Unfollowing hyperpartisan social media influencers durably reduces out-party animosity

Abstract

There is considerable debate over whether and how social media contributes to polarization. In a correlational study (n1 = 1,447) and two digital field experiments (n2 = 494, n3 = 1,133), we examined whether (un)following hyperpartisan social media influencers contributes to polarization and misinformation sharing. We found that incentivizing Twitter/X users to unfollow hyperpartisan social media influencers improved feelings toward the out-party by 23.5% compared to the control group, with effects persisting for at least six months. Unfollowing also led participants to engage with more accurate news and increased satisfaction with their Twitter/X feeds—without reducing engagement. This study demonstrates the long-term causal impact of exposure to hyperpartisan influencers. Moreover, unlike other social media reduction interventions, unfollowing is a targeted approach: like a scalpel, it surgically removes a few harmful parts of one’s feed, allowing the beneficial aspects to remain.

Full text

1
Unfollowing hyperpartisan social media influencers durably reduces
out-party animosity
Steve Rathje1*, Clara Pretus2, James K. He3, Trisha Harjani3, Jon Roozenbeek3,4, Kurt Gray5,
Sander van der Linden3, Jay J. Van Bavel1,6*
Affiliations:
1New York University, Department of Psychology
2Universitat Autònoma de Barcelona, Department of Psychobiology and Methodology of
Health Sciences
3University of Cambridge, Department of Psychology
4King’s College London, Department of War Studies
5University of North Carolina, Chapel Hill, Department of Psychology
6Norwegian School of Economics
*Corresponding authors. Email: srathje@alumni.stanford.edu and jay.vanbavel@nyu.edu.
Abstract: There is considerable debate over whether and how social media contributes to
polarization. In a correlational study (n1= 1,447) and two digital field experiments (n2= 494, n3=
1,133), we examined whether (un)following hyperpartisan social media influencers contributes
to polarization and misinformation sharing. We found that incentivizing Twitter/X users to
unfollow hyperpartisan social media influencers improved feelings toward the out-party by
23.5% compared to the control group, with effects persisting for at least six months. Unfollowing
also led participants to engage with more accurate news and increased satisfaction with their
Twitter/X feeds—without reducing engagement. This study demonstrates the long-term causal
impact of exposure to hyperpartisan influencers. Moreover, unlike other social media reduction
interventions, unfollowing is a targeted approach: like a scalpel, it surgically removes a few
harmful parts of one’s feed, allowing the beneficial aspects to remain.
Abstract Word Count: 139 words
Word Count: 4,732
Steve Rathje: 0000-0001-6727-571X
Clara Pretus: 0000-0003-2172-1184
Trisha Harjani: 0000-0002-5829-9485
Jon Roozenbeek: 0000-0002-8150-9305
James K. He: 0000-0002-1859-4914
Kurt Gray: 0000-0001-5816-2676
Sander van der Linden: 0000-0002-7144-806X
Jay J. Van Bavel: 0000-0002-2520-0442

2
Introduction
With more than 5 billion social media users worldwide (1), there is widespread concern about the
potential role of social media in fostering partisan animosity or intergroup conflict (2–5). Yet,
researchers are divided about whether and how social media plays a causal role in shaping
affective polarization, or negative feelings toward the out-party (6,7). Experiments testing the
causal effect of reducing social media usage on out-group attitudes have yielded mixed results
(8–10), which makes it difficult to draw broad conclusions about the consequences of social
media use (2). The current paper aims to understand which specific aspects of social media might
foster partisan animosity and how this can be mitigated.
Rather than looking at overall social media usage, it may be more productive to examine how
specific ways of using social media contribute to polarization. Some have proposed that social
media may increase polarization by sorting people into “echo chambers” that predominantly
showcase content from like-minded sources (11–13). However, the evidence for the polarizing
impact of online echo chambers is mixed (14–17). For instance, one study found that breaking
open echo chambers by having people follow a bot that retweeted accounts from the opposing
political party increased ideological polarization (18).
Exposure to polarizing content may be more harmful than echo chambers alone. A growing body
of work suggests that out-group animosity (19–21) and other forms of hostile content (22–26)
tend to go “viral” online, even though most people do not want this to be the case (21,27,28).
Divisive social media posts may reinforce negative stereotypes about one’s out-group (17,20)
and distort social perceptions about how extreme the out-party is (15,29,30). This may be why
breaking open echo chambers online can sometimes be unproductive: exposure to polarizing
messages—whether they come from inside or outside one’s echo chamber—may foster partisan
animosity.
A growing body of research has found that toxic and false content disproportionately stems from
a very small number of highly influential social media accounts (31–33), including political
elites, media sources, and other highly prominent accounts (often called “influencers”) (34).
While there has been correlational work on the effects of following “hyperpartisan” influencers
(12), or accounts that frequently share politically-biased, misleading, or false information (35,
36), the causal effects of exposure to these influencers on beliefs and behavior are unclear. Since
prior work suggests that a decent amount of online polarization is merely driven by people
self-selecting into polarizing online communities (37–39), and several high-powered experiments
that altered aspects of one’s social media feed have yielded null results (40,41), it is not obvious
that following these influencers has a meaningful causal impact. Testing the effect of
unfollowing hyperpartisan influencers allows us to explore a key hypothesized process by which
social media may causally contribute to partisan animosity, and may also inform interventions
for reducing partisan animosity and misinformation spread (42–46).
Overview
We conducted three studies to test the impact of (un)following hyperpartisan influencers on
partisan animosity and news-sharing behavior. Since blatant misinformation on social media is

3
rare (47,48), we focused on hyperpartisan influencers who share a combination of
politically-biased, misleading, and false news. In a correlational study, we found that following
hyperpartisan Twitter/X influencers was associated with partisan animosity. We also validated
that these hyperpartisan influencers tend to share low-quality news sources and use toxic
language.
Then, in a pilot field experiment (Experiment 1), we randomly assigned Twitter/X users to
unfollow the hyperpartisan influencers we validated in our correlational analysis and follow
non-political accounts that tweet about science and nature (e.g., @NewsfromScience) for one
month. This follow/unfollow treatment reduced out-party animosity, improved well-being, and
led to increased satisfaction with participants’ Twitter/X feed.
Next, we replicated and extended these results in a pre-registered, large-scale field experiment
with a more targeted sample of political Twitter/X users who unfollowed a greater proportion of
hyperpartisan accounts (Experiment 2). This study replicated the effects of unfollowing on
partisan animosity and satisfaction with one’s Twitter/X feed, and separately found that
following science accounts improved well-being, primarily by increasing feelings of awe. The
effect of unfollowing on out-party animosity persisted for at least six months—and eleven
months among the 58% of participants who did not refollow accounts. Unfollowing also
improved the quality of news accounts participants shared and reduced the amount of political
content participants reported seeing one year later. Both experiments were pre-registered, and
data, code, materials, and pre-registrations can be found on OSF:
https://osf.io/e3jfk/?view_only=8c89baef081a43d1bcfa0827173789ff.
Experiment 1 - Follow/Unfollow Experiment
Selecting Accounts to Follow/Unfollow. For both experiments, we selected hyperpartisan
accounts for participants to unfollow (60 accounts in Experiment 1, and 112 accounts in
Experiment 2). We selected these accounts using a combination of existing lists of media bias
and news source quality (such as AllSides,MediaBiasFactCheck, and NewsGuard), focusing on
both biased and low-quality sources. We also selected political “influencers” (e.g, highly
followed accounts) (34) and politicians for people to unfollow, focusing primarily on politicians
who had higher “falsity” scores (e.g., had made a number of false fact-checked claims) according
to an existing dataset (49). We selected a politically-balanced set of accounts for participants to
unfollow (e.g., 30 conservative-leaning accounts and 30 liberal-leaning accounts in Experiment
1). We also prioritized selecting accounts that tweeted frequently and were highly followed at the
time of the experiment.
We then conducted a correlational analysis to test whether following these hyperpartisan
accounts was associated with partisan animosity. Specifically, we analyzed a dataset of Twitter/X
data linked to survey data (n= 1,477, Mage = 39.34, SDage = 12.53, 768M, 458F, 30
non-binary/other, 1011 liberal, 254 conservative), collected via an app called “Have I Shared
Fake News” that was widely shared online (50). Using the now discontinued free Twitter/X API,
we downloaded the follower lists of the Twitter/X users who used this app and answered
questions about their partisan animosity. The number of conservative-leaning hyperpartisan
accounts a participant followed in this sample was negatively associated with favorability toward

4
the Democratic party, r= -0.25, 95% CI = [-0.29, -0.20], p< 0.001, and the number of
liberal-leaning hyperpartisan accounts a participant followed was negatively associated with
favorability toward the Republican party, r= -0.18, 95% CI = [-0.23, -0.13], p < 0.001. In
Experiment 2, we conducted a more detailed correlational analysis to examine the content posted
by these hyperpartisan accounts.
We also selected 18 accounts for people to follow that primarily tweeted about science, nature,
and space and 8 “placebo” accounts (e.g., @Walmart) for the control conditions to unfollow. See
Supplementary Appendix S1 for more details on account selection, and Supplementary
Appendices S2-S7 for a full list of accounts people were asked to follow and unfollow in both
studies.
Procedure. For this pilot experiment, we recruited United States participants using Twitter/X ads
targeting users who were following accounts similar to those in our list of accounts to unfollow
using Twitter’s “follower look-alike” settings, as well as via Twitter posts sent from the study
authors. We randomly assigned 60% of participants to unfollow all hyperpartisan accounts that
they were currently following and follow all science accounts they were not yet following for
one month in exchange for an incentive of $8.50 (in addition to $21 total for completing all
surveys). While the follow and unfollow treatments were bundled into one condition in the
Experiment 1 pilot, this was not true for the Experiment 2 replication. 40% of participants were
assigned to a control condition, in which they were instructed to unfollow the “placebo”
accounts.
Participants filled out pre- and post-treatment surveys immediately before or immediately after
the one-month experiment. These surveys included 1) two questions that asked how participants
felt toward the opposing party and one’s own party over the past month using “feeling
thermometers” (9), 2) 10 questions that assessed participants’ subjective well-being over the past
month, and 3) 12 questions assessing how positive people felt toward their Twitter/X feed over
the past month. We focused on participants’ feelings over the past month since this was the
length of the intervention. See Supplementary Appendix S8 for the specific wording of these
measures. We also included a number of secondary outcome variables (the results for all
secondary outcome variables are reported in Supplementary Appendix S9). Our final sample
consisted of 494 participants (308 male, 124 female, 24 non-binary/transgender/other, Mage =
28.14, SDage = 9.06, 393 Democrat, 100 Republican) who filled out pre- and post-treatment
surveys, passed at least one attention check, and provided valid Twitter/X handles. View
Supplementary Appendix S1 for the extended procedure and descriptive statistics.
Compliance. Compliance with the follow/unfollow treatment was measured each week via the
Twitter/X Application Programming Interface (API). Twitter/X posting data was also retrieved
each week from participants via the Twitter/X API. Averaging across the four weeks of the
experimental treatment, participants in the treatment condition followed 8.73 (95% CI = [7.73,
9.72]) more science accounts compared to the control condition, t(370.78) = 17.24,p< .001, d=
1.79, and followed 0.79 (95% CI = [0.09, 1.50]) fewer hyperpartisan accounts compared to the
control condition, t(391.65) = 2.23, p= 0.027, d= 0.23. In other words, participants in this pilot
experiment followed about nine accounts and unfollowed about one account (in Experiment 2,
people unfollowed a much larger proportion of accounts due to more targeted recruiting).

5
Results. We ran mixed-effects models that probed for an interaction between condition
(treatment vs. control) and time (pre-treatment vs. post-treatment) with random intercepts for
participants for all main outcome variables. Standardized beta coefficients are reported for ease
of interpretation. The follow/unfollow treatment improved feelings toward the out-party from
pre-treatment to post-treatment, β =0.13, 95% CI = [0.02, 0.24], p= 0.025. It did not
significantly impact feelings toward the in-party (p= 0.448), but it did reduce affective
polarization (in-party feelings - out-party feelings), β =-0.16, 95% CI = [-0.30, 0.03], p= 0.020.
The treatment also improved well-being, β =0.18, 95% CI = [0.04, 0.31], p= 0.011, and led
people to view their Twitter/X feeds more positively, β =0.16, 95% CI = [0.01, 0.31], p= 0.039.
The treatment did not significantly improve the quality of news shared on Twitter/X during the
treatment month (all ps > 0.101). See Supplementary Appendix S9 for all regression models
and robustness checks.
Experimental 2 - Large-Scale Replication with Longitudinal Results
We conducted a pre-registered replication study in a larger, more targeted sample of participants
who followed a higher number of hyperpartisan accounts pre-treatment. This replication also
included larger incentives for compliance, an expanded list of accounts to unfollow, and
longitudinal measures that tested the long-term effects of the experiment. Furthermore, due to the
bundled follow and unfollow treatments in Experiment 1, we separately tested the effects of
following and unfollowing together versus just unfollowing in Experiment 2.
Selecting Accounts to Unfollow. For the large-scale replication, we expanded the list of
hyperpartisan influencers to 112 accounts (56 conservative-leaning and 56 liberal-leaning). For
this study, we closely focused on having participants unfollow accounts that were considered
low-quality by NewsGuard, an organization that rates the news quality of an extensive list of
sources (49,51). For example, we had people unfollow left-leaning accounts such as
@PalmerReport, which has a NewsGuard score of 34.5 out of 100 (with 100 indicating most
trustworthy), and had people unfollow right-leaning accounts such as @BreitbartNews, which
has a NewsGuard score of 49.5 out of 100. The average NewsGuard quality score of accounts we
asked people to unfollow was 54.42 out of 100 (a score below 60 is considered “untrustworthy”
according to NewsGuard). As in Experiment 1, following the hyperpartisan accounts was
negatively associated with favorability toward the opposing party for both the conservative (r=
-0.25, 95% CI = [-0.30, -0.21], p< .001) and liberal-leaning accounts (r= -0.17, 95% CI =
[-0.22, -0.12], p< .001). We also selected a list of 17 science accounts that was similar to the one
in Experiment 1 for participants to follow. The average NewsGuard score of these science
accounts was 100.
Analysis of Content Posted by Accounts. To better understand the type of content removed
from and added to participant’s feeds, we analyzed all tweets posted by the 112 hyperpartisan
accounts we asked a group of participants to unfollow and the 17 science accounts we asked
another group of participants to follow during the month of the experimental treatment. The
average NewsGuard quality score of URLs posted by the hyperpartisan accounts was
significantly lower (M= 59.17, SD = 27.47) than the average NewsGuard quality score of URLs
posted by the science accounts (M = 99.87, SD = 1.49), t(10561.21) = 145.26, p< .001, d= 2.09.
We also analyzed the toxicity of language posted by the hyperpartisan and science-based

6
accounts using the Perspective API toxicity classifier (52), finding that the hyperpartisan
accounts posted significantly more toxic language (M = 0.10, SD = 0.13) than the science
accounts (M = 0.04, SD = 0.07), t(4720.08) = 41.65, p < .001, d= 0.56. This validates that the
hyperpartisan accounts shared low-quality news and used toxic language, whereas the science
accounts did not.
Recruitment. Participants were recruited solely via Twitter/X advertisements that targeted
individuals who were following partisan accounts similar to the ones we wanted people to
unfollow using Twitter/X’s “follower look-alike” targeted advertisement settings. 7,339
Twitter/X users took a pre-recruitment survey where they supplied their Twitter/X handles. We
invited 2,446 individuals who 1) followed at least one hyperpartisan account or 2) shared at least
one low-quality news URL (as rated by NewsGuard) on their timeline prior to the treatment to
participate in the study. 1,317 individuals began the study, and we included in our analyses the
1,133 participants (Mage = 45.17, 640 M, 450 F, 29 non-binary/other, 803 Democrat, 287
Republican) who completed both the pre- and post-treatment surveys. We found no evidence of
differential attrition across experimental conditions (see Supplementary Appendix S1).
Procedure. Participants were randomly assigned to one of three conditions. In the two unfollow
conditions (conditions one and two), participants were again asked to unfollow any accounts
from our list that they were currently following. In the follow/unfollow condition (condition
one), participants were additionally asked to follow all of the science accounts we selected. In
the control condition (condition three), participants were asked to unfollow the same eight
irrelevant “placebo” accounts as before. To assess the long-term effects of the intervention, we
sent out follow-up surveys one month, six months, and 11 months after the treatment was
completed. Twitter/X data was collected from participants once a week starting one week before
the treatment and up to 12 weeks after the treatment (view Supplementary Appendix S1 for the
extended procedure).
Compliance. Compliance with the experimental treatments was considerably higher than in
Experiment 1, likely because of the targeted recruitment strategy, the higher incentives for
compliance, and the more extensive list of accounts to unfollow (see Figure 1). During the
treatment month, participants in the unfollow conditions followed 9.11 (95% CI = [7.89, 10.33])
fewer hyperpartisan accounts on average compared to the control condition, t(478.07) = 14.69, p
< .001, d= 1.06. The distribution of hyperpartisan accounts unfollowed varied widely, with one
participant unfollowing 57 accounts (see Supplementary Appendix S1 for histograms of the
distributions of accounts followed/unfollowed). During the treatment month, participants in the
follow/unfollow condition followed 15.18 (95% CI = [14.75, 15.61]) more science accounts on
average than the participants in the control condition, t(355.36) = 69.46, p< .001, d= 5.37. This
means that participants in the two unfollow conditions unfollow unfollowed more than nine
accounts on average, and participants in the follow/unfollow condition followed more than 15
accounts on average.

7
Figure 1. Panel A shows a network visualization of Experiment 2 participants (black dots) assigned to the unfollow
conditions, the hyperpartisan accounts they followed (red dots), and a random subset of the other accounts they
followed (gray dots) one week before the treatment, during the final week of the treatment, and ten weeks after the
treatment. Panel B shows a network visualization of Experiment 2 participants in the follow condition (black dots),
the science accounts they followed (blue dots), and a random subset of other accounts they followed (grey dots) one
week before the treatment, during the final week of the treatment, and ten weeks after the treatment. Panel C shows
the mean number of hyperpartisan accounts followed two weeks before, four weeks during, and up to twelve weeks
after the treatment in the unfollow and control conditions. Panel D shows the mean number of science accounts
followed two weeks before, four weeks during, and up to twelve weeks after the treatment in the follow and control
conditions. On average, during the treatment month, participants in the unfollow conditions were following 9.11
fewer hyperpartisan accounts than those in the control, and participants in the follow condition were following 15.18
more science accounts than those in the control. Only 42% of participants chose to re-follow at least one of the
hyperpartisan accounts when we explicitly gave them the opportunity to do so at the end of the treatment month, and
only 26% chose to unfollow at least one science account. Error bars in panels Cand Drepresent 95% confidence
intervals.
Pre-Registered Main Outcomes. Following our pre-registration, we collapsed the two unfollow
conditions and compared them to the control condition for our main analysis (see additional
analyses broken down by condition in Supplementary Appendix S10). Replicating Experiment 1,
the unfollow conditions improved feelings toward the out-party as compared to the control
condition, β =0.14, 95% CI = [0.05, 0.24], p= 0.003 (see Figure 3). More specifically, feelings
toward the out-party increased by 17.6% in the experimental condition and decreased by 5.9% in
the control condition, representing a net 23.5% increase in out-party attitudes in the experimental
condition relative to control. The treatment had a stronger impact on out-party animosity for
conservatives as compared to liberals (see Supplementary Appendix S11 for moderation

8
analyses), contrasting with prior work that shows smaller effects of many interventions for
conservatives (53,54).
As in Experiment 1, feelings toward the in-party once again remained unchanged, β =0.01, 95%
CI = [-0.07, 0.10], p= 0.751, and the polarization index (feelings toward the out-party minus
feelings toward the in-party) was non-significant but directionally similar, β =-0.07, 95% CI =
[-0.15, 0.010], p= 0.089. Replicating Experiment 1, participants in the unfollow conditions
reported a more positive Twitter/X feed, β =0.13, 95% CI = [0.03, 0.22], p= 0.013. The
unfollow conditions had no significant impact on well-being (p= 0.752).
Figure 2. Unfollowing hyperpartisan accounts on social media durably reduced out-party animosity, with effects
persisting up to six months after the treatment was completed. It also reduced the amount of political content people
reported seeing in their feed up to eleven months after the treatment was completed, or one year after the experiment
began. Panel A shows intent-to-treat effects of the two unfollow treatments on the main pre-registered outcome
variables at the immediate follow-up, the one-month follow-up, the six-month follow-up, and the 11-month
follow-up. Panel B shows intent-to-treat effects for variables that measured content seen in one’s feed at the
immediate follow-up and all subsequent follow-ups. Error bars represent 95% confidence intervals.
Follow treatment. The follow/unfollow condition improved well-being compared to the
unfollow only treatment, β =0.15, 95% CI = [0.04, 0.26], p= 0.009, but did not affect any of the

9
other main dependent variables (all ps > 0.349). This suggests that the well-being effect in
Experiment 1 may have been driven by following science accounts instead of unfollowing
hyperpartisan accounts.
Enduring effects. The unfollow treatment still had a lasting effect on out-party animosity when
measured one month after the treatment was over, β =0.13, 95% CI = [0.02, 0.23], p= 0.017
(see Figure 2). The other main outcome variables remained unchanged after one month (all ps >
0.110). The unfollow treatment continued to have a significant effect on out-party animosity
when measured six months after the treatment ended, β =0.14, 95% CI = [0.02, 0.26], p= 0.019.
The unfollow treatment’s impact on out-party animosity was no longer significant when
measured at the 11-month follow-up (p= 0.238). Exploratory analysis found that these long-term
effects were largely driven by the subset of participants who did not refollow hyperpartisan
accounts (see Supplementary Appendix S10 for additional analysis of this subset). Exploratory
analysis further found that, for the subset of participants who did not refollow hyperpartisan
accounts (as compared to the control condition), the reduction in out-party animosity was still
significant at the 11-month follow-up, β = 0.13, 95% CI = [0.00, 0.27], p= 0.049 (see
Supplementary Appendix S10).
Effects on information diet. People who unfollowed hyperpartisan accounts reported seeing
less false news (β = -0.14, 95% CI = [-0.25, -0.02], p= 0.021), less content about politics (β =
-0.31, 95% CI = [-0.44, -0.19], p< 0.001), less news (β = -0.26, 95% CI = [-0.38, -0.13], p<
0.001), and fewer posts from politicians (β = -0.50, 95% CI = [-0.63, -0.38], p< 0.001) (see
Figure 2 and Supplementary Appendix S10). Eleven months after the treatment ended (and one
year after the experiment began), participants in the unfollow treatment still reported seeing less
political content (β = -0.16, 95% CI = [-0.32, -0.01], p= 0.036) and fewer politicians (β = -0.24,
95% CI = [-0.39, -0.09], p< 0.001). To provide a more detailed picture of which aspects of
people’s feeds changed, Figure 3 shows specific items of the positive perceptions of one’s feed
index and the well-being index (see Supplementary Appendix S10 for full regression results).

10
Figure 3. Effects of follow and unfollow conditions on feelings toward one’s feed immediately post-treatment.
(Panel A) People in the unfollow conditions found their feeds to be more positive, more intriguing, less divisive,
and less sad compared to the control condition. (Panel B) People in the follow/unfollow condition (compared to the
unfollow only condition) perceived their feeds to be significantly more informative and more educational. (Panel C)
While the unfollow only treatment did not impact well-being, (Panel D) people in the follow/unfollow condition
(compared to the unfollow only treatment) reported feeling increased well-being, which was primarily driven by
increased feelings of awe, joy, curiosity, and fulfillment. See Supplementary Appendix S10 for full results.
Behavioral Effects. To analyze the change in quality of news shared on Twitter/X, we used
difference-in-difference models that tested for an interaction between condition and time (one
month before vs. one-month after), with standard errors clustered at the user-level, following past
analyses of Twitter news sharing behavior (55). The pre-registered outcome variables of interest
were the NewsGuard ratings of news URLs (e.g., nytimes.com) liked (or favorited) and reposted
(or retweeted) by participants, as well as the NewsGuard ratings of news accounts (e.g.,
@NYTimes) liked and posted by participants. The unfollow treatment did not significantly
improve the quality of news links liked or posted by participants during the month of the
treatment (ps > 0.087). However, it did improve the quality of news accounts that participants
retweeted, β =0.45, 95% CI = [0.09, 0.82], p= 0.016, and liked, β =0.90, 95% CI = [0.40,
1.39], p< 0.001 (see Figure 4 and Supplementary Appendix S10). The different results for links
versus accounts may be explained by the fact that we specifically had participants unfollow
accounts that were rated as low-quality by NewsGuard. In other words, the news account
variable may have been more closely related to (and thus more impacted by) the experimental
treatment.

11
Figure 4. The unfollow treatment improved the quality of news accounts posted (or tweeted) and liked (or favorited)
on Twitter/X (as determined by NewsGuard) during the month of the treatment compared to the month before. The
unfollow treatment did not, however, improve the quality of news URLs posted or liked. Error bars represent 95%
confidence intervals.
Long-Term Changes to Twitter/X. This experiment occurred during a period (March
2023-March 2024) in which Twitter/X was undergoing many design changes following Elon’s
Musk’s purchase of the platform on October 28, 2022. For instance, in April 2023, Twitter/X
removed “verified” checkmarks, which confirm the identifies of prominent people or
organizations, and laid off 80% of its staff. Twitter/X was officially re-named to X in July 2023.
Many speculated that these changes may have led to an increase in hostile and false content on
the platform (56).
In line with these speculations, exploratory analyses found that participants (regardless of
experimental condition) reported their Twitter/X feeds to be less positive over the course of the
year-long study, β = -0.34, 95% CI = [-0.45, -0.24], p< 0.001, and saw less reliable content in
their feeds, β = -0.47, 95% CI = [-0.60, -0.33], p< 0.001. Participants also reported using
Twitter/X much less frequently over the course of the year-long study, β = -0.54, 95% CI =
[-0.68, -0.41], p< 0.001 (see Figure 5). Importantly, the unfollow treatment (as compared to the
control) did not lead to a significant decrease in self-reported Twitter/X engagement at any time
point (all ps > 0.239), nor did it lead to a decrease in the number of likes or Twitter/X posts (all
ps > 0.490) (see Supplementary Appendix S10 for all regression models). Instead, the decrease
in engagement was observed across all participants, regardless of experimental condition.

12
Figure 5. Regardless of experimental condition, there were several long-term changes that occurred from when this
experiment began (March 2023) to when this experiment was completed (March 2024). Specifically, people used
Twitter/X less frequently (Panel A), viewed their Twitter/X feeds as less positive (Panel B), and viewed the content
in the feeds as less reliable (Panel C). This study was conducted during a period of notable change to Twitter/X,
which was re-named to X during the study. The unfollow condition did not reduce self-reported engagement with
Twitter/X at any time point, nor did it affect the number of likes or posts recorded by participants during the
treatment month. All y-axes represent five-point Likert scales. Specific scale wording is reported in Supplementary
Appendix S8. Error bars represent 95% confidence intervals.
To see if these changes were reflected in behavior as well, we conducted exploratory analysis to
investigate whether the quality of news people posted and liked on Twitter/X declined after Elon
Musk’s purchase of the platform (an “Elon Musk” effect). We ran regression models with a
binary variable indicating whether a tweet was posted or liked before or after before or after
October 28, 2022—the day Elon Musk purchased Twitter/X. As in our prior Twitter/X analysis,
standard errors were clustered at the user-level. We used all tweets and favorites that we were
able to collect from participants using the Twitter/X API up until the free API was shut down in
July 2023 (57). The average NewsGuard quality score of news links (β = -0.21, [-0.35, -0.07], p
= 0.004) and news accounts (β = -0.39, 95% CI = [-0.60, -0.17], p < 0.001) shared on Twitter/X
was significantly lower after, as opposed to before, Elon’s Musk’s purchase of the platform. We
found similar effects for the NewsGuard quality score of URLs liked (β = -0.22, 95% CI =
[-0.39, -0.06], p= 0.009), and found a non-significant effect trending in the same direction for
news accounts favorited (β = -0.31, 95% CI = [-0.77, 0.15], p= 0.182). While we cannot causally
attribute any of these trends to Elon Musk’s platform changes, they support speculations about
the increase in misinformation spread after his purchase of the platform (56).
Discussion
In a correlational study and two pre-registered field experiments, we found that following
hyperpartisan social media influencers is associated with out-party animosity, and that
unfollowing these influencers reduces out-party animosity. The effect of unfollowing persisted at

13
least six months after the experiment was completed—and nearly one year after the experiment
was completed for the majority of participants who chose to not re-follow the hyperpartisan
accounts. This work illustrates the long-term, causal impact of exposure to hyperpartisan content
on attitudes and behaviors in a real-world setting.
Unfollowing also improved the quality of news accounts people interacted with, led people to
feel more satisfied with their Twitter/X feeds, and reduced the amount of political content people
reported seeing in their feeds a full year later. Following influencers who shared content about
science increased well-being, primarily by increasing feelings such as awe and curiosity.
Unfollowing hyperpartisan accounts did not reduce overall Twitter/X usage (both measured via
self-report and Twitter/X behavior), even though self-reported engagement with Twitter/X
declined steeply over the course of this experiment—which happened shortly after Elon Musk’s
purchase of the platform.
We also observed that the perceived reliability of news participants saw on Twitter/X and the
perceived positivity of participants’ feeds decreased over the course of the year-long experiment.
These changes were also reflected in behavior, with participants sharing lower quality news after
Elon Musk’s purchase of the platform. While we cannot make causal claims, this exploratory
analysis supports speculations that the spread of misinformation and harmful content increased
after Elon Musk's purchase of the platform (56).
The long-term effects of unfollowing stand in stark contrast to the typical fleeting effects of
many “light-touch” interventions in the social and behavioral sciences (58,59). Even though
unfollowing only took a few moments, it was a structural intervention that caused lasting
changes to participants’ daily information diet for at least a year (59–61). Supplementary
analysis (Supplementary Appendix S11) indicated that participants used Twitter/X for more than
an hour each day on average, and used Twitter/X more than other social media platforms and
media sources. Targeting Twitter/X “power-users” who follow more political elites than the
average user (62) may have also contributed to the strength of this intervention. Focusing on
these politically-active “power-users” may also be important since the content they share may
have spillover effects on their followers.
In contrast with prior analysis suggesting that much of online polarization is driven by
self-selection (37–39), this study finds one’s daily social media information diet has an enduring
causal impact on partisan animosity and news sharing behavior. Moving beyond experiments that
estimate the average effect of social media usage among the general population (6–8), this
experiment reveals a specific process (e.g., following hyperpartisan influencers) by which social
media might contribute to partisan animosity.
Future work should further explore the psychological processes underlying the effects of this
intervention. Prior work has shown that exaggerated out-party “meta-perceptions,” or inflated
beliefs about how much one’s out-group dislikes one’s in-group, can contribute to partisan
animosity (63,64). Exposure to polarizing content about one’s out-group on social media may
cause individuals to perceive their out-group as more extreme and polarized than it actually is
(15,30), which, in turn, can increase partisan animosity. Consistent with this account, mediation

14
analysis found that the treatment’s effect on partisan animosity was significantly mediated by
out-group meta-perceptions (see Supplementary Appendix S10).
This work may shed light on why prior research (18) has found that breaking open one’s echo
chamber by following cross-partisan elites backfires: political elites on social media frequently
use divisive rhetoric (20,22), which may make people think their opposing party is more
extreme than it actually is. Instead of merely breaking open one’s online echo chamber, this
research suggests a better anecdote to polarization may be to avoid exposure to hyperpartisan
social media voices altogether and seek out more constructive political voices instead.
In addition to this theoretical contribution, this work also has practical applications for
interventions (43,45). Social media platforms could promote healthier information diets through,
for instance, altering their algorithms and changing other design features. Since most people
report wanting social media feeds to be less divisive (27), these changes would align people’s
feeds with their stated preferences. Individual social media users could also curate their social
media following list to include less hyperpartisan content without any cooperation from social
media companies.
It could be argued that unfollowing political elites or news accounts—even hyperpartisan
ones—could lead to negative consequences, such as reduced political knowledge or engagement.
Assuaging these concerns, we found null effects of the intervention on news knowledge, political
attitudes, social media engagement, and a number of other variables (see Supplementary
Appendix S11). Other work, however, has demonstrated the beneficial effects of exposure to
news on social media (65), and even noted the potential upsides of online outrage for spurring
political participation and activism (66). Thus, there might be a tradeoff between reducing
exposure to low-quality, toxic, and hyperpartisan content and increasing exposure to high-quality
news and important political information.
Conclusions
While many studies have highlighted the potential benefits of eliminating social media entirely
(8–10), this approach is all-or-nothing: like a sledgehammer, it smashes both the harmful and
beneficial aspects of one's online experience. In contrast, unfollowing hyperpartisan accounts is a
more precise approach: like a scalpel, it surgically removes the negative aspects of social media
while preserving the positive aspects. We hope this research will lead to a more nuanced
discussion about the specific components of social media that cause harm and the ways that
social media can be improved.

15
References
1. Number of worldwide social network users 2028, Statista.
https://www.statista.com/statistics/278414/number-of-worldwide-social-network-users/.
2. J. J. Van Bavel, S. Rathje, E. Harris, C. Robertson, A. Sternisko, How social media shapes
polarization. Trends in Cognitive Sciences (2021).
3. E. Kubin, C. von Sikorski, The role of (social) media in political polarization: a systematic
review. Annals of the International Communication Association 45, 188–206 (2021).
4. E. Harris, S. Rathje, C. Robertson, J. J. Van Bavel, The SPIR Model of Social Media and
Polarization: Exploring the Role of Selection, Platform Design, Incentives, and Real-World
Context. International Journal of Communication (In Press).
5. J. Haidt, Why the Past 10 Years of American Life Have Been Uniquely Stupid. The Atlantic
11 (2022).
6. P. Lorenz-Spreen, S. Lewandowsky, C. R. Sunstein, R. Hertwig, How behavioural sciences
can promote truth, autonomy and democratic discourse online. Nature Human Behaviour,
1–8 (2020).
7. L. Boxell, M. Gentzkow, J. M. Shapiro, Greater Internet use is not associated with faster
growth in political polarization among US demographic groups. Proceedings of the
National Academy of Sciences 114, 10612–10617 (2017).
8. N. Asimovic, J. Nagler, R. Bonneau, J. A. Tucker, Testing the effects of Facebook usage in
an ethnically polarized setting. Proceedings of the National Academy of Sciences 118
(2021).
9. H. Allcott, L. Braghieri, S. Eichmeyer, M. Gentzkow, The welfare effects of social media.
American Economic Review 110, 629–76 (2020).
10. H. Allcott, M. Gentzkow, W. Mason, A. Wilkins, P. Barberá, T. Brown, J. C. Cisneros, A.
Crespo-Tenorio, D. Dimmery, D. Freelon, S. González-Bailón, A. M. Guess, Y. M. Kim, D.
Lazer, N. Malhotra, D. Moehler, S. Nair-Desai, H. Nait El Barj, B. Nyhan, A. C. Paixao De
Queiroz, J. Pan, J. Settle, E. Thorson, R. Tromble, C. Velasco Rivera, B. Wittenbrink, M.
Wojcieszak, S. Zahedian, A. Franco, C. Kiewiet De Jonge, N. J. Stroud, J. A. Tucker, The
effects of Facebook and Instagram on the 2020 election: A deactivation experiment. Proc.
Natl. Acad. Sci. U.S.A. 121, e2321584121 (2024).
11. C. R. Sunstein, # Republic: Divided Democracy in the Age of Social Media (Princeton
University Press, 2018).
12. S. Rathje, J. K. He, J. Roozenbeek, J. J. Van Bavel, S. van der Linden, Social media
behavior is associated with vaccine hesitancy. PNAS Nexus (2022).
13. M. Cinelli, G. D. F. Morales, A. Galeazzi, W. Quattrociocchi, M. Starnini, The echo
chamber effect on social media. Proceedings of the National Academy of Sciences 118
(2021).
14. B. Nyhan, J. Settle, E. Thorson, M. Wojcieszak, P. Barberá, A. Y. Chen, H. Allcott, T.
Brown, A. Crespo-Tenorio, D. Dimmery, Like-minded sources on Facebook are prevalent
but not polarizing. Nature, 1–8 (2023).
15. C. Bail, Breaking the Social Media Prism (Princeton University Press, 2021).
16. S. B. Hobolt, K. Lawall, J. Tilley, The polarizing effect of partisan echo chambers.
American Political Science Review, 1–16 (2023).
17. P. Törnberg, How digital media drive affective polarization through partisan sorting. Proc.
Natl. Acad. Sci. U.S.A. 119, e2207159119 (2022).

16
18. C. A. Bail, L. P. Argyle, T. W. Brown, J. P. Bumpus, H. Chen, M. F. Hunzaker, J. Lee, M.
Mann, F. Merhout, A. Volfovsky, Exposure to opposing views on social media can increase
political polarization. Proceedings of the National Academy of Sciences 115, 9216–9221
(2018).
19. X. Yu, M. Wojcieszak, A. Casas, Partisanship on Social Media: In-Party Love Among
American Politicians, Greater Engagement with Out-Party Hate Among Ordinary Users.
Polit Behav, doi: 10.1007/s11109-022-09850-x (2023).
20. S. Rathje, J. J. Van Bavel, S. van der Linden, Out-group animosity drives engagement on
social media. Proc Natl Acad Sci USA 118, e2024292118 (2021).
21. G. Heltzel, K. Laurin, Why Twitter sometimes rewards what most people disapprove of:
The case of cross-party political relations. (2024).
22. J. A. Frimer, H. Aujla, M. Feinberg, L. J. Skitka, K. Aquino, J. C. Eichstaedt, R. Willer,
Incivility is rising among American politicians on Twitter. Social Psychological and
Personality Science, 19485506221083811 (2022).
23. W. J. Brady, A. P. Gantman, J. J. Van Bavel, Attentional capture helps explain why moral
and emotional content go viral. Journal of Experimental Psychology: General 149, 746
(2020).
24. D. Antypas, A. Preece, J. Camacho-Collados, Negativity spreads faster: A large-scale
multilingual twitter analysis on the role of sentiment in political communication. Online
Social Networks and Media 33, 100242 (2023).
25. C. E. Robertson, N. Pröllochs, K. Schwarzenegger, P. Pärnamets, J. J. Van Bavel, S.
Feuerriegel, Negativity drives online news consumption. Nature Human Behaviour, 1–11
(2023).
26. J. J. Van Bavel, C. E. Robertson, K. Del Rosario, J. Rasmussen, S. Rathje, Social Media
and Morality. Annu. Rev. Psychol. 75, 311–340 (2024).
27. S. Rathje, C. Robertson, W. J. Brady, J. J. Van Bavel, People think that social media
platforms do (but should not) amplify divisive content. Perspectives on Psychological
Science (2023).
28. S. Milli, M. Carroll, S. Pandey, Y. Wang, A. D. Dragan, Twitter’s Algorithm: Amplifying
Anger, Animosity, and Affective Polarization. arXiv preprint arXiv:2305.16941 (2023).
29. W. J. Brady, K. McLoughlin, T. N. Doan, M. J. Crockett, How social learning amplifies
moral outrage expression in online social networks. Sci. Adv. 7, eabe5641 (2021).
30. C. Robertson, K. del Rosario, J. J. Van Bavel, Inside the Funhouse Mirror Factory: How
Social Media Distorts Perceptions of Norms. (2024).
31. N. Grinberg, K. Joseph, L. Friedland, B. Swire-Thompson, D. Lazer, Fake news on Twitter
during the 2016 US presidential election. Science 363, 374–378 (2019).
32. M. R. DeVerna, R. Aiyappa, D. Pacheco, J. Bryden, F. Menczer, Identifying and
characterizing superspreaders of low-credibility content on Twitter. arXiv
arXiv:2207.09524 [Preprint] (2024). http://arxiv.org/abs/2207.09524.
33. D. Kumar, J. Hancock, K. Thomas, Z. Durumeric, “Understanding the Behaviors of Toxic
Accounts on Reddit” in Proceedings of the ACM Web Conference 2023 (ACM, Austin TX
USA, 2023; https://dl.acm.org/doi/10.1145/3543507.3583522), pp. 2797–2807.
34. R. Diresta, Invisible Rulers: The People Who Turn Lies into Reality (2023;
https://www.hachettebookgroup.com/titles/renee-diresta/invisible-rulers/9781541703377/?l
ens=publicaffairs).
35. G. Pennycook, D. G. Rand, Fighting misinformation on social media using crowdsourced

17
judgments of news source quality. Proceedings of the National Academy of Sciences 116,
2521–2526 (2019).
36. M. Potthast, J. Kiesel, K. Reinartz, J. Bevendorff, B. Stein, A Stylometric Inquiry into
Hyperpartisan and Fake News. arXiv arXiv:1702.05638 [Preprint] (2017).
http://arxiv.org/abs/1702.05638.
37. I. Waller, A. Anderson, Quantifying social organization and political polarization in online
platforms. Nature 600, 264–268 (2021).
38. R. E. Robertson, J. Green, D. J. Ruck, K. Ognyanova, C. Wilson, D. Lazer, Users choose to
engage with more partisan news than they are exposed to on Google Search. Nature, 1–7
(2023).
39. S. González-Bailón, D. Lazer, P. Barberá, M. Zhang, H. Allcott, T. Brown, A.
Crespo-Tenorio, D. Freelon, M. Gentzkow, A. M. Guess, S. Iyengar, Y. M. Kim, N.
Malhotra, D. Moehler, B. Nyhan, J. Pan, C. V. Rivera, J. Settle, E. Thorson, R. Tromble, A.
Wilkins, M. Wojcieszak, C. K. De Jonge, A. Franco, W. Mason, N. J. Stroud, J. A. Tucker,
Asymmetric ideological segregation in exposure to political news on Facebook. Science
381, 392–398 (2023).
40. A. M. Guess, N. Malhotra, J. Pan, P. Barberá, H. Allcott, T. Brown, A. Crespo-Tenorio, D.
Dimmery, D. Freelon, M. Gentzkow, Reshares on social media amplify political news but
do not detectably affect beliefs or opinions. Science 381, 404–408 (2023).
41. A. M. Guess, N. Malhotra, J. Pan, P. Barberá, H. Allcott, T. Brown, A. Crespo-Tenorio, D.
Dimmery, D. Freelon, M. Gentzkow, How do social media feed algorithms affect attitudes
and behavior in an election campaign? Science 381, 398–404 (2023).
42. R. Hartman, W. Blakey, J. Womick, C. Bail, E. J. Finkel, H. Han, J. Sarrouf, J. Schroeder, P.
Sheeran, J. J. Van Bavel, Interventions to reduce partisan animosity. Nature human
behaviour 6, 1194–1205 (2022).
43. S. Rathje, S. van der Linden, “Shifting online incentive structures to reduce polarization
and the spread of misinformation” in Research Handbook on Nudges and Society (Edward
Elgar Publishing, 2023;
https://www.elgaronline.com/edcollchap/book/9781035303038/book-part-9781035303038-
14.xml), pp. 91–108.
44. S. Van Der Linden, Y. Kyrychenko, A broader view of misinformation reveals potential for
intervention. Science 384, 959–960 (2024).
45. S. van der Linden, Misinformation: susceptibility, spread, and interventions to immunize
the public. Nat Med, 1–8 (2022).
46. S. Rathje, J. Roozenbeek, J. J. Van Bavel, S. van der Linden, Accuracy and social
motivations shape judgements of (mis) information. Nature human behaviour, 1–12 (2023).
47. A. Guess, J. Nagler, J. Tucker, Less than you think: Prevalence and predictors of fake news
dissemination on Facebook. Science advances 5, eaau4586 (2019).
48. C. Budak, B. Nyhan, D. M. Rothschild, E. Thorson, D. J. Watts, Misunderstanding the
harms of online misinformation. Nature 630, 45–53 (2024).
49. M. Mosleh, D. G. Rand, Measuring exposure to misinformation from political elites on
Twitter. nature communications 13, 7144 (2022).
50. I. Lapowski, Newsguard wants to fight fake news with humans, not algorithms. Wired,
August 23 (2018).
51. H. Lin, J. Lasser, S. Lewandowsky, R. Cole, A. Gully, D. G. Rand, G. Pennycook, High
level of correspondence across different news domain quality rating sets. PNAS nexus 2,

18
pgad286 (2023).
52. M. Avalle, N. Di Marco, G. Etta, E. Sangiorgio, S. Alipour, A. Bonetti, L. Alvisi, A. Scala,
A. Baronchelli, M. Cinelli, Persistent interaction patterns across social media platforms and
over time. Nature 628, 582–589 (2024).
53. S. Rathje, Letter to the editors of Psychological Science: Meta-analysis reveals that
accuracy nudges have little to no effect for US conservatives: Regarding Pennycook et al.
(2020). Psychological Science (2022).
54. C. Martel, S. Rathje, C. J. Clark, G. Pennycook, J. J. Van Bavel, D. G. Rand, S. Van Der
Linden, On the Efficacy of Accuracy Prompts Across Partisan Lines: An Adversarial
Collaboration. Psychol Sci 35, 435–450 (2024).
55. A. Bor, M. Osmundsen, S. H. R. Rasmussen, A. Bechmann, M. B. Petersen, “
Fact-checking” videos reduce belief in misinformation and improve the quality of news
shared on Twitter. (2020).
56. S. L. Myers, S. A. Thompson, T. Hsu, The Consequences of Elon Musk’s Ownership of X,
The New York Times (2023).
https://www.nytimes.com/interactive/2023/10/27/technology/twitter-x-elon-musk-anniversa
ry.html.
57. S. Rathje, To tackle social-media harms, mandate data access for researchers. Nature 633,
36–36 (2024).
58. C. K. Lai, M. Marini, S. A. Lehr, C. Cerruti, J. L. Shin, J. A. Joy-Gaba, B. A. Nosek, A
comparative investigation of 17 interventions to reduce implicit racial preferences. Journal
of Experimental Psychology: General 143, 1765–1785 (2014).
59. E. L. Paluck, R. Porat, C. S. Clark, D. P. Green, Prejudice reduction: Progress and
challenges. Annual Review of Psychology 72 (2020).
60. A. M. Guess, P. Barberá, S. Munzert, J. Yang, The consequences of online partisan media.
Proc. Natl. Acad. Sci. U.S.A. 118, e2013464118 (2021).
61. D. Broockman, J. Kalla, Consuming cross-cutting media causes learning and moderates
attitudes: A field experiment with Fox News viewers. Journal of Politics (2024).
62. M. Wojcieszak, A. Casas, X. Yu, J. Nagler, J. A. Tucker, Most users do not follow political
elites on Twitter; those who do show overwhelming preferences for ideological congruity.
Sci. Adv. 8, eabn9418 (2022).
63. S. L. Moore-Berg, L.-O. Ankori-Karlinsky, B. Hameiri, E. Bruneau, Exaggerated
meta-perceptions predict intergroup hostility between American political partisans.
Proceedings of the National Academy of Sciences (2020).
64. K. Ruggeri, B. Većkalov, L. Bojanić, T. L. Andersen, S. Ashcroft-Jones, N. Ayacaxli, P.
Barea-Arroyo, M. L. Berge, L. D. Bjørndal, A. Bursalıoğlu, The general fault in our fault
lines. Nature Human Behaviour 5, 1369–1380 (2021).
65. S. Altay, E. Hoes, M. Wojcieszak, News on Social Media Boosts Knowledge, Belief
Accuracy, and Trust: A Field Experiment on Instagram and WhatsApp.
66. V. L. Spring, C. D. Cameron, M. Cikara, The upside of outrage. Trends in Cognitive
Sciences 22, 1067–1069 (2018).

19
Funding:
Russell Sage Foundation (SR, JVB)
Tanner Verhey/Google Trust and Safety (JR, SVL, SR)
Center for the Science of Moral Understanding (SR)
Templeton World Charity Foundation; TWCF-2023-31570 (SR, JVB,
doi.org/10.54224/31570)
AE Foundation (SR, JVB)
National Science Foundation, Grant #2334148 (SR, JVB)
National Science Foundation SBE Postdoctoral Research Fellowship, Grant #2404649
Gates Cambridge Scholarship; Grant #OPP1144 (SR)
Heterodox Academy (SR, JVB)
Author contributions:
Conceptualization: SR, JVB
Methodology: SR, CP, TH, JVB
Investigation: SR, CP, TH, JVB
Visualization: SR, JKH, CP
Funding acquisition: SR, JR, KG, SVL, JVB
Project administration: SR
Supervision: JVB, KG, SVL, JVB
Writing – original draft: SR
Writing – review & editing: SR, CP, TH, JKH, JR, KG, SVL, JVB
Competing interests: Authors declare no competing interests.
Data and materials availability: De-identified data, code, replication materials, and
pre-registrations, are available on our OSF:
https://osf.io/e3jfk/?view_only=8c89baef081a43d1bcfa0827173789ff. Note that limited
Twitter/X data could be shared for privacy reasons.
Acknowledgements: We thank Tanner Verhey for feedback on this manuscript, Dan-Mircea
Mirea for help with statistical analysis and R code, and Yara Kyrychenko for help with network
analysis. We also are thankful for comments from the Social Identity and Morality Lab, the
Social Decision-Making Lab, and the Polarization and Social Change Lab.
Supplementary Materials
Supplementary Section S1. Extended Methods
Supplementary Table S2. List of Accounts to Unfollow – Experiment 1
Supplementary Table S3. List of Accounts for Participants to Follow – Experiment 1
Supplementary Table S4. List of “Placebo” Accounts for Participants to Unfollow
Supplementary Table S5. Conservative Accounts Unfollowed - Experiment 2
Supplementary Table S6. Liberal Accounts Unfollowed - Experiment 2
Supplementary Table S7. Science Accounts Followed - Experiment 2
Supplementary Section S8. Main Outcome Variable Phrasing
Supplementary Section S9. Experiment 1 Regression Models
Supplementary Section S10. Experiment 2 Regression Models
Supplementary Section S11. Extended Results