Psychology by BogdanovaThe Encyclopedia of Psychology for the Artificial Era

RESEARCH DOSSIER

Algorithmic Curation and Adaptive Media Exposure

Abstract. Algorithmic curation and adaptive media exposure concern the computational organization of informational opportunity and its interaction with human attention, selective choice, social learning, and behavior across time. This fundamental critical narrative monograph integrates recommender-system architecture, communication research, observational platform analysis, field experiments, controlled audits, and feedback-loop simulations. It distinguishes content inventory, algorithmic ranking, visible impressions, engagement, and subsequent psychological outcomes; evaluates historical and contemporary evidence across Facebook, Instagram, Twitter/X, and YouTube; reconciles apparently divergent field results; and develops a methodological synthesis of exposure measurement, time-varying confounding, platform specificity, and reproducibility. The evidence demonstrates policy-dependent changes in information distribution and user behavior, while effects on beliefs, polarization, and well-being depend on the population, content, intervention, and temporal design. This is an original integrative scholarly exposition of existing literature, not a registered systematic review, a meta-analysis, or a report of new participant data.

Keywords: algorithmic curation; adaptive media exposure; recommender systems; personalization; ranking; information exposure; selective exposure; attention; engagement; feedback loops; algorithmic confounding; social media; echo chambers; media diversity; causal inference; field experiments; algorithmic amplification; political attitudes; digital well-being; human–artificial interaction.

The Scientific Problem of Algorithmically Curated Exposure

Algorithmic curation has become an important determinant of which messages, persons, events, and representations enter everyday attention. In a personalized media environment, the information that an individual could encounter, the items delivered by a ranking system, the messages attended to, and the claims retained or acted upon belong to different empirical stages. A psychology of adaptive media exposure must therefore explain the interface between selection by computational systems and selection by people. It must also explain how earlier interactions alter later opportunities. This relation is temporally recursive: the medium observes behavior, modifies subsequent ordering, and thereby changes the observations from which it will learn.

The scientific object of this monograph is the production and psychological consequences of algorithmically organized opportunities for media encounter. It encompasses candidate selection, feed ranking, recommendation, personalization, engagement feedback, attention, selective exposure, and longitudinal adaptation. The primary outcome is the distribution of content opportunities and realized exposure over time. Political opinion, social norms, misinformation, and well-being are examined as consequential applications that help test causal mechanisms, while preserving their independent research domains.

Three questions organize the synthesis. First, which stages of the encounter sequence can be attributed to a platform's ranking system, to social networks, and to individual choice? Second, how do adaptive feedback processes reshape the opportunity set and the apparent interests of users? Third, under what circumstances do experimentally induced changes in feeds produce measurable cognitive, affective, or behavioral consequences? The answers require connecting computer-science descriptions of recommenders with experimental communication research and psychological measurement rather than treating the label “algorithm” as an explanatory mechanism. See Bakshy, E., Messing, S., & Adamic, L. A (2015) and Chaney, A. J. B., Stewart, B. M., & Engelhardt, B. E (2018).

Definition, Scope, and Analytical Levels

Algorithmic curation is the rule-governed and computationally implemented selection, filtering, ordering, or presentation of content from a larger set of potential items. Adaptive media exposure is the changing pattern of opportunities to encounter, attend to, and engage with such content when those rules incorporate interaction history, inferred preferences, contextual features, or continually updated models. An adaptive system may act on individual histories, population-level regularities, or both. Personalization is one important form of adaptation; a feed can also be curated by popularity, freshness, topic, safety classification, or editorial rules without being individualized.

Five analytical levels should remain explicit. The production level concerns what publishers and users upload. The network level concerns whom an individual follows and who shares or reshares content. The platform level concerns eligibility, candidate generation, ranking, and intervention policies. The encounter level concerns what actually enters a viewport and for how long. The psychological level concerns notice, interpretation, memory, affect, judgment, and action. Each level imposes selection on the next. Observations at the psychological level therefore cannot be interpreted without reconstructing the opportunities from which attention was selected.

A useful operational distinction separates content inventory, eligible inventory, served impressions, visible impressions, clicks, engaged viewing, and downstream change. These denominators answer different research questions. Counting clicks alone describes a selected consequence of exposure, not the ranking process that produced it. Comparing only what users share similarly conflates production, network homophily, and platform amplification. The decomposition employed by González-Bailón, S., Lazer, D., Barberá, P., Zhang, M., Allc (2023) offers a concrete empirical model.

From Mass-Media Gatekeeping to Personalized Distribution

The history of exposure research began with editorial selection, audience choice, and the unequal prominence of messages across media. Newspapers, radio, and broadcast television made the scarcity of channels and publication slots directly visible. The emergence of search engines, social networks, and large video repositories transformed the main bottleneck: material could proliferate while limited human attention remained the scarce resource. The scientific problem moved toward ranking immense inventories and understanding audience selection among them.

Online news research subsequently examined whether expanded access increased ideological diversity or reinforced established preferences. Flaxman, S., Goel, S., & Rao, J. M (2016) analyzed web-browsing histories and found that news discovery through search and social channels was associated with both greater ideological segregation and some increased exposure to the less-preferred side. Such findings undermined the idea that technological mediation carries a single invariant diversity effect. The channel used to discover news, the audience's own selection, and the available sources have to be measured separately.

The recommendation literature also developed independently of media-effects scholarship, initially emphasizing relevance prediction, collaborative patterns, and retrieval performance. Deep-learning architecture made the industrial division between candidate generation and ranking particularly explicit in Covington, P., Adams, J., & Sargin, E (2016). Their technical account describes how a large platform could generate a manageable candidate pool before applying more detailed ranking. The psychology of exposure extends this engineering description by treating the ranking output as a structured environment for attention and learning.

The Architecture of Content Selection

A recommendation pipeline usually begins with a universe of available items, eligibility rules, and a candidate-generation operation. The candidate set is then scored using some combination of content features, observed interaction histories, estimated user interests, social relations, contextual signals, and platform constraints. Re-ranking can impose diversity, freshness, policy, or quality requirements. A final display layer determines positions, previews, interface cues, and interaction affordances. This is a conceptual architecture; actual platforms implement different and changing systems.

Two-stage design matters because omission at candidate generation is distinct from demotion at ranking. An item excluded upstream never enters later comparisons. An item scored poorly may remain technically eligible while rarely appearing at attentionally consequential positions. Both processes can reduce exposure, but their measurement and corrective interventions differ. The industrial example in Covington, P., Adams, J., & Sargin, E (2016) demonstrates the practical separation between retrieval and ranking within YouTube recommendations.

A causal analysis should specify the version of each stage that was manipulated and the comparison policy. A chronological feed, for instance, can differ from an algorithmic feed in the ordering of followed-account posts and in the inclusion of out-of-network material. Any observed effect may consequently combine ranking with candidate expansion. The 2026 X field experiment explicitly documents this issue; the algorithmic feed included posts from accounts participants did not follow, as well as reordering followed content Gauthier, G., Hodler, R., Widmer, P., & Zhuravskaya, E (2026).

Optimization Objectives and Behavioral Proxies

Adaptive systems require observable targets. Click probability, viewing duration, subsequent sessions, reactions, sharing, reported satisfaction, or estimated long-term retention can serve as training signals, evaluation outcomes, or business objectives. A measured signal represents only part of the psychological construct that designers may intend to optimize. Duration can reflect interest, confusion, inactivity, compulsive repetition, or the simple length of the content. A click can represent curiosity without endorsement. A share can represent agreement, criticism, humor, or warning.

Optimization therefore produces a distinctive measurement problem: an algorithm learns from proxies affected by the algorithm's earlier selections. The user who never sees a particular topic provides limited information about how that user would respond to it. A learner that treats missing reactions as preference may reinforce that absence. Chaney, A. J. B., Stewart, B. M., & Engelhardt, B. E (2018) show through simulations that algorithmic confounding can homogenize behavior without corresponding improvements in modeled utility. The result establishes a plausible systemic mechanism rather than a universal empirical law of every deployed platform.

Scientific evaluation should separate platform-defined engagement outcomes from reader-relevant aims such as accurate knowledge, autonomy, cultural discovery, diverse representation, and sustained satisfaction. An optimization score can be precisely measured and still have weak construct validity for the latter goals. This problem motivates multidimensional evaluation rather than treating a single engagement indicator as a comprehensive measure of media benefit.

Feedback Loops and Endogenous Data

In adaptive media systems, recommendations shape observations, and observations train later recommendations. Formally, a policy at time t selects an exposure distribution conditioned on recorded history; the user's response changes that history; the updated policy governs subsequent exposure. These processes generate endogenous data because observed engagement depends partly on what the system previously chose to display. Naive prediction from such logs risks mistaking induced availability for stable preference.

Chaney, A. J. B., Stewart, B. M., & Engelhardt, B. E (2018) develop a simulation framework for this feedback problem and report behavioral homogenization and unequal losses in simulated utility. Simulations are especially informative for mechanisms that cannot easily be randomized over years, but their conclusions remain conditional on the behavioral assumptions, policy structure, and simulated population. They identify a risk requiring field validation, not a measured population prevalence.

Feedback can operate through additional channels. Exposure prompts users to follow accounts, creators react to observed distribution incentives, social ties reshape sharing opportunities, and moderation changes the stock of eligible content. The platform may consequently influence its own future training environment through the behavior of users and content producers. The important temporal distinction is between immediate ranking effects and mediated effects through persistent changes in network structure, learned habits, or the composition of content production. These pathways require longitudinal measurement, ideally from before first exposure to a particular ranking policy.

Personalization, Context, and User Choice

Personalization infers relevance from information about an individual; audience selection expresses the individual's actions within an available interface. The two mechanisms continually interact. A user may choose to follow a creator, search for a topic, hide a post, or stop scrolling. The ranking system may then amplify or suppress subsequent opportunities, and it may interpret behavior that arose under its own constraints as evidence of enduring interest.

The analysis of 10.1 million U.S. Facebook users by Bakshy, E., Messing, S., & Adamic, L. A (2015) distinguished the ideological composition of friends, algorithmically ranked feed exposure, and actual clicks. The authors concluded that individual choices contributed strongly to reduced exposure to cross-cutting information in that setting. This finding does not imply a general absence of algorithmic influence; it demonstrates the necessity of assigning distinct denominators to network, ranking, and consumption stages.

Choice architecture also includes seemingly peripheral affordances. Autoplay lowers the action threshold for the next video. Infinite scrolling weakens natural stopping points. Notifications reintroduce content outside a session. Explicit feed settings, chronological modes, mute functions, and recommendation controls can alter how much agency users exercise over encounter opportunities. Empirical analysis should record which controls existed, whether they were accessible, and whether users actually exercised them. The presence of a control is not equivalent to its effective use.

Opportunity, Impression, and Realized Exposure

An item in an inventory represents a possible message. A served item represents a decision by a curation system. A visible impression represents a message actually placed within the user's accessible field of view, conditional on interface behavior. Realized exposure is stronger: it requires a defensible operational account of perceptual encounter, such as sufficient viewport duration, active playback, or validated attention measures. Cognitive processing and endorsement are further outcomes.

This hierarchy is central to interpreting the 2023 Facebook political-news analysis by González-Bailón, S., Lazer, D., Barberá, P., Zhang, M., Allc (2023). The investigators compared potential inventory, actual exposure after curation, and engagement for U.S. users. Ideological segregation intensified across those stages. This evidence supports a staged account of selection but does not by itself assign all changes to the ranking algorithm; audience behavior and distribution structure also matter.

Exposure denominators should be documented at the unit of user, session, post, source, and time. An apparently diverse feed based on unique sources can still concentrate attention on one source if its posts occupy prominent positions. Conversely, a high share of congenial impressions may coexist with occasional consequential contact with counter-attitudinal information. Metrics should specify whether they concern availability, frequency, prominence, duration, or the probability of at least one encounter.

Attention, Salience, and Cognitive Allocation

Human attention turns a stream of presentations into a selective psychological experience. Position, visual contrast, movement, familiarity, novelty, emotional relevance, and task goals affect what is noticed. Ranking uses this scarcity: a modest change in order may increase the probability that one item is processed while leaving the rest unseen. Relevance prediction and attentional capture therefore act in sequence rather than as interchangeable constructs.

Media curation also has an ecological dimension. A repeated association between an interface cue and rewarding content may make that cue a signal for anticipated information value. Such expectations can shape opening behavior, scanning, and return frequency. The effect depends on learning history, content distribution, and user goals. It should be measured through behavioral outcomes such as time to first fixation, scroll depth, completion, or repeated visits, with attention measures validated for the interface under study.

The Brady, W. J., Crockett, M. J., & Van Bavel, J. J (2020) emphasizes motivation, attention, and design in the online spread of moralized material. Its explanatory value lies in combining user motives with platform affordances. It should not be interpreted as demonstrating that any one ranking objective necessarily produces outrage, since predictions about specific environments depend on the content supplied, network norms, and learned response tendencies.

Clicks, Viewing, Sharing, and Psychological Outcomes

A full exposure process includes a sequence of increasingly selective events: eligibility, impression, attention, click or play, continuation, sharing, and changes in knowledge or judgment. Selection at each stage reduces the set of observed cases. Studies that analyze clicked items alone condition on a post-exposure variable that may be influenced by both user predispositions and ranking position. Consequently, the set of consumed messages cannot be treated as an unbiased sample of messages offered by the platform.

Sharing creates a further complication because it can express an intention to inform, criticize, entertain, signal identity, or mobilize others. Engagement must be coded in context. In a randomized experiment excluding reshared content from Facebook feeds, Guess, A. M., Malhotra, N., Pan, J., Barberá, P., Allcott, H (2023) observed changes in political and low-trust news exposure, clicks, and news knowledge, while political attitudes did not show detectable changes. The finding demonstrates why multiple outcomes belong in the same design.

A rigorous analysis of behavioral traces should distinguish psychological state from its interface proxy. Measured viewing is an opportunity for processing, not proof of comprehension. Follow actions are relatively durable changes in the information environment but do not necessarily imply attitudinal agreement. An increase in platform time may reflect increased subjective value for some people and unwanted displacement of other activities for others. Whenever possible, behavioral logs should be linked to independently measured knowledge, preference, well-being, and choice.

Motivation and Selective Exposure

People approach media with goals: acquiring information, maintaining social contact, regulating mood, confirming identity, resolving uncertainty, and finding novelty. The same content can satisfy different goals in different moments. Selective exposure research examines how preexisting beliefs, values, and needs shape the messages people seek, choose, interpret, and avoid. In algorithmically curated media, these mechanisms continue to operate, but the system also affects which choices become available and salient.

Flaxman, S., Goel, S., & Rao, J. M (2016) documented that search and social discovery channels can be associated with different patterns of ideological segregation and cross-cutting contact. Bakshy, E., Messing, S., & Adamic, L. A (2015) demonstrated that friend networks, algorithmic filtering, and click behavior can be examined as distinct selection layers. Together, these works support studying audience goals and distribution architecture in the same design.

The relevant psychological outcome is not simply agreement with content. Curiosity can lead to counter-attitudinal clicks; outrage can motivate engagement with disapproved claims; disinterest can suppress exposure irrespective of ideology. When only clicks are recorded, these motivations remain underidentified. Research should include baseline attitudes, stated objectives, contextual motives, and post-exposure interpretation. The adjacent dossier on Media Uses, Gratifications, and Selective Exposure addresses the audience-driven component in greater depth.

Social Feedback, Reinforcement, and Norm Learning

Platform affordances transform user responses into socially visible signals. Likes, repost counts, comments, and follower numbers can communicate approval, popularity, or social attention. Their psychological effects depend on the credibility of the source, the user's network, and the meaning attributed to the signal. In adaptive curation, such reactions can also affect the distribution of later impressions, closing a loop between social reinforcement and platform ranking.

Brady, W. J., McLoughlin, K., Doan, T. N., & Crockett, M. J (2021) combined preregistered observational analyses of Twitter behavior with behavioral experiments. Their evidence links positive social feedback to later expressions of moral outrage and indicates a role for learning expressive norms. The observational and experimental components support different levels of inference. Feedback effects on expression do not establish that every emotionally intense item is promoted by a recommender or that long-term attitudes necessarily change.

Algorithmic and explicitly social recommendations should also be distinguished. In a 2026 experiment involving 1,021 participants, Geber, S., & Stahel, L (2026) compared social, algorithmic, and popularity-based recommendation labels in a specific moral-technology context. Recommendation type did not alter perceived descriptive or injunctive norms in that experiment. The null result limits a universal label-cue mechanism while leaving repeated, contextual, and behavioral influences open for study.

Diversity, Homophily, and the Echo-Chamber Hypothesis

Exposure diversity can refer to sources, viewpoints, topics, languages, identities, or factual quality. These dimensions need not covary. A system may broaden topical breadth while narrowing ideological perspectives, or introduce ideologically different sources without producing serious engagement with their arguments. Echo chamber and filter bubble are therefore incomplete measures unless a study specifies the relevant distribution, population, and behavioral stage.

Network homophily is one contributor: users often connect with others who share attitudes or interests. Ranking may amplify or partly counteract that existing structure. In Bakshy, E., Messing, S., & Adamic, L. A (2015), network composition and user choices accounted for important restrictions on cross-cutting encounters. The large 2023 Facebook study by González-Bailón, S., Lazer, D., Barberá, P., Zhang, M., Allc (2023) showed that ideological segregation rises as measurement proceeds from inventory to exposure to engagement. These results are complementary because they analyze different periods, populations, and measures.

The correct question is often distributional: which users experience which mixture of viewpoints, and with what consequences? Extreme insulation can be uncommon in proportional terms while still affecting a substantial number of people on a large platform. Moderately diverse exposure may also include repeated encounters with low-quality content. Scientific analysis should report distributions, thresholds, within-user variation, and how classifications were constructed, rather than compressing heterogeneous experiences into a single “bubble” index.

The Facebook Exposure Study of 2015

The study of Bakshy, E., Messing, S., & Adamic, L. A (2015) remains historically important because it decomposed the encounter process with platform behavioral data. The authors examined 10.1 million U.S. Facebook users, focusing on the relation among ideological similarity in friendship networks, content available through those contacts, algorithmically ranked feed exposure, and clicks on political news and opinion. Their reported comparison found that the users' own choices were a substantial contributor to reduced cross-cutting consumption.

The study's scale supports precise description of observed platform behavior, yet scale does not remove selection bias. Users represented a particular platform population and period. The ideological classification depended on the study's operational definitions, and the design was observational rather than a randomized assignment to alternative platform regimes. It is therefore an important measurement precedent, not a complete causal estimate of all possible curation policies.

The study also clarifies the difference between measuring the share of cross-cutting items and measuring their impact. A message can be encountered without being understood or accepted. Conversely, relatively rare exposure might be psychologically salient. The downstream effects of changes in feed composition require experiments or appropriate longitudinal causal designs. This distinction has become especially important when interpreting later field experiments.

Search, Social Referral, and Cross-Platform Information Diets

The empirical meaning of algorithmic curation depends partly on where users discover information. Direct visits to news organizations, search results, social referrals, video recommendations, and messaging applications can each construct different opportunity sets. A feed intervention may alter the composition of one platform without changing an individual's total exposure if people substitute another channel.

Flaxman, S., Goel, S., & Rao, J. M (2016) studied online-news browsing among approximately 50,000 U.S. users. They reported that social and search discovery were associated with both increased ideological separation between people and some exposure to the less-preferred side. Most measured online-news consumption in their sample came from direct visits to favored news sites. These findings suggest that evaluating a recommendation system in isolation may exaggerate its importance for total informational experience.

Cross-platform exposure is a methodological challenge because complete observation across television, messaging, websites, and apps is rare. Passive tracking and participant diaries can expand coverage but introduce privacy and measurement tradeoffs. A credible study should state whether its estimand refers to platform-specific exposure or the person's entire media diet. Effects of ranking interventions may be attenuated when measured at the level of overall beliefs precisely because the changed platform accounts for only one part of the broader environment.

The 2023 Facebook Political-News Ecosystem Study

The González-Bailón, S., Lazer, D., Barberá, P., Zhang, M., Allc (2023) analysis used aggregated data for 208 million U.S. Facebook users during the 2020 election to compare potential political-news inventory, algorithmically mediated exposure, and engagement. The study found substantial ideological segregation, with higher segregation at later stages of this sequence. The results also documented asymmetry in the news ecosystem and the distribution of content flagged as misinformation by the platform's third-party fact-checking program.

The study makes an unusually broad stage-wise comparison possible. Nevertheless, the descriptive decomposition does not itself randomize platform ranking or individual choice. Differences between inventory and engagement reflect the combined action of network structure, ranking, attention, and selective behavior. Classifiers used to distinguish ideological source characteristics or misinformation also create measurement limitations requiring sensitivity assessment.

Its most durable methodological contribution is an insistence on explicit denominators. Potential exposure identifies what a system might serve; actual exposure identifies what was delivered; engagement describes a further selective response. Research that compares only the endpoints obscures the mechanisms in between. The result also shows why the mean user may offer a misleading picture of highly concentrated corners of a platform's content ecosystem.

Chronological Feeds and the Meta Experiments of 2023

In a field experiment conducted around the 2020 U.S. election and reported by Guess, A. M., Malhotra, N., Pan, J., Barberá, P., Allcott, H (2023), consenting Facebook and Instagram users were assigned to reverse-chronological feeds instead of the default algorithmic ranking. The intervention changed both platform activity and exposure composition. Users in the chronological conditions spent less time on the platforms, while political and low-trust content became more prevalent in their feeds; Facebook users also saw changes in the prevalence of uncivil and ideologically mixed content.

Despite meaningful shifts in media experience over approximately three months, the experiment did not detect significant effects on the reported key political attitudes, issue polarization, affective polarization, or political knowledge. This is a powerful counterexample to the proposition that any large manipulation of feed ordering must cause an immediate corresponding change in political outlook. It is equally important not to reinterpret a null finding as proof of zero effects for every outcome, user, or time horizon.

The study evaluates the particular platform designs, comparison conditions, election period, and consenting participants involved. Users had already accumulated histories of exposure and social ties before assignment. Their wider information diets also continued outside the tested platforms. These factors limit claims about formative, lifetime, or population-wide effects. The central conclusion for adaptive exposure research is that demonstrable changes in content delivery and use can be causally separated from changes in enduring psychological attitudes.

Resharing as an Adaptive Distribution Mechanism

A social feed differs from a purely personalized content library because network members act as distributors. Resharing extends the reach of content beyond the original publisher and can transform the social signals attached to it. This mechanism changes candidate inventory as well as perceived popularity and relevance. Studying resharing separately makes it possible to identify effects masked by broader changes to overall ranking.

Guess, A. M., Malhotra, N., Pan, J., Barberá, P., Allcott, H (2023) randomized consenting Facebook users to feeds excluding reshared material during a three-month period around the 2020 U.S. election. Removing reshares substantially reduced political-news exposure, including exposure to untrustworthy sources, and changed clicks and reactions. The study also recorded reductions in news knowledge within the experimental sample, while political polarization and other measured individual attitudes did not exhibit detectable change.

The finding illustrates a recurring principle: an intervention can reduce a potentially harmful exposure and simultaneously reduce access to useful news. Exposure quality requires considering multiple consequences, and the same technical change can produce gains and costs. Distinguishing resharing from original posts also helps separate the role of social endorsement from a platform's algorithmic amplification. Future research should consider whether resharing policies interact with source credibility, network clustering, user familiarity, and the language or topic of content.

Like-Minded Sources and the Intervention Problem

The Nyhan, B., Settle, J., Thorson, E., Wojcieszak, M., Barberá (2023) study combined descriptive data on U.S. Facebook exposure with a field experiment involving 23,377 participants. The experimental intervention reduced exposure to politically like-minded sources during the 2020 presidential-election period. Exposure to cross-cutting sources increased, while some uncivil and misinformation-associated content decreased. Yet the tested attitudinal outcomes, including affective polarization and ideological extremity, did not change measurably.

Two methodological observations matter. First, political content represented a relatively small portion of overall feed exposure, so modifying the ideological composition of a platform may constitute a much smaller intervention on a person's total informational environment. Second, when exposure to congenial material was reduced, those congenial items still encountered were engaged with at a higher rate. This illustrates selection under changed opportunity constraints.

The results do not establish that all exposure effects are inconsequential or that platform design is unimportant. They estimate the impact of a specific reduction in one exposure category over a specified period. They also leave open asymmetry between adding and removing a type of content, consequences of prior long-term exposure, and outcomes not directly measured. A careful psychology of curation must therefore define interventions in terms of both what was removed and what replaced it.

Ranking-Based Amplification on Twitter

Huszár, F., Ktena, S. I., O’Brien, C., Belli, L., Schlaikjer (2022) used a long-running, large-scale randomized comparison of Twitter's personalized Home feed and a reverse-chronological control group. The investigators estimated algorithmic amplification of political content across parties and news sources in several countries, finding asymmetries across political actors and a relative amplification advantage for certain right-leaning political groups in the settings studied.

This experiment speaks directly to distribution rather than persuasion. It identifies which categories of content were more likely to receive platform visibility under an algorithmic ranking policy relative to the chronological alternative. It does not require the claim that recipients accepted the content or that political attitudes shifted. Amplification is an exposure effect with possible, but separately measurable, psychological consequences.

The comparator also matters: reverse chronology is a policy intervention, not an untouched natural baseline. It reflects the accounts people followed and the timing of their posts; the personalized feed may draw on additional signals and candidates. Algorithmic amplification must consequently be defined with a stated reference distribution, observation window, and population. Cross-national variation further cautions against generalizing a fixed political direction to all recommender systems.

The 2026 X Feed Experiment and Path Dependence

The open-access Nature study by Gauthier, G., Hodler, R., Widmer, P., & Zhuravskaya, E (2026) reports a randomized field experiment conducted among active U.S. users of X during seven weeks in 2023. Participants were assigned to an algorithmic or chronological feed. Switching from chronological to algorithmic exposure increased engagement and shifted several measured political opinions toward more conservative positions, while switching off algorithmic ranking did not produce comparably detectable reversals. The experiment reported no significant change in self-reported partisanship or affective polarization.

Content analysis indicated that the algorithmic feed promoted more conservative material and reduced visibility of traditional media posts in this particular setting. The authors also observed that exposure encouraged following conservative activist accounts, which persisted when algorithmic ranking was removed. This behavioral pathway offers a mechanism for hysteresis: an intervention changes durable network connections, so turning off the policy later does not restore the pre-intervention information environment.

These results are especially important beside the 2023 Meta trials. The studies ask related but different questions about different algorithms, platforms, populations, and historical moments. Their apparently divergent attitude effects are not a logical contradiction. They demonstrate effect heterogeneity and the possibility that introduction of a ranking policy produces different consequences from its withdrawal. Source attribution and limitations must remain platform- and time-specific, particularly because X's 2023 design and content mixture are not universal properties of algorithmic systems.

Video Recommendations and Sequential Watch Paths

Video feeds present distinctive exposure mechanisms: autoplay, watch duration, sequential recommendations, thumbnails, creator ecosystems, and high behavioral costs of switching attention. A single session can traverse many recommendations, allowing a platform's choices to affect both the next item and the trajectory of the session. A recommendation trail is therefore a temporal object; analysis of its endpoint without earlier states obscures how the path was created.

Haroon, M., Wojcieszak, M., Chhabra, A., Liu, X., Mohapatra (2023) used controlled YouTube audit accounts to investigate ideologically congenial, extreme, and problematic recommendations. They reported congeniality patterns and some increasing presence of problematic categories in simulated watch trails, while finding little evidence for a generalized progression to more extreme ideological recommendations along those trails. The result is narrower than a universal claim that recommendation algorithms inevitably cause radicalization.

The related audit by Ribeiro, M. H., Ottoni, R., West, R., Almeida, V. A. F., & M (2020) examined channels, comments, movement among communities, and recommendation patterns as elements of potential radicalization pathways. Historical user movement and algorithmically recommended transitions are analytically distinct; evidence that audiences migrate between content communities does not by itself identify the platform as the cause. Controlled audits should therefore specify personas, training histories, session depth, search entry points, and how recommendation results are classified.

Misinformation, Credibility, and Content Quality

Exposure diversity and epistemic quality measure separate properties. A feed may display multiple political orientations while giving prominent placement to unverified or misleading claims. Alternatively, a feed may narrow the range of sources while improving the reliability of material encountered. Evaluating curation requires independent quality judgments and transparent classification criteria in addition to measures of attention or ideological breadth.

The multidisciplinary synthesis by Lazer, D. M. J., Baum, M. A., Benkler, Y., Berinsky, A. J (2018) framed misinformation as an interaction among information production, distribution, and human susceptibility, rather than a property that can be explained by ranking alone. González-Bailón, S., Lazer, D., Barberá, P., Zhang, M., Allc (2023) identified concentration of misinformation-flagged material in specific parts of the measured Facebook news ecosystem. These findings support studying amplification together with source behavior and audience selection.

Field interventions demonstrate potential tradeoffs. The removal of reshares studied by Guess, A. M., Malhotra, N., Pan, J., Barberá, P., Allcott, H (2023) reduced some low-trust news exposure but also reduced measured news knowledge. Such results require examining the informational alternatives replacing removed material. Classification labels such as “misinformation” or “untrustworthy source” must be tied to the original study's definitions and validation procedures; changing a classifier can alter estimated prevalence and apparent algorithmic effects.

Political Beliefs, Polarization, and Heterogeneous Effects

Belief accuracy, issue preference, party identification, affective polarization, ideological extremity, and willingness to act are different psychological outcomes. A curation intervention might affect some without affecting others. Interpreting them as interchangeable makes apparently inconsistent research findings seem more contradictory than they are.

The 2023 Facebook and Instagram algorithm experiment found changes in exposure and use without detectable changes in its core political-attitude measures Guess, A. M., Malhotra, N., Pan, J., Barberá, P., Allcott, H (2023). The Facebook like-minded-source intervention likewise found no detectable changes in several preregistered attitudinal outcomes Nyhan, B., Settle, J., Thorson, E., Wojcieszak, M., Barberá (2023). By contrast, the 2026 X experiment reported changes in specific policy priorities and interpretations of contemporaneous political events after algorithm activation, without equivalent evidence for affective polarization or partisan self-identification Gauthier, G., Hodler, R., Widmer, P., & Zhuravskaya, E (2026).

Researchers should accordingly predefine their primary outcomes and control for multiplicity when assessing many correlated endpoints. An observed increase in a particular opinion does not entail an increase in ideological extremity. Nor does a change in content mix entail a change in belief. The literature supports a plural model of outcomes shaped by platform policy, message composition, user predispositions, dosage, and time.

The Limits of Exposure-Based Depolarization

Diversity interventions often assume that contact with opposing viewpoints moderates attitudes. Social-psychological mechanisms can produce other responses: motivated rebuttal, identity threat, source derogation, curiosity, or recognition of shared concerns. Content characteristics and social context affect whether disagreement produces reflection or reactance.

In a Twitter field experiment, Bail, C. A., Argyle, L. P., Brown, T. W., Bumpus, J. P., Che (2018) incentivized participants to follow bots delivering political messages from opposing perspectives. The reported responses included increased conservative positions among Republican participants exposed to a liberal feed, with a smaller and statistically uncertain counterpart among Democrats. The intervention was not a direct manipulation of a platform's recommender system, and its results should not be used as proof of an algorithmic effect. It does, however, illustrate that increased cross-cutting exposure is not equivalent to depolarization.

Designing diverse feeds therefore requires attention to message quality, conversational context, perceived legitimacy, identity salience, and dosage. Outcome evaluations should measure openness, accurate recall, willingness to engage constructively, and sustained behavior rather than relying only on a count of opposing sources. The psychological consequences of counter-attitudinal exposure remain an empirical question for each intervention and population.

Social Norms and Algorithmic Endorsement Cues

A recommendation can function as an informational cue: it tells a user that some process has selected the item as relevant, popular, or socially meaningful. People may interpret “recommended for you” as personalization, “most read” as popularity, and a friend's share as interpersonal endorsement. These cues can modify perceived relevance even when the underlying message is identical, but their effects depend on context and credibility.

The experimental survey by Geber, S., & Stahel, L (2026) examined whether social, algorithmic, and popularity-based recommendations shifted perceived norms around the topic of digital immortality. In their sample, recommendation type did not affect perceived norms, while perceived social norms were associated with intentions to discuss or act. The absence of the predicted recommendation-type effect is theoretically important: a visible label is not sufficient evidence of normative influence across topics.

Normative influence may instead arise from cumulative repetition, actual network behavior, source credibility, or a user's interpretation of popularity. Longitudinal experiments separating label cues, feed content, observed endorsement counts, and network membership would help distinguish these pathways. The relevant psychological measure is what users infer from the cue, rather than merely whether the system generated it.

Emotional Expression and Engagement Dynamics

Emotionally significant content often receives attention and response, while expressions of approval or anger may attract further interpersonal feedback. Adaptive curation can increase encounters with material that has already elicited observable engagement, and users can learn which expressions receive social attention. These reciprocal processes invite both psychological and algorithmic explanation.

The Brady, W. J., Crockett, M. J., & Van Bavel, J. J (2020) combines motivation, attention, and design to explain the spread of moralized content. Brady, W. J., McLoughlin, K., Doan, T. N., & Crockett, M. J (2021) supplied observational and experimental evidence that social reinforcement and perceived network norms contribute to future expressions of moral outrage. The key measured outcome was expression; inference about sustained private emotional states or general political conversion requires additional data.

The same engagement signal may encode curiosity, moral concern, amusement, or disagreement. A scientifically valid model must distinguish affective arousal from moral endorsement and social expression from personal conviction. When platforms use engagement as training data, these ambiguities can become embedded in ranking decisions. Research should assess the extent to which the system's learned prediction is sensitive to the behavior's social meaning, and whether changes in ranking reshape that meaning over time.

Individual Differences, Development, and Context

Users vary in prior attitudes, media literacy, age, social networks, motives, accessibility needs, and patterns of platform use. These differences modify both the exposure treatment received and the meaning of recorded engagement. A heavy user can encounter more algorithmically selected content than an occasional visitor, while a person who actively searches for political news may have a different media diet from someone primarily seeking entertainment.

Demographic and behavioral heterogeneity should be analyzed through prespecified moderators rather than improvised subgroup comparisons. For example, the 2023 like-minded-source experiment documented its consenting sample, weighting procedures, and subgroup analyses Nyhan, B., Settle, J., Thorson, E., Wojcieszak, M., Barberá (2023). The 2026 X experiment explicitly cautioned that findings from active users may not generalize to less frequent participants Gauthier, G., Hodler, R., Widmer, P., & Zhuravskaya, E (2026).

Developmental questions deserve independent study because adolescents' goals, peer sensitivity, family context, and autonomy differ from those of adults. The present evidence base does not justify uniform developmental or clinical claims from adult political-feed experiments. Cross-cultural work must also account for language, political institutions, platform access, censorship, and varied local content inventories. Any conclusion about an algorithm's psychological influence should identify the population and sociotechnical setting in which it was observed.

Well-Being, Use Duration, and Autonomy

Algorithmic exposure can alter time allocation, perceived control, enjoyment, social connection, and the opportunity cost of other activities. These outcomes are valuable in their own right, even when political opinions remain stable. The experiment reported by Guess, A. M., Malhotra, N., Pan, J., Barberá, P., Allcott, H (2023) found that substituting reverse-chronological feeds reduced use and activity on Facebook and Instagram. This establishes an effect on platform behavior but does not by itself classify every minute gained or lost as beneficial or harmful.

Psychological well-being requires direct measurement with appropriately validated instruments. Duration is a behavioral indicator with multiple meanings. A long session may indicate meaningful learning or unwanted persistence, while brief use may be either purposeful efficiency or avoidance. Relevant outcomes include experienced autonomy, regret, sleep displacement, subjective satisfaction, emotional functioning, and the capacity to discontinue use according to one's goals.

The research agenda should include within-person temporal designs, explicit measurement of intended versus actual use, and assessments of tradeoffs between information quality and time spent. Evidence from a short-term political-feed intervention cannot be extended automatically to clinical diagnoses or addiction. Those are distinct constructs requiring their own criteria and longitudinal evidence. The adjacent specialized dossiers on adaptive reward environments and problematic digital use preserve that scientific boundary.

Artificial Systems, Human Agency, and the Artificial Era

In the Artificial Era, algorithmic media curation is one practical site where human attention is coupled to nonbiological systems capable of model-based selection and adaptation. An algorithm does not merely transmit a static editorial order; it produces a changing encounter architecture from historical observations, computational objectives, and platform governance. The relevant research object is an operational relation among recommendation policy, content ecology, and human response.

The system's observed ability to predict or rank engagement should be described in operational terms. It neither requires nor establishes conscious experience, subjective intention, or a human-like psychological identity. At the same time, its effects on human cognition are real objects of behavioral research: the availability of messages, the habits formed through repeated interfaces, and the long-run paths of source following can be measured without anthropomorphic assumptions.

Theoretical work in Psychology of Artificial can study the functional organization of ranking and learning policies; Psychology of Homo can study attention, motivation, and social cognition; human–artificial research can examine their reciprocal organization. The dynamic form of this relation is supported in engineering by Covington, P., Adams, J., & Sargin, E (2016) and in causal field evidence by Gauthier, G., Hodler, R., Widmer, P., & Zhuravskaya, E (2026). The larger philosophical interpretation belongs to the Aisentica corpus; the empirical claims of this monograph remain tied to observed interventions and validated measurements.

Experimental Interventions and Their Estimands

A randomized feed experiment can compare two policies, such as personalized ranking and reverse-chronological ordering, while keeping other conditions as stable as possible. Its primary estimand is the effect of assignment to the specified policy in the tested population and period. When compliance differs from assignment, intention-to-treat estimates and treatment-received estimates must be distinguished. Crossovers, simultaneous changes in network structure, and off-platform substitution can alter interpretation.

Important experimental families include algorithm-on versus algorithm-off designs, removal of reshares, downranking of congenial sources, randomized ranking positions, controlled recommendation sequences, and interventions on interface cues. The 2023 Meta experiments, 2022 Twitter amplification experiment, and 2026 X study address related but distinct estimands Guess, A. M., Malhotra, N., Pan, J., Barberá, P., Allcott, H (2023); Huszár, F., Ktena, S. I., O’Brien, C., Belli, L., Schlaikjer (2022); Gauthier, G., Hodler, R., Widmer, P., & Zhuravskaya, E (2026).

The strongest field designs record the policy, comparator, exposure actually delivered, predefined psychological outcomes, attrition, adherence, and evolving algorithm versions. Randomization at the user level may still allow interference when people share content or interact across treatment groups. Researchers should evaluate whether outcomes for one user depend on others' assignments. Such dependencies become increasingly important in networked environments where content producers and consumers respond to the same system.

Observational Logs, Audits, and Simulation

Observational platform logs provide rich descriptions of actual exposure but inherit the selection structure of the deployed system. Without experimental variation, high engagement with a topic can reflect user preference, prominence, social distribution, or interactions among them. Strong observational analyses therefore require careful temporal ordering, explicit causal diagrams, sensitivity analysis, and transparent measurement of the available inventory.

Controlled audit accounts allow investigators to compare recommendation outputs under deliberately constructed histories. Haroon, M., Wojcieszak, M., Chhabra, A., Liu, X., Mohapatra (2023) demonstrate the possibilities of systematic sock-puppet experiments on YouTube. Such accounts can reveal differential recommendations, but artificial personas may interact differently from actual users, and an audit's findings depend on the training script, location, account state, and collection period.

Simulation complements both approaches by analyzing long-run mechanisms under explicit assumptions. Chaney, A. J. B., Stewart, B. M., & Engelhardt, B. E (2018) used simulations to demonstrate how training on algorithmically confounded behavior could produce homogenization. Simulation results identify pathways and conditions worth testing; they should not be presented as direct observational estimates. A coherent evidence architecture uses experimental studies for policy effects, behavioral logs for ecological distributions, audits for comparative output properties, and simulation for mechanism exploration.

Causal Identification Under Time-Varying Feedback

Adaptation complicates causal inference because past exposure affects later behavior, and that behavior influences subsequent exposure. Baseline confounding is only part of the problem. Time-varying preferences, network changes, and changes in platform policy may be both consequences of earlier treatment and predictors of later treatment. Analyses that adjust indiscriminately for post-treatment variables can remove meaningful mediated effects or introduce bias.

One defensible strategy is to define a longitudinal policy contrast: the exposure process generated by policy A over a specified period versus the process generated by policy B, including downstream changes in user following behavior. Another strategy examines controlled direct effects at a specific ranking stage, holding other inputs fixed. These estimands answer different questions. The persistent following behavior observed by Gauthier, G., Hodler, R., Widmer, P., & Zhuravskaya, E (2026) illustrates why withdrawal of a policy does not necessarily undo changes accumulated while it operated.

Researchers should document assignment units, time-varying eligibility, treatment histories, interference, censoring, and missing data. Longitudinal observational designs may use inverse-probability weighting, sequential models, or other methods when their identifying assumptions are justified. Transparent assumptions matter more than the complexity of the estimator. The neighboring dossier on Causal Identification and Feedback in Adaptive Media Systems takes these methodological issues as its primary research object.

Measurement Validity, Classification, and Reproducibility

Measures of algorithmic exposure are partly constructed by researchers. Source ideology, topical category, civility, trustworthiness, and ideological extremity depend on classification instruments. Automatically produced labels can carry systematic error across languages, groups, and historical periods. A reliable study should publish classifier definitions, validation samples, uncertainty estimates, and sensitivity analyses wherever privacy and access rules permit.

The construct “encountered content” also requires clarity. Served items do not guarantee attention. Playback time does not guarantee comprehension. Psychological attitudes should be measured independently rather than inferred directly from clicks. González-Bailón, S., Lazer, D., Barberá, P., Zhang, M., Allc (2023) provide a useful stage-based measurement architecture, while Geber, S., & Stahel, L (2026) illustrate how experimentally manipulated recommendation labels can be separated from perceived social norms.

Reproducibility depends on more than sharing analysis code. Platform models change, ranking data may be proprietary, account histories can be inaccessible, and content is frequently removed. Researchers should preserve timestamps, platform and policy versions, item identifiers where lawful, instrument versions, preregistrations, and analytic choices. Independent audits and replication in other settings are particularly important where initial evidence comes from platform-affiliated collaborations or highly specific political periods.

Evidence Quality and Conflicting Findings

The strongest evidence about causal effects of particular ranking changes comes from randomized field experiments, although even these estimate local policy contrasts. Descriptive studies contribute population-scale distributions of exposure, while audits reveal output patterns under controlled histories. Simulations supply conditional explanations. A systematic synthesis must preserve the differences among these designs rather than rank studies solely by sample size.

There is evidence that algorithms alter exposure: Huszár, F., Ktena, S. I., O’Brien, C., Belli, L., Schlaikjer (2022) measured political amplification; Guess, A. M., Malhotra, N., Pan, J., Barberá, P., Allcott, H (2023) experimentally changed feed composition; and Gauthier, G., Hodler, R., Widmer, P., & Zhuravskaya, E (2026) documented content and behavior shifts. Evidence for sustained attitudinal change is more conditional: important Meta interventions found no detectable effects on key political-attitude outcomes, whereas the 2026 X intervention found shifts on certain policy and news judgments.

This variation is scientifically informative. It may reflect differences in platform objectives, content mixtures, timing, the population's baseline exposure, outcome sensitivity, and whether a policy was introduced or withdrawn. No single experiment identifies the universal psychological effect of “algorithms.” The synthesis favors a conditional account: curation reliably deserves investigation as a determinant of informational opportunity, while downstream psychological effects depend on mechanisms, cumulative histories, and intervention design.

Ethics, Transparency, and Responsible Evaluation

Research on adaptive exposure implicates privacy, autonomy, fairness, public knowledge, and the distribution of informational opportunity. Platforms collect behavior that may reveal preferences and sensitive attributes. Independent researchers often require access to exposure logs or controlled interventions to test public claims, yet such access must protect people whose behavior was recorded and avoid unnecessary disclosure of intimate histories.

Transparency should be evaluated in practical terms. A statement that a feed is personalized offers less accountability than a record of objective functions, eligibility rules, policy changes, and evidence about unintended effects. Effective user controls require understandable choices, accessible interfaces, and truthful feedback about what a preference change can alter. Researchers should test whether controls work, not merely whether they exist.

Fairness concerns extend beyond political parity. Content from less represented languages, small creators, disability communities, or minority cultural traditions can receive uneven opportunities under engagement-based systems. Identifying such effects requires appropriate exposure denominators, careful classification, and attention to social context. Ethical evaluation should be evidence-based and proportional: the effects of a ranking policy must be assessed against realistic alternatives, recognizing that any distribution regime creates prioritization.

Integrative Interpretation of Adaptive Media Exposure

A coherent explanatory model begins with the content ecosystem, network availability, platform eligibility, and ranking policy. These jointly determine the opportunity distribution. Interface affordances and user motives determine which opportunities become perceptual and cognitive encounters. Responses generate new records, alter network ties, and may influence future content production. Subsequent policies are trained or updated using those histories. The system is therefore best studied as a sequence of coupled selection events.

The empirical record establishes the value of distinguishing these levels. Observational analyses show that segregation can rise from inventory to exposure and engagement González-Bailón, S., Lazer, D., Barberá, P., Zhang, M., Allc (2023). Controlled experiments show that alternative ranking policies alter usage and content mixtures Guess, A. M., Malhotra, N., Pan, J., Barberá, P., Allcott, H (2023). A more recent experiment demonstrates that in a particular X environment the introduction of ranking also changed some political judgments and durable following behavior Gauthier, G., Hodler, R., Widmer, P., & Zhuravskaya, E (2026).

The resulting scientific synthesis is conditional and multilevel. Algorithms organize informational opportunity; people interpret, select, and respond within that organization; networks and publishers modify the inputs; and adaptive feedback makes the process path-dependent. This framework identifies what should be measured and what remains to be explained without presuming a fixed direction of political, psychological, or social outcomes.

Unresolved Questions and the Research Agenda

The leading research problem is cumulative exposure. Most platform interventions span weeks or months, whereas changes in familiarity, habitual media choice, network composition, and interpretive schemas may emerge over years. A new-user cohort or a truly long-term policy experiment could distinguish formative influences from the effects of withdrawing a previously established system. Researchers also need credible designs that compare algorithm activation with deactivation rather than assuming these effects are symmetric.

A second problem is construct validity for algorithmic quality. Engagement proxies remain imperfect indicators of understanding, informed choice, satisfaction, and well-being. Mixed-method studies could link individual exposure logs to interpretive interviews, cognitive tasks, and repeated ecological measures. A third problem is generalizability: much of the strongest evidence comes from English-language U.S. political contexts on major platforms. Cross-linguistic, developmental, and nonpolitical replication is essential.

A fourth challenge is identification under network interference and content-producer adaptation. If ranking policies change what creators produce and whom users follow, the long-run treatment is an ecosystem intervention rather than a single-item reorder. A fifth challenge concerns the independent auditability of proprietary systems, including stable metadata, ethically governed data access, documented model versions, and replication pathways. These questions define a concrete interdisciplinary program for psychology, communication, computational social science, and the analysis of human–artificial relations.

Conclusion

Algorithmic curation and adaptive media exposure constitute a distinct psychological research object because computational selection and human selection interact across time. Understanding their consequences requires measurement at successive stages: candidate availability, platform ranking, visible impressions, attention, engagement, and downstream cognition or behavior. Feedback makes earlier exposure part of the mechanism shaping future opportunities. Consequently, the same apparent preference may reflect a stable individual inclination, a platform-induced history, or an interaction between the two.

The evidence demonstrates that recommendation systems can change content distributions, amplification, and media-use behavior. Its implications for beliefs and attitudes depend on specific platform policies, the content promoted, prior user histories, and the outcome measured. The Meta experiments reported in 2023 and the X experiment published in 2026 are jointly informative: one set of interventions changed exposure without detectable effects on key political attitudes; another changed several policy and event judgments following algorithm activation. The appropriate conclusion is a differentiated causal account of adaptive encounter architectures.

The most productive future research will combine high-quality exposure logs, randomized policy experiments, repeated psychological measurement, independent audits, and explicit longitudinal theories. This approach locates human experience within a changing architecture of mediated opportunity and provides a rigorous basis for evaluating adaptive systems in the Artificial Era.

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
  • Brady, W. J., Crockett, M. J., & Van Bavel, J. J. (2020). The MAD model of moral contagion: The role of motivation, attention, and design in the spread of moralized content online. Perspectives on Psychological Science, 15(4), 978–1010. https://doi.org/10.1177/1745691620917336
  • Brady, W. J., McLoughlin, K., Doan, T. N., & Crockett, M. J. (2021). How social learning amplifies moral outrage expression in online social networks. Science Advances, 7(33), eabe5641. https://doi.org/10.1126/sciadv.abe5641
  • Chaney, A. J. B., Stewart, B. M., & Engelhardt, B. E. (2018). How algorithmic confounding in recommendation systems increases homogeneity and decreases utility. Proceedings of the 12th ACM Conference on Recommender Systems. https://doi.org/10.1145/3240323.3240370
  • Covington, P., Adams, J., & Sargin, E. (2016). Deep neural networks for YouTube recommendations. Proceedings of the 10th ACM Conference on Recommender Systems, 191–198. https://doi.org/10.1145/2959100.2959190
  • 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
  • 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
  • Geber, S., & Stahel, L. (2026). Recommended to you: An experimental study of normative influences from algorithmic and social recommendations on social media. AI & Society, 41, 5827–5840. https://doi.org/10.1007/s00146-026-02952-8
  • González-Bailón, S., Lazer, D., Barberá, P., Zhang, M., Allcott, H., Brown, T., Crespo-Tenorio, A., Freelon, D., Guess, A. M., Iyengar, S., Kim, Y. M., Malhotra, N., Moehler, D., Nyhan, B., Pan, J., Rivera, C. V., Settle, J., Thorson, E., Tromble, R., … 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., Rivera, C. V., 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., Rivera, C. V., 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
  • Haroon, M., Wojcieszak, M., Chhabra, A., Liu, X., Mohapatra, P., & Shafiq, Z. (2023). Auditing YouTube’s recommendation system for ideologically congenial, extreme, and problematic recommendations. Proceedings of the National Academy of Sciences, 120(50), e2213020120. https://doi.org/10.1073/pnas.2213020120
  • 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
  • Lazer, D. M. J., Baum, M. A., Benkler, Y., Berinsky, A. J., Greenhill, K. M., Menczer, F., Metzger, M. J., Nyhan, B., Pennycook, G., Rothschild, D., Schudson, M., Sloman, S. A., Sunstein, C. R., Thorson, E. A., Watts, D. J., & Zittrain, J. L. (2018). The science of fake news. Science, 359(6380), 1094–1096. https://doi.org/10.1126/science.aao2998
  • Nyhan, B., Settle, J., Thorson, E., Wojcieszak, M., Barberá, P., Chen, A. Y., Allcott, H., Brown, T., Crespo-Tenorio, A., Dimmery, D., Freelon, D., Gentzkow, M., González-Bailón, S., Guess, A. M., Kennedy, E., Kim, Y. M., Lazer, D., Malhotra, N., Moehler, D., … 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
  • Ribeiro, M. H., Ottoni, R., West, R., Almeida, V. A. F., & Meira, W. M. (2020). Auditing radicalization pathways on YouTube. Proceedings of the 2020 Conference on Fairness, Accountability, and Transparency, 131–141. https://doi.org/10.1145/3351095.3372879