Cameron Martel’s research while affiliated with Massachusetts Institute of Technology and other places

Individuals preferentially reciprocate connections with copartisans versus counter-partisans online. However, the mechanisms underlying this partisan bias remain unclear. Do individuals simply prefer viewing politically congenial content, or do they additionally prefer socially connecting with copartisans? Is this driven by preference for in-party ties or distaste for out-party ties? In a Twitter (now called X) field experiment, we created bot accounts varying by partisanship and whether they identified as bots or humans. We randomly assigned Twitter users (N = 3,013) to be followed by one of these accounts. We found evidence for social motivation—users were much more likely to reciprocate links to copartisan relative to counter-partisan accounts when the accounts identified as humans versus bots. We also found evidence for both in-party preference and out-party dispreference—users were as likely to follow back copartisan accounts as they were unlikely to followback counter-partisan accounts, compared to politically neutral accounts. A follow-up survey experiment (N = 990) provides further evidence for distinct roles of issue polarization, out-party animosity, and in-party affinity in moderating follow-back decisions online.

Warning labels from professional fact-checkers are one of the most widely used interventions against online misinformation. But are fact-checker warning labels effective for those who distrust fact-checkers? Here, in a first correlational study (N = 1,000), we validate a measure of trust in fact-checkers. Next, we conduct meta-analyses across 21 experiments (total N = 14,133) in which participants evaluated true and false news posts and were randomized to either see no warning labels or to see warning labels on a high proportion of the false posts. Warning labels were on average effective at reducing belief in (27.6% reduction), and sharing of (24.7% reduction), false headlines. While warning effects were smaller for participants with less trust in fact-checkers, warning labels nonetheless significantly reduced belief in (12.9% reduction), and sharing of (16.7% reduction), false news even for those most distrusting of fact-checkers. These results suggest that fact-checker warning labels are a broadly effective tool for combatting misinformation.

Rapid advancements in generative AI may transform many aspects of society. However, little research has examined the public’s attitudes and opinions towards generative AI. This is critical for developing democratically informed policy. In a nationally representative survey of Americans (n = 1,497; July-August 2023), we query participants across three topics relevant to new AI tools: (a) their familiarity and usage with these technologies, (b) whether they perceived these tools to be positive or negative across a variety of applications, and (c) the extent of their agreement with several important policy positions towards AI. First, we find overall low reported usage of generative AI technologies – 60% of participants reported they have never used any AI tool. Second, we observe substantial differences in the perceived positivity of AI on society across specific applications. Americans were positive towards using AI to predict extreme weather events and write software, but thought that using AI to write news articles, draft legal documents, and classify misinformation would be net negative for society. Democrats and individuals with greater familiarity and usage of AI perceived generative AI impacts as more positive across applications. Third, we find overall low trust in new AI – and we also find overall support, on average, for policies endorsing transparency and caution when proceeding with developing new AI tools. Americans’ views towards generative AI vary across potential use cases and individuals’ familiarity with AI, demographics, and politics – all of which should be considered when developing new policies towards developing and regulating AI.

Debates on how tech companies ought to oversee the circulation of content on their platforms are increasingly pressing. In the U.S., questions surrounding what, if any, action should be taken by social media companies to moderate harmfully misleading content on topics such as vaccine safety and election integrity are now being hashed out from corporate boardrooms to federal courtrooms. But where does the American public stand on these issues? Here we discuss the findings of a recent nationally representative poll of Americans’ views on content moderation of harmfully misleading content.

Content moderation is a critical aspect of platform governance on social media and of particular relevance to addressing the belief in and spread of misinformation. However, current content moderation practices have been criticized as unjust. This raises an important question – who do Americans want deciding whether online content is harmfully misleading? We conducted a nationally representative conjoint survey experiment (N=3,000) in which U.S. participants evaluated the legitimacy of hypothetical content moderation juries tasked with evaluating whether online content was harmfully misleading. These moderation juries varied on whether they were described as consisting of experts (e.g., domain experts), laypeople (e.g., social media users), or non-juries (e.g., computer algorithm). We also randomized features of jury composition (size, necessary qualifications) and whether juries engaged in discussion during content evaluation. Overall, participants evaluated expert juries as more legitimate than layperson juries or a computer algorithm. However, modifying layperson jury features helped increase legitimacy perceptions – nationally representative or politically balanced composition enhanced legitimacy, as did increased size, individual juror knowledge qualifications, and enabling juror discussion., Maximally legitimate layperson juries were comparably legitimate to expert panels. Republicans perceived experts as less legitimate compared to Democrats, but still more legitimate than baseline layperson juries. Conversely, larger lay juries with news knowledge qualifications who engaged in discussion were perceived as more legitimate across the political spectrum. Our findings shed light on the foundations of procedural legitimacy in content moderation and have implications for the design of online moderation systems.

There is strong political assortment of Americans on social media networks. This is typically attributed to preferential tie formation (i.e., homophily) amongst those with shared partisanship. Here we demonstrate an additional factor beyond homophily driving assorted networks: preferential prevention of social ties. In two field experiments on Twitter, we created human looking bot accounts that identified as Democrats or Republicans, and then randomly assigned users to be followed by one of these accounts. In addition to preferentially following-back co-partisans, we found that users were 11 times more likely to block counter-partisan accounts compared to co-partisan accounts in the first experiment, and 5 times more likely to block counter-partisan accounts relative to a neutral account or a co-partisan account in the second experiment. We then replicated these findings in a survey experiment and found evidence of a key motivation for blocking: wanting to avoid seeing any content posted by the blocked user. Additionally, we found that Democrats preferentially blocked counter-partisans more than Republicans, and that this asymmetry was likely due to blocking accounts who post low-quality or politically slanted content (rather than an asymmetry in identity-based blocking). Our results demonstrate that preferential blocking of counter-partisans is an important phenomenon driving political assortment on social media.

The spread of misinformation is a pressing societal challenge. Prior work shows that shifting attention to accuracy increases the quality of people’s news-sharing decisions. However, researchers disagree on whether accuracy-prompt interventions work for U.S. Republicans/conservatives and whether partisanship moderates the effect. In this preregistered adversarial collaboration, we tested this question using a multiverse meta-analysis ( k = 21; N = 27,828). In all 70 models, accuracy prompts improved sharing discernment among Republicans/conservatives. We observed significant partisan moderation for single-headline “evaluation” treatments (a critical test for one research team) such that the effect was stronger among Democrats than Republicans. However, this moderation was not consistently robust across different operationalizations of ideology/partisanship, exclusion criteria, or treatment type. Overall, we observed significant partisan moderation in 50% of specifications (all of which were considered critical for the other team). We discuss the conditions under which moderation is observed and offer interpretations.

False news can manipulate public opinion, stir up fear and hatred, and undermine the credibility of legitimate news sources. Although many studies have examined false news sharing, there has been no comprehensive, comparative, and computational investigation of the interventions that can reduce this harmful behavior. To do so, we introduce a novel experimental method, the Dynamic Semi-Integrative Approach (DSIA). DSIA involves testing multiple interventions, individual- and item-level moderators, and computational choice modeling (drift–diffusion modeling) in a single framework. By applying DSIA to false news, we find that warning labels and media literacy interventions are particularly effective at increasing news sharing accuracy, followed by a social norm intervention. Accuracy prompts were least effective. Intervention effects were consistent across individual- and item-level characteristics, such as age, analytical thinking, and the political-lean of news items, suggesting wide applicability. The interventions operated via different decision-making processes, suggesting that each intervention engages distinct mental processes to attenuate false news sharing. By developing and applying DSIA, we provide a uniquely detailed insight into false news interventions and establish DSIA as a promising, scalable approach for future experimental research.

The spread of misinformation is a pressing societal challenge. Prior work shows that shifting attention to accuracy increases the quality of people’s news sharing decisions. However, researchers disagree whether accuracy prompt interventions work for U.S. Republicans/conservatives and whether partisanship moderates the effect. In this pre-registered adversarial collaboration, we tested this question using a multiverse meta-analysis (k=21; N=27,828). In all 70 models, accuracy prompts improved sharing discernment among Republicans/conservatives. We observed significant partisan moderation for single-headline “evaluation” treatments (a critical test for one research team), such that the effect was stronger among Democrats than Republicans. However, this moderation was not consistently robust across different operationalizations of ideology/partisanship, exclusion criteria, or treatment type. Overall, we observed significant partisan moderation in 50% of specifications (all of which were considered critical for the other team). We discuss the conditions under which moderation is observed and offer interpretations.

Warning labels from professional fact-checkers are one of the most widely used interventions against online misinformation. Prior work suggests that such warning labels are effective at reducing the belief and spread of false content on average. However, there is substantial distrust of fact-checkers, particularly among those on the political right. Does this distrust undermine the effectiveness of fact-checks? In the current work, we investigate this question empirically. In a first correlational study (N=1,000), we establish and validate a measure of trust in fact-checkers. We confirm that more Republican-favoring participants are less trusting of fact-checkers, and also find that skill-based traits such as procedural news knowledge and analytic thinking exacerbate this partisan asymmetry. Next, we conduct meta-analyses across 21 experiments (total N=14,133) in which participants evaluated actual true and false news posts and were randomized to either see no warning labels (control) or to see warning labels on a high proportion of the false posts. We find that warning labels were on average effective at reducing belief in (27.6% reduction), and sharing of (24.7% reduction), false headlines. While warning effects were somewhat smaller for participants with less trust in fact-checkers, warning labels nonetheless significantly reduced belief in, and sharing of, false news even for those most distrusting of fact-checkers. Together, these results suggest that fact-checker warning labels are a broadly effective tool for combatting misinformation.