infrequently, 7 = very frequently).³

Results

An independent sample t-test revealed decreased perceptions of the tweet generator in the bot condition (M = 2.12, SD = 1.40) relative to the human condition (M = 3.57, SD = 1.50), t(497) = 11.14, p < .001, d = 1.00, CI = [1.19, 1.70]. Furthermore, participants in the bot condition engaged less with the tweet (M = 2.12, SD = 1.30) than those in the human condition (M = 3.50, SD = 1.72), t(497) = 10.08, p < .001, d = 0.91, CI = [1.11, 1.64]. Finally, participants in the bot condition expressed less trust in the message of the tweet (M = 1.72, SD = 1.19), than those in the human condition (M = 3.02, SD = 1.60), t(497) = 10.33, p < .001, d = 0.94, CI = [1.06, 1.55].

To test the robustness of these relationships, we then reanalyzed the data in a series of analyses of covariance, testing whether condition affected our primary measures while controlling for outgroup animosity (from the feeling thermometer), partisan strength, self-reported ideology, and Twitter usage. These analyses further tested for the main effect of party affiliation, as well as the interaction between party affiliation and condition, to test whether Democrats or Republicans were higher by default, or were uniquely influenced by the experimental manipulation. These results are presented in Table 1. We find robust evidence that our manipulation of bot attribution affected perceptions of the tweet generators, engagement with the tweet, and trust in the tweet’s message. These findings do not change when only analyzing participants who were aware of social media bots prior to participating in the study (see Supplemental Materials) Furthermore, no significant main effect of party affiliation or interaction with condition emerged for any primary dependent variable.

Table 1

The Effects of Bot Condition, Party Affiliation, and Control Variables on Tweet Generator Perceptions, Disregard of Tweets, and Trust of Tweets

	Dependent variable
Independent variable	Perceptions	Engagement	Trust
Condition	119.03***	189.95***	109.32***
Party	0.00	7.26	0.03
Condition*party	2.01	6.12	0.00
Outgroup feeling thermometer	53.11***	109.71***	58.61***
Partisan strength	0.01	2.10	0.39
Ideology	1.91	1.75	0.04
Twitter usage	4.28*	5.34	5.32*
R ²	0.31	0.27	0.28

Note. F statistics are presented.

p < .05. ***p < .001.

Study 4 demonstrated simple, yet powerful, consequences of bot attribution: When people are led to believe they are viewing a bot-generated tweet, they dementalize the source of the tweet, take it less seriously, and distrust its message. These cognitive factors represent the underpinnings of how psychological processes online carry weight for political thought and behavior generally (e.g., Mooijman et al., 2018). Study 4 also showed that the effect of bot attribution persisted when controlling for partisan animosity generally, demonstrating that these related psychological processes are conceptually distinct.

General Discussion

Across four nationally representative studies, we document a political bot bias such that people’s ideology guides their beliefs about whether actors online are human or nonhuman. American partisans were more likely than their comparable out-party members to attribute tweets to bots versus humans when the tweets were counter to their ideology than when they were consistent with their ideology. This partisan bias appears to apply to judgments of both real bot tweets or human tweets, and influences perceptions of recent major political events. We, in addition, showed that this bias corresponded to indicators of political conflict, including out-party animosity and perceived extremity of the out-party, and that bot attribution directly influences how people perceive, engage with, and trust online discourse. These findings are important given ongoing discussion on whether social media and bots contribute to polarization or not (Bail et al., 2020; Haidt & Bail, ongoing). It is, therefore, possible that while the bots themselves do not meaningfully sway political opinion and drive discord, their mere presence influences attitudes toward their political foes. Our findings are broadly relevant to online life. People now encounter technology-generated content in nearly all online platforms, not just Twitter: Everything from spam emails to robocalls, deepfakes, and generative artificial intelligence (Heilweil, 2023) that produces artwork (e.g., DALL-E) and conversations (e.g., chatGPT) present people with potentially politicized information from nonhuman sources. A vast and growing amount of information encountered through technology is, therefore, prone to the political bot bias.

Our findings suggest the effort to eliminate bots altogether presents a considerable challenge. After bot interference in the 2016 U.S. presidential election (Twitter Public Policy, 2018) prompted Twitter to shut down nearly 70 million fake accounts (Timberg & Dwoskin, 2018), bots continued to influence discourse around subsequent elections (Ferrara et al., 2020). Elon Musk’s dispute over his purchase of Twitter also centered on whether Twitter has sufficiently eliminated and identified bots, with Musk suggesting Twitter has overly relied on human reviewers to detect bots (Levine, 2022). Whether Musk’s assertions are accurate, research clearly shows that social media bots are notoriously difficult for people to detect (Cresci et al., 2017; Pew Research Center, 2018). This makes sense given that the anonymous nature of much online participation (Bargh & McKenna, 2004; Lapidot-Lefler & Barak, 2012) can obscure whether a user is human at all, and even automated tools for bot detection involve subjective judgments (Davis et al., 2016). Our research indicates an additional reason that contributes to the difficulty of bot detection, which is that people’s political leanings bias their judgments on this matter.

While we find relationships between the political bot bias and indicators of political conflict, the causality of these relationships should be interpreted cautiously. Although Study 4 demonstrates that bot attribution influences judgments related to conflict, is possible that these factors increase in a cyclical manner, where, for example, partisan animosity strengthens bot attribution, which in turn reinforces animosity. Future research should examine the causal relationships among these variables to disentangle them.

Future research can also examine whether these findings speak to how interactions with machines affect interactions with humans. For example, it is possible that exposure to deceptive technology (e.g., social media bots) generates a distrust in technology that contributes to distrust in humans as well, which occurs most markedly for political outgroup members. It is also possible that if the bias documented here causes people to perceive human-generated tweets to be bot-generated, given the results of Study 4, people might dementalize, disregard, and distrust humans online because they incorrectly believe they are bots. Given that existing studies show sometimes trust in technology is positively associated (Li et al., 2012) and sometimes negatively associated (Molina & Sundar, 2022) with trust in humans, additional studies are needed to understand how general experiences with technology contribute to political conflict in this way.

Despite considerable concern over the impact of political social media bots online, much research has focused on how bots themselves sway discourse rather than how their presence provides an opportunity to discount information online. Here, we document a consistent partisan bias in bot attribution and demonstrate an unfortunate corollary, in that it is associated with aspects of extreme political polarization.

Footnotes

Handling Editor: Danny Osborne.

Declaration of Conflicting Interests

The author(s) declared no potential conflicts of interest with respect to the research, authorship, and/or publication of this article.

Funding

The author(s) received no financial support for the research, authorship, and/or publication of this article.

ORCID iD

Shane Schweitzer

Supplemental Material

The supplemental material is available in the online version of the article.

Notes

Author Biographies

Shane Schweitzer is a postdoctoral research associate in the NUS Business School at National University of Singapore. His current research concerns perceptions of advanced, human-like technologies.

Kyle S. H. Dobson is a postdoctoral research fellow at The University of Texas at Austin in the Population Research Center. He studies authenticity and how we form connections with others.

Adam Waytz is the Morris and Alice Kaplan Professor in Ethics and Decision Management at the Northwestern University’s Kellogg School of Management. He looks at when we attribute or deny mental states to other entities, and the moral and ethical implications of these processes.

References

Ahler

D. J.

(2014). Self-fulfilling misperceptions of public polarization. The Journal of Politics, 76(3), 607–620.

Anderson

Brion

Moore

D. A.

Kennedy

J. A.

(2012). A status-enhancement account of overconfidence. Journal of Personality and Social Psychology, 103(4), 718–735.

Badawy

Addawood

Lerman

Ferrara

(2019). Characterizing the 2016 Russian IRA influence campaign. Social Network Analysis and Mining, 9(1), 1–11.

Bail

C. A.

Guay

Maloney

Combs

Hillygus

D. S.

Merhout

Freelon

Volfovsky

(2020). Assessing the Russian Internet Research Agency’s impact on the political attitudes and behaviors of American Twitter users in late 2017. Proceedings of the National Academy of Sciences, 117(1), 243–250.

Barberá

(2020). Social media, echo chambers, and political polarization (Social Media and Democracy: The State of the Field, Prospects for Reform, 34). Cambridge University Press.

Bargh

J. A.

McKenna

K. Y.

(2004). The internet and social life. Annual Review of Psychology, 55, 573–590.

Bessi

Ferrara

. (2016). Social bots distort the 2016 US Presidential election online discussion. First monday, 21(11–7).

Bigman

Y. E.

Gray

(2018). People are averse to machines making moral decisions. Cognition, 181, 21–34.

Bigman

Y. E.

Wilson

Arnestad

M. N.

Waytz

Gray

. (2023). Algorithmic discrimination causes less moral outrage than human discrimination. Journal of Experimental Psychology: General, 152(1), 4–27.

10.

Blake

K. R.

O’Dean

S. M.

Lian

Denson

T. F.

(2021). Misogynistic tweets correlate with violence against women. Psychological Science, 32(3), 315–325.

11.

Bonezzi

Ostinelli

(2021). Can algorithms legitimize discrimination? Journal of Experimental Psychology: Applied, 27(2), 447–459.

12.

Brady

W. J.

Wills

J. A.

Jost

J. T.

Tucker

J. A.

Van Bavel

J. J

. (2017). Emotion shapes the diffusion of moralized content in social networks. Proceedings of the National Academy of Sciences, 114(28), 7313–7318.

13.

Cheney

(2017). House Democrats fail to muster support to challenge Trump’s Electoral College win. Politico. https://www.politico.com/story/2017/01/no-trump-electoral-college-challenge-233294

14.

Cinelli

Morales

G. D. F.

Galeazzi

Quattrociocchi

Starnini

(2021). The echo chamber effect on social media. Proceedings of the National Academy of Sciences, 118(9), 1–8.

15.

Cresci

Di Pietro

Petrocchi

Spognardi

Tesconi

(2017, April). The paradigm-shift of social spambots: Evidence, theories, and tools for the arms race. In R. Barrett, & R. Cummings (Eds.), Proceedings of the 26th International Conference on World Wide Web Companion (pp. 963–972). ACM.

16.

Davis

C. A.

Varol

Ferrara

Flammini

Menczer

(2016, April). Botornot: A system to evaluate social bots. In J. Bourdeau, J. Hendler, & R. N’Kambou (Eds.), Proceedings of the 25th International Conference Companion on World Wide Web (pp. 273–274). International World Wide Web Conferences Steering Committee.

17.

Dovidio

J. F.

Ten Vergert

Stewart

T. L.

Gaertner

S. L.

Johnson

J. D.

Esses

V. M.

Riek

B. M.

Pearson

A. R.

(2004). Perspective and prejudice: Antecedents and mediating mechanisms. Personality and Social Psychology Bulletin, 30(12), 1537–1549.

18.

Dunning

Meyerowitz

J. A.

Holzberg

A. D.

(1989). Ambiguity and self-evaluation: The role of idiosyncratic trait definitions in self-serving assessments of ability. Journal of Personality and Social Psychology, 57(6), 1082–1090.

19.

Ferrara

Chang

Chen

Muric

Patel

(2020). Characterizing social media manipulation in the 2020 US presidential election. First Monday, 25(11).

20.

Ferrara

Varol

Davis

Menczer

Flammini

(2016). The rise of social bots. Communications of the ACM, 59(7), 96–104.

21.

Finkel

E. J.

Bail

C. A.

Cikara

Ditto

P. H.

Iyengar

Klar

Mason

McGrath

M. C.

Nyhan

Rand

D. G.

Skitka

L. J.

Tucker

J. A.

Van Bavel

J. J.

Wang

C. S.

Druckman

J. N.

(2020). Political sectarianism in America. Science, 370(6516), 533–536.

22.

Gamez-Djokic

Waytz

(2020). Concerns about automation and negative sentiment toward immigration. Psychological Science, 31(8), 987–1000.

23.

Gerber

A. S.

Huber

G. A.

Washington

(2010). Party affiliation, partisanship, and political beliefs: A field experiment. American Political Science Review, 104(4), 720–744.

24.

Gerhart

(2021). Election results under attack: Here are the facts. The Washington Post. https://www.washingtonpost.com/elections/interactive/2020/election-integrity/

25.

Granulo

Fuchs

Puntoni

(2019). Psychological reactions to human versus robotic job replacement. Nature Human Behaviour, 3(10), 1062–1069.

26.

Haidt

Bail

. (2022). Social media and political dysfunction: A collaborative review. Unpublished manuscript, New York University.

27.

Haslam

Loughnan

(2014). Dehumanization and infrahumanization. Annual Review of Psychology, 65, 399–423.

28.

Heilweil

(2023). What is generative AI, and why is it suddenly everywhere? Vox. https://www.vox.com/recode/2023/1/5/23539055/generative-ai-chatgpt-stable-diffusion-lensa-dall-e

29.

Iyengar

Lelkes

Levendusky

Malhotra

Westwood

S. J.

(2019). The origins and consequences of affective polarization in the United States. Annual Review of Political Science, 22, 129–146.

30.

Jackson

J. C.

Castelo

Gray

(2020). Could a rising robot workforce make humans less prejudiced? American Psychologist, 75(7), 969–982.

31.

Kahan

D. M.

(2013). A risky science communication environment for vaccines. Science, 342(6154), 53–54.

32.

Kemp

(2022). Digital 2022: Another year of bumper growth. We Are Social. https://wearesocial.com/us/blog/2022/01/digital-2022-another-year-of-bumper-growth-2/

33.

Kozak

M. N.

Marsh

A. A.

Wegner

D. M.

(2006). What do I think you’re doing? Action identification and mind attribution. Journal of Personality and Social Psychology, 90(4), 543–555.

34.

Kunda

(1990). The case for motivated reasoning. Psychological Bulletin, 108(3), 480–498.

35.

Lapidot-Lefler

Barak

(2012). Effects of anonymity, invisibility, and lack of eye-contact on toxic online disinhibition. Computers in Human Behavior, 28(2), 434–443.

36.

Levine

(2022). Elon wants to fight the bots. Bloomberg. https://www.bloomberg.com/opinion/articles/2022-07-18/elon-wants-to-fight-the-bots

37.

Rong

Thatcher

J. B.

(2012). Does technology trust substitute interpersonal trust?: Examining technology trust’s influence on individual decision-making. Journal of Organizational and End User Computing (JOEUC), 24(2), 18–38.

38.

Luceri

Deb

Badawy

Ferrara

(2019, May). Red bots do it better: Comparative analysis of social bot partisan behavior. In Liu

White

(Eds.), Companion proceedings of the 2019 World Wide Web conference (pp. 1007–1012).

39.

Martherus

J. L.

Martinez

A. G.

Piff

P. K.

Theodoridis

A. G.

(2021). Party animals? Extreme partisan polarization and dehumanization. Political Behavior, 43(2), 517–540.

40.

Matthes

Knoll

Valenzuela

Hopmann

D. N.

Von Sikorski

(2019). A meta-analysis of the effects of cross-cutting exposure on political participation. Political Communication, 36(4), 523–542.

41.

Minson

J. A.

Chen

F. S.

Tinsley

C. H.

(2020). Why won’t you listen to me? Measuring receptiveness to opposing views. Management Science, 66(7), 3069–3094.

42.

Molina

M. D.

Sundar

S. S

. (2022). Does distrust in humans predict greater trust in AI? Role of individual differences in user responses to content moderation. New Media & Society. https://doi.org/10.1177/14614448221103534

43.

Mooijman

Hoover

Lin

Dehghani

(2018). Moralization in social networks and the emergence of violence during protests. Nature Human Behaviour, 2(6), 389–396.

44.

Pariser

(2011). The filter bubble: What the Internet is hiding from you. Penguin.

45.

Pew Research Center. (2017). The partisan divide on political values grows even wider. https://www.pewresearch.org/politics/2017/10/05/the-partisan-divide-on-political-values-grows-even-wider/

46.

Pew Research Center. (2018). Most Americans have heard about social media bots, many think they are malicious and hard to identify. https://www.pewresearch.org/journalism/2018/10/15/most-americans-have-heard-about-social-media-bots-many-think-they-are-malicious-and-hard-to-identify/

47.

Quattrociocchi

Scala

Sunstein

C. R.

(2016). Echo chambers on Facebook. https://ssrn.com/abstract=2795110

48.

Reschke

Minson

Bowles

H. R.

De Vaan

Srivastava

S. B.

(2020). Mutual receptiveness to opposing views bridges ideological divides in network formation. https://ssrn.com/abstract=3703958

49.

Stella

Ferrara

De Domenico

(2018). Bots increase exposure to negative and inflammatory content in online social systems. Proceedings of the National Academy of Sciences, 115(49), 12435–12440.

50.

Sunstein

C. R.

(2002). Risk and reason: Safety, law, and the environment. Cambridge University Press.

51.

Sunstein

C. R.

(2018). Legal reasoning and political conflict. Oxford University Press.

52.

Timberg

Dwoskin

(2018). Twitter is sweeping out fake accounts like never before, putting user growth at risk. The Washington Post. https://www.washingtonpost.com/technology/2018/07/06/twitter-is-sweeping-out-fake-accounts-like-never-before-putting-user-growth-risk/

53.

Twitter Public Policy. (2018). Update on Twitter’s review of the 2016 US election. https://blog.twitter.com/official/en_us/topics/company/2018/2016-election-update.html

54.

Van Bavel

J. J.

Pereira

. (2018). The partisan brain: An identity-based model of political belief. Trends in Cognitive Sciences, 22(3), 213–224.

55.

Warner

B. R.

Villamil

. (2017). A test of imagined contact as a means to improve cross-partisan feelings and reduce attribution of malevolence and acceptance of political violence. Communication Monographs, 84(4), 447–465.

56.

Waytz

Young

L. L.

Ginges

. (2014). Motive attribution asymmetry for love vs. hate drives intractable conflict. Proceedings of the National Academy of Sciences, 111(44), 15687–15692.

57.

Williams

M. L.

Burnap

Javed

Liu

Ozalp

(2020). Hate in the machine: Anti-Black and anti-Muslim social media posts as predictors of offline racially and religiously aggravated crime. The British Journal of Criminology, 60(1), 93–117.

58.

Woolley

S. C

. (2016). Automating power: Social bot interference in global politics. First Monday, 21(4).

59.

Yeomans

Minson

Collins

Chen

Gino

(2020). Conversational receptiveness: Improving engagement with opposing views. Organizational Behavior and Human Decision Processes, 160, 131–148.

60.

Zmigrod

Rentfrow

P. J.

Robbins

T. W.

(2019). Cognitive inflexibility predicts extremist attitudes. Frontiers in Psychology, 10, 989.