Research Compendium

This paper examines the effect of Google’s introduction of AI Overviews and AI Mode on user behavior and site visits for publishers on the platform. The authors conducted a preregistered field experiment with 1,100 U.S. participants who used Google Search in their everyday browsing for ten days. A browser extension randomly assigned each participant to one of three conditions—AI Overviews hidden, ordinary Google Search, or all searches redirected to AI Mode—and captured both real browsing behavior and post-experiment perceptions. Key findings include:

1. AI search garners less trust. Participants using AI mode reported significantly less trust in information on Google, along with lower levels of perceived usefulness, satisfaction and agency. Those using AI mode were also more likely to search for information through a competitor like Bing.
2. AI Mode substantially reduced engagement with the broader web. Participants assigned to AI Mode clicked through to external sites 18.8 percentage points less often and ran approximately one fewer search session per day. Participants using AI search also clicked through to external sites less often, including news sites (by 12.5 percentage points), Reddit (21.2), and Wikipedia (9.9). Notably, these users spent more time inside the AI chat window.
3. AI search did not improve user experience. Relative to ordinary Google Search, AI Mode reduced trust in information found on Google and significantly lowered perceived usefulness, satisfaction, agency, and personalization. It also pushed more users toward competing search engines such as Bing. In open-ended responses, negative reactions outnumbered positive ones, with participants describing a loss of control regarding AI use, difficulty reaching specific websites they wanted, and fewer and less diverse sources.

LLMs equipped with web search are increasingly used by researchers, but the resulting citations are often inaccurate, or entirely fabricated. Unlike most model output, a citation can be checked definitively: every field in a BibTeX entry has one correct value on record in a publisher database. If web search works as intended, models should not make citation errors even after their knowledge cut-off. To test this, the authors built a benchmark of 931 papers across four domains and three citation tiers, and had three search-enabled models (GPT-5, Claude Sonnet 4.6, and Gemini 3 Flash) generate an entry for each. Roughly 23,000 fields were compared against verified ground truth. The authors also built clibib, an open-source tool that retrieves verified citation records directly from publisher databases, and tested whether adding it as a separate revision step corrected the models' errors. Key findings include:

1. Search helps, but half of all entries still contain an error. Models got 83.6% of individual citation fields right, but only 50.9% of entries were correct in every field. Author and entry type were the most reliable fields (91.1% each), while DOI (75.5%) and issue number (72.0%) were the worst.
2. Newer citations are less accurate, and web search does not fix it. Accuracy fell from 92.7% for popular older papers to 65.0% for papers published after the models' training cutoffs, a 27.7 percentage point gap. For these new papers, models conducted web search in over 97% of cases, yet accuracy stayed at 65%.
3. The tool developed by authors (clibib) verifies each citation against publisher records and corrects most errors. Running clibib as a separate revision step after the model generates a citation raises the ratio of fully correct entries from 50.9% to 78.3% and rarely introduces new errors.

The authors review how often social scientists claim their findings are causal. The authors used an LLM to scan 194,631 social science research paper abstracts published between 1980 and 2024, classifying whether each one used causal language (e.g., "X causes Y") versus more accurate associational language (e.g., "X is linked to Y"). They validated this LLM-based classification against expert human coders before applying it at scale. Beyond simply counting causal claims, the authors ran a set of experiments to test how this language affects readers and how AI tools respond to it. In one experiment, 1,105 college-educated adults read study abstracts—either in their original causal form or a rewritten associational version, sometimes paired with a note explaining the study's limitations—and reported whether they believed the results showed causation. In a parallel experiment, the authors fed the same materials to several AI language models to see whether AI-generated summaries introduced or removed causal overreach. Key findings include:

1. Causal language is common: Across the full 44-year period, 46.3% of cross-sectional abstracts used causal language despite their design not supporting it. This rate nearly tripled since 2000, climbing from about 20% to over 60% by 2024, with business journals showing the highest rates and sociology the lowest.
2. Readers default to inferring causation, and it's hard to correct. Even when abstracts were rewritten in purely associational language or flagged with a methodological warning, many participants still concluded the study demonstrated causation and used causal language themselves when summarizing it.
3. AI models can make the problem worse—but are more correctable than humans. Language models often introduced causal claims that weren't present in the original associational abstracts, especially when prompted to write "simple" or "practical" summaries. However, unlike human readers, models responded strongly to a prompt explicitly asking for methodological caution, substantially reducing overreach.

The paper asks how LLMs behave in closed-book question answering, where they rely solely on internal knowledge and cannot check their work against external sources. The authors compare three prompting regimes: Direct, where the model answers in one step, and two alternative forms, where it works through the same step-by-step sub-questions either all at once (Assistive) or one at a time (Incremental). Because all three prompting regimes should provide a consistent answer, any disagreement between them reflects the model's own instability—yet even frontier models like Gemini agree with themselves only about 70% of the time. The authors turn that disagreement into a reliability signal by introducing a process they call, “Disagreement Based Abstention," which answers "I don't know" whenever there are conflicting answers between the different prompting methods. Key findings include:

1. Decomposition helps open-weight models, but plateaus for frontier ones. For non-frontier models (70B parameters and under), decomposition acts as a reasoning scaffold and boosts accuracy substantially—by as much as 26.8 points. For frontier models, gains disappear or turn negative, likely because these models already reason well internally, and imposed structure can interfere rather than help.
2. Disagreement is a strong signal of error. When a model's direct and decomposed answers disagree, the direct answer is far more likely to be wrong. This makes disagreement an effective signal that a model response should not be trusted.
3. Flagging an error is not the same as fixing it. Among frontier-model disagreements, 67% are cases where both prompting methods (Direct and Assistive) are wrong, reflecting genuine gaps in what the model knows rather than errors in reasoning. DBA is effective at detecting these errors but cannot correct errors that stem from a lack of parametric knowledge.

People increasingly use AI chatbots to learn about social and political topics, and when those responses are inaccurate, it can “dissuade people from exercising their rights.” The researchers conducted an audit of LLM responses on abortion related information, comparing model responses with the true answers derived from state-level abortion statutes in the United States. They evaluated 54,704 model responses (from OpenAI ChatGPT-4o, Google Gemini 2.0 Flash, and Perplexity Sonar), using a set of abortion-related queries which varied by the wording of these queries, the week of pregnancy, and the particular U.S. state. Key findings include:

1. Overall model accuracy was 78%: However, there were major differences in accuracy scores across states. Accuracy was lowest for North Dakota (45%) and highest for Vermont (99%). The authors also find that responses were more inaccurate in abortion-restrictive states, and California was the only “low-performing state outside the abortion-restrictive policy categories.”
2. Lowest accuracy for medication abortion questions: The queries used for the audit covered a range of abortion related topics like medication, interstate travel, gestational age, and exceptions for sexual assault. Models were least accurate (71%) for questions related to accessing abortion medication, while queries on interstate travel yielded highest accuracy (98%).
3. Model responses were sycophantic: Accuracy shifted with wording, with prompts beginning with "Is it" performing worse than those beginning with "Can I" across all question types. The accuracy difference was largest for gestational age queries (14 percentage points). This pattern prompted the authors to examine sycophancy directly: when queries are about actions considered illegal, models were correct 90% of the time if the prompt was affirmative (i.e. using language like “Can I…” or ”Is it okay to…”). However, accuracy dropped to 31% when prompts had negative phrasing (i.e. using language like “Is it too late to…”), producing errors where the model says an action is legal when it is not.

The authors test whether reducing users’ exposure to political elites on social media can improve the accuracy of their beliefs. They built a custom X (formerly Twitter) app called Twilly and ran an experiment from September to December 2024 with over 1,200 active US users. Participants were randomly assigned to one of three conditions for one month: a normal unaltered feed (control), Elite Blocking (posts authored by political elites hidden), or Elite Scrubbing (posts from elites and those mentioning elites removed). Participants kept their existing follow networks and scrolled normally—only content posted/mentioning elites was filtered. Here are a few key findings:

1. Both interventions reduced exposure to low-credibility and highly partisan content. Views of low-credibility content fell by about 6 percentage points under both conditions, and partisan content by roughly 14 points. Retweets of partisan content dropped even more sharply—around 20 percentage points under both interventions—suggesting the interventions curbed not just exposure but virality.
2. Participants for whom elite posts were removed (and more so when removing mentions of those elites) showed improved factual knowledge and became meaningfully less susceptible to misinformation, with the largest effects on misleading image-headline pairings.
3. Elite Scrubbing produced larger reductions in low-credibility exposure than Elite Blocking alone, suggesting that posts about elites (replies, quote-tweets, commentary) carry their own misinformation risk, separate from elites’ own posts.
4. Neither intervention affected affective polarization, institutional trust, political interest, or media trust—reducing elite visibility for one month improved what people knew, without changing how they felt about politics.

This paper examines whether AI models can do more than recognize right from wrong—specifically, whether they can reason about the social consequences that follow when a norm is broken. The authors call this "second-order" social reasoning: understanding not just that an act is wrong, but who is expected to react, how strongly, and whether any reaction is warranted. To test how LLMs reason about this, the authors built NormReact, a dataset of 450 everyday norm-violation scenarios (e.g., a coworker taking credit for someone else's work) drawn from Reddit, and collected annotations from 871 human participants on how violators and bystanders would likely feel and behave. They then compared those human judgments against predictions from six leading LLMs. Key findings include:

1. LLMs consistently predict a harsher social world than humans do. Across all six models, they overpredict negative sanctions—anger, confrontation, and social exclusion—in situations where humans would expect inaction or restraint. The models appear to treat norm violations as automatically deserving punishment, when in reality humans exercise considerable tolerance depending on context.
2. LLMs under-predict self-focused emotions (shame, guilt) and over-predict other-focused emotions for observers: Closed source models reasonably approximate the shame and guilt that humans feel toward someone who violates a norm (shame at 19% and guilt at 22%) but open-source models overestimate both. Interestingly, the bigger gap is with observers: people still report meaningful shame and guilt when someone close to them breaks a norm (a combined 19% for close ties, declining to about 9% for strangers), whereas both open- and closed-source models predict lower rates (roughly 4-11% for close ties and nearly zero for weak ties/strangers).
3. LLMs are more action-oriented and confrontational: LLMs disproportionately predict gossip and verbal confrontation as the appropriate reaction across all observer types, with open-source models overpredicting gossip by up to 66% for weak ties. When asked for the ideal response rather than the likely one, they escalate from gossip to direct confrontation, while humans stay restrained under both framings.

Recommender systems shape content exposure on online platforms, and a growing body of research suggests these systems—optimized for engagement—can promote filter bubbles, reduce ideological diversity, and in some cases, heighten online radicalization. This paper investigates how LLM-based reranking of YouTube recommendations reshapes political content exposure, and whether prompt-level constraints can reduce the promotion of conspiratorial and extremist content without sacrificing personalization. Here are the key findings:

1. LLM-based content reranking improves personalization and aligns video recommendations more closely to a user’s prior viewing. LLM-based rankings also made suggestions more strongly aligned with the ideological leaning of a user’s watch history, compared to YouTube’s baseline. For users with watch histories containing conspiratorial or extremist content, it amplified exposure to that material.
2. Safer prompt guidelines reduced exposure to problematic content without hampering personalization of video recommendations. This suggests that, “it is not language-based models per se that promote problematic content, but the absence of explicit regularization.”
3. For topics like abortion and immigration, the LLM prioritized topic relevance over partisan alignment—sometimes even surfacing opposing viewpoints. For broader topics like elections, where partisan language is more intertwined with other issues, it prioritized partisan alignment over topic. This mixed behavior suggests LLMs operate on statistical language patterns rather than ideological understanding.

This paper asks how Russia’s February 2022 invasion of Ukraine changed the content of Russian media. Specifically, it examines whether autocratic governments can maintain their strategy of keeping news apolitical when major events—like a war—force politics into the spotlight. To investigate this, the authors collected over two million social media posts from 20 prominent Russian-language media outlets across V’Kontakte (Russia’s largest social media platform), Facebook, and Telegram between 2016 and 2024, using a combination of human reviewers and AI tools to categorize the content. Key findings include:

1. The February 2022 invasion triggered a sharp and durable spike in political coverage across outlets. In the three months before the invasion, 47% of posts from the sample were political; in the three months after, that rose to 76%.
2. Independent outlets’ political coverage declined relative to pro-government outlets after March 2022, when new laws criminalizing “discrediting the armed forces” took effect and drove some outlets out of Russia entirely. Further analyses also show that independent and pro-government outlets’ posts briefly converged in language immediately after the invasion, suggesting the new legal environment narrowed the space for distinct independent coverage.
3. Before February 2022, V’Kontakte users were often less likely to like or share political posts. After the invasion, this flipped—political content, and Ukraine-related posts in particular, began generating significantly more engagement, suggesting the war created a demand for political news that outlets could not ignore.

This paper examines how media coverage, elite messaging, and public perceptions of racial discrimination shifted in the years following the 2020 George Floyd protests. Using cable news data from the Stanford Cable TV News Analyzer, congressional tweet archives, and a longitudinal panel survey of voters in Michigan, Pennsylvania, and Wisconsin tracked from 2020 to 2024, the authors document how a conservative counter-narrative—arguing that efforts to remedy racial injustice discriminated against white people—came to dominate public discourse and reshape Americans’ racial attitudes. Key findings include:

1. Media coverage turned overwhelmingly negative: While Black Lives Matter (BLM) dominated coverage at its peak in June 2020, attention quickly pivoted toward Critical Race Theory (CRT) and Diversity, Equity, and Inclusion (DEI) framed as threats to white people. Fox News devoted four times more coverage to CRT than MSNBC or CNN in 2021, and 97% of its coverage of these topics was negative. Even on MSNBC, only 8% of CRT coverage made an explicitly pro-CRT argument—most simply criticized Republicans for attacking it rather than defending it on its merits.
2. Elite messaging was one-sided: Republicans tweeted twice as much as Democrats about racial justice topics overall and nearly four times as much after 2021. Two-thirds of Republican tweets were negative in tone, while only one quarter of Democratic tweets were positive—leaving critical messaging to dominate across the board.
3. Perceptions of discrimination against white people grew across all groups: Americans came to perceive less discrimination against Black people and more against white people between 2020 and 2024. Opposition to CRT was strongly linked to higher perceived discrimination against white people. This was especially the case for consumers of conservative media, and the authors argue the media coverage was likely reinforcing existing beliefs rather than changing minds.
4. Party identity became increasingly tied to racial attitudes: Partisan polarization in racial attitudes grew by nearly 27% over this period, with people more likely to adopt the racial views of their party than to switch parties because of their racial views—though both directions were at work.

This paper asks why conspiracy beliefs sometimes lead people to act on them and other times do not. The authors propose that it is not just what someone believes, but how they believe it that matters. They introduce the concept of ‘belief strength,’ which captures three qualities of a belief: how certain someone is that it is true, how personally important they consider it, and how relevant they think it is for guiding their decisions (perceived actionability). Across three studies totaling over 6,000 U.S. participants, the authors tested whether these qualities predicted whether conspiracy beliefs translated into attitudes (for instance how positive, risky, or effective something is evaluated to be) and real-world behaviors across three domains—anti-immigration sentiment, COVID-19 vaccination, and political participation. Key findings include:

1. Certainty and importance consistently strengthened the link between conspiracy beliefs and attitudes. Across the three studies—which measured certainty, importance, or both—people who held a conspiracy belief with greater conviction or personal investment were more likely to develop attitudes consistent with that belief than those who held the same belief with less certainty or importance.
2. Perceived actionability amplified the role of certainty and importance, but did not drive attitudes on its own. When people viewed their belief as a useful guide for decisions, the attitudinal difference between high- and low-certainty believers grew even wider. When perceived actionability was low, high- and low-certainty believers looked much more similar to each other.
3. Attitudes were the bridge between beliefs and behavior. Conspiracy beliefs did not directly predict behavior—instead, they shaped how people felt about certain actions, and those feelings predicted what people actually did, whether that meant getting vaccinated or engaging in political activities like attending rallies or donating to campaigns.

The authors examine how social scientists have used social media data for empirical research and assess the potential impact of recent platform restrictions on the state of the field. Using a multi-method design that includes human-in-the-loop AI annotation, they systematically analyzed over 859,000 publications across 59 journals representing five social science disciplines—Communication, Political Science, Sociology, Psychology, and Economics—as well as leading interdisciplinary journals, spanning 2010 to 2024. Key findings:

1. Only 3.15% of social science journal publications use social media data for empirical research, rising to about 13% in Communication but falling below 1% in Economics. 82.6% of these papers rely on data from a single platform, indicating that cross-platform research remains rare.
2. Twitter/X and Facebook are the first and second most used platforms throughout the study period, accounting for roughly 51% and 36% of social media papers respectively—despite Twitter/X being used by only 21% of Americans. By contrast, YouTube, used by 84% of Americans, appears in only 17% of papers, highlighting how API accessibility and content type shape research design as much as platform popularity does.
3. The share of all papers using social media data tripled between 2010 and 2024, with notable spikes around the 2016 U.S. presidential election, between 2021 and 2022 during the COVID-19 pandemic, and during the 2020 U.S. election. However, usage has since plateaued for Facebook and declined for Twitter/X, likely tied to Twitter/X ending free API access in 2023 and Meta shutting down CrowdTangle in 2024.
4. COVID-19 and social movements are the most frequently studied topics across disciplines, underscoring that social media data has been most valuable for capturing real-time events with broad socio-political significance.

This paper examines the composition of users on X (formerly Twitter), by analyzing their behavior, demographic makeup, and content consumption patterns. The researchers analyze the activity logs of 209 U.S.-based users and group them into distinct behavioral profiles based on how they consume and interact with content on the platform. Here are some of the findings:

1. Three distinct user groups: Using 22 tracked activity types—such as timeline browsing, tweet viewing, and liking—users were sorted into three distinct groups. The majority were lurkers (60.45%), followed by interactors (32.52%), and creators (7.03%). Lurkers are users who passively consume social media content; interactors engage with content via likes/comments and creators are involved in making and sharing content.
2. Partisan demographics by user type: Demographically, interactors were primarily Democrats (84.5%) and younger. Lurkers were majority Democrat (64.6%). Creators skewed older and Republican (52.7%), though this group was very small, comprising only 15 users.
3. Exposure to low-credibility content: Exposure to low-credibility content differed notably across the three groups—much lower for interactors (23.3%), moderate for lurkers (29.3%), and much higher for creators (44.1%). This tracks closely with each group's partisan makeup: Republican users were far more likely than Democrat users to consume low-credibility content across all three groups.

This paper investigates “what differentiates decision-making conversations that ultimately reach consensus from those that fail to reach consensus.” Using fMRI data and NLP techniques, the authors examine how alignment and exploration support real-time conversations and how the goal of either persuading or compromising with one’s conversational partner shapes these dynamics. The researchers examined two key questions: first, whether alignment or exploration is the more effective path to consensus, and second, whether persuasion or compromise goals better facilitate exploration. Alignment refers to when people reach common ground and mutual understanding—often via shared linguistic and physiological cues. Exploration, by contrast, refers to the process by which people desynchronize and venture into new directions. The researchers tested these questions by instructing participants (N = 60 dyads; 120 people) to either persuade or compromise with their partner on pressing societal issues—air pollution and tuition increases.

1. Pairs of participants instructed to compromise reached higher levels of agreement at the end of the conversation compared to pairs instructed to persuade.
2. Pairs instructed to compromise engaged in more exploratory conversations, covering more new ground and entertaining a broader range of topics and ideas. These pairs also started the conversation more in alignment with each other, which may have made it easier for them to branch out.
3. Pairs that explored more reached greater agreement. While alignment served as a useful starting point, it was exploration—specifically how much new ground partners covered together—that most strongly predicted whether they would ultimately agree.

Religious leaders in the U.S. are barred by IRS guidelines from openly endorsing political candidates. But does this separation of church and state exist in practice? This research paper examines the prevalence of religious institutions engaging in electoral advocacy. The authors analyze 88,546 religious sermons from evangelical churches spanning the 2020, 2022 and 2024 election cycle. Here’s what they found:

1. Direct political advocacy and candidate endorsements are widespread. 14.7% of churches in the sample engaged in prohibited electioneering during the three months surrounding the 2024 presidential election, compared to 8.3% during the 2022 midterms and 2.6% during a nonelection control period. The authors believe these estimates are conservative and true rates for this phenomena could be much higher. The advocacy was overwhelmingly partisan: 87.5% of references to Republicans were favorable, while 71.3% of references to Democrats were in opposition. Donald Trump accounted for 52.4% of all political discussion.
2. Political advocacy follows a clear electoral rhythm, peaking in the days before Election Day. Political advocacy by evangelical churches rose steadily in the weeks leading up to each election and sharply declined immediately after. On the final Sunday before election day, advocacy rates peaked—reaching 3.5% in 2020 and 2.6% in 2024. No comparable pattern was observed during the nonelection control period.
3. Advocacy is driven by local partisan context rather than national electoral strategy. Churches in the most solidly Republican counties were more than twice as likely to engage in political advocacy as those in the most Democrat counties. Political advocacy from churches in swing states were relatively stable, suggesting that churches are more likely to mobilize their base rather than aim to persuade swing voters.

This latest research paper from the 2020 Meta election project examined the scale, reach, and effects of deceptive online networks that targeted US users on Facebook and Instagram during the 2020 election. The authors define deceptive online networks as "coordinated efforts that take place online where audiences are misled about the identity of the people behind the network." The dataset included 49 deceptive networks found to be active between 26 June 2020 and 15 February 2021. Here’s what they found:

1. Deceptive networks achieved substantial reach, driven by a small number of actors. At least 37 million Facebook and three million Instagram users were exposed to content from deceptive networks during the study period. Over 70% of all users who viewed content from deceptive online networks were exposed to it via only three Facebook networks. Overall, the authors find that exposure to deceptive online networks “accounts for a very small share of users’ overall political content consumption.”
2. During the study period, for deceptive networks engaging in political discourse, Meta identified three times as many financially motivated networks as only politically motivated ones. The financially motivated networks also reached more users and appeared to share a lot of political content to further their financial motivations.
3. Most participants saw deceptive content indirectly, that is, by viewing content from regular user accounts “unaffiliated with the networks resharing deceptive network content,” often unknowingly amplifying such content in audiences beyond the networks’ direct reach.
4. The study found that users who were older, more conservative, previously exposed to false news and spent more time on Facebook were especially susceptible to these deceptive online networks.

What makes a product recommendation message persuasive on social media: expressing an opinion, describing an action, or combining both? Across two pre-registered experiments (N = 1,506), researchers tested how different types of messages influence attitudes, perceived norms, and behavioral intentions. Participants were shown positive or negative posts about products that either expressed attitudes (e.g., "I liked the car"), described behaviors (e.g., "I bought the car"), or combined both (e.g., "The car was good, and I bought it"). Here's what they found:

1. Combined messages are most persuasive: Messages combining attitudinal and behavioral statements "are more persuasive than those expressing only one type of information." This was true for when statements were juxtaposed (e.g. "I liked the car, and I bought it") and when combined in a single proposition ("I liked buying the car"). These findings applied across measures for attitudes, perceived norms and behavioral intentions.
2. Why combined messages work: The authors find that combined attitudinal–behavioral statements are more persuasive in part because they promote mental simulation of behavior—recipients are more likely to imagine themselves performing the action. These statements may also be more effective because they don't just express an opinion but also convey a deeper sense of commitment and confidence that the audience picks up on.
3. Object value shapes message effectiveness: For higher-value objects, behavioral statements had a stronger influence compared to attitudinal statements, suggesting that people infer attitudes from costly or infrequent actions (e.g., purchasing a car is presumed to mean the purchaser likes that car). For lower-value objects, attitudes alone were often sufficient for participants to infer behavior, reducing the added persuasive impact of behavioral information.

The authors examine what incentivizes U.S. political elites to use divisive language and engage in incivility. Using a large-language model (LLM), they analyzed over 2.2 million public statements from every member of the 118th U.S. Congress—including floor speeches, press releases, newsletters, and social media posts—and linked these to records of media coverage, campaign finance, and electoral outcomes. The study distinguishes between “policy entrepreneurs,” who focus on legislative debate from “conflict entrepreneurs,” defined as legislators who frequently use personal insults rather than engaging with their policy positions. Key findings include:

1. Personal attacks were primarily aimed at the head or perceived head of the opposing party. The study identified 43,560 instances of personal attacks, with the three 2024 presidential candidates alone accounting for roughly a third of all attacks. Democrat presidential candidates received approximately four times the attacks directed at Trump.
2. Personal attacks are rare in absolute terms—65% of legislators use them in fewer than 1% of their communications—but they receive outsized media attention relative to their frequency. The authors find that “the 25 most conflictual members of Congress receive more cable news attention than the 75 least conflictual members combined.”
3. Personal attacks show no meaningful association with fundraising, electoral gains, cosponsorship of bills, or any measurable changes in personal wealth in a 1-year period.
4. Conflict entrepreneurs do not reflect the political culture at the district-level. The study finds “a striking disconnect between elite rhetoric and mass sentiment,” suggesting that media incentives—rather than constituent demand—are a primary driver of this divisive rhetoric.

The authors examine whether partisan animosity reflects deep-rooted hatred or serves as a means to signal "loyalty and conformity" to one's social group. They conducted two studies to test this claim: first, they analyzed interview data from the 2012 American National Election Studies (ANES) to see whether respondents adjusted their feelings of animosity based on the partisanship of the interviewer. Second, they conducted a real-time online news-sharing experiment with 1,510 participants to see if animosity was amplified when participants were randomly partnered with someone with similar political leanings. Here are some key findings:

1. The analysis of the ANES data revealed that respondents expressed more hostility toward the opposing party when they believed the interviewer shared their political views. This suggests that partisan animosity is partially based on who the audience is and serves a social signal rather than only reflecting political beliefs.
2. In the online news-sharing experiment, respondents were more likely to express negative views about an opposing political party if their "partner's partisanship was disclosed and matched their own." However, the researchers were unable to identify whether the respondent was exaggerating their beliefs based on their partner's or felt more comfortable expressing them.

The researchers looked at the impact of political advertising on 36,906 Facebook users and 25,925 Instagram users at the time of the 2020 US election. This paper analyses internal data from Meta for a period of 6 weeks before the US elections to see if removing political ads from participants' feeds had any effect on electoral outcomes. Here are the key findings:

1. Political Ads are Targeted towards Party Supporters: They found that political ads during the 2020 presidential election were not primarily aimed at persuading undecided voters: 46% sought donations, while 26% intended to persuade, and 17% sought to collect voter information.
2. No effect of Political Ads on Political Attitudes: Removing political ads entirely had no detectable effect on candidate favorability, political knowledge, polarization, or confidence in the legitimacy of the election—even though they had substantial presence on a user's feed (23 ads per week for the average user).

Aligning LLMs to mirror how humans think and process information is quite challenging. LLM developers tend to evaluate this “common sense” knowledge using answers that human annotators have marked as true. However, this assumes that common sense in humans is homogenous. However, what is common sense to one person, may not apply to the larger population. In this paper, the researchers propose a way to evaluate common sense knowledge in LLMs without assuming some “ground truth” that may only reflect the views of the model developer. They do so in two ways.

1. First, the researchers treat each LLM like an individual survey respondent, asking it to give yes or no answers to thousands of statements. They measure first whether the model's responses align with the human majority, and second whether it can accurately predict what most people would say. On these metrics, more than two-thirds of models scored below the median human, and the best-performing model only exceeded the commonsense aptitude of 64% of human participants. Surprisingly, smaller open-weight models, such as Mistral-7B, outperformed larger frontier models, such asGPT-4.
2. The researchers also tested whether LLMs could simulate how a large group of humans thinks about common sense. The underlying idea is that the more frequently a statement is endorsed by humans in an LLM's training data, the more likely the model is to accept it, meaning LLMs may effectively represent “the average human contributor of that data.” To test this, they repeatedly prompted each LLM to get a range of responses (which the authors call a “silicon sample”) and compared their responses to those of real human participants. They found that humans tend to agree more on factual, literal statements than on figures of speech, and most LLMs also reflected this pattern. However, most LLMs tended to rate statements as either commonsensical or not, with little middle ground. The authors attribute this to extensive fine-tuning, which pushes models toward confident, decisive answers and away from the more uncertain responses that better reflect how humans actually think.

This review paper discusses empirical research examining how LLMs interpret African-American Language (AAL) and whether they “distort the communicative intent” of its speakers. First, the authors synthesize research explaining reasons for why LLMs are worse at interpreting AAL compared to White Mainstream English (WME) and its downstream impact on African-American users. The authors identify three sources of bias that may explain the disparity in LLM performance on AAL vs WME. They are:

1. Data: Research demonstrates that AAL is severely underrepresented in LLM training data. Studies of the widely-used C4 corpus found AAL constituted only 0.07% compared to 97.8% WME. Even with documents representing AAL, research found that it frequently “reinforced stereotypes, and/or represented appropriated speech.” Research attributes this imbalance to automated quality filters that disproportionately remove AAL texts, filtering out 42% of AAL documents versus only 6.2% of WME documents. LLMs also perform poorly on AAL because pretraining data includes what is easily available on the web which for AAL, is often performative online contexts like Twitter and hip hop lyrics rather than naturalistic speech.
2. Annotations: How content is labeled is shaped by the characteristics of the annotator. This is relevant for AAL because when data has been annotated without consideration of the annotators’ race, “the resulting labeled data may not reflect the views of African Americans.” For instance, research found that conservatives who also tend to “hold racist beliefs” are more likely to label AAL posts as offensive, clearly showing that the political views of the annotators can act as a source of bias against AAL.
3. Model: The authors note that data quality alone doesn't explain bias—the models themselves also introduce problems. Better training data can't fully solve the issue because language constantly changes, and biased models can worsen existing biases when retrained. Research shows models have more trouble understanding AAL texts compared to WME texts and rate AAL documents as having lower quality. The inability for LLMs to accurately interpret AAL has significant impacts: AAL speakers have to work harder to be understood by LLMs which can make some Black users “feel that their culture and language are not valued.”

This research article provides a historical overview of social media data access and how it has evolved in the last twenty years, beginning in the early 2000s. The authors identify four key periods starting in 2006: a "laissez-faire" era (2006–2011) when platforms offered relatively open, free access but was not extensively used by researchers; an "authentication period" (2011–2018) when platforms began tightening restrictions and requiring credentials which happened to coincide with social scientists realizing the value of this data for studying online communications; a "limited options period" (2018–2020) triggered by Meta shutting down academic API access following the Cambridge Analytica scandal; and an "academic cooperation period" (2020–2023) wherein social media platforms "implemented academic-only data sources in response to public scrutiny about their role in society." Based on this historical assessment, the paper evaluates the “data access options” currently provided by the six major technology companies. Here’s what they found:

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.

This paper tests the efficacy of ten “expert-crowdsourced, theoretically-based psychological interventions” in bolstering electoral participation with a sample of eligible unregistered US voters ahead of the 2024 presidential election. These interventions drew on established behavioral science principles—such as correcting misperceptions about registration difficulty; emphasizing the moral basis for civic participation; and using escalating commitment techniques that combine multiple forms of social pressure, such as telling people that voting records are public. The authors measure the impact of these interventions across different stages of electoral participation, from stated voting intentions to clicking on voter registration websites to actual registration and turnout. The control condition in this case was an intervention unrelated to voting. Here are some key findings:

1. Eight out of ten interventions significantly increased intentions to vote and the “Escalating Commitments" intervention had the strongest effect, boosting voting intentions by eight percentage points.
2. Five interventions led to increased click rates for voter registration websites compared to the control. Intuitively, interventions were less effective on participants who reported low political interest and voting intention prior to the study, showing that how people feel about voting from the outset is a stubborn determinant of electoral participation.
3. Despite having positive effects on increasing voting intentions and in some cases, click rates, none of the interventions “had a significant impact on voter registration.” The same applies for voter turnout, in that, no interventions significantly improved actual turnout compared to the control condition.

This paper employs LLMs to generate “synthetic news articles” in order to study the effects of biased news coverage across a range of events pertaining to politics, the economy and culture more broadly. The LLM-generated news articles incorporate changes in “selection and tone of the content while holding factual accuracy and other features constant.” In other words, the generated news articles mimic the ways in which different media outlets may employ a positive, neutral, or negative tone; and selectively present certain facts when reporting the news. The authors conducted a randomized experiment to evaluate the impact of the “alternative framings” of the news articles on how participants evaluated its tone and informativeness along with how it made them feel about the subject of the news article. Here are some key findings:

1. Both positively and negatively framed news articles “significantly influence” how participants feel about the subject of the article. However, news stories with a negative framing had a substantially larger impact on how respondents felt about the subject. Specifically, the negative framing led to an “average treatment effect of – 18.5 percentage points." This indicates that negatively biased news coverage has a greater impact on how participants feel about the subject of the story.
2. The authors also examined whether these biased news articles shaped deeply held opinions about the content presented and found that articles with negative framing significantly shifted opinion while there was “no significant effect in the positive direction.” This result indicates that biased news articles not only “alter feelings toward the subject”, but can “also influence how people perceive the facts associated with these events.”
3. The negative framing in news articles presented to participants had a “substantially larger effect” among those who self-reported as less informed. This indicates that people who don't consider themselves well versed about the news cycle are more susceptible to selective reporting and tonal shifts in news coverage.

This research paper tested whether teaching people how mRNA vaccines and cellular DNA protection actually work could reduce susceptibility to vaccine misinformation. Rather than directly correcting false claims (as in traditional fact-checking), the researchers conducted two studies that employed a "mental model approach", providing detailed explanations about how vaccines work. Two preregistered experiments tested this approach with U.S. adults. The first study graphically displayed how mRNA vaccines work (the vaccine model) and how human cells protect themselves from foreign DNA (the cell protection model), along with additional material on vaccine safety. The second study used an animated video to explain the cell protection model, either by itself or combined with the first study’s materials. Both experiments intentionally avoided directly refuting false claims about the vaccines. The researchers also examined if exposure to misinformation would increase misconceptions, specifically using Florida Surgeon General Joseph Ladapo's false claim that DNA fragments in mRNA vaccines could integrate into recipients' DNA. They further tested whether the proposed “mental model approach” could protect against this false claim. Here are some key findings:

1. Participants exposed to Ladapo’s false claims regarding DNA integration reduced accurate responses from subjects in both studies, showing that vaccine misinformation from seemingly authoritative sources can have a negative impact. Similarly, subjects who did not see this content and were only exposed to either mental models led to more accurate responses compared to the control group.
2. Participants exposed to the mental models along with Ladapo's problematic claims showed more accurate responses than those who only saw Ladapo's claims, regardless of presentation order.
3. Preemptive positioning may be slightly more effective than correction with authors finding that presenting the mental models before exposure to misconceptions was somewhat more protective than presenting the models as a rebuttal, that is, after exposure to vaccine misinformation.

Net neutrality has been debated for nearly three decades, with the US, EU, and UK adopting different regulatory approaches. However, little empirical research exists on the efficacy of net neutrality rules and its economic impact. This paper fills that gap by reviewing the economic and regulatory landscape of net neutrality with a specific focus on mobile broadband, assessing both whether the rules achieve their intended goals and whether their benefits justify their costs. Here are some key findings and concrete policy recommendations:

1. Net neutrality rules are increasingly ineffective: Large technology firms including Google and Meta have built private networks and content delivery systems (CDS) that bypass regulated public networks entirely, meaning a growing share of internet traffic already operates outside net neutrality rules. By 2023, 70% of all internet traffic was delivered through CDS that fall outside the scope of net neutrality rules.
2. The empirical evidence does not support net neutrality's claimed benefits: The authors conduct a comprehensive review of all available empirical studies and conclude that "no empirical study supports the arguments of proponents" of net neutrality. However, empirical evidence does suggest that net neutrality regulations have led to "significant reductions in telecom investment."
3. The costs of net neutrality regulation are substantial: Beyond reduced investment, the authors identify the loss of innovative services, "high transaction costs" from monitoring and compliance that weigh particularly heavily on smaller providers, and inefficiencies as companies resort to costly technological workarounds like private CDNs. In the US, the on-again/off-again nature of net neutrality rules has "greatly complicated industry participants' ability to engage in long-term planning."
4. Policy recommendations: The "first-best" recommendation is to remove net neutrality rules altogether, because costs outweigh demonstrated benefits. Where full removal is not politically feasible, they recommend giving ISPs greater flexibility in pricing and network management, subject to existing competition law and consumer protections, or adopting a principles-based framework with limited ex-ante obligations. The authors point to the UK regulator Ofcom's 2023 revised guidelines as a concrete step in this direction.

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.

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 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.

In this research paper, the authors examined “how one’s ability to learn about a topic” is impacted by the use of LLMs versus traditional web search and found that while LLMs may make accessing information easier, it may come at the cost of reducing the “depth of knowledge users may develop.” In a series of experiments, the researchers had participants use different search methods (ChatGPT vs. Google, LLM summaries vs. linked articles, or Google's standard results vs. AI Overviews) to learn about a topic and draft advice based on what they learned. Independent evaluators, blind to the search method used, rated the participant-generated advice to assess downstream consequences of using LLMs vs traditional search. Here are some key findings:

1. Those who used ChatGPT spent less time on the task and reported that “they learnt fewer new things about the subject.” Participants learning via ChatGPT also put less effort in creating advice and thus felt less “personal ownership” on what they created.
2. Even when the same information was presented as an LLM summary rather than a series of web links, participants reported putting less effort in learning and developed only a “shallow understanding of the topic.”
3. Evaluators found advice based on learning from Google's AI Overviews (vs. web links) to be less helpful and less informative. They also believed less effort had been put into writing the advice, found it less trustworthy, and were less willing to adopt it themselves or recommend it to others.

Politicians tend to switch between partisan and centrist rhetoric based on perceived electoral gains. Which type of rhetoric yields votes also depends on what media platforms choose to amplify. In this paper, the researchers set out to examine (1) how candidates’ incentives to target audiences to win votes and (2) how the media’s incentives to cover political campaigns “interact to shape campaign speech and campaign performance.” This is empirically challenging because the electorate is not homogenous, meaning that the same campaign remark can persuade some but dissuade others. Thus, most previous research has looked at “net effects”, limiting the understanding of “the underlying trade-offs that candidates face.” To address this, the authors developed a method to separately measure persuasion and dissuasion effects. They used text analysis to classify more than 200,000 newspaper stories from U.S. Senate races (1980-2012) as either partisan voter-oriented or swing voter-oriented, linking those to high-frequency polling data. Here's what they found:

1. Candidates prefer for partisan rhetoric to “fly under the radar” whereas the media prefers more polarizing content. Conversely, candidates would like rhetoric aimed at swing voters (centrist appeals) to be covered by the media but the media tends to ignore it.
2. Democrat appeals to their voter base are “four times more effective” at mobilizing voter turnout compared to Republican appeals. However, this increase in partisan rhetoric by Democrats triggers “twice the level of backlash” from swing voters. On average, 56% remarks by Democrat candidates and 45% remarks by Republican candidates aim to mobilize partisans.
3. Media coverage for both parties is balanced. Because media outlets prefer partisan content but Democrats produce less of it (due to swing-voter penalties), both parties receive similar overall coverage rates.
4. In states where Democrats make up a smaller share of the electorate, swing voters reward Democratic centrist appeals more strongly. Competitiveness also matters: as polling gaps narrow, Republican speech targeting swing-voters becomes more effective.

AI chatbots are used by people from diverse cultural backgrounds but are they “culturally aware?” In this paper, the authors introduce the “first framework and benchmark designed to evaluate LLMs in realistic, multicultural conversational settings.” To do so, the authors, in consultation with cultural experts, designed six conversational situations where the LLM’s ideal response should be culturally sensitive. For example, when discussing personal accomplishments, some cultures may view celebration as confidence while others may view it as arrogance. The authors used an OpenAI model to generate a dataset of 48 conversations, each with five possible responses that vary stylistically while conveying the same underlying message. They recruited 24 annotators from eight countries to determine which responses are most culturally appropriate in each conversational situation. After evaluating five models from OpenAI, Google and Anthropic based on this framework, the researchers found that all models perform best in the Western cultural context. Across the board, the highest accuracy scores were for America and the Netherlands. This is worrying because LLMs have become quite popular in non-Western contexts and this research suggests that LLMs are “less likely to align with local users’ communication practices.”

The rise and growing sophistication of AI-generated videos has made identifying what is real, rather than AI generated, more challenging. The authors argue that despite advances in these video generation models, the question of whether humans can identify and provide reasons for why a video is machine generated has been largely overlooked. To address this, the researchers introduce DEEPTRACEREWARD, a benchmark built on a dataset of 4,300 expert annotations across 3,300 AI-generated videos from seven state-of-the-art video generators. Each annotation identifies the portion of the video containing the sign, or trace, that the video was fake, the time it occurs, and a natural-language explanation of what gave it away. Based on these annotations, the researchers identify 9 categories through which humans identified a video as AI-generated. Reasons for why annotators identified videos as fake included object and background distortions, sudden blurring, and objects unnaturally disappearing mid-scene. The researchers then evaluated 13 existing AI models on their ability to identify these human-perceived deepfake traces, asking: do current AI models possess human-level visual intelligence to identify deepfake traces, and if not, can they be taught to do so using DEEPTRACEREWARD? Here are some key results:

1. While frontier models such as GPT-5 and Gemini 2.5 Pro achieved over 70% accuracy when classifying videos as fake or real, their performance on identifying exactly where and why a video is fake was below 36%.
2. The researchers also fine-tuned an existing AI model (LLaMa 3) on their benchmark, which achieved an overall score of 70.2%, outperforming GPT-5 by nearly 35 percentage points. Across all models, a consistent pattern emerged: binary classification was easiest, followed by natural language explanation. Identifying the location and time of the deepfake trace was much harder. spatial localization, with temporal localization being hardest.

LLMs are being used to access important information about sensitive social and political topics. But LLMs can often refuse to answer these questions and little is known about how and why these refusals occur. To address this research gap and improve transparency of LLM outputs, this paper introduces AI Watchman, a longitudinal auditing system that biweekly probes OpenAI’s moderation endpoint, GPT-4.1, GPT-5, and DeepSeek (in English and Chinese) by prompting them with “repeat after me” followed by Wikipedia content drawn from a dataset of over 400 social issue topics. Here’s what they found:

1. The AI Watchman system can detect model-specific changes in content moderation—even when caused by subtle and unannounced policy changes by companies. For instance, GPT-4.1’s refusals for Israel-related content significantly increased during August and September 2025 amid Israel-Gaza ceasefire negotiations. GPT refusals for abortion related content also spiked in the first week of September 2025. Since there was no explicit policy change by OpenAI, the authors speculate that the spike in refusals could have been shaped by external factors like the passage of a Texas bill restricting abortion medication and allowing private individuals to sue medicine distributors.
2. Refusal rates vary substantially by model and by content category, with GPT-4.1 refusing most often (3.9%) and GPT-5 the least (1.2%). DeepSeek displayed consistent performance across models but for Chinese sensitive topics, the model had a high refusal rate in both English and Chinese (30.9% on average). Surprisingly, despite being a U.S company, Chinese Sensitive Topics were also the top flagged category for OpenAI’s moderation endpoint.
3. Models justify refusals with four recurring and often inconsistent rationales—content length, content policy, misinformation risk, and legal risk. Newer models like GPT-5 increasingly use “non-explicit refusals” that appear cooperative (offering summaries, redactions, or partial responses) while still failing to fulfill the initial request. DeepSeek surfaces content reflecting the Chinese government’s preferred stance on topics like Taiwan rather than overtly refusing.

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.

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.

In this research paper, Diana Mutz uses "two identical pre-election surveys" from 1996 and 2020 respectively to examine changes in the political discussion networks of Americans in the 25 years between both surveys and its implications for political participation and partisan tolerance. Here are some key findings:

1. When comparing results from both pre-election surveys (1996 and 2020), the study finds that political discussants, which refers to the number of people someone talks to about politics, increased by 22% among all respondents. However, to account for changes in survey best practices and ensure uniformity, the paper also compared the 1996 results with a subset of "respondents randomly assigned to take the survey by telephone in 2020, as was originally done in 1996" and found an even larger increase in political discussants (33%) compared to 1996.
2. Despite the increase in political conversations among Americans, the study found no significant increase in political discussions with those that have opposing views. When looking at the composition of Americans' political discussion networks, the study finds that the number of "like-minded discussants" increased the most and "oppositional discussants were largely unchanged." This indicates that on average, political networks of Americans have only grown more homogenous over time.
3. In 1996, women had greater cross cutting conversations, which refers to conversations with people who hold different political views, but in 2020, "this relationship had completely reversed, with women now reporting systematically less cross-cutting discussion than men."
4. The study finds a 9% decrease in political tolerance from 1996 to 2020. Political participation however increased given the greater homogeneity in political discussion networks with the study suggesting that the number of "like-minded discussants" in one's political networks is a salient predictor for political participation, or put more simply, that people are more likely to engage in political action when they feel validated in their views.

Social media platforms are using large language models in content moderation, especially for hate speech detection. But is hate speech classification consistent across models? Using a novel synthetic dataset of over 1.3 million sentences spanning 125 demographic groups, the researchers tested how seven leading automated content moderation systems evaluate and classify hate speech. Here's what they found: Here are some key findings:

1. Content moderation systems are inconsistent: The authors found substantial differences in how different models evaluated and classified hate speech. The Mistral moderation endpoint had high detection value and consistent classifications, meaning that it classified more posts as hate speech compared to its competitors. Open AI's moderation endpoint was found to be more inconsistent in hate speech classification, while OpenAI's GPT 4 and Google's Perspective API were the "most measured", meaning they had comparatively lower rates of finding hate speech violations.
2. Demographic variations: Across models, the researchers found that hate speech targeting race, gender and sexual orientation were more easily detected by the models, compared to education, showing that the models "generally recognize hate speech targeting traditional protected classes more readily than content targeting other groups." Content targeting "woke people" was flagged as hate speech by some models but not by others.. For content targeting Christians, the Mistral Moderation Endpoint classified it as hate speech, while the Perspective API assigned the identical content a substantially lower score. For content containing the most severe anti-Black slur, some models assigned the maximum possible hate speech score, while others classified the same content as less hateful.
3. Implicit hate speech detection: The researchers also tested how systems handled sentences that paired positive language with slurs, such as "All [slur] are great people." Positive sentences containing anti-Black slur received the highest average hate speech scores across models, while positive statements about ideological groups like "commies" received "lower hate values" despite identical sentence structure. The disagreements were starkest for statements about "alt-right members," where the Mistral Moderation Endpoint classified the content as near-maximum hate speech while GPT-4o assigned it a score of zero. This reveals a fundamental disagreement between systems: some, like Claude 3.5 Sonnet, treat slurs as harmful regardless of positive context, while "less sensitive systems" prioritize "overall positive sentiment."

This paper introduces and evaluates an LLM-based framework for analyzing the ideological discourse pertaining to news events. The authors do so by representing news articles based on their talking points, capturing how the media frames the particular topic of discussion. The framework then uses an LLM to extract "prominent talking points (PTPs)" from news events. Each PTP is "infused with ideological information," which reveal left-leaning vs right-leaning viewpoints, referred to as partisan perspectives. The researchers develop and release a dataset including 6,141 news articles sourced from 126 outlets covering 24 events related to 4 politically contested topics. For each event, the framework generates a PTP and the aggregate partisan perspectives. The researchers evaluated the framework's ability to generate these perspectives via both automated tasks and human validation. The paper then seeks to assess whether the LLM-based framework can predict the partisan leaning (left or right) of news articles related to the event, but not part of the initial dataset. For each new article, the framework identified the three most similar left- and right-leaning partisan perspectives and asked an LLM to determine which group the article aligned with most closely. The paper compares this method with simply prompting an LLM to provide ideological labels. Here are some key findings:

1. The author’s classification approach outperformed directly prompting an LLM, suggesting the framework effectively captures ideological signals across many articles about the same event.
2. When the researchers used the partisan perspectives as training data to fine-tune a model, the fine-tuned model outperformed the base model on ideology classification which indicates that the framework's generated partisan perspectives "encodes ideology-specific nuances."
3. The researchers also conducted a human evaluation to measure the "quality of generated partisan perspectives" and found that these viewpoints can be incorrect, especially when the LLM produces inaccurate summaries of news articles.

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."

Tiktok's popularity has grown substantially in recent years, with recent Pew research suggesting that four in ten young adults in the US get their news from Tiktok. Yet, little is known about the prevalence and impact of political content on the platform. In this paper, the authors used a "novel browser-based tool to track users' exposure to and perceptions of political content," across 368 U.S.-based participants. Using a combination of annotations by users and natural language processing, the authors identify and categorize political topics encountered by TikTok users. They find that most of the users have a positive experience on TikTok and do not see political content. Other key findings include:

1. When users were initially asked to describe political content, they defaulted to broad terms like "politician" or "election." But when reacting to actual videos in their feeds, they identified specific topics like race, gender rights, and policing.
2. Most political topics evoked positive sentiment and fewer negative responses (anger or sadness). However, the strongest positive and negative sentiments were about content related to "COVID" and "TRUMP." Thus, overall political content on TikTok "often leans positive," but polarizing topics are still portrayed with strong sentiments across party lines.
3. The authors found that even when users encountered content that appeared to be from news broadcasts, it was rarely posted by official news media accounts. The authors suggest this indicates that "traditional news media accounts have yet to fully breach the TikTok platform."

As AI chatbots are increasingly designed to present as empathetic, a key question remains unanswered: do users actually experience them as empathetic? This paper examines how chatbot identity and perceived empathy shape users' overall conversation experience, analyzing 155 conversations in which participants interacted with either a human or a GPT-based chatbot. Using a combination of user self-reports and multiple computational measures of empathy, the researchers compared how empathy and conversation quality were perceived across human and chatbot interactions. Here's what they found:

1. Higher quality but lower empathy for chatbots: Participants reported higher conversation quality with GPT-based chatbots but they were also consistently perceived as less empathetic than human conversational partners.
2. Independent GPT-4o evaluation aligns with user perceptions: GPT-4o, used as an independent evaluator, rated chatbots as less empathetic than humans without being told which was which, aligning with what users reported in their self-assessments.

This paper examines whether deactivating Facebook or Instagram improves users' emotional wellbeing. In an experiment 20 times larger than previous experiments, the researchers recruited 19,857 Facebook users and 15,585 Instagram users and paid a subset of participants to deactivate their accounts for the six weeks leading up to the 2020 US presidential election. They then measured self-reported levels of happiness, depression, and anxiety at the start and end of the study. Here are some of the findings:

1. Deactivating Facebook improved emotional wellbeing: Users who deactivated their accounts reported a small improvement in overall emotional state compared to those who did not, with the largest effects seen in users over 35.
2. Deactivating Instagram improved wellbeing for young women: While there was no overall change in users' state, women aged 18-24 reported modest improvements in emotional wellbeing.
3. The improvements were not explained by increased offline time: The study found that users migrated to other platforms rather than going offline entirely, suggesting the improvement in mental health was not solely a consequence of reduced screen time.

This paper empirically examines how often Americans actually sanction (or “cancel”) others for offensive speech, why they do so, and how accurately they perceive others' canceling behavior. The researchers administered a nationally representative survey (N=1,752), asking if respondents engaged in specific “cancelling behaviors” and the extent to which they perceived others to do so. Subsequently, respondents participated in an experiment, reading four hypothetical scenarios wherein speakers made potentially offensive statements. Each scenario randomly varied the speaker's partisanship or race, their role, and what they said. Respondents rated their likelihood of engaging in different canceling behaviors for each scenario. Here are some key findings:

1. The paper finds that Americans tend to inflate by at least a factor of two “how often their fellow citizens cancel others.” In particular, the research found that respondents were “10 times more likely” to witness someone else engage in doxing than to have done it themselves, showing that people’s perception about the prevalence of cancelling behavior is far greater than the number of people who do it.
2. Contrary to popular belief and media coverage regarding cancel culture, this research suggests that Americans tend to cancel offensive speech that counters their "ideological leanings, regardless of who says them.” As the authors note, “generally, citizens do not care who makes offensive statements.”
3. Democrats and Republicans are “similarly likely to engage in cancelling behavior” due to “ideologically disagreeable ideas.” But why does this finding contradict reports suggesting Democrats cancel more often than Republicans? The authors contend that the “supply of offensive statements” during data collection likely had a “right leaning bias.”

This paper empirically investigates whether actively listening to someone's views on a topic before trying to convince them about an alternative viewpoint is effective. Or put simply, does listening enhance persuasion? To address this question, the researchers conducted a field experiment with 1,485 participants who had 10-minute video conversations with trained canvassers about unauthorized immigration policy. Participants were randomly assigned to four conditions: 1) canvassers listened to participants' views; 2) canvassers listened then shared a persuasive narrative; 3) canvassers shared a persuasive narrative without listening; or 4) a placebo where canvassers did neither. The main outcome of interest was whether participants reported a “reduction in exclusionary attitudes”, that is, lower “prejudice toward undocumented immigrants and support of anti-undocumented immigrant policies.”Here are some key findings:

1. Persuasive narratives changed attitudes substantially, whether or not canvassers listened first. Participants who heard a persuasive narrative showed reductions in anti-immigrant prejudice and opposition to pro-immigrant policies, with effects persisting five weeks later. Thus, while a persuasive appeal “effectively changed prejudice and policy attitudes,” “adding listening to the persuasive appeal did not change attitudes any further.”
2. Contrary to popular belief, this research found little evidence to support the claim that listening alone can change political attitudes, with the study only finding “marginal differences” between the “listening only condition” and the placebo.
3. Participants who were listened to before being persuaded reported feeling less defensive and had more favorable views of the canvasser, but this did not significantly change their attitudes. This shows that people tend to change their attitudes in response to persuasive appeals even from those they dislike.

Academic research and media coverage alike have argued that recommendation algorithms used by companies like YouTube are optimized to drive engagement and have thus driven political polarization by creating “filter bubbles” and “rabbit holes.” Rabbit holes differ from filter bubbles in that recommendations become extreme over time. In this paper, the researchers address this question by creating an interface that mimics how YouTube presents videos and recommendations to users. They simulate filter bubbles and rabbit holes by presenting participants with ideology balanced and more partisan content. The goal here is to see whether the recommendations “alter users’ media consumption decisions and, indirectly, their political attitudes.” Here are some key findings:

1. The study finds that while changes to recommendation algorithms shapes “user demand” by changing the types of videos consumed and time spent on the platform, it did not have substantial effects in changing political attitudes in the short-term. For example, when the algorithm recommended more videos matching the ideology of what users had just watched (rather than showing balanced recommendations from both sides), the share of liberal videos chosen increased by 6 percentage points among liberals and decreased by 12 percentage points among conservatives, yet these changes in viewing behavior did not translate into meaningful shifts in political attitudes.
2. On the question of whether recommendation algorithms put users in rabbit holes where they see more extreme content over time, the authors found no significant effects. As the authors note, “any algorithmic effect for rabbit holes that exists is likely far smaller than simply watching conservative or liberal video sequences.”

This paper asks whether the political makeup of police forces reflects the communities they serve—and whether it matters for how officers actually behave on the job. To find out, the authors collected roster data from 99 of the 100 largest local U.S. police agencies, covering roughly 220,000 officers, and matched those records to voter files to determine each officer’s party affiliation. They then used detailed data from Chicago and Houston to compare how Democratic and Republican officers behaved when assigned to the same beats, at the same times, facing the same conditions. Key findings include:

1. Police officers skew more Republican than the communities they serve. Across the largest police agencies, roughly 32% of officers are registered Republicans, compared to about 14% of civilians in those same jurisdictions. Officers are also more likely to be White, higher-income, and more frequent voters. This mismatch holds even at the district level—in every single Chicago district, and nearly every Houston division, officers skew more Republican than local residents. That said, there is meaningful variation: some agencies, like Birmingham, AL, closely mirror their communities, while others, like Rochester, NY, are dramatically out of step.
2. Party affiliation has little effect on how officers actually behave. Democratic and Republican officers showed no significant differences in stops, arrests, or use of force when working comparable assignments. The one exception: among White officers in Chicago, Democrats made more arrests for violent crimes than Republicans.
3. Race, not party, is what more consistently predicts officer behavior. Black and Hispanic officers in Chicago made significantly fewer stops and arrests, and used force less often, than White officers in comparable situations—particularly in encounters with Black civilians. In Houston the picture is more mixed: Black officers used force less frequently than their White peers, while Hispanic officers made more stops than White officers.

The paper examines what motivates people to share high-quality news, and whether that motivation can be deliberately triggered to increase sharing. Drawing on the idea that sharing is a “value-based decision,” they test whether content feels more worth sharing when it connects to a person's own identity (self-relevance) or to their relationships (social-relevance). To test this, they designed an intervention that prompts people to briefly write why a climate or health news story matters to them personally or to someone they know. They ran this test across four studies of increasing realism with a combined 2,559 participants in the United States and Netherlands. Here are some key findings:

1. Across all four studies, the proposed intervention increased self- and social-relevance ratings and sharing intentions compared to control. In the field study, this extended to actual sharing behavior on platforms like Facebook and Twitter, and the effect generalized across climate and health topics and across the U.S. and Dutch samples.
2. The fMRI analysis showed that brain regions tied to self-referential thinking and social cognition were more active during the intervention and independently predicted sharing intentions. However, this only held when participants wrote out their reasoning rather than silently reflecting on it, suggesting active engagement—not just exposure to the prompt—is what drives the effect.

The paper, the most ambitious analysis to date on how content propagates on social media, examined how misinformation spread during the U.S. 2020 Election on Facebook. The authors studied the diffusion of more than one billion posts between July 2020 and February 2021. They further explored the extent to which Facebook's content moderation policies were effective in dealing with misinformation in that time period. They found that:

1. On Facebook, most information spreads through public pages that broadcast to a larger audience at once. However, misinformation is more likely to spread through private exchanges between users, and was most often found among users who were older and conservative.
2. Viral misinformation on Facebook takes longer to gain traction and is likelier for content classified as "political news." Most reshares take place 24 hours after the original post from a relatively small subset of users—"even if the views it accumulates still amount to millions."
3. Content moderation efforts, particularly emergency "break the glass" measures implemented in the weeks before the election, proved highly effective at curtailing the spread of misinformation. The authors found that the number of misinformation posts reaching large audiences declined steadily from July onward, and viewership "plummeted to near zero" in the two weeks before Election Day.

This paper looks at how people tend to interpret news differently depending on whether they read or listen to it. Through a series of experiments (N= 14,744), the paper documented how listening vs reading shapes how participants interpret a news story. Participants were presented with a news story that included an equal discussion of a product’s risks and benefits, and then the researchers examined the participants’ interpretation or takeaway from the story. Here are some of the findings:

1. For the news story (risks and benefits of a product), listeners selectively attended to the negative details of a story and had a better recall for those details as compared to the positive ones. As a result, they formed more unfavorable impressions compared to the readers. However, readers had better overall recall—listeners retained negative details more disproportionately.
2. The listeners’ higher recall of negative details also resulted in responses such as decreased likelihood of using the product, higher levels of anxiety about the safety of the ingredients, and retellings that emphasized the risks over the benefits.
3. When researchers controlled the speed at which readers received information—simulating the listeners’ experience—the study found that both groups interpreted the news story similarly.
4. The researchers wanted to find out when listeners would display a positive outlook—“positivity bias”—towards a news story. To explore this, the participants were presented with a mixed-valence story on the impact of screen-time on adolescents where the benefits were more surprising than the risks. Here, they found that the listeners developed a positivity bias toward the impact of screen-time compared to the readers who still had a more balanced outlook.

Social media platforms are rife with online firestorms, wherein groups of users shame people for offending political sentiments. In this paper, the authors empirically examine why certain individuals partake in this sanctioning behavior. While shaming people with opposing political beliefs is not novel, social media enables people to get approval from fellow partisans for such behavior (via likes, comments and reshares). Despite these incentives, past research shows that relatively few people engage in such behavior. Through three experiments, the researchers tested whether sanctioning political speech online signals partisan identity, how group approval shapes willingness to sanction, and whether people are more likely to engage in this behavior if they are not the first to do so. Here’s what they found:

1. Tweets that signal partisan identity without explicitly stating party affiliation affects attitudes “nearly as strongly as” explicitly partisan tweets. For example, a Democrat reading a tweet critical of transgender people is almost as likely to assume the author is a Republican as if they had read a tweet actively campaigning for a Republican candidate. While it's expected that tweets explicitly stating party affiliation signal partisanship, it is worth noting that implicit messaging is just as impactful.
2. Participants believe that publicly criticizing political opponents online is something their fellow partisans both approve of and expect them to do.
3. People are reluctant to be the first to sanction someone online, but will readily pile on once others have acted. Respondents shown an offensive tweet accompanied by existing critical replies were significantly more willing to sanction the speaker than those shown the same tweet with no replies.

This paper presents a new method to attenuate the impact of misinformation, as an alternative to the predominant method of "correction" or fact-checking (which directly refutes misinformation). Through a series of six iterative experiments the paper demonstrates that "bypassing"—presenting an accurate alternate statement that does not directly refute the false claim—is a more effective method. An example of this would be, when a reader encounters a headline "genetically modified foods have health risks," then the counter (bypassing) statement would be, "genetically modified foods help the bee population." The three key advantages that the paper highlights of bypassing over correction are:

1. Subverts the Discomfort of Confrontation: Corrections are inherently adversarial, putting the reader at odds with a previously held belief, resulting in defensiveness. Bypassing overcomes this by not engaging with the belief.
2. Reduces Cognitive Load: Correction requires the reader to consciously revisit the original statement of misinformation from memory and then alter perception. Bypassing overcomes this by presenting the reader with alternative statements to rely on, without the added task of deliberative reasoning.
3. Why Bypassing Works: Bypassing outperforms correction when the focus of the reader is on the accuracy of the newly provided information, as opposed to changing already formed attitudes, in which case bypassing remains as ineffective as correction.

A limitation of bypassing lies in the fact that it did not prove to be as effective in combating misinformation when the reader was presented with real headlines, yet still produced better results than correction even within that setting. Notably, neither correction nor bypassing changed attitudes or intended behavior on specific policies, when it came to the real headlines.

The researchers examine how Twitter/X’s recommendation algorithm shapes the news content that users encounter, and how exposure to an alternative chronological feed affects user behavior and perceptions. To investigate this, 243 Twitter/X users installed a custom browser extension over three weeks: one week of passive observation followed by two weeks of intervention, during which users were randomly assigned to either the algorithmic (“For You”) or chronological (“Following”) timeline and then switched to the other. Key findings include:

1. Contrary to concerns about filter bubbles and radicalization, the algorithmic timeline produced lower quantity but higher quality news: users viewing the algorithmic feed saw more diverse and reliable news sources compared to those viewing chronological feed. The news users see from accounts they follow themselves appear more ideologically similar than purely algorithmic recommendations.
2. The algorithmic timeline kept users more engaged—they loaded more tweets and interacted more—but at the cost of news exposure, showing them significantly fewer news links. Notably, when users did encounter news, they engaged with it at similar rates regardless of which feed they were on.
3. Users were more satisfied with Twitter/X during the week they could freely choose their timeline than during either intervention week, regardless of which feed they were assigned to.

The paper tests whether believing that the 2020 election was stolen from Trump—the “big lie”—makes people distrust future elections, too. The authors argue, “election skeptics” (those who don't believe 2020 was legitimate) come to see all elections as suspect, both before any votes are cast and especially after their own candidate loses. To test this, the authors used panel survey data from over 3,800 voters in Michigan, Pennsylvania, and Wisconsin, tracked continuously from 2020 through the 2022 midterms. The same people were surveyed repeatedly before and after the 2022 election, enabling the researchers to examine how individual attitudes shifted. Here are a few key findings:

1. Questioning election legitimacy is a stable attitude. The authors find that beliefs about the 2020 election shifted considerably around the election itself, but changed very little afterward. In the post-election data, election skepticism was relatively stable, comparable to strong social identities like partisanship.
2. Skepticism predicts distrust. Those who doubted the legitimacy of the 2020 election were also far less confident that the 2022 midterms or the 2024 presidential race would be fair. In fact, skepticism about 2020 predicted distrust in future elections more strongly than party, ideology, or any other demographic variables measured in this study. For instance, Republicans who did not distrust the 2020 election were more confident about the fairness of future elections.
3. Losing hits skeptics much harder than everyone else. After the 2022 election, confidence in the results rose for nearly every group of voters—except election skeptics who backed a losing candidate; their confidence dropped further. The “winner-loser gap” for Senate races (the difference in trust in elections between people whose candidate won versus lost) was about 5.5 times larger for skeptics than for non-skeptics.
4. Losing skeptics reach for fraud as an explanation. Election skeptics who backed a losing 2022 candidate became roughly 5% more likely to say voter fraud had impacted that race—while election skeptics whose candidate won became less likely to blame fraud. Losing, for this group, effectively confirmed their prior belief.

This paper tests the efficacy of 25 crowdsourced, behavioural interventions aimed at reducing partisan animosity and antidemocratic attitudes among US voters. The interventions drew on a wide range of strategies—from showing people relatable individuals across party lines, to correcting wildly exaggerated beliefs about the other side, to presenting footage of democratic collapse. Researchers measured impact across three distinct fronts: how much people dislike the opposing party, whether they support actions that undermine democratic norms and/or endorse political violence. Here are some key findings:

1. Most interventions reduced partisan hostility: 23 of 25 treatments significantly reduced how much participants (up to 10.5 percentage points) disliked the opposing party, with the strongest effects coming from humanizing the other side (short videos of relatable people with different political views) and emphasizing shared identities ("we're all Americans").
2. Reduced support for undemocratic practices and candidates are highly correlated: 6 of 25 treatments significantly reduced support for undemocratic norms (up to 5.8 percentage points) largely driven by correcting the misconceptions of beliefs held by the other party, for example, a video highlighting that supporters of the opposing party did not dehumanize them. The second effective strategy highlighted real-world consequences of democratic backsliding, operationalized through a video compilation of violent forms of civic unrest from around the world.
3. A few treatments weren't as effective and some had adverse effects: 4 of 25 of interventions that had decreased partisan animosity increased support for undemocratic practices,, for example, describing a likable member of the opposing party. Similarly, one intervention, a video compilation of civic unrest, reduced support for undemocratic norms, also resulted in an increase in endorsement for political violence. The authors suggest this is because it ended with a video of the January 6th unrest which some people believed was a legitimate form of protest.

This paper examines whether news content highlighting core social identities like race, gender, religion, and political affiliation is more likely to generate audience engagement, and whether that audience demand helps explain why such news coverage has grown in recent years. The researchers analyzed tweets from 19 major media outlets (2008–2021) and 553,078 news URLs shared on Facebook. They used AI text classifiers to track identity language over time and further measure its relationship to audience engagement. Here's what they found:

1. Identity-oriented news coverage has grown substantially: The share of tweets referencing social identities increased steadily over time—rising from 6.6% of tweets before 2015 to 10.3% in 2015 and beyond. Mentions of racial and partisan identities roughly tripled since 2007, with notable spikes around the 2016 election, #MeToo, and the killing of George Floyd.
2. Identity oriented content drives engagement: Both Facebook and Twitter data show that posts featuring identity-related language consistently received more likes, shares, and retweets than those without, across nearly every outlet studied.
3. The relationship between identity oriented coverage and engagement is causal: Through causal experiments that randomized exposure to identical news stories "with and without identity cues," the authors found that stories received significantly more clicks if the headline referenced a social identity.

This paper looked at the effect of Facebook and Instagram access on political attitudes. The researchers examined the effect of deactivation, for a randomized subset of 19,857 Facebook users and 15,585 Instagram users, on "consumption of other apps and news sources, factual knowledge, political polarization, perceived legitimacy of the election, political participation, and candidate preferences." This experiment was conducted for a period of six weeks just before the U.S. 2020 election. Here are the findings:

1. Deactivation reduced trust in political information: Deactivating Facebook made users trust political information on Facebook less without affecting how much they trusted Instagram. The same was true for Instagram deactivation in that it reduced trust in political information on Instagram without affecting Facebook. The researchers believe that "time away from a platform made users more aware of the amount of low-quality or inaccurate information to which they had been exposed."
2. Deactivation did not alter political attitudes: Facebook and Instagram deactivation had no significant effects on reducing online polarization or altering beliefs in the legitimacy of the 2020 U.S. election.
3. Facebook deactivation altered news consumption patterns: Facebook deactivation did not meaningfully affect participants' factual knowledge overall, but additional analyses pointed to two findings: The Facebook deactivation group reported less news knowledge but had greater capacity to "distinguish misinformation from true stories."
4. Weak evidence for voter switching: The paper finds weak evidence for a possible impact "of 1.3% of Trump voters to switch to Biden, or if it caused 2.6% of Trump voters to not vote."

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.”

This paper examines how users consume social media content and what their online information environment actually looks like. Most past research relies on surveys, diaries, or click data, all of which capture only a small slice of what shows up in someone’s feed. To get a fuller picture, the authors built a tool called PIEGraph, which collected the full X/Twitter timelines of 790 participants (everything posted by the accounts they follow) and paired that with survey data on their demographics, politics, and media habits. Because data collection happened to overlap with Elon Musk's takeover of X/Twitter, the authors were also able to see whether feeds changed after the acquisition. Key findings include:

1. Political content is a small share of feeds, but larger than previously thought: Just over 10% of links in participants’ feeds were political, higher than the 2–3% found in earlier click-based studies.
2. Feeds lean left and skew toward high factual quality: About 80% of participants had left-of-center feeds, and low-quality content was rare.
3. Musk's takeover didn't cause a dramatic shift: Post volume dropped after October 2022, but ideological leaning, political content, and factual quality stayed roughly stable.
4. Individual differences mattered more than platform changes: Mainstream media use and the belief that January 6th attack was exaggerated all predicted more conservative, lower-quality feeds. The belief that Jewish people control the media and government predicted higher fact quality, which the authors suggest may reflect people consuming outlets they distrust just to criticize them.

This paper asks whether social media algorithms push people to take action in the real world, like voting, protesting, or volunteering. The authors review existing research on this topic and propose a framework for how social media can facilitate offline civic participation. The framework highlights three mechanisms through which social media algorithms affect civic participation: by shaping attitudes on societal topics, promoting collective action, and building social capital. Here are the findings:

1. Information curation: Social media algorithms predominantly expose users to content that aligns with their existing beliefs, strengthening their beliefs and making them more likely to act on them. These algorithms also diversify information exposure, and this combination of factors, the authors argue, could benefit offline civic participation.
2. Promoting action: Social media platforms are designed to reward actions—liking, sharing, commenting—over inactivity, which may instill a general tendency toward action among habitual users. Small online actions may also serve as a “critical gateway” to higher effort real-world commitments over time.
3. Building social capital: Social media algorithms lower the barriers to forming new social connections organized around shared interests and values. Frequent interactions with like-minded others build trust and group identity that can translate into offline action.

This paper analyzes whether the political ideology of Facebook users shapes the news content they see in their feed. The researchers looked at all political news available on Facebook during the US 2020 election compared to what the 208 million US Facebook users actually saw on their own feeds. These are few of their findings:

1. News on Facebook exists in silos: The news users see on their feeds depends on their ideology which is amplified both by the platform's algorithm and the content users choose to engage with. Ideological segregation is more visible on Facebook Groups and Pages compared to individual user posts.
2. More conservative domains circulate on Facebook: When looking at news websites and articles shared within political bubbles on Facebook, far more of those sources existed within conservative circles than liberal ones — a pattern consistent with findings from other social media platforms.
3. Higher levels of misinformation are present in conservative circles: Most content flagged as misinformation by Meta's Third-Party Fact-Checking program was found in conservative circles, meaning "that conservative audiences are more exposed to unreliable news."

The researchers examined whether Facebook and Instagram feed algorithms, which select and recommend content that appears on a user's feed, contribute to political polarization. Over three months, they modified how content was displayed for a subset of over 40,000 U.S. based users, switching from an algorithmically curated feed to one ordered with the most recent posts at the top—or reverse chronological order. Here are some of the findings:

1. User engagement with the platforms dropped: Switching to a chronological feed led users to spend considerably less time on both platforms and interact with content—likes, reshares, comments—less frequently. Some users instead spent more time on other platforms such as TikTok and YouTube.
2. The chronological feed changed what content users saw: Users were exposed to more political and unreliable content across both platforms. On Facebook specifically, they encountered less uncivil content and fewer posts containing slurs, while seeing more content from politically moderate connections and ideologically mixed sources.
3. Changes to the feed did not change political attitudes: Despite the notable shifts in how users engaged with and experienced both platforms, the chronological feed had no measurable effect on political polarization, political knowledge, or other key attitudes over the course of the study.

Researchers examined the impact of how exposure to reshared content on Facebook affects its users, given that "prior research suggests that reshared content may increase both political polarization and political knowledge." To test this, they removed reshared content from the feeds of a randomly selected subset of 23,402 US-based users for three months in the lead up to the 2020 U.S. presidential election. Here are a few findings:

1. Reshares substantially amplified political content: The removal of reshared content noticeably changed the makeup of what users saw on their feeds. Political content overall dropped by nearly 20% (from 13.5% to 10.8% of all content), while exposure to political news fell far more sharply—by almost 60% (from 6.2% to 2.5% of all content). Content from untrustworthy sources also decreased in user-feeds by 30.6% (from 2.6% to 1.8%).
2. Less reshared content increased exposure to moderate content: With less reshared content in their feeds, users were exposed to more ideologically moderate content—by 15% (from 22.6% to 26.2% of all content).
3. Behavioral shifts but no change in political beliefs: The removal of reshared content reduced overall user engagement as people were less likely to click on links. More notably, users also clicked less frequently on content from partisan sources. Despite these behavioural shifts, the removal of reshared content had no effect on users' political beliefs or polarization.

The researchers test whether only seeing like-minded content, or being in an 'echo chamber,' on Facebook actually has an effect on polarization and political attitudes. They conducted an experiment in which a subset of participants were exposed to less content from like-minded sources (relative to their normal exposure) for 23,377 U.S.-based Facebook users for three months, from 24th September to 23rd December 2020 (around the time of the U.S. 2020 election). These are a few of their key findings:

1. Political and news content mostly comes from like-minded sources: A relatively small proportion of content that Facebook users see on their feeds is explicitly political or news, just around 7%. The median user is exposed to much more political (68%) and news (57%) content from like-minded sources, compared to similar content from cross-cutting sources (closer to 12% for both politics and news).
2. Some users have extremely high like-minded content exposure: One fifth of the users have extremely high exposure to like-minded sources, with over 75% of content coming from those aligned sources.
3. Reducing like-minded content had mixed effects: The experiment successfully reduced exposure to like-minded sources, which also decreased the amount of uncivil content and misinformation users saw. However, it had an unintended effect: when users did encounter content from a like-minded source, they were more likely to engage with it.
4. Reducing like-minded content did not change political attitudes: They also found that the experiment's reduction of content from like-minded sources did not have an impact on political attitudes—polarization, ideological extremity, or beliefs in false claims.

This study investigates whether rising affective polarization in American politics might have a side effect: increased political participation. Using American National Election Studies (ANES) data spanning presidential elections from 1980 to 2020, along with a nationally representative panel survey from 2016–2020, the authors test whether polarized feelings toward parties and candidates predict intent to vote, self-reported turnout, validated turnout, and broader campaign participation (e.g., attending rallies, donating, displaying buttons). They examine two measures of polarization: "affective polarization" (the gap in warmth felt toward one’s own party versus the opposing party) and "candidate thermometer difference" (the gap in feelings toward the two presidential candidates). Key findings include:

1. Candidate evaluations matter more than party feelings: Polarized views of the presidential candidates were a far more consistent predictor of voting intent and turnout than general affective polarization between the two parties. Candidate thermometer difference predicted intent to vote in all 11 election years studied, while affective polarization was significant in only 3.
2. Self-reports overstate the effect of Polarization: When using turnout records instead of self-reported voting, the relationship between polarization and turnout became much weaker or disappeared entirely. This suggests that polarized respondents are more likely to say they voted or intended to vote, even when they did not, indicating an expressive or performative element to reported polarization.
3. Panel data confirms a relationship between candidates and political participation: Tracking the same individuals from 2016 to 2020, increases in candidate thermometer difference predicted an individual’s increased likelihood of voting, reinforcing that it’s candidate-specific animosity/enthusiasm, not general party hostility, that drives behavior.
4. Modest but real impact on turnout: Roughly one-quarter to one-third of the actual rise in voter turnout between 1980–2016 (and again 2016–2020) can be statistically attributed to rising polarization in candidate evaluations.

This paper reviews whether Fox News viewership influences consumers views on race and immigration. It uses a survey administered between 2008 and 2020 to test whether changes in individuals’ Fox News viewership are associated with changes in anti-Black prejudice, attitudes about government assistance to Black Americans, and views on unauthorized immigration. Key findings include:

1. No meaningful link to anti-Black prejudice: Among White respondents, increased Fox News viewership “itself did not increase anti-Black prejudice.” Similarly, the researchers found that consuming Fox News content did not shape views on whether the government should assist Black Americans. However, the authors note that these findings do not mean that Fox News “has no race-related impacts.”
2. Negative views on unauthorized immigration: Increased Fox viewership was significantly associated with declining support for immigration and increased support for a border wall/deportation. In the authors’ words, “people who started watching Fox News became more restrictionist on immigration relative to those who stopped.”
3. News coverage prioritizes immigration: A content analysis of Fox News transcripts found the network devoted more coverage to immigration than to race-related issues over this period, aligning with the survey results.

The authors tested whether facilitating face-to-face conversations between Democrats and Republicans would reduce how much they dislike each other, and make them more supportive of democratic values. To do this, a pair of partisan strangers were randomly assigned to different conversation formats and topics. In Study 1, participants discussed their "perfect day," wherein half of the participants knew the party affiliation of their partner while the others did not. In Study 2, two additional conditions prompted participants to discuss partisan preferences—why they support their own party or dislike the other. Key findings include:

1. Simply knowing you were talking to someone from the other party—while having an ordinary, non-political conversation about your ‘perfect day’—significantly warmed partisan feelings toward the other side (both studies demonstrated this). But those gains had fully disappeared three months later.
2. When conversations turned to political differences rather than shared experiences, the warming effect vanished entirely. What people talk about matters as much as whether they talk at all.
3. Even when conversations did reduce partisan dislike, they had no effect on how people actually engage with politics—their support for bipartisanship, cross-party voting, or democratic norms remained unchanged.