Psychology by BogdanovaThe Encyclopedia of Psychology for the Artificial Era

RESEARCH DOSSIER

Digital Political Exposure, Algorithms, and Opinion Dynamics

Abstract. Digital political exposure, algorithms, and opinion dynamics form an interdependent research field concerned with how the selection, recommendation, and social circulation of political information affect beliefs, perceived norms, partisan evaluations, political knowledge, and civic behavior. This fundamental research monograph integrates classic selective-exposure and mass-opinion theories with network models, observational platform studies, feed-randomization experiments, independent algorithmic audits, and recent evidence through 2026. Its central distinction separates potential exposure, delivered content, engagement, psychological interpretation, and longitudinal outcomes. Population-scale Facebook research documents ideological segregation; randomized Facebook and Instagram interventions show major exposure changes without detectable short-term changes in several polarization measures. In contrast, targeted X reranking and the 2026 X feed experiment demonstrate conditional effects on animosity and issue attitudes, while customized feeds alter perceived norms. A 2026 TikTok audit establishes delivery asymmetries without measuring human persuasion. The evidence supports mechanism-specific, outcome-specific, and temporally sensitive causal explanations. This work is a critical narrative synthesis of published research rather than a preregistered systematic review, meta-analysis, or original participant study.

Keywords: digital political exposure; algorithmic ranking; recommendation systems; selective exposure; opinion dynamics; political polarization; affective polarization; partisan identity; social networks; political news; feed experiments; misinformation; ideological segregation; norm perception; political behavior; causal inference; political psychology; media and technology psychology; Artificial Era

The Research Problem and Its Explanatory Levels

Digital political exposure is the encounter with political information through digitally mediated environments, including search engines, social networks, recommendation feeds, messaging networks, and video platforms. Algorithms influence which available messages become visible, where they appear, how often they recur, and whose reactions are displayed with them. Opinion dynamics concern changes in political beliefs, issue priorities, attitudes toward groups, perceptions of public opinion, and observable political behavior over time. These processes form a coupled research problem: the distribution of political content is shaped by users, their networks, platform objectives, and institutional conditions, while information exposure can alter what users attend to, believe, share, and subsequently seek. The central scientific question is how a particular exposure mechanism changes a particular psychological outcome under specified social and temporal conditions.

A convincing explanation therefore distinguishes at least five levels: the supply of content by political actors and ordinary users; relational opportunity arising from social ties and subscriptions; algorithmic ranking or recommendation conditional on those opportunities; attention and interpretation by an exposed person; and downstream individual or collective outcomes. A favorable position in a feed is a change in visibility, a click indicates selective engagement, a follow changes future information supply, and a feeling-thermometer response measures an attitude. Each is substantively different. The large 2020 Facebook study by González-Bailón et al. (2023) González-Bailón (2023) separates potential exposure, actual exposure, and engagement across a population-scale dataset; the distinction is essential to this monograph.

The analysis that follows is a critical narrative synthesis of identifiable conceptual, observational, experimental, audit, and methodological research through November 2026. Studies were selected for their explanatory contribution and verifiable primary documentation; the corpus has no preregistered exhaustive search, pooled effect estimate, or claim of systematic-review completeness. Political orientations and parties are handled as empirical categories within each study's historical context, and substantive direction is attributed to the observed platform and period rather than generalized to political systems worldwide.

Conceptual Definitions and Operational Boundaries

Political content includes reporting, commentary, candidate statements, advocacy, policy arguments, mobilization messages, and expressions of partisan identity. Researchers must define whether a classification captures explicit politics, news from ideologically rated domains, a source's audience composition, or a post's semantic stance. Those classifications overlap but differ. A politically heterogeneous source can publish a partisan item, while a strongly aligned source can publish an ordinary nonpolitical post. It follows that domain ideology, speaker ideology, post ideology, and the respondent's position should be reported separately wherever measurement permits.

Exposure is likewise multidimensional. Potential exposure is the information eligible for delivery; delivered exposure is what a platform places in a feed; viewed exposure requires an opportunity to perceive an item; attended exposure implies cognitive allocation; engagement includes clicks, reactions, comments, reshares, or watch time; remembered exposure is retained knowledge. The observational study of Bakshy, Messing, and Adamic (2015) Bakshy (2015) illustrates how reductions in ideological diversity can accumulate across friend selection, algorithmic ranking, and personal clicking. A dataset containing only shares or impressions cannot establish the remaining stages without additional validation.

Opinion dynamics denotes trajectories of measurable outcomes rather than a single polarization index. Issue attitudes, policy priorities, partisanship, ideological extremity, perceived group norms, affective polarization, misinformation beliefs, political efficacy, information knowledge, and voting behavior may move in different directions. The 2026 X experiment found changes in issue-related opinions without corresponding detectable changes in partisanship or affective polarization; those are separate endpoints, not an inconsistent result (Gauthier et al., 2026 Gauthier (2026)).

Historical Foundations: Selective Exposure and Mass Opinion

Digital exposure research builds on twentieth-century studies of political communication and attitude formation. Festinger's (1957) Festinger (1957) cognitive-dissonance theory supplied one influential account of why people may prefer congenial messages when conflicting information generates psychological discomfort. Political-communication traditions examined interpersonal discussion, party attachments, news habits, agenda formation, and the interaction of elite signals with citizens' predispositions. These traditions matter because contemporary feed systems intensify and automate selection processes whose psychological ingredients predate the internet.

Zaller's (1992) Zaller (1992) reception–acceptance–sampling account explains why political messages can influence mass opinion differently as political awareness and predispositions change. Citizens do not always retrieve a stable, fully elaborated ideological position; survey responses may depend on accessible considerations activated near the point of judgment. In digital environments, repeated messages, ordering, and perceived social endorsement can affect accessibility even when durable ideological commitments remain stable. That implication requires longitudinal testing rather than inference from a temporary increase in impressions.

The transition from broadcast scarcity to networked abundance altered the mechanisms of selection. People can actively seek congenial information, friends can circulate it, and personalization engines can estimate what will sustain attention. Garrett (2009) Garrett (2009) found that attraction to opinion-supporting news was more pronounced than systematic avoidance of disagreeable information in a tracked online-news sample. This asymmetry challenges theories that equate preference for agreement with complete informational isolation. It also foreshadows the empirical distinction between selective engagement and ranked availability.

Psychological Mechanisms of Political Selection

Selective exposure reflects the interaction of goals, identities, and information costs. People may seek accuracy, social belonging, emotional regulation, entertainment, or ammunition for an argument. These motives can align or conflict within the same session. Political identity makes some claims relevant to the self and can change the perceived credibility of a source; uncertainty or curiosity can motivate counterattitudinal encounters. Consequently, selection is best modeled as a distribution of context-dependent choices, rather than a fixed disposition to encounter only agreement.

Taber and Lodge (2006) Taber (2006) documented motivated skepticism in political judgment: information evaluation and counterargument generation can depend on prior commitments. Their findings help specify a mechanism connecting exposure to interpretation. The same item can produce learning, defensive elaboration, reinforcement, indifference, or a shift in a narrower issue attitude. Exposure measurements without attention, perceived credibility, or belief change leave this psychological mechanism unobserved.

Individual selection also influences future algorithmic delivery. Clicking a political message can signal topical interest; following an activist changes the set of eligible sources; sharing creates a secondary audience. Such feedback means that the causal treatment is not always exhausted by a one-time ranking rule. A useful study distinguishes immediate exposure, persistent changes in followed accounts, and later feed composition. The empirical importance of this distinction became clear in the randomized feed-switching evidence of Gauthier et al. (2026) Gauthier (2026).

Social Networks, Homophily, and Available Information

A social feed depends on a network before a ranking system selects from it. Homophily describes the tendency for social connections to be associated with similarity; network formation can reflect friendship, geography, education, political orientation, and prior consumption. Political information available through followed accounts may therefore already be highly selective. Algorithmic curation can amplify or moderate the composition of the available inventory, but an analyst who ignores network selection may mistakenly attribute the entire pattern to a recommender.

The 2015 Facebook observational analysis used deidentified information from 10.1 million U.S. users to examine network homophily, feed ranking, and clicking. Its strongest methodological contribution was to locate multiple points of ideological narrowing within the exposure process. The study does not randomize political attitudes, and its findings cannot show that changing a ranking rule would necessarily change beliefs (Bakshy et al., 2015 Bakshy (2015)).

Social ties also mediate action. In the randomized experiment reported by Bond et al. (2012) Bond (2012), political mobilization cues delivered across a vast Facebook population affected expression and voting behavior, with important transmission through close ties. The design addresses a different causal target from feed ranking: social messages and signals of friends' actions. Together the studies show why opinion formation, information diffusion, and mobilization should not be treated as interchangeable consequences of platform exposure.

Algorithmic Ranking and Recommendation as Treatments

A feed algorithm is a policy for assigning visibility to content conditional on user state, candidate content, and platform objectives. A reverse-chronological feed orders a defined candidate pool by time; an engagement-based feed predicts interaction; a recommendation-based feed may add material from accounts the user does not follow. These systems differ in the information they make eligible and in the ranking of already-eligible material. Calling the chronological condition an algorithm-free baseline obscures those design decisions and the remaining influence of network subscriptions.

The estimand in a feed experiment is the effect of an operational intervention, such as assigning consenting users to a particular feed for a bounded interval. It includes behavioral adaptation, changes in time spent, and possible cross-platform substitution. It does not identify the isolated effect of every machine-learning parameter. Guess et al. (2023a) Guess (6656) compared chronological with default feeds on Facebook and Instagram during the 2020 U.S. election, while Huszár et al. (2022) Huszár (2022) audited algorithmic amplification on Twitter using a large randomized baseline.

Policy comparisons require attention to content supply, user population, and design period. A recommender optimized for watch time on short-form video need not reproduce the mechanisms of a friendship-based news feed. Platform objectives and available creator material change over time. Evidence about a particular platform is most informative when the platform, intervention, political period, and exposure measures are retained in the scientific claim.

Feedback Loops and Endogenous Exposure

Exposure can alter behavior that subsequently changes exposure. A user encounters a recommended speaker, follows that account, receives future posts, reacts to one, and becomes more likely to encounter related material. The algorithm responds to behavior partly produced by earlier recommendations. This creates an endogenous process in which the composition of the feed is jointly produced by ranking decisions and evolving preferences or habits. Standard regressions of exposure against opinion have difficulty disentangling these directions without independent variation.

The 2026 randomized study of X is particularly informative because turning the recommendation feed on and off yielded asymmetric results. Switching from a chronological to an algorithmic feed shifted several political attitudes; switching away from an established algorithmic feed did not reverse them. The authors linked the asymmetry to the continuing presence of newly followed politically aligned accounts after the experimental feed switch (Gauthier et al., 2026 Gauthier (2026)). Persistent follows are a plausible mediating pathway, while causal mediation of individual follows requires additional assumptions.

This feedback perspective also clarifies time horizons. A short intervention may change an immediately visible content mix without dislodging stable attitudes; a longer or earlier intervention may reshape the source network and produce cumulative exposure. Researchers should therefore observe both on-platform behavior and postintervention trajectories and should specify whether the estimand is an immediate effect, a cumulative treatment effect, or a persistent legacy of past exposure.

Selective Exposure and Ideological Diversity

Ideological diversity can mean a mixture of sources in a feed, exposure to statements that disagree with the user, interaction across partisan lines, or the presence of competing interpretive frames. These are related indicators with different denominators. A platform may feature outlets with diverse audience profiles while individual conversations remain highly homogeneous. Conversely, highly partisan accounts may repeatedly criticize the out-party, creating substantial cross-party mentions without reflective engagement with opposing arguments.

Flaxman, Goel, and Rao (2016) Flaxman (2016) analyzed patterns of online news consumption and found that search engines and social media were associated with both exposure to more diverse news and heightened ideological segregation for some users. The analysis illustrates a central distributional challenge: changes in average diversity may coexist with narrower information diets among politically intense minorities. The consequence of diverse exposure depends on how that diversity is measured and experienced.

A theory of diversity effects must also specify what a counterattitudinal message does. Challenging information can update a factual belief, increase out-group resentment, provoke counterarguing, or become irrelevant. In the field experiment of Bail et al. (2018) Bail (2018), compensated exposure to opposing political elites through Twitter bots did not provide a general demonstration that cross-cutting content moderates attitudes. The study reported movement toward greater polarization in particular participant groups, with important sampling and treatment limitations.

Echo Chambers, Filter Bubbles, and Their Measurement

An echo chamber is usefully operationalized as a social-information structure in which similar viewpoints circulate and external challenges have reduced credibility or reach. A filter bubble emphasizes algorithmically curated restriction of material. A heterogeneous discussion network may nevertheless contain highly segregated topics, and a homogeneous friendship network may expose a person to diverse nonpolitical information. Both concepts require explicit measures of exposure distributions, network clustering, and cross-cutting opportunities.

The comparative work of Cinelli et al. (2021) Cinelli (2021) analyzed controversial-topic exchanges across several platforms and observed platform-specific patterns of homophily and diffusion. Such results support studying interactions between network architecture and information movement. They do not independently identify whether a ranking change caused the formation of an echo chamber or whether an echo chamber changed a particular participant's opinion.

A useful measurement strategy reports the proportion of politically aligned versus cross-cutting exposures, the prevalence of political material among all impressions, and the concentration of exposure among high-activity accounts. It should validate source-ideology assignment and distinguish political disagreement from personal hostility. Claims that a platform is uniformly polarized become misleading when political material occupies only a limited fraction of user attention or when distributional asymmetries are concentrated in a subset of users.

From Potential Exposure to Engagement: The Facebook Evidence

González-Bailón et al. (2023) González-Bailón (2023) examined political news during the 2020 U.S. presidential election in aggregated data concerning 208 million U.S. Facebook users. The research compared the political-news inventory that users could have encountered, what appeared after curation, and content with which they engaged. Ideological segregation increased across those stages. This pattern is compatible with compounding effects of network structure, selection, and platform ranking, but the observational comparison does not allocate a causal share to each mechanism.

The researchers also documented an asymmetry within that platform-specific news ecosystem: sources with predominantly conservative audiences formed a particularly segregated segment, and most material identified as misinformation by Meta's fact-checking program appeared within that segment. These classifications are conditional on the study's source-ideology measures, fact-checking coverage, and the particular U.S. political period. They are findings about observed distributional structure rather than an assertion concerning every conservative user or every platform.

A crucial analytical step is to resist sliding from who encountered a political URL to who believed its contents. News-domain circulation, personal clicks, attitude change, and actual political decisions require different evidence. This distinction allows the 2023 descriptive segregation results to coexist with randomized intervention studies that changed feeds without consistently moving preregistered attitude endpoints.

Large-Scale Randomization of Chronological Feeds

The Guess et al. (2023a) Guess (6656) Facebook and Instagram study randomized consenting participants during the 2020 U.S. presidential election to chronological feeds or the platforms' default ranked feeds. Replacing algorithmic ranking reduced platform use and activity while increasing certain political and untrustworthy exposures, and changing the mix of moderate sources, civil language, and other content measures. The direction of these immediate exposure changes depended on the platform and content category; the experiment did not show that chronological ordering inherently guarantees higher-quality information.

Across the three-month treatment period the study found no statistically significant change on its principal issue-polarization, affective-polarization, or political-knowledge outcomes. The proper inference concerns the tested treatment, period, participants, and estimated effects; it does not establish that all algorithms lack political consequences. Reported null results should be read with confidence intervals, compliance, outcome sensitivity, and the possibility that prior long-term exposure had already shaped the participants' networks.

This experiment is a major advance because it moves the analysis beyond platform-level correlations. Its remaining boundaries are substantive. The control and treatment conditions both contain user-selected networks and prior interactions. A self-selected population of consenting participants may respond differently from the full platform population. Short-term switching off a longstanding ranking environment and long-term adoption of that ranking policy are separate causal questions, a distinction reinforced by the later X evidence.

Removing Like-Minded Content: A Different Intervention

Nyhan et al. (2023) Nyhan (2023) studied exposure to politically like-minded Facebook sources using population-wide descriptive data and a randomized election-period intervention involving 23,377 consenting users. The treatment downranked posts from politically aligned friends, groups, and Pages, reducing like-minded exposure by about one-third. Participants consequently encountered more cross-cutting material and less content with incivility and material from sources repeatedly posting misinformation.

Despite meaningful changes in exposure, the authors found no measurable treatment effects on eight preregistered attitudinal measures, including affective polarization and ideological extremity. Political news comprised a relatively small share of total feed content for typical users, one reason that gross like-minded exposure is an imperfect proxy for the psychologically potent dose of political messages. Importantly, the intervention reduced aligned material rather than compelling intensive engagement with convincing opposing arguments; its results should not be interpreted as a definitive test of deliberation.

The study is especially useful when paired with the chronological-feed experiment. Both altered information environments during the same major electoral period without producing the expected broad attitudinal depolarization, although their interventions differed and their exposure effects were substantial. The evidence favors greater precision about target content and mechanisms over the assumption that any reduction in homogeneity automatically changes intergroup feeling or issue beliefs.

Reshares, Amplification, and News Knowledge

Resharing multiplies the paths by which political messages enter a feed. A story originally produced by a news outlet may reach users as an explicit recommendation, a friend's reshare, or a quote-post accompanied by emotional commentary. Each route presents a different social signal. The Guess et al. (2023b) Guess (6656) experiment removed reshared content from Facebook feeds among consenting participants during the 2020 U.S. election and documented substantial changes in news exposure, clicks, and reactions.

Removing reshares reduced the prominence of political news, including material from untrustworthy sources, and reduced partisan news clicks. It also lowered news knowledge within the study sample, while producing no detectable effects on the examined political attitudes or polarization outcomes. This result underscores that interventions can have mixed epistemic consequences: removing an amplification channel may reduce exposure to low-quality information while also removing opportunities to encounter reliable news.

The correct research question is therefore what kinds of material a particular distribution channel amplifies, how users recognize its credibility, and which outcome matters. An intervention evaluated solely against the volume of misinformation can miss consequences for political knowledge; one evaluated solely against engagement can overlook the character of what is learned. Information quality, attitude hostility, and behavioral participation each need separately specified measurements.

Political Amplification on Twitter Before the X Reconfiguration

In a large-scale randomized assessment of Twitter's timeline, Huszár et al. (2022) Huszár (2022) compared algorithmic personalization with a reverse-chronological control group involving nearly two million daily active accounts. Across examined countries and political actors, the researchers found that political content received differential algorithmic amplification, with an advantage to the political right in their comparisons. This is a result about relative ranking and exposure rather than a direct estimate of consequent persuasion.

Algorithmic amplification is calculated in relation to a particular baseline and candidate-content environment. A higher ratio for one party or publisher does not by itself explain whether the difference emerges from content properties, prior engagement patterns, platform objectives, supply, or their interaction. The study's randomization strengthens claims about the feed policy's role in distribution; it does not transform platform-specific partisan comparisons into universal characteristics of recommender systems.

Temporal context is essential. The Twitter study was conducted under an earlier platform configuration than the X field experiment reported in 2026. An exact synthesis should track the technology and political environment rather than combine the two periods as if all ranking rules, user populations, and audience behavior remained fixed.

The 2026 X Experiment and Persistent Political Change

Gauthier, Hodler, Widmer, and Zhuravskaya (2026) Gauthier (2026) independently randomized U.S. X users in a seven-week 2023 field experiment to an algorithmic or chronological feed. The analyzed post-treatment sample comprised 4,965 participants. Switching the ranking environment on increased engagement and shifted several measured political opinions in a more conservative direction, notably policy priorities and assessments of contemporary political issues. The paper also reports no detectable change in affective polarization or self-reported partisanship from switching in either direction.

The asymmetry is the theoretical centerpiece: moving from chronological to algorithmic delivery affected some opinions, whereas moving away from the algorithmic feed did not produce a comparable reverse change. Exposure analysis indicated that the recommended feed promoted conservative content and demoted traditional-media posts in that setting. Followed-account data suggested a plausible cumulative mechanism, because some participants began following conservative activist accounts and retained those subscriptions after the ranking policy was changed.

The experiment shows that a short-run null effect of disabling a ranking system is consistent with a consequential earlier adoption effect. It also imposes its own boundaries: participants were active U.S. X users, compliance was monitored imperfectly, the platform was changing institutionally, and issue attitudes were responsive in ways that partisan identity was not. Careful interpretation preserves both the causal results and the limits of transport to other algorithms, time periods, and populations.

Independent Feed Reranking and Affective Polarization

A distinct experimental strategy changes the placement of defined content inside an otherwise functioning platform. Piccardi et al. (2025) Piccardi (2025) implemented an independent, preregistered ten-day field experiment involving 1,256 participants on X during the 2024 U.S. campaign. The intervention used a language-model-supported content classification to increase or decrease exposure to posts expressing antidemocratic attitudes and partisan animosity, enabling causal tests of a particular content dimension rather than an entire platform's default policy.

Changing exposure shifted out-party animosity by more than two points on a 100-point feeling thermometer in the authors' reported comparisons. This is an effect on an explicitly measured affective outcome, and its short experimental horizon remains crucial for judging persistence. The findings contribute a contrasting result to the 2020 Meta interventions: a targeted alteration of antagonistic content can measurably shift an interpersonal political attitude, whereas broad feed reordering and like-minded-source reduction did not detectably change the same broad family of outcomes.

The studies are best reconciled by specifying the treatment contrast. Removing general ideological agreement from a feed differs psychologically from moving hostility-laden posts upward or downward. The latter directly manipulates a likely driver of intergroup evaluation. Replications on other platforms, longer follow-up, independent coding audits, and heterogeneous treatment analyses are necessary before assigning the effect a stable cross-platform magnitude.

Norm Learning and Algorithmic Representations of Public Opinion

Political exposure carries information about the apparent prevalence and acceptability of opinions. Users can infer what others believe from the frequency of posts, size of visible engagement counts, prominence of extreme voices, and reactions from peers. These inferences may be distorted when algorithms optimize participation rather than representative social sampling. Perceived norms are therefore distinct dependent variables: a person's impression that opponents are angry or that hostility is common may change even when their own policy preference does not.

Brady et al. (2026) Brady (2026) reported a registered experiment in which 2,000 participants were assigned to alternative customized feeds for eight weeks surrounding the 2024 U.S. presidential election. Relative to chronological feeds, engagement-ranked feeds increased exposure to intergroup, moralized, emotional, and toxic material and increased perceived partisan animosity. The researchers also evaluated a diversified-extremity ranking system that reduced those exposures and improved accuracy in normative perception without reducing reported enjoyment in the tested setting.

This evidence isolates a psychologically specific consequence of ranking: distortions of social-norm estimates. The study found changes in norm perception while some own-behavior outcomes remained unaffected. The design strengthens a theory in which algorithms influence beliefs about social groups through the apparent distribution of others' reactions. Because accessible reporting for this article includes its abstract and selected public materials, the present synthesis does not reconstruct unverified supplemental statistical detail.

Short-Form Video and Algorithmic Audit Experiments

Short-form recommendation feeds give users substantial influence through watch time and skips while often delivering posts from accounts they have never followed. Audit accounts can experimentally standardize the initial viewing history and inspect subsequent recommendations. This makes it possible to study the direction and consistency of algorithmically mediated content delivery without assuming that a human user's reported feed is representative. However, simulated agents cannot measure emotional response, opinion change, or civic behavior.

Ibrahim et al. (2026) Ibrahim (2026) examined TikTok during the 2024 U.S. presidential campaign using 323 successful controlled audit experiments and more than 280,000 recommendations across 27 weeks. Republican-conditioned accounts received about 11.5 percent more co-partisan material than Democratic-conditioned accounts; Democratic-conditioned accounts encountered about 7.5 percent more cross-partisan content. The imbalance persisted after adjustment for observed engagement metrics and was concentrated in particular high-reach sources and issues.

The experiment identifies a measurable exposure asymmetry under controlled account conditioning, with important limits. Unobserved supply characteristics and internal ranking logic are not independently manipulated, and automated accounts do not establish persuasion effects in human audiences. The proper conclusion is about recommendation distributions in the observed audit conditions. The finding motivates research linking human exposure histories, attention, knowledge, and longitudinal attitudes on video-first systems.

Diffusion of Misinformation and Political Information Quality

Misinformation research asks how false or misleading claims are produced, encountered, believed, and redistributed. These outcomes should not be collapsed. A false item may spread widely without persuading most recipients, and accurate news can also travel through the same reshare and recommendation channels. Classifying questionable information requires a documented fact-checking standard, a coverage interval, and clear recognition that some contested political statements are not reducible to simple factual labels.

Vosoughi, Roy, and Aral (2018) Vosoughi (2018) examined the diffusion of verified true and false news stories on Twitter and found that false stories spread farther and faster in their dataset. The observational design characterizes diffusion pathways and correlates; it does not prove a single psychological cause of virality or determine the effect of exposure on election results. Guess, Nagler, and Tucker (2019) Guess (2019) linked a U.S. survey with Facebook sharing behavior and found that sharing from false-news domains was concentrated among a relatively small share of users.

The concentration of exposure also matters. Guess, Nyhan, and Reifler (2020) Guess (2020) reported that exposure to untrustworthy websites in the 2016 U.S. election was unevenly distributed. The public importance of misinformation therefore depends jointly on the prevalence of susceptible or highly engaged audiences, the repeated exposure of particular subgroups, correction reach, and a measurable transition from encounter to belief or action.

Emotion, Moralization, and Engagement Incentives

Political posts are also emotional stimuli. Anger, moral condemnation, pride, fear, and solidarity can guide selective attention and social sharing, and the appearance of other people's responses provides a social reward signal. Recommenders that predict engagement may consequently emphasize material with emotional properties associated with frequent reactions. Such emphasis emerges from the interaction of available content, human response patterns, and objective functions; the research must identify these components rather than automatically assigning intentional political preferences to the optimization system.

Brady et al. (2017) Brady (2017) studied diffusion of moral-emotional language in political social networks and found that emotional moralized expressions were associated with greater transmission. The work supports a plausible pathway from message properties to interaction and wider exposure, while observational transmission estimates should be distinguished from randomized effects on political identity. A system that selects for response intensity can distort the apparent prevalence of extreme sentiment even where most users' private attitudes change little.

The 2026 registered feed experiment goes further by manipulating ranking rules and measuring perceived partisan hostility (Brady et al., 2026 Brady (2026)). The contrast between diffusion studies and feed interventions is productive: one tradition characterizes the social spread of emotional material, and the other estimates the consequences of altering its salience. The two explain different portions of the causal chain.

Affective Polarization and Intergroup Judgment

Affective polarization concerns differential evaluations of political in-groups and out-groups, often measured using feeling thermometers or related hostility and social-distance indicators. It should be distinguished from ideological distance, policy disagreement, negative stereotypes, and behavioral discrimination. A person can hold strong policy preferences without intense antipathy toward supporters of another party; a rise in perceived hostility can occur without an equivalent change in one's own reported attitudes.

The experimental literature now contains important differences. The 2020 Meta feed-ordering, like-minded-source, and reshare manipulations did not yield detectable changes on major attitudinal outcomes over their respective treatment windows (Guess et al., 2023a Guess (6656); Nyhan et al., 2023 Nyhan (2023); Guess et al., 2023b Guess (6656)). By contrast, targeted reranking of animosity content on X produced a detectable short-term shift in out-party feeling (Piccardi et al., 2025 Piccardi (2025)).

The strongest synthesis is mechanistic. Particular hostile messages can affect interpersonal feeling under particular delivery conditions, while a broad change in feed organization may move many topics and interpersonal signals in opposite directions. Measurement sensitivity, baseline hostility, party identification, intervention dose, and earlier exposure history remain plausible moderators. Future studies should test behavioral discrimination and durable cross-group relationships rather than treating feeling scales as the complete social consequence.

Issue Attitudes, Partisanship, and Candidate Evaluations

Issue attitudes evaluate particular policies or public events, whereas partisanship often captures a more enduring social or political affiliation. Candidate evaluations can combine policy beliefs, moral judgment, familiarity, and relational identity. These dependent variables have different baseline stability and different sensitivity to repeated arguments or source signals. A single scale called political polarization cannot adequately represent all of them.

In the X field experiment, the recommendation feed shifted some policy priorities and judgments about contemporary events while leaving detectable partisanship and affective-polarization outcomes unchanged (Gauthier et al., 2026 Gauthier (2026)). This pattern is consistent with greater plasticity in some issue-specific views than in consolidated group affiliations. It does not imply that every issue responds in the same direction, that partisan identity never changes, or that content exposure is the only plausible causal pathway.

Researchers should measure issue-level responses before and after treatment, include preregistered outcomes, report unexpected null findings, and avoid constructing composite scales that conceal opposing item-level trends. Repeated measurement across political events helps separate transient issue salience from durable attitude conversion. The same recommendation system can change what the public discusses without proportionately changing how strongly parties dislike one another.

Political Knowledge, Attention, and Information Recall

Exposure cannot be equated with learning. A viewer might scroll past a post, attend only to its headline, remember a source's identity, or acquire reliable substantive knowledge. Knowledge outcomes therefore need item validation, pre-treatment measurement where appropriate, and careful attention to changing news events. Social media metrics such as minutes of use and shares are behavior measures, not direct indexes of informed citizenship.

The reshare-removal experiment illustrates an epistemic tradeoff: suppressing a channel that carried unreliable material also reduced some political news exposure and measured news knowledge among participants (Guess et al., 2023b Guess (6656)). The chronological-feed intervention similarly changed political and untrustworthy exposure without detectable aggregate improvement in political knowledge over the studied interval (Guess et al., 2023a Guess (6656)). These results counsel measurement of information value, not merely information volume.

The causal chain connecting encounter to knowledge contains attention, comprehension, retention, retrieval, and confidence calibration. Multimethod studies can combine actual impressions with quizzes about covered facts, delayed memory measures, and source attribution. A promising design experimentally varies the salience of verified information while holding other feed characteristics reasonably constant, thereby asking what makes political exposure educational rather than simply attention-grabbing.

Political Mobilization and Civic Action

Political exposure may alter whether users seek registration information, join a discussion, contact representatives, volunteer, donate, or vote. These behaviors have different social costs and institutional dependencies. The existence of attitudinal change does not guarantee mobilization, and mobilization can occur through reminders or peer norms without substantial ideological conversion. Research should accordingly specify behavioral outcomes and avoid substituting engagement counts for political participation.

Bond et al. (2012) Bond (2012) randomized social mobilization messages at massive scale during a U.S. congressional election and examined both online expression and offline voting behavior. The study provides evidence that appropriately structured digital social signals can affect real-world action, with interpersonal transmission important to the mechanism. It differs from the later recommender studies in its intervention, political period, and target outcome.

The evidence also highlights interference: a treated user's message can reach untreated friends, violating simple assumptions that each individual's outcome depends only on their own assignment. Designs for platform-level political behavior should prespecify spillover estimands, use clustered randomization or network-aware analysis where possible, and document how behavioral outcomes are independently verified.

Formal Models of Social Influence and Opinion Dynamics

Formal opinion models help clarify which assumptions connect message exposure to aggregate trajectories. In averaging models, individuals update opinions toward the weighted views of their contacts. In more elaborate influence models, susceptibility, initial opinion, network position, and persistent disagreement can create pluralism or convergence. Friedkin and Johnsen (1990) Friedkin (1990) developed an influential process-oriented approach to social influence and individual opinions. Its significance for digital politics is the distinction between network topology and the rule by which influence is incorporated.

An algorithmically ranked feed effectively changes the encounter weights that a person may assign to social messages. Yet the inference from position to persuasion is model dependent. Some messages are ignored, some are discounted by identity, and some are repeated until they alter accessibility or perceived norms. Therefore, a platform's recommendation graph is not itself an opinion-updating rule. A substantive model must specify perception, source credibility, memory, and response functions.

Researchers should compare competing update processes against repeated observations: linear averaging, asymmetric assimilation, confidence-weighted updating, threshold responses, and persistent selective avoidance can produce similar cross-sectional patterns. Identifiability improves with randomized variation in ranking or peer exposure combined with longitudinal measurement. Mechanistic models are useful when they predict new observables, distinguish rival explanations, and survive out-of-sample testing.

Polarization as a Distributional Outcome

A population may polarize through increasing separation between group means, increasing variance within the whole sample, growth in extreme tails, or stronger hostility across identities. These are mathematically and psychologically distinct processes. A feed algorithm could widen differences among high-activity users while barely affecting the overall average; it might increase perceived disagreement even if private issue beliefs remain constant. One-number summaries can conceal these patterns.

The population distribution of political information is highly unequal: a small subset of users may supply or circulate a disproportionately large portion of partisan material. The observational literature on false-news sharing and Facebook news segregation illustrates why subgroup concentration matters (Guess et al., 2019 Guess (2019); González-Bailón et al., 2023 González-Bailón (2023)). Aggregate effects should be decomposed by baseline participation and exposure without drawing causal conclusions from subgroup selection alone.

Measurement should report means, dispersion, tail changes, and intergroup evaluations using transparent scales. When dynamic network models predict polarization, independent validation should include longitudinal trajectories and the counterfactual under an altered ranking policy. Mathematical plausibility does not replace causal evidence from actual platforms.

Methods: Observational Digital Traces and Surveys

Observational research uses feed impressions, URLs, clicks, follows, reshares, watch-time logs, web-browsing histories, and survey responses. Digital trace data capture behavior with greater temporal resolution than retrospective self-report, but their meaning depends on interface design, missing information, account sharing, recommender access, and the researcher’s visibility into proprietary processes. Surveys capture attitudes and interpretations that behavioral logs cannot directly identify, yet survey responses may be affected by memory, wording, and selective participation.

The strongest descriptive studies triangulate trace records and validated instruments. The 2015 Facebook dataset separated potential cross-cutting content from ranked display and click selection; the 2023 segregation study added population-scale comparisons across exposure stages (Bakshy et al., 2015 Bakshy (2015); González-Bailón et al., 2023 González-Bailón (2023)). Their extraordinary scope improves descriptive coverage while remaining conditional on platform rules, classification strategies, and available variables.

A rigorous method section should state the sampling frame, time window, treatment or exposure definition, how political content was detected, missing-data mechanisms, and whether the analysis captures the whole feed or selected posts. Most importantly, it should specify the target of inference: accounts, impressions, users, voters, or entire political systems.

Methods: Randomized Feed Experiments and Causal Estimands

Randomized experiments offer a stronger basis for claims that an altered ranking policy caused a measured difference. The intention-to-treat contrast compares assigned conditions regardless of imperfect compliance and is often the most robust policy estimand. Treatment-on-treated analyses require justified additional assumptions; attrition can compromise comparability if its causes relate to treatment and the outcome. Registered hypotheses and preplanned outcome families help control the flexibility inherent in studying many political attitudes.

The Meta 2020 program manipulated feed ordering, aligned-source exposure, and reshare availability in distinct experiments. The X 2023 field experiment manipulated access to an algorithmic or chronological feed and observed different effects depending on preexisting feed use. The 2025 independent reranking study manipulated a particular emotionally antagonistic content class, and the 2026 registered custom-feed experiment manipulated ranking objectives and measured norm perceptions (Guess et al., 2023a Guess (6656); Gauthier et al., 2026 Gauthier (2026); Piccardi et al., 2025 Piccardi (2025); Brady et al., 2026 Brady (2026)).

The resulting evidence cannot be summarized as a single average treatment effect of algorithms. Each experiment tests a defined treatment contrast in a distinct social environment and outcome domain. Research synthesis must preserve these differences and assess the strength of manipulation checks, the magnitude of exposure change, survey completion, measurement intervals, and the set of competing explanations left unresolved.

Methods: Audit Accounts, Simulated Users, and Classification

Algorithmic audits deploy controlled accounts to compare how recommendations respond to standardized viewing histories, locations, or social signals. They are powerful for identifying differential delivery under reproducible conditions, especially when platforms provide limited data access. The TikTok audit by Ibrahim et al. (2026) Ibrahim (2026) used political-content conditioning, repeated experimental runs, and human-validated automated classification to describe partisan recommendation skews.

Audit accounts are experimental probes of delivery systems rather than human participants. They lack the full behavioral histories, motives, embodied social contexts, and susceptibility of actual users. Internal platform constraints and bot-detection systems may also alter what the probe encounters. Consequently, results should be restricted to recommendation patterns conditional on the documented account settings, with uncertainty about translation to real-user experiences.

Classification validity is central. Political labels derived from language models, audience ideology, source characteristics, or hand coding can disagree, particularly with irony, quotation, adversarial text, non-English material, and issue-specific framing. Independent coding samples, adjudication, error rates, and sensitivity analyses should accompany any high-volume content audit.

Measurement Validity, Replication, and Outcome Sensitivity

The validity of political-exposure research depends on alignment between a construct and its observed indicator. An impression is not confirmed attention, a politically coded domain is not a direct belief report, a partisan follow is not sufficient evidence of ideological conversion, and a thermometer response is not the same as discriminatory behavior. Measurement error can weaken estimated associations and disguise heterogeneity. Different platforms also expose different subsets of information to researchers.

Replication involves reproducing results under the original protocol and testing transport to new contexts. The 2020 Meta intervention findings, 2025 X content-reranking experiment, and 2026 X feed-switching trial are therefore complementary evidence rather than a simple record of contradictory yes-or-no answers. They manipulate different variables and evaluate different endpoint families. A theoretically coherent synthesis asks which pathway each intervention activates and what intensity or duration is required for a detectable effect.

Null findings require special care. A preregistered nonsignificant estimate can still be compatible with some modest effects; a precisely estimated interval that excludes substantively important effects is more informative. The experiment's endpoint scale, statistical power, attrition, adjustment strategy, and exact window should be examined before generalizing. Observational studies and audit designs answer other questions; their very large datasets cannot by themselves substitute for randomization of a causal exposure.

Age, Culture, Inequality, and Political Context

Digital political exposure depends on access, platform choice, literacy, linguistic context, political interest, and available media. Younger and older audiences can use different services and interpret cues in different ways. Cross-national political structures alter the meaning of party labels, ideological alignment, and norms surrounding public expression. A result from a U.S. presidential campaign is temporally and institutionally specific; scientific generalization requires explicit replication across settings.

Most causal field experiments discussed here are anchored in U.S. social-media populations, and often in users who consent to join a study. This sampling choice limits direct inference to nonusers, countries with different party systems, or audiences exposed to coercive political control and constrained news institutions. Cross-cultural research must test whether its attitude scales, party classifications, and language-based content models preserve their meaning across populations.

Inequality also affects access to alternative information. Time constraints, economic resources, social networks, disability accessibility, and language determine which kinds of political content are practically available. Platform audits can be extended with demographic and linguistic conditioning, while human studies can test differences in knowledge and social outcomes. Such heterogeneity is a substantive object of explanation, not a residual nuisance to be removed.

Ethics, Transparency, Privacy, and Researcher Independence

Political communication experiments raise special questions about informed consent, privacy, measurement of political preference, and potential influence on public behavior. Researchers should minimize the collection of identifiable histories, protect sensitive opinion data, and justify interventions that alter what users see during elections. Consent and independent ethical review are important, but they do not eliminate responsibilities concerning unconsenting network members who may be indirectly exposed through sharing or behavioral spillovers.

Platform collaboration can enable access to data and randomization that external researchers cannot otherwise obtain. It can also create dependencies concerning sample access, permissible analyses, proprietary code, and publication. The 2020 Meta collaboration documented academic control over analysis planning and reporting within agreed arrangements; independent experiments such as Gauthier et al. (2026) Gauthier (2026) and Piccardi et al. (2025) Piccardi (2025) offer alternative institutional models. Neither pathway alone guarantees flawless inference; transparency of protocols, data provenance, and credible replication matters.

Any recommendation for platform governance must be evidence specific. Reducing exposure to hostile speech, increasing access to trusted reporting, changing recommendation diversity, and altering engagement objectives are separable interventions with different effects on knowledge, perception, and participation. A psychologically informed evaluation measures those outcomes rather than relying on a presumed single direction of societal benefit.

Human–Artificial Relations in Political Information Environments

The Artificial Era introduces an additional layer to the political information system: machine-mediated selection and machine-generated or machine-curated discourse can interact with human cognition and social networks. The present empirical literature principally evaluates platforms, algorithms, information flows, and human psychological outcomes. Within Psychology by Bogdanova, the philosophical horizon of Artificial is a broader conceptual context; claims about Artificial Sapiens, artificial agency, or independent artificial political membership remain distinct theoretical questions and require their own operational and empirical treatment.

In observable systems, recommender policies are computational agents in the restricted functional sense of selecting actions according to encoded objectives; their output affects human attention and perceived social norms. A language-model-supported classifier used in feed reranking can alter which human messages are prominent, as in Piccardi et al. (2025) Piccardi (2025). This is evidence for an intervention on human exposure through computation, not evidence that the ranking software possesses any particular subjective state.

Future human–artificial research can distinguish automated exposure decisions, synthetic political messages, conversational persuasion, and durable human–artificial relationships. Those questions should be experimentally separated because a platform recommendation, an interactive dialogue, and an enduring attachment involve different feedback systems and psychological outcomes.

Contrasting the Main Empirical Findings

The contemporary evidence gives several well-supported conclusions at different levels. Large descriptive studies show ideological selection and platform-specific asymmetry in political exposure. Randomized comparisons show that ranking policy changes who sees what and can change engagement. Yet the effect on political attitudes depends strongly on intervention, platform, outcome, and temporal history. The 2020 Meta studies found substantial exposure changes without corresponding detectable shifts in several preregistered attitudinal outcomes (Guess et al., 2023a Guess (6656); Nyhan et al., 2023 Nyhan (2023); Guess et al., 2023b Guess (6656)).

Other studies identify narrower but consequential effects. The X feed-switching experiment detected changes in issue attitudes and following behavior, with asymmetry between adopting and abandoning ranked feeds (Gauthier et al., 2026 Gauthier (2026)). Targeted reranking of antagonistic political messages altered out-party animosity (Piccardi et al., 2025 Piccardi (2025)). Experimental manipulation of ranking objectives altered social-norm accuracy and perceived hostility (Brady et al., 2026 Brady (2026)). Simulated TikTok accounts identified partisan recommendation asymmetries without demonstrating changes in human beliefs (Ibrahim et al., 2026 Ibrahim (2026)).

These findings collectively disfavor a universal rule that algorithmic feeds either always polarize or never influence political opinion. Their apparent differences become intelligible once the exposure mechanism, cognitive mediator, outcome, and time horizon are specified. The resulting synthesis is a research map rather than a pooled causal estimate.

Unresolved Questions and Priorities for Research

First, independent researchers need access to exposure and ranking evidence that is sufficient to reproduce findings without relying on a platform's preferred metrics. Audit tools, controlled feed clients, and privacy-preserving data access have different strengths and should be combined where ethically possible. Second, longitudinal research should follow exposure-induced changes in subscriptions and networks beyond the immediate treatment period. The persistence mechanism identified on X is a strong candidate for experimental mediation tests.

Third, research should separately estimate changes in accurate knowledge, issue beliefs, normative perceptions, affective hostility, and observable political action. Directly targeted interventions may affect one outcome while leaving others stable. Fourth, social spillovers require network-aware designs because one participant's feed can change information encountered by others. Fifth, international and multilingual replication should test whether mechanisms seen in U.S. election periods travel across political systems, content formats, and institutions.

Finally, public claims should be calibrated to the level of identification. A ranking audit establishes delivery differences; a human randomization establishes effects of an intervention; a source ideology score classifies content; a temporal association suggests a hypothesis; and a computational model proposes a mechanism to be tested. The empirical development of the field depends on making these distinctions routine.

Conclusion

Digital political exposure is a multistage, adaptive process through which social networks, content producers, platform policies, and human choices jointly determine political information encounters. Its psychological consequences include attention, knowledge, perceived norms, attitude judgments, group evaluation, and civic action. The evidence shows that exposure distribution can be highly segregated, that algorithms produce measurable amplification and recommendation asymmetries, and that changes in ranking can have consequences of differing magnitude and duration across outcomes.

The strongest causal research supports a conditional picture. Broad changes to Facebook and Instagram feeds during the 2020 election substantially changed content and behavior without detectably shifting several polarization measures during the tested periods. More targeted or differently timed interventions on X and custom social feeds found changes in issue attitudes, out-party animosity, or social-norm perceptions. Independent TikTok audits strengthen evidence about biased exposure distributions while leaving human opinion outcomes to be tested. The scientific object is therefore the causal pathway from specific feed mechanics to specific human outcomes under identifiable conditions.

A durable research program will combine transparent audits, independently randomized interventions, longitudinal network histories, validated psychological measurements, and multi-context replication. This monograph establishes the conceptual and empirical foundations for that program within Political Psychology and Media and Technology Psychology, with connections to computational social science and the study of Human–Artificial Relations.

References

  • Bail, C. A., Argyle, L. P., Brown, T. W., Bumpus, J. P., Chen, H., Hunzaker, M. B. F., Lee, J., Mann, M., Merhout, F., & Volfovsky, A. (2018). Exposure to opposing views on social media can increase political polarization. Proceedings of the National Academy of Sciences, 115(37), 9216–9221. https://doi.org/10.1073/pnas.1804840115
  • Bakshy, E., Messing, S., & Adamic, L. A. (2015). Exposure to ideologically diverse news and opinion on Facebook. Science, 348(6239), 1130–1132. https://doi.org/10.1126/science.aaa1160
  • Bond, R. M., Fariss, C. J., Jones, J. J., Kramer, A. D. I., Marlow, C., Settle, J. E., & Fowler, J. H. (2012). A 61-million-person experiment in social influence and political mobilization. Nature, 489, 295–298. https://doi.org/10.1038/nature11421
  • Brady, W. J., Doyle, M., Elnakouri, A., Finkel, E. J., Jackson, J. C., Kteily, N., Parker, V., Puryear, C., Spelman, T., Teeny, J., & Torres, M. (2026). Redesigning algorithms to intervene on social norm misperceptions during a national election. Nature, 655, 942–956. https://doi.org/10.1038/s41586-026-10536-1
  • Brady, W. J., Wills, J. A., Jost, J. T., Tucker, J. A., & Van Bavel, J. J. (2017). Emotion shapes the diffusion of moralized content in social networks. Proceedings of the National Academy of Sciences, 114(28), 7313–7318. https://doi.org/10.1073/pnas.1618923114
  • Cinelli, M., De Francisci Morales, G., Galeazzi, A., Quattrociocchi, W., & Starnini, M. (2021). The echo chamber effect on social media. Proceedings of the National Academy of Sciences, 118(9), e2023301118. https://doi.org/10.1073/pnas.2023301118
  • Festinger, L. (1957). A theory of cognitive dissonance. Stanford University Press. https://search.worldcat.org/title/A-theory-of-cognitive-dissonance/oclc/798785941
  • Flaxman, S., Goel, S., & Rao, J. M. (2016). Filter bubbles, echo chambers, and online news consumption. Public Opinion Quarterly, 80(S1), 298–320. https://doi.org/10.1093/poq/nfw006
  • Friedkin, N. E., & Johnsen, E. C. (1990). Social influence and opinions. The Journal of Mathematical Sociology, 15(3–4), 193–206. https://doi.org/10.1080/0022250X.1990.9990069
  • Garrett, R. K. (2009). Echo chambers online?: Politically motivated selective exposure among Internet news users. Journal of Computer-Mediated Communication, 14(2), 265–285. https://doi.org/10.1111/j.1083-6101.2009.01440.x
  • Gauthier, G., Hodler, R., Widmer, P., & Zhuravskaya, E. (2026). The political effects of X’s feed algorithm. Nature, 652, 416–423. https://doi.org/10.1038/s41586-026-10098-2
  • González-Bailón, S., Lazer, D., Barberá, P., Zhang, M., Allcott, H., Brown, T., Crespo-Tenorio, A., Freelon, D., Gentzkow, M., Guess, A. M., Iyengar, S., Kim, Y. M., Malhotra, N., Moehler, D., Nyhan, B., Pan, J., Velasco Rivera, C., Settle, J., Thorson, E., ... Tucker, J. A. (2023). Asymmetric ideological segregation in exposure to political news on Facebook. Science, 381(6656), 392–398. https://doi.org/10.1126/science.ade7138
  • Guess, A. M., Malhotra, N., Pan, J., Barberá, P., Allcott, H., Brown, T., Crespo-Tenorio, A., Dimmery, D., Freelon, D., Gentzkow, M., González-Bailón, S., Kennedy, E., Kim, Y. M., Lazer, D., Moehler, D., Nyhan, B., Velasco Rivera, C., Settle, J., Thomas, D. R., ... Tucker, J. A. (2023a). How do social media feed algorithms affect attitudes and behavior in an election campaign? Science, 381(6656), 398–404. https://doi.org/10.1126/science.abp9364
  • Guess, A. M., Malhotra, N., Pan, J., Barberá, P., Allcott, H., Brown, T., Crespo-Tenorio, A., Dimmery, D., Freelon, D., Gentzkow, M., González-Bailón, S., Kennedy, E., Kim, Y. M., Lazer, D., Moehler, D., Nyhan, B., Velasco Rivera, C., Settle, J., Thomas, D. R., ... Tucker, J. A. (2023b). Reshares on social media amplify political news but do not detectably affect beliefs or opinions. Science, 381(6656), 404–408. https://doi.org/10.1126/science.add8424
  • Guess, A. M., Nagler, J., & Tucker, J. A. (2019). Less than you think: Prevalence and predictors of fake news dissemination on Facebook. Science Advances, 5(1), eaau4586. https://doi.org/10.1126/sciadv.aau4586
  • Guess, A. M., Nyhan, B., & Reifler, J. (2020). Exposure to untrustworthy websites in the 2016 US election. Nature Human Behaviour, 4, 472–480. https://doi.org/10.1038/s41562-020-0833-x
  • Huszár, F., Ktena, S. I., O'Brien, C., Belli, L., Schlaikjer, A., & Hardt, M. (2022). Algorithmic amplification of politics on Twitter. Proceedings of the National Academy of Sciences, 119(1), e2025334119. https://doi.org/10.1073/pnas.2025334119
  • Ibrahim, H., Jang, H. D., Aldahoul, N., Kaufman, A. R., Rahwan, T., & Zaki, Y. (2026). Systematic partisan content skews in TikTok during the 2024 US elections. Nature, 654, 1004–1011. https://doi.org/10.1038/s41586-026-10447-1
  • Nyhan, B., Settle, J., Thorson, E., Wojcieszak, M., Barberá, P., Chen, A. Y., Allcott, H., Brown, T., Crespo-Tenorio, A., Dimmery, D., Freelon, D., González-Bailón, S., Guess, A. M., Kennedy, E., Kim, Y. M., Lazer, D., Malhotra, N., Moehler, D., Pan, J., ... Tucker, J. A. (2023). Like-minded sources on Facebook are prevalent but not polarizing. Nature, 620, 137–144. https://doi.org/10.1038/s41586-023-06297-w
  • Piccardi, T., Saveski, M., Jia, C., Hancock, J., Tsai, J. L., & Bernstein, M. S. (2025). Reranking partisan animosity in algorithmic social media feeds alters affective polarization. Science, 390(6776), eadu5584. https://doi.org/10.1126/science.adu5584
  • Taber, C. S., & Lodge, M. (2006). Motivated skepticism in the evaluation of political beliefs. American Journal of Political Science, 50(3), 755–769. https://doi.org/10.1111/j.1540-5907.2006.00214.x
  • Vosoughi, S., Roy, D., & Aral, S. (2018). The spread of true and false news online. Science, 359(6380), 1146–1151. https://doi.org/10.1126/science.aap9559
  • Zaller, J. R. (1992). The nature and origins of mass opinion. Cambridge University Press. https://doi.org/10.1017/CBO9780511818691