PaperlayerView exact revision on arXiv

Source-linked AI summary

Exploring Lightweight Interventions at Posting Time to Reduce the Sharing of Misinformation on Social Media

Farnaz Jahanbakhsh, Amy X. Zhang, Adam J. Berinsky, Gordon Pennycook, David G. Rand, David R. Karger

arXiv:2101.11824v5cs.HC

TL;DR

Social media’s engagement-focused design and low-friction sharing can leave users sharing without considering accuracy. This paper develops a taxonomy of reasons for believing or disbelieving claims and tests posting-time accuracy and rationale prompts. The prompts reduce sharing of false content more than true content, while also reducing true sharing.

  • Problem

    Engagement-focused systems and low barriers to sharing can leave users sharing posts without much attention to accuracy or potential consequences.

    Source: §1 · p. 2

  • Method

    The paper develops a taxonomy of self-reported reasons for claim accuracy and tests lightweight posting-time prompts for accuracy assessment and rationale.

    Source: §10 · p. 26

  • Results

    Accuracy assessment and rationale prompts reduce sharing of both true and false content, but sharing of false content decreases more.

    Source: §7 · p. 24, §10 · p. 26

  • Takeaways & Limitations

    Structured user assessments could be surfaced as credibility signals to friends encountering a post.

    Source: §10 · p. 26

    The taxonomy study lacked a checklist-only condition, so it could not determine whether the checklist alone could replace free-text rationale.

    Source: §9 · p. 26

Abstract

from arXiv · show

When users on social media share content without considering its veracity, they may unwittingly be spreading misinformation. In this work, we investigate the design of lightweight interventions that nudge users to assess the accuracy of information as they share it. Such assessment may deter users from posting misinformation in the first place, and their assessments may also provide useful guidance to friends aiming to assess those posts themselves. In support of lightweight assessment, we first develop a taxonomy of the reasons why people believe a news claim is or is not true; this taxonomy yields a checklist that can be used at posting time. We conduct evaluations to demonstrate that the checklist is an accurate and comprehensive encapsulation of people's free-response rationales. In a second experiment, we study the effects of three behavioral nudges -- 1) checkboxes indicating whether headings are accurate, 2) tagging reasons (from our taxonomy) that a post is accurate via a checklist and 3) providing free-text rationales for why a headline is or is not accurate -- on people's intention of sharing the headline on social media. From an experiment with 1668 participants, we find that both providing accuracy assessment and rationale reduce the sharing of false content. They also reduce the sharing of true content, but to a lesser degree that yields an overall decrease in the fraction of shared content that is false. Our findings have implications for designing social media and news sharing platforms that draw from richer signals of content credibility contributed by users. In addition, our validated taxonomy can be used by platforms and researchers as a way to gather rationales in an easier fashion than free-response.

1 Introduction

The paper addresses sharing systems that prioritize engagement and allow posts to circulate without much attention to accuracy. It develops a rationale taxonomy and investigates lightweight posting-time prompts, finding that accuracy assessment and rationale reduce sharing of false content more than true content.

1 Introduction

  • Engagement-focused systems and low barriers to sharing leave the platform design that contributes to misinformation’s spread unaddressed.A corrective comment can itself disseminate an incorrect post when the system treats engagement as a positive signal.Source: §1 · p. 2
  • The paper explores lightweight posting-time prompts: assess accuracy, select a reason from a checklist, or write a short rationale.The prompts are intended to encourage reflection without requiring extensive training or research before sharing.Source: §1 · p. 2

1.1 Research Questions

  • The first study develops and iteratively refines categories for people’s reasons for believing or disbelieving claims; the second tests how accuracy and rationale prompts affect sharing intentions.The taxonomy is intended to provide a checklist that can replace free-text rationales.Source: §1 · pp. 2–3, §1.1 · p. 3
  • Accuracy prompts and free-text rationales reduce sharing of both true and false headlines, with false sharing reduced more; adding a checklist does not further improve the false-to-true sharing ratio over free text.Structured reasons may still help platforms capture credibility signals.Source: §1.1 · p. 3
  • The taxonomy is also proposed as a way to collect structured assessments that could be aggregated and shown to a poster’s friends.The paper gives an example of displaying how many friends think a post is true or false and the reasons they report.Source: §1.1 · p. 3

2 Related Work

Prior work examines misinformation’s spread, responses to it, and factors shaping perceived accuracy. This paper extends that work by studying lightweight accuracy and rationale prompts and by developing a structured taxonomy of people’s stated reasons.

2.1 Spread of Misinformation

  • False content has been found to diffuse farther, faster, deeper, and more broadly than truth, while accuracy reminders can reduce intentions to share false information.The paper extends this line of work by testing accuracy assessments and rationales together as potential posting-time interventions.Source: §2.1 · p. 4

2.2 Measures to Counter Misinformation

  • Existing countermeasures include corrections, fact-checking, moderation, and media-literacy efforts, while platform interventions can place platforms in a truth-arbitration role.The authors instead explore friend-sourced assessments, with platforms delivering assessments rather than making them.Source: §2.2 · pp. 4–5

2.3 Why People Believe News

  • Perceived accuracy can be influenced by prior exposure, source cues, and topic or author information, while warning flags may raise credibility ratings for unflagged stories.The paper focuses on people’s self-reported reasons for believing or disbelieving claims and iteratively revises a taxonomy for untrained users.Source: §2.3 · p. 5
  • The paper tests whether structured rationale prompts improve discernment compared with unguided free-form reasoning, while noting that structured reasons could be integrated as credibility signals.The improved-discernment comparison is presented as a hypothesis, not an established finding in this section.Source: §2.3 · p. 6

3 Terminology and Methods

The paper evaluates posting-time interventions by separating changes in sharing of true and false content and by focusing on the ratio of their sharing rates. It distinguishes objective veracity from users’ perceived accuracy.

3.1 Performance Metrics for Sharing Interventions

  • The paper prioritizes changes in the false-to-true sharing-rate ratio over changes in the absolute difference between rates.An intervention can leave the difference unchanged yet improve the composition of the shared output stream.Source: §3.1 · p. 6
  • For a balanced input stream, reducing true and false sharing rates from 20% and 10% to 15% and 5% changes the shared output from two-thirds true to three-quarters true.The example illustrates that the ratio can improve even when the absolute difference between the sharing rates does not.Source: §3.1 · p. 6
  • Reducing the false-to-true sharing ratio carries a cost because it also reduces sharing of true content; platforms may invoke an intervention on only some shares.The paper notes that the ratio of false to true content in the output depends on the input ratio and the intervention’s false-to-true discernment ratio.Source: §3.1 · p. 6

3.2 Veracity and Perceived Accuracy

  • The minimal intervention asks users whether a claim is accurate, requiring enough deliberation to answer; the paper expects greater deliberation generally to improve assessments, bounded by users’ knowledge and rationality.The authors state that available or inaccurate knowledge and limits in rationality prevent perfect assessment.Source: §3.2 · pp. 6–7
  • Veracity means whether a claim is accurate independently of users’ knowledge, while perceived accuracy is a user’s initial judgment and may imperfectly correlate with objective accuracy.Average perceived accuracy is the fraction of users who perceive a claim as accurate.Source: §3.2 · p. 7

4 Experimental and Study Design

The study comprises a taxonomy study and a nudge study of accuracy and rationale prompts. Participants assessed claims presented in a headline-like format, and the claim pool varied in veracity, domain, partisanship, and source.

4 Experimental and Study Design

  • The taxonomy study develops categories for reasons people believe or disbelieve claims, while the nudge study tests whether accuracy assessments and rationale prompts affect sharing intentions.The nudge study compares free-text reasoning with structured reason categories.Source: §4 · p. 7

4.1 Claims

  • Claims came mainly from Snopes, varied in veracity, domain, partisanship, and source, and were presented with an image, lede, and source.Snopes claims used the site’s fact-checker ground truth; the few mainstream-media claims were checked by investigating their referenced sources.Source: §4.1 · p. 7, Figure 1 · p. 8
  • Political claims were refreshed across taxonomy-study iterations, and the nudge study used a separate set that substantially overlapped with the final taxonomy iteration.The authors selected claims with an identified originating source or that Snopes described as widespread rumors, and excluded nonfactual items such as satire.Source: §4.1 · p. 7

4.2 Participants

  • The studies recruited U.S.-based Mechanical Turk participants: 317 contributed non-spam taxonomy responses and 1,668 were non-spam nudge participants.Of participants providing demographic information across both studies, 42% were female, with a median age of 35.Source: §4.2 · p. 8

5 RQ1: Developing a Taxonomy of Reasons People Believe or Disbelieve Claims

The study developed and iteratively refined a first-person taxonomy for labeling reasons people believe or disbelieve news claims. The resulting categories distinguish evidence-based judgments, plausibility, and speculation, while separating some credibility signals from accuracy judgments.

5.1 Procedure

  • At the third iteration, agreement reached 0.63 across 729 idea units from 48 claims, exceeding the recommended threshold.The researchers refined categories through multiple rounds, including requiring an explanation for each selected category; the study spanned 13 months.Source: §5.1 · p. 9

5.2 Results

  • The taxonomy groups reasons as evidence for or against accuracy, plausibility or implausibility, and speculation labeled “Don’t know.”The “Don’t know” elaborations were omitted in the final iteration because participants tended to repeat the category labels.Source: §5.2 · p. 9

5.2.1.3 The Claim Appears Motivated or Biased.

  • The taxonomy separates misleading presentation and distrust of a source from accuracy judgments, while perceived agendas often served as grounds for refuting partisan claims.Participants sometimes inferred an agenda from prior familiarity with a topic or source.Source: §5.2 · pp. 9–10, §5.2.1.3 · p. 11, §5.3 · pp. 13–14

5.2.1.1 The Claim Is (Not) Consistent with My Past Experiences and Observations.

  • Consistency with past experience covered personal observations, perceived patterns, customary expectations, and familiarity; familiarity was merged into this category when participants could not distinguish them.Repeated exposure could make a claim seem plausible even when its accuracy had not been established.Source: §5.2.1.1 · p. 10

5.2.1.2 The Claim Appears to Be Inaccurate Based on Presentation.

  • Sensational language, grammatical errors, clickbait, and poor artifacts served as presentation-based reasons to refute claims; participants saw only headlines, which may explain why these cues were not used to affirm accuracy.Source: §5.2.1.2 · p. 10

5.2.2 Common Misalignments and Transformation of the Categories.

  • Category use could diverge from intended meanings: confidence affected whether prior exposure counted as familiarity or experience, and participants’ interpretations prompted revisions to expertise and firsthand-knowledge categories.Source: §5.2.2.1 · p. 13, §5.2.2.3 · p. 13

5.2.3 Demographics.

  • The distribution of rationales differed statistically across participant genders.Source: §5.2.3 · p. 13

5.3 Discussion of Taxonomy

  • Consistency with past experience is weak evidence for truth, though participants often treated it as sufficient; designers might instead classify it as uncertainty, and the taxonomy could benefit from further refinement.Treating lack of refutation as accuracy may also create opportunities for manipulation.Source: §5.3 · p. 14

6 RQ2: Effects of Providing Accuracy Reasons on Sharing Behavior (Nudge study).

This section examines how accuracy assessment and reasoning interventions affect sharing intentions for true and false headlines, including checklist-based reasoning and differentiation by perceived accuracy. It discusses sharing intentions as a proxy for perceived accuracy while noting that other factors also shape sharing.

6.1 Method

  • Among 1,668 participants, the study analyzed sharing intentions for true and false headlines across accuracy-assessment and reasoning conditions using mixed-effects models.Participants saw up to 10 claims drawn from 54 headlines; the veracity model included participant and claim random effects, while the perceived-accuracy model used conditions 2–4.Source: §6.1 · p. 14, §6.2 · p. 16, §6.2.1.1 · p. 16

6.2.2.1 Effect of Providing Accuracy Assessments.

  • Accuracy assessment reduced sharing of true headlines by 18% and false headlines by 37%, lowering the ratio of false to true shared headlines.The accuracy-by-veracity interaction was not significant (χ2(1) = 3.22, p = 0.07).Source: §6.2.2.1 · p. 17, Figure 5 · p. 18

6.2.2.2 Effect of Providing Reasoning.

  • Providing reasoning reduced sharing of false headlines by 27% and true headlines by 14%, but did not reduce sharing of headlines participants initially judged false.Reasoning reduced sharing of headlines participants judged true; sharing of those judged false was already low (6% in condition 2, 5% in condition 3).Source: §6.2.2.2 · pp. 17–18, Figure 5 · p. 18

6.2.2.3 Effect of Reasoning Format.

  • Checkbox reasoning reduced sharing versus free-text reasoning by 17% for false headlines and 18% for true headlines, without further differentiating sharing by objective veracity.Checkboxes reduced sharing of headlines participants perceived as true, but not those perceived as false; perceived-false sharing was already rare (5% in condition 3, 4% in condition 4).Source: Figure 5 · p. 18, §6.2.2.3 · pp. 18–19, Figure 6 · p. 19

6.2.3 Reasons for Sharing Content Perceived as False.

  • Among 427 explanations for sharing headlines perceived as false, common reasons included entertainment (22%), fact-checking first (20%), and sparking debate (17%).Other responses included warning friends that a claim was false (11%), fact-checking it (10%), and sharing in case it was true (9%); a few knowingly shared claims to advance a view or had changed their accuracy judgment.Source: §6.2.3 · pp. 19–20

6.2.4.1 Confidence.

  • Reasoning and checklist prompts affected sharing across confidence levels for false headlines misjudged as true, while for correctly judged true headlines their effect appeared at higher confidence.Participants reported confidence before giving reasons, so reflecting afterward may have lowered confidence.Source: Figure 7 · p. 20, §6.2.4.1 · p. 21

6.2.4.2 Average Perceived Accuracy.

  • Sharing increased with a headline’s average perceived accuracy, and treatment groups showed steeper slopes and narrower confidence intervals than control; demographic effects were explored in Appendix B.The authors note that slopes were below 1, so sharing likelihood rose more slowly than perceived accuracy.Source: Figure 8 · p. 22, §6.2.4.2 · p. 22, §6.3 · p. 23

6.3 Discussion of Nudge Study Findings.

  • The authors note a possible Hawthorne effect but cite prior studies, including a Twitter intervention, as evidence that accuracy nudges may generalize to social platforms.They also suggest structured checklists can capture and surface reasons even though they did not further distinguish sharing of true from false content.Source: §6.3 · p. 23

6.3 Discussion of Nudge Study Findings

  • Prior work found a correlation between hypothetical survey-reported news sharing and actual social-media sharing, supporting implementation of these nudges to encourage informed news propagation and consumption.The authors argue that their results generalize to actual social-media platforms.Source: §6.3 · p. 24
  • Differences in spam rates may have affected the data, but the authors consider similar spamming unlikely on social media, where spam itself could signal a sharer’s credibility.The spam-rate discussion concerns the makeup of data collected on a paid-worker platform.Source: §6.3 · p. 24
  • Dubious credibility was the primary stated reason for not sharing, although topic interest, sensitivity, platform habits, information overload, and partisan considerations also influenced intentions.Participants’ accuracy assessments were collected before the reasoning interventions, so the discussion compares sharing intentions—not those assessments—to understand intervention effects.Source: §6.3 · p. 24
  • Sharing intentions, rather than accuracy assessments, are used to understand intervention effects; Figure 8 shows share probability rises with perceived claim accuracy.The control group did not provide accuracy assessments, and the treatment groups’ assessments came before the reasoning interventions.Source: §6.3 · p. 24

7 Discussion

The discussion argues that posting-time accuracy and rationale prompts can reduce the share of false content, while also reducing true sharing. It proposes low-friction user-level interventions and structured rationale signals as ways to support content filtering and credibility assessment.

7 Discussion

  • Accuracy and rationale prompts reduce the ratio of false to true content participants intend to share, at the cost of curbing some true sharing.The authors judge this preferable to leaving misinformation unchecked, while noting possible engagement or profit trade-offs.Source: §7 · p. 24

7.1 Design Implication

  • Platform moderation often acts after misinformation spreads, while platform incentives may not align with users’ interests.Policy moderation also faces gray areas that can prompt ad-hoc decisions and criticism.Source: §7.1 · p. 25
  • The interventions add a low barrier to sharing and redirect attention toward accuracy rather than social feedback and engagement.The authors describe them as usable on existing platforms and alternative publishing models such as WikiTribune.Source: §7.1 · p. 25
  • Structured rationale tags could serve as filterable credibility metadata, and the taxonomy was reliably usable by people untrained in expert credibility signifiers.The authors propose that users might filter for posts friends with relevant domain knowledge have evaluated as true.Source: §7.1 · p. 25

8 Future Work

The future-work agenda extends the study to users’ reactions to rationale tags, optional rather than required assessments, and reasons people share stories they may not believe. It also identifies questions about tailoring nudges to users and testing them in social-network settings.

8 Future Work

  • A social-network field study could examine how readers judge rationale tags while accounting for the sharer’s influence on perceived credibility.The authors also identify credibility judgments and persuasiveness of tagged rationales as open questions.Source: §8 · p. 25
  • Future work could test optional accuracy and reasoning signals, rather than requiring users to provide them.The paper compares this possibility to how likes and upvotes are captured.Source: §8 · p. 25
  • Future work could test whether checklists of sharing intentions affect behavior and how readers use those signals when consuming posts.The authors note that people may share claims they do not believe, including to entertain or seek help fact-checking.Source: §8 · p. 26
  • Exploratory observations raise whether nudges should target habitual misinformation sharers or be applied only occasionally.The analysis was not planned a priori; the authors point to prior studies reporting similar party-related findings.Source: §8 · p. 26

9 Limitations

The taxonomy study’s checklist condition also required free-text explanations, leaving the checklist’s standalone replacement value unresolved.

9 Limitations

  • Because checklist users also explained their choices in free text, the study cannot determine whether the checklist alone matches free-text discernment.A checklist-only condition would have enabled that comparison.Source: §9 · p. 26
  • The study therefore does not establish whether requiring free text is necessary for users to interpret taxonomy categories as intended.Condition 4 used free-text explanations to assess participants’ use and understanding of the categories.Source: §9 · p. 26
  • Whether optional free text alongside the checklist better supports adoption remains an open question.The authors suggest testing a condition in which free text is optional rather than required.Source: §9 · p. 26

10 Conclusion

The conclusion reports that accuracy and reasoning nudges reduce sharing of both true and false content, with a larger decrease for false content. It also describes a reliably usable rationale taxonomy and reports exploratory demographic analyses, including a gender association that warrants caution because the tests were underpowered.

10 Conclusion

  • Repeated taxonomy-study iterations produced categories participants could reliably use to label rationales for claim accuracy.The taxonomy was developed to capture reasons for believing or disbelieving news claims in a structured format.Source: §10 · p. 26
  • Sharing of both true and false content fell under accuracy and reasoning nudges, with a larger decrease for false content.The conclusion suggests taxonomy-guided assessments could help people encountering posts filter their newsfeeds by sharers’ stated reasons.Source: §10 · pp. 26–27

A Association Between Participants’ Demographics and the Taxonomy Categories.

  • The exploratory demographic analysis used the last study iteration and excluded cases missing gender, party, or ethnicity data, leaving 953 datapoints.Of those, 645 included free-text elaboration.Source: A Association Between Participants’… · p. 30
  • Participants were grouped into Democratic or Republican categories and White or Not White categories for the demographic analysis.Education was grouped as college degree, including Associate’s degree, versus no degree.Source: A Association Between Participants’… · p. 30
  • Party and ethnicity were not significantly associated with rationale categories, while gender was associated; the tests were underpowered.Reported results were χ2(22) = 30.03, p= 0.12 for party; χ2(22) = 21.80, p= 0.47 for ethnicity; and χ2(22) = 40.87, p= 0.009 for gender.Source: A Association Between Participants’… · p. 30
  • Figure 9 compares rationale-category ratios by gender for claims perceived as accurate and inaccurate, normalized by each gender’s total rationales.The figure caption identifies the comparison and normalization, but reports no outcome.Source: A Association Between Participants’… · p. 30, Figure 9 · p. 31

B Effects of the Demographics of the Nudge Study Participants

The exploratory demographic analysis found that sharing varied with partisan concordance, party, gender, and education. The authors caution that these analyses were not planned in advance and their p-values should not be treated as definitive.

B.0 Effects of the Demographics of the Nudge Study Participants

  • The demographic analyses were exploratory, were not planned in the experiment design, and their p-values are not indicative of true significance.The model excluded 638 datapoints with incomplete demographics and 46 datapoints from participants identifying as neither male nor female; demographic-factor interactions were omitted because the sample lacked sufficient power.Source: B Effects of the Demographics of the… · pp. 31–33
  • The analysis modeled share intention with headline veracity, treatments, concordance, and demographic factors, while accounting for participant and claim as random effects.Concordance measured alignment between participants’ party and their headline-partisanship rating on a 1–5 scale.Source: B Effects of the Demographics of the… · p. 32

B.0.0.1 Concordance

B.0.0.2 Party

B.0.0.3 Gender and B.0.0.4 Education

C Cumulative Link Models

Cumulative link mixed models, used as an ordinal-response check on the linear models, yielded consistent findings. Accuracy assessment and reasoning reduced sharing likelihood, and checkbox reasons reduced it more than free-text reasons.

C Cumulative Link Models

  • The authors tested the interventions with cumulative link mixed models for ordered “No,” “Maybe,” and “Yes” sharing responses, including participant and claim as random effects.They used likelihood-ratio tests and post-hoc estimated marginal means; results were consistent with the linear models.Source: C Cumulative Link Models · pp. 35–36
  • Objective truth and perceived truth each predicted greater sharing likelihood: E(False) = 1.23 versus E(True) = 1.48, and E(Perceived as false) = 1.05 versus E(Perceived as true) = 1.58.Both effects were statistically significant in their respective models.Source: C Cumulative Link Models · p. 36
  • Participants were less likely to share when asked to assess accuracy, provide reasoning, or select checkbox reasons rather than provide free-text reasons.For the veracity model, estimated means were 1.44 versus 1.28 for accuracy not provided versus provided, 1.41 versus 1.30 for reasoning not provided versus provided, and 1.40 versus 1.32 for free-text versus checkbox reasons.Source: C Cumulative Link Models · p. 36
  • The accuracy-by-veracity interaction was statistically significant but not practically meaningful, with assessment lowering estimated sharing for both false and true headlines.Estimated means were 1.32 versus 1.15 for false headlines and 1.57 versus 1.40 for true headlines, without versus with accuracy assessment.Source: C Cumulative Link Models · p. 36

D.1 Makeup of Data and Impact of Removing Spams.

The Nudge study’s spam responses varied across conditions, raising a possible participant-composition concern. Including spam datapoints weakened some findings for heavier interventions, so the main analyses excluded them.

D.1 Makeup of Data and Impact of Removing Spams

  • Participants were assigned 10 claims, but 30–50 per condition did not finish all headlines, with dropout counts similar across conditions.The authors also considered whether condition-dependent spam rates could change which participants contributed data.Source: §D.1 · p. 36
  • Spam counts differed significantly across conditions, although spam share rates were similar at approximately 0.72 regardless of headline veracity.The authors note that differential spam rates could reflect different types of participants completing more laborious conditions.Source: §D.1 · pp. 36–37
  • Including spam datapoints made some results for heavier interventions no longer statistically significant because positive sharing responses diluted the differences between conditions.The main analyses excluded spam datapoints; adding them increased sample size but also added many positive sharing responses for both true and false headlines.Source: §D.1 · p. 37

D.2 Deliberation Priming.

In the control condition, asking participants why they would or would not share may itself have prompted deliberation about accuracy. A separate test found no significant effect of repeated accuracy judgments on whether participants’ assessments matched veracity.

D.2 Deliberation Priming.

  • The control condition’s post-decision rationale question may itself have primed deliberation, though the estimated first-versus-last interaction was nonsignificant.The veracity association was 0.16 for the last decision and 0.08 for the first; the authors describe this as some support for a priming hypothesis.Source: §D.2 · p. 37

D.3 Investigating Potential Learning Effects of Repeated Accuracy Assessments

  • Repeated accuracy judgments did not significantly change whether participants’ assessments matched headline veracity (χ2(1) = 0.07, p = 0.80).The analysis compared participants’ first and last treatment-condition datapoints.Source: §D.3 · p. 37

E Spams

Including spam datapoints, requesting accuracy assessments reduced sharing of both false and true headlines, with a larger reduction for false headlines. Reasoning and its format showed limited overall effects, while sharing patterns varied for headlines participants perceived as true or false.

E Spams

  • With spam datapoints included, participants were more likely to share objectively true than false headlines and perceived-true than perceived-false headlines.The veracity and perceived-accuracy effects were both significant in their respective models.Source: E Spams · pp. 37–38

E.1 Effect of Providing Accuracy Assessments

  • Sharing false headlines fell 29% and true headlines fell 17% when participants were asked to assess accuracy.The comparison is between conditions 1 and 2.Source: §E.1 · p. 38, Figure 18 · p. 39

E.2 Effect of Providing Reasoning

  • Reasoning had no significant overall effect on sharing, though requesting it reduced sharing of headlines perceived as true; perceived-false headline sharing changed little.The interaction between reasoning and perceived accuracy was significant, but the interaction with veracity was not.Source: §E.2 · p. 38

E.3 Effect of Reasoning Format

  • Checkbox reasoning increased sharing of headlines participants initially perceived as false compared with free-text reasoning, while reasoning format had no significant overall effect.The interaction between reasoning format and perceived accuracy was significant, whereas its interaction with veracity was not.Source: §E.3 · p. 38, Figure 19 · p. 40

F Headlines Used in the Study of Behavioral Nudges

Related papers

Loading 2101.11824v5…