This Research Compendium curates and translates University of Pennsylvania research on the quantitative study of the information ecosystem and its impact on democracy. It brings together empirical work examining issues such as information integrity, media ecosystems, political speech, and democratic resilience, with the aim of making rigorous academic research more accessible and actionable.
The compendium features research from seven schools and centers at Penn—including the Annenberg School for Communication, the Annenberg Public Policy Center, the Wharton School, SAS Political Science, SEAS Computer and Information Science, the School of Social Policy and Practice, and the Carey Law School—and is designed to serve journalists, media leaders, civil society organizations, and policymakers seeking evidence-based insights into the functioning and governance of digital information systems.
Center Newsletter
Subscribe to our (occasional) newsletter for updates with the latest research, events, and announcements from the Center.
Read an introduction to the Center from Faculty Directors Duncan Watts and Christopher Yoo – December 2025.
Read the inaugural Center newsletter from Executive Director Alex Engler – October 2025.
| Title | Penn Faculty Lead | Date | Summary and Key Insights |
|---|---|---|---|
| Computational Social Science: Past, Present, and Future | Duncan Watts | 12/11/2025 | Overview:This introductory chapter lays out Duncan Watts and David Lazer's "subjective appraisal of the field of computational social science (CSS)." They first trace the field's history beginning in the late 90s and its growth in popularity during the Web 2.0 revolution which allowed researchers access to unprecedented amounts of online data because of the growing reliance on digital tools. In 2007, Duncan Watts shared that if handled well, data about online communications could "revolutionize our understanding of collective human behavior." Incidentally, 2007 was also the year where people began interpreting CSS as a field that could "produce an understanding of the global network on which many global problems exist: SARS and infectious disease, global warming, strife due to cultural collisions, and the livability of our cities.” What is the current state of CSS? Watts and Lazer argue that computational social science (CSS) has produced a substantial number of empirical insights that would have been difficult or impossible in the pre-digital era. Examples include large-scale validation of network science theories using online social networks, measuring “structural virality” across billions of social media cascades and predicting poverty levels from mobile phone metadata. But has the field's impact extended beyond the realm of scientific publishing? The authors conclude that potential exists but the evidence thus far is "inconclusive." Looking to the future, the authors identify 5 key challenges that would shape the field in the next decade. First, the field's overreliance on industry data creates a power imbalance wherein platforms get to decide which data to share with researchers and what data remains "locked up." Second, CSS researchers will continue grappling with "the difference between what is measured and what would ideally be measured." Third, CSS should pursue solution-oriented research intended to tackle real-world problems. Finally, embracing open science and developing ethical frameworks for digital data will be essential to ensure research is reliable, transparent, and socially responsible. Takeaway:Computational social science is a "a collection of compelling and consequential problems, grounded in the rise of socio-technical systems: systems of devices, platforms, algorithms, and data that are equally social and computational, and that cannot be understood or managed through either lens in isolation of the other." |
| Brain activity explains message effectiveness: A mega-analysis of 16 neuroimaging studies | Emily B. Falk | 11/4/2025 | Overview:This paper uses a large-scale mega-analysis of 16 fMRI studies to examine the neural mechanisms underlying why some messages are more persuasive than others across domains such as health, marketing, and political communication. By pooling raw neuroimaging data across 16 separate studies, the authors test whether the neural coorelates for persuasive messaging applies at the individual level and among large groups "of message receivers who did not undergo neuroimaging." Here are some key findings: 1. Messages that elicit greater activation in brain systems associated with reward and social cognition are more likely to be effective. Moreover, message effectiveness at scale was corelated with greater activation in the VTA, a part of the dopaminergic reward system, related to "anticipation and receipt of personal rewards" and social conformity. 2. Mentalizing, which refers to the process by which "people understand themselves and the minds of others," was associated with strong effects in the "dorsomedial prefrontal cortex, and cerebellar regions." 3. Supplimentary analyses found that brain regions associated with language processing and emotions had a positive effect on message effectiveness. Takeaway:This study shows that neural indicators of reward, language and emotions processing are associated with the persuasiveness of messages both for individuals and at scale. Why is this important?Persuasive messaging shapes outcomes in public health, advertising and political communications, yet research on what makes messages effective is often siloed by domain or method. This paper contributes approach of pooling available data from 16 different neuroimaging studies shows how "certain basic mechanisms may be active across different messaging contexts and may inspire novel strategies targeting these mechanisms." |
| Identity-related Speech Suppression in Generative AI Content Moderation | Danaé Metaxa | 11/4/2025 | Overview:This paper examines how automated content-moderation systems may incorrectly suppress identity-related speech. Using both traditional short-form user-generated text and “longer generative-AI-focused data” introduced in the paper, the researchers created a benchmark to measure speech suppression for nine identity groups. Here’s what they found: 1. Identity-related speech is more frequently suppressed than other forms of speech across both traditional and gen-AI based automated moderation services. 2. Identity-related speech has a higher likelihood of being suppressed for both marginalized and non-marginalized groups, except for those identified as “straight” and/or “Christian.” 3. The reasons for speech suppression by the content moderation systems tested in this paper differ based on the stereotypes and text associations for specific identity groups. Eg: non-Christian content had a higher likelihood of being incorrectly flagged as hateful. Takeaway:This paper provides a methodological framework to test automated moderation systems for incorrect speech suppression and in doing so, shows that there will be a tradeoff between “filtering out undesired content and ensuring that other speech is allowed.” Why is this important?Gen AI is being rapidly integrated into platform moderation. Companies like Meta and Tiktok have laid off trust and safety workers and contracted moderators in favor of automated systems, potentially to cut costs and improve efficiency while also reducing the reliance on human moderators that experience severe psychological distress for being exposed to streams of problematic content. Given these incentives, it is crucial to understand the gaps and potential pitfalls of relying on LLM based content moderation. While a lot of AI safety research has looked into how to prevent these systems from producing “undesired outcomes”, less attention has been paid to “making sure appropriate text can be generated.” This paper addresses this research gap by providing “the first comprehensive bias audit of generative AI speech suppression across five automated content moderation APIs” and shows how identity-based stereotypes may permeate LLMs and inadvertently suppress permissible speech. |
| Why Depolarization is Hard: Evaluating attempts to decrease partisan animosity in America | Yphtach Lelkes | 9/23/2025 | Overview:This paper introduces two new experiments and provides a meta analysis that looks into the efficacy of online interventions designed to reduce “partisan animosity.” Here’s what they found: 1. Depolarization is difficult to implement and scale. In their meta-analysis, the researchers find that these interventions have small effects, “a 5-point shift on a 101-point scale,” and become weaker with time. 2. Efficacy of depolarization does not improve with repeated exposure, indicating that solutions to online polarization require going beyond user-level interventions. Takeaway: Online interventions aimed at “depolarization” by themselves are not a “scalable solution for reducing societal conflict” and efforts must be made to study “elite behaviors and structural incentives that fuel partisan conflict.” Why is this important?Partisan divides on both traditional and online media platforms are often framed as a consequence of “echo chambers” wherein people are less likely to view/engage with disagreeable content. The tendency for homogeneity in social discourse has led some to tout depolarization interventions as an effective solution to reduce partisan divides. But does exposure to contrary views bring people closer together? This is an empirical question with important policy implications. This paper challenges this assumption, providing empirical evidence suggesting that the gains of depolarization may be overstated and suggests shifting focus toward understanding the societal level incentives that encourage online partisanship. |
| Changing beliefs or changing behavior? Understanding the belief-to-behavior process and intervening to curb the impact of misinformation | Dolores Albarracín | 9/10/2025 | Overview:Prior research has shown that the relationship between beliefs and behavior is often weak, variable, and highly context-dependent. This paper examines this "belief-to-behavior inference model" and challenges the assumption that "misinformation must be corrected to change behavior." The authors look at both individual and societal level interventions to examine if and to what extent belief change can prompt behavior change, especially when it comes to combating misinformation. Here are a few important findings from their review: 1. Beliefs are more likely to drive behavior change when specific goals are activated and "inferential paths are short." Based on the author's proposed framework, interventions that combat misinformation like prebunking will be better at changing behaviors by focusing on outcome beliefs or include "direct calls to action." 2. The authors also provide promising alternatives to prebunking and fact-checking for combating misinformation with behavior change in mind. For instance, they find that self affirmations reduce defensiveness, especially when misinformation targets a social identity and "bypassing" which highlights alternative viewpoints without directly confronting misinformation to be effective. The important thing to keep in mind is that interventions must "shift the locus of change from belief accuracy to behavior change." 3. Legal and administrative sanctions with the goal of increasing trust in institutions have "negligible effects" on behavior change while broader societal interventions like providing support networks, and upholding social norms are important. Takeaway:To effectively combat misinformation, interventions must go beyond describing what is true or false. If the goal is to change behavior in this regard, it is critical to recognize "when, how, and why a belief matters for behavior, and when behavior must be addressed in a direct way." Why is this important?Combating misinformation is a priority for journalists, platforms and policymakers. Strategies like fact-checking is often framed as a solution to the problem. But as this paper shows, if the goal is to target behavior change, merely educating people about what is true or false may not be enough. Through this paper, the authors show that interventions are more likely to shape behavior if "beliefs are behaviorally engaged, inferentially accessible, and contextually relevant". Overall this paper provides the groundwork to ensure that future interventions to combat misinformation are "better calibrated to the realities of human. |
| More platforms, less attention to news? A multi-platform analysis of news exposure across TV, web, and YouTube in the United States | Sandra González-Bailón | 5/29/2025 | Overview:This paper explores how news consumption is shaped by a "multi-platform media environment" and whether exposure to multiple media platforms "alleviates or exacerbates observed inequalities in attention to news." To do so, the researchers tracked and analyzed news exposure across TV, the web and YouTube for 55000 unique panelists across 39 months. Results from this study provide insights into the fraction of news consumers specific to each platform and their demographic profiles. Further analyses looks at whether multi-platform news exposure affects time spent engaging with news sources. Here are some key findings: 1. TV has the largest reach when it comes to news exposure, with 80% of panelists reporting seeing at least one news channel. Less than half of the panelists reported accessing news via the web while only 5% of overall visits to YouTube were to access news sources. 2. When looking at time spent accessing news, the study finds that the "the skewness of the distribution becomes more prominent as we move from TV to the web and to YouTube." This trend indicates the possibility of "news fatigue" on TV and web searchers while YouTube news consumption is on the rise despite being a very small component of overall YouTube traffic. 3. Young people are comparatively less interested in news consumption and multi-platform news consumers tend to be older and more educated. 4. Lastly, exposure to more media platforms only increases news consumption for the "the unrepresentative minority of news consumers, who generate most engagement with online news." Takeaway:This study shows that in a world where most people tend to avoid news, a "committed minority" of news consumers have a disproportionate influence on how news is disseminated. Why is this important?Based on their findings, the researchers importantly conclude that "online media is amplifying the already high levels of interest of cross-platform consumers, setting them farther apart from the average citizen." Moreover, with results showing that TV is still the most common avenue to access news, this research makes the case for why the amount of scholarly interest on online news exposure is disproportionate to the relative impact these sources have compared to TV." |
| Causally estimating the effect of YouTube’s recommender system using counterfactual bots | Duncan Watts | 2/13/2024 | Overview:This research paper seeks to causally estimate the effect of YouTube’s recommendation algorithm on consumption of “partisan content” on the platform. The authors accomplish this by comparing bots that replicate the YouTube consumption patterns of real users with what they call, “counterfactual bots” wherein consumption preferences “rely exclusively on recommendations” from YouTube’s algorithm. Here are some key findings: 1. Relying solely on Youtube’s recommendation system “results in a more moderate experience on YouTube relative to the real user.” 2. When Youtube users shift from consuming partisan content to more moderate content, the sidebar is quick to reflect the change in content preferences while “homepage recommendations react more slowly.” Takeaway:While recommendation algorithms may shape content exposure and user preferences on online platforms like YouTube, this paper suggests that narratives about widespread algorithmic manipulation may be overstated. Why is this important?With over 2.5 billion active monthly users, YouTube is one the biggest online platforms in the world. While in some ways, the platform democratized video sharing and consumption, it's also been criticized for hosting radical content, much like other Big Tech companies like Meta. But is the consumption and proliferation of radical content a consequence of user choice or recommendation algorithms that are optimized to drive engagement? Disentangling the effects of algorithmic amplification from user intentions is difficult and this paper provides a framework to do so. By causally estimating the role of recommendation algorithms in driving partisan content consumption, this paper’s findings have important implications for policymakers seeking to hold platforms accountable for the content they host. |