Psychology by BogdanovaThe Encyclopedia of Psychology for the Artificial Era

RESEARCH DOSSIER

Digital Social Influence and Algorithmic Consensus

Abstract. Digital social influence arises when social information delivered through ranked and interactive digital systems changes attention, evaluation, public expression, and interpretations of collective opinion. Algorithmic consensus is the appearance of agreement inferred from selected social signals and publicly visible engagement. This critical narrative research synthesis integrates the history of informational and normative influence, norm perception, cascades, online rating experiments, artificial cultural markets, network experiments, and randomized feed interventions. The 2023 Meta experiments, a 2026 randomized X feed study, and a 2026 registered report on norm-correcting rankings demonstrate that exposure effects, political attitudes, and norm accuracy must be evaluated with separate measurements and platform-specific causal claims. Methodological analysis addresses delivered-treatment fidelity, sequential feedback, network interference, experimental ethics, and population generalization. The synthesis develops an evidence-grounded account of technologically mediated social evidence, its distortions, and research priorities for human–artificial social environments. This manuscript is a scholarly narrative synthesis of existing studies, not a protocol-registered systematic review or an original experiment.

Keywords: digital social influence; algorithmic consensus; social proof; informational social influence; normative social influence; descriptive social norms; injunctive social norms; norm misperception; algorithmic ranking; engagement optimization; social media; information cascades; herding; online ratings; political attitudes; perceived group hostility; selective exposure; causal inference; network effects; Artificial Era

Scientific Object and Research Questions

Digital social influence comprises changes in attention, interpretation, evaluation, communication, and behavior arising from social information delivered through computationally organized environments. A feed can expose an individual to messages from friends, strangers, organizations, automated accounts, and curated sources; it can also display counts of approvals, shares, and comments that summarize the conduct of an incompletely observed population. Algorithmic consensus denotes an apparent distribution of agreement inferred from these selectively exposed signals. The central scientific problem is how a person moves from a platform-generated sample of social activity to beliefs about what others think, approve, and do. This monograph examines experimentally manipulated cues, belief updating, social-norm judgments, affective interpretation, and behavioral outcomes in a critical narrative synthesis of independently identifiable research Deutsch & Gerard (1955) Brady et al. (2026).

The research questions follow from this definition. Which properties of a digital cue make it informative, credible, or socially consequential? When does seeing a highly rated item change a private judgment, and when does it merely increase the probability of clicking or contributing a public response? How does an engagement-based ranking alter the visible sample of social behavior relative to a chronological baseline? How accurately do observers infer population attitudes and norms from their feeds? Which changes survive a platform switch, persist into offline decisions, or spread through subsequent exposure? These questions require separating selection, delivered exposure, attention, interpretation, reported belief, observable action, and network spillovers. The empirical literature supports causally identified effects at particular links while leaving the full chain incompletely characterized Muchnik et al. (2013) Guess et al. (2023) Gauthier et al. (2026).

The analytical unit can be an individual encounter, a sequence of encounters, or a feedback loop involving a population and ranking system. A single user-level effect does not establish a collective consensus, and a high percentage of visually similar posts does not establish agreement among all users. The object here is technologically mediated social evidence and the resulting psychology of perceived consensus. Mechanisms governing who becomes connected to whom and how peer-to-peer diffusion propagates independent of interface exposure belong primarily to Social Network Selection, Influence, and Diffusion. The engineering of recommendation models belongs primarily to Algorithmic Curation and Adaptive Media Exposure. Here technical systems matter as manipulable conditions that structure the evidence available to social perception Acemoglu & Ozdaglar (2011) Eslami et al. (2015).

A Historical Continuum of Social Influence Research

The psychological foundations of digital influence preceded networked media. Deutsch and Gerard's 1955 experiments distinguished informational influence, which treats others as evidence about reality, from normative influence, which responds to expectations of acceptance and disapproval. Their distinction remains indispensable when an online user agrees with a popular claim because popularity appears evidential versus when the user publicly endorses it because dissent seems socially costly. These motives can coexist, and a like button can reflect attention, amusement, approval, conformity, or strategic self-presentation. Consequently, experiments must observe more than a single aggregate popularity counter before attributing a shift to social learning or social pressure Deutsch & Gerard (1955).

The next conceptual line concerns the interpretation of norms. Cialdini, Reno, and Kallgren distinguished descriptive norms concerning prevalent behavior from injunctive norms concerning perceived approval, while experiments demonstrated the importance of which norm becomes salient. Later work by Kallgren, Reno, and Cialdini investigated the circumstances in which norm activation affects conduct. Digital systems blend these signals: a high share count appears descriptive, an approving reaction appears injunctive, and the interface's placement can make either salient. Their combination is historically new in delivery and scale, although its psychological components are recognizable in long-established experimental theories Cialdini et al. (1990) Kallgren et al. (2000).

Formal models of social learning supplied another foundation. Bikhchandani, Hirshleifer, and Welch analyzed informational cascades in which observers rationally follow previous choices while disregarding private signals, making collective behavior sensitive to initial observations. Acemoglu and Ozdaglar subsequently reviewed Bayesian and non-Bayesian opinion dynamics, explaining how communication networks can aggregate or distort dispersed information. These theoretical models clarify why visible agreement and accurate collective knowledge are distinct outcomes. A digital interface can accelerate sequential observation, but claims that a particular platform generates a formal cascade require tests of the model's information assumptions, order effects, dependence structure, and decision thresholds Bikhchandani et al. (1992) Acemoglu & Ozdaglar (2011).

Norms, Public Agreement, and Private Belief

A scientifically useful vocabulary distinguishes four observable variables: the distribution of actual individual attitudes, the public expression of those attitudes, the subset of expressions made visible to a given user, and that user's perception of the distribution. Consensus in the first variable would be a population property; consensus in the third is an exposure property; perceived consensus in the fourth is an inference. A platform's count of interactions is drawn from behaviors conditional on seeing content and being willing to act. It is therefore a noisy measure of approval even when the count is authentic. Researchers should independently ask respondents what they personally believe, what they think most others believe, what they think others publicly endorse, and whether they expect sanctions for dissent Cialdini et al. (1990) Brady et al. (2023).

Perceived norms may regulate conduct without changing underlying factual beliefs. A person can privately reject a misleading claim yet repeat it jokingly, signal group belonging, or avoid contradicting a perceived majority. Conversely, a user can update an opinion after observing genuine high-quality evidence distributed by trusted peers. Both cases create public traces that may be recirculated as social evidence. Experimental paradigms should therefore distinguish private survey responses, public platform actions, later recall, and costly decisions. If a study reports more clicks after a popularity cue, the result establishes a change in selection; persuasion and enduring consensus require their own outcome measures Messing & Westwood (2014) Geber & Stahel (2026).

Algorithmic consensus is a working analytical term for a psychological outcome, not a claim that all users converge on one position. It can take the form of an exaggerated belief that a view is common, that a type of expression is normal, that an out-group is hostile, or that a post represents the community's judgment. A precise operationalization compares a user's norm estimate with an independently measured benchmark for a specified population and time. An experiment in which ranking increases apparent agreement can then test whether a corrected benchmark reduces the belief. This distinction provides a direct connection between historical social influence theory and recent randomized tests of social-norm misperceptions Brady et al. (2023) Brady et al. (2026).

Platforms as Selective Social-Information Environments

In a typical feed, the material available for display is already selected by user follows, content production, recommendation eligibility, moderation, and previous interactions. A ranking model then allocates visibility among eligible items, while scroll position and attention determine what is actually processed. Later reactions and follows can change future selections. Causal attribution is difficult because a user with strong preferences may simultaneously select congenial contacts, click congenial content, and receive more of it from the ranking system. Bakshy, Messing, and Adamic's analysis of Facebook distinguished connections, ranked exposure, and clicking; the relative contributions were specific to their population, product, and measurement period Bakshy et al. (2015).

A ranking is a form of choice architecture with distributive consequences. Items at the top receive earlier opportunities to be noticed; their apparent significance can also be inferred from the fact that the platform selected them. Search engines and social feeds differ in purpose and content, but both can alter which perspectives appear available. Epstein and Robertson experimentally manipulated search-result rankings and measured changes in preferences under controlled conditions. This establishes a plausible ordering mechanism while leaving open whether similar effects generalize to continuously personalized feeds, naturalistic repeated exposure, or durable actions outside a study Epstein & Robertson (2015).

The interface creates additional uncertainty when its mechanism is opaque. Eslami and colleagues' user-centered study documented difficulties reasoning about Facebook's invisible feed filtering. Perceived silence from a friend can be mistaken for the friend's inactivity or disinterest when the relevant post was simply withheld by ranking. The same epistemic problem extends to collective interpretation: absence of dissent can be caused by selective sampling, fewer dissenting posts, muted participation, or altered attention. Testing a perceived consensus requires the researcher to distinguish these processes rather than assuming that the displayed feed is a transparent photograph of social reality Eslami et al. (2015) Cinelli et al. (2021).

Causal Pathways and Levels of Explanation

A general causal map begins with the eligible content inventory and continues through assignment of ranking or engagement cues, delivered impressions, noticed signals, social inference, beliefs and emotions, expressed responses, and feedback into the inventory. Each link has its own empirical meaning. An algorithm can increase exposure to inflammatory posts while leaving users' stated attitudes unchanged over the observation period. Another algorithm can change who users follow, creating lasting effects after a ranking setting has been removed. Mechanistic explanation therefore requires time-indexed measurements and comparisons among plausible pathways, not simply a before-and-after estimate of time spent online Guess et al. (2023) Gauthier et al. (2026).

Individual and collective effects also have different identification requirements. One treated user's changed rating may influence later raters, violating the independent-unit assumption of conventional randomized experiments. Experiments that randomize entire groups, streams, or network positions can estimate additional spillovers and recursive amplification. Bond and colleagues' large Facebook experiment combined randomized social messages with evidence of effects transmitted through close friends, illustrating the distinction between a directly delivered stimulus and downstream social influence. For research on apparent consensus, the corresponding design would measure how an induced shift in publicly visible reactions changes later norm estimates in untreated observers Bond et al. (2012) Muchnik et al. (2013).

The monograph uses source-based historical reconstruction, comparative analysis of randomized and observational studies, methodological criticism, and synthesis of findings rather than a protocol-registered systematic review. Findings from laboratory interfaces, naturalistic field trials, platform-internal audits, and observational corpora are evaluated according to what their design actually identifies. Causal conclusions are confined to the manipulated conditions and measured outcomes. This evidentiary discipline is especially important because researchers use terms such as amplification, polarization, learning, herding, and consensus for outcomes that may be measured in quite different ways. The following chapters specify mechanisms before considering their practical significance Flaxman et al. (2016) Milli et al. (2025).

Popularity Signals and Social-Proof Heuristics

Visible endorsement counts turn distributed actions into compact summaries. Their informational value depends on who acted, whether users had independent evidence, whether responses are authentic, and whether the metric was sampled from the population relevant to the observer. A high count can be useful when many independent knowledgeable observers evaluated the same item; it becomes misleading when a few early evaluations generate further visibility and imitation. Because click, like, and share labels are semantically flexible, even perfect counter accuracy would not provide a direct estimate of truth or approval. The reviewer should ask what each reaction measures and whether the population of responders matches the population about which a user is making an inference Bikhchandani et al. (1992) Muchnik et al. (2013).

In the randomized web-rating experiment of Muchnik, Aral, and Taylor, an initial positive assessment changed subsequent rating behavior and generated accumulating positive herding, while manipulated negative ratings prompted partly corrective responses. The reported asymmetry shows that a metric can influence decisions through several routes: altered evaluation, altered propensity to participate, and adjustment to perceived consensus. It also illustrates how source-group relationships condition effects; the meaning of approval from friends can differ from approval attributed to opponents. A popularity count is therefore an intervention on the social evidence participants observe, and its effect depends on credibility and contextual inference Muchnik et al. (2013).

An adequate replication should vary the direction and magnitude of the count, whether the count is attributed to friends or anonymous others, the informational quality of the underlying item, and the order in which a participant encounters evidence. Outcomes should include private preference, willingness to register a public rating, remembered consensus, and behavior after the cue disappears. Researchers must also distinguish an increase in the mean rating from a change in rating variance and participation volume. Such design separates convergence, positive-feedback amplification, and genuine informational learning, which can create superficially similar visible results Salganik et al. (2006) Muchnik et al. (2013).

Ranking, Attention, and Apparent Representativeness

A ranking system decides which voices are prominent and which are peripheral. People often lack a calibrated model of the selection mechanism and can treat the most visible material as typical of the wider population. Amplification arises when the probability of seeing a post differs systematically from the proportion of people who would produce or endorse it under representative sampling. Repeated exposure can then make a minority position seem ordinary. The relevant psychological mechanism is inferential sampling: the observer constructs a belief about a group from a curated stream without knowing the stream's selection probabilities. Studying it requires both the eligible inventory and delivered impressions, preferably linked to independently sampled population attitudes Eslami et al. (2015) González-Bailón et al. (2023).

Engagement-based ranking introduces feedback because strongly reacting audiences can improve future visibility of a post. The optimization target may be clicks, watching, replies, shares, or dwell time, none of which is equivalent to reflective endorsement. In a preregistered audit, Milli and colleagues compared engagement-based ranking against chronological and stated-preference baselines and found differential amplification of divisive and emotionally charged political content. Their analysis demonstrates that changes in a user's visible environment can depend on the objective function, although its measured content contrasts should not automatically be interpreted as changes in population beliefs Milli et al. (2025).

The prominence of a post also influences the practical sampling frame. A user scrolling only the first few screens receives a different distribution from someone who searches deliberately, and someone participating in comments receives yet another. A simple study that randomizes items into a fixed top position can identify a rank effect, but its generalization to adaptive systems depends on whether the experiment preserves the surrounding feed, repeated exposure, and opportunities to react. The concept of apparent representativeness thus joins attention psychology to social judgment: rankings determine what is encountered, while beliefs about others depend on how encounters are interpreted Epstein & Robertson (2015) Flaxman et al. (2016).

Normative Pressure, Expression, and Self-Presentation

Public digital action is frequently observable to an imagined or known audience. Users may endorse a comment because it communicates affiliation, avoid dissent because it risks disapproval, or refrain from responding despite strong private views. Normative social influence provides a framework for these actions, but it cannot be inferred from agreement alone. A design that compares an anonymous private judgment with a public response after exposure to a majority signal can begin to estimate expressive pressure. Follow-up questions about expected approval, feared sanction, and perceived authenticity of others' reactions can help identify mediating interpretations, though self-reports also have validity limits Deutsch & Gerard (1955) Kallgren et al. (2000).

Digital interfaces vary in audience visibility. An anonymous platform may encourage counter-normative expression, whereas a profile linked to colleagues or friends can increase the reputational stakes of a response. Even the same number of visible likes can mean different things according to whether they come from a community of trusted peers, a large unknown audience, or suspected automated accounts. A strong empirical test manipulates perceived audience and norm cues separately; otherwise a change in posting can reflect concern about identity or reputation rather than movement in private belief. Experimental observations of social endorsements in news selection underscore the importance of context for any claim that platforms induce uniform conformity Messing & Westwood (2014) Geber & Stahel (2026).

Consensus measurement must account for expressive asymmetry. People who post most frequently may hold more intense positions than observers who rarely post, and algorithms can amplify frequent contributors. As a result, counting comments as equally representative persons is generally unjustified. A population benchmark based on independently recruited participants or calibrated probability sampling can reveal whether the visible consensus overstates the prevalence of an attitude. The conceptual distinction between descriptive norms and injunctive norms further requires researchers to avoid using the number of visible comments as a proxy for what most people consider acceptable Cialdini et al. (1990) Brady et al. (2026).

Emotional Displays and Meta-Perceptions

Emotion is both a target of digital influence and a cue for social inference. A feed containing visibly angry expressions may alter a reader's immediate affect, but it may also change the reader's belief about how angry a community is. These are different constructs: emotional contagion concerns affective response, whereas norm perception concerns attribution to others. Kramer, Guillory, and Hancock's large-scale Facebook experiment manipulated the emotional composition of feeds and observed changes in users' subsequent expressive language. It established experimentally responsive posting under its conditions; the paper also generated a serious ethics debate and an editorial expression of concern over the research context Kramer et al. (2014).

Brady and colleagues examined a different mechanism by comparing authors' own reports of moral outrage with observers' interpretations. Their field survey and preregistered behavioral experiments found overperception of outrage and downstream distortions of perceived collective hostility. This matters because users can infer a hostile community even without undergoing an equivalent change in their own emotions. If ranking increases the visibility of intensely moralized expressions, overperception and selection may reinforce one another. The resulting hypothesis is causal but must be tested jointly: evidence for an overperception mechanism alone does not establish the magnitude of ranking-induced distortion in each platform Brady et al. (2023) Milli et al. (2025).

Normative reactions to emotional content also depend on audience, topic, and the distinction between experiencing an emotion and performing it communicatively. An angry statement may be genuine, sarcastic, strategically amplified, or received as more angry than its author intended. Experiments that use observers' ratings of isolated text should compare these ratings with authors' momentary reports when possible. In platform settings, the stronger design measures changes in exposure, perceived emotions, perceptions of group hostility, and users' own subsequent behavior, allowing an investigator to determine which links in the proposed mechanism were actually altered Brady et al. (2023) Brady et al. (2026).

Social Endorsement, Identity, and Source Credibility

Individuals do not encounter likes and recommendations in an informational vacuum. Source identity can determine whether popularity signals expertise, shared values, risk, or group competition. Messing and Westwood's experimental research found that social endorsements increased selection of news content and could reduce partisan selective exposure under their experimental conditions. This finding cautions against treating all popularity signals as necessarily polarizing. When endorsements come from socially heterogeneous contacts, they may carry information that encourages crossing familiar source boundaries. The direction of influence therefore depends on the perceived credibility and composition of the endorsing group Messing & Westwood (2014).

The same social metric can also function as an identity cue. A post recognized as popular among an out-group may be rejected even when it is encountered frequently; a post endorsed by trusted in-group members may gain credibility before its underlying argument is evaluated. In the web-rating experiment, heterogeneous reactions to positive and negative cues showed that participants need not passively copy the apparent majority. These observations are compatible with motivated reasoning, credibility assessment, and informational learning, but identifying the relative contributions requires designs that independently manipulate group identity, cue strength, and available evidence Muchnik et al. (2013) Bail et al. (2018).

Algorithmic recommendations add a second potential authority signal: the content was selected by a computational system. Geber and Stahel's 2026 experimental survey compared social, algorithmic, and popularity-based recommendations in a moral-issue setting. Recommendation type did not change perceived social norms in that study, although perceived norms were associated with engagement intentions. Such a null treatment result is theoretically important. Users may discount algorithmic endorsements, interpret them as convenience features, or rely on preexisting views; the research design identifies the tested comparison rather than a universal immunity to recommendation effects. It demonstrates the need to treat source cues, perceived norms, and intentions as separate variables Geber & Stahel (2026).

Artificial Cultural Markets and Path Dependence

Salganik, Dodds, and Watts designed an artificial music market in which participants encountered unfamiliar songs under varying information conditions. The study included 14,341 participants and compared more independent choice with environments where participants could see preceding choices. Greater social influence increased inequality in song success and made outcomes across otherwise comparable social-influence worlds less predictable. The experiment offers a particularly clear demonstration of the feedback between observing others and generating signals for later observers. It does not imply that quality was irrelevant: song appeal still constrained which outcomes were plausible. The central insight is that the same initial pool of options can yield divergent public rankings when popularity becomes part of the choice environment Salganik et al. (2006).

This design has direct implications for digital consensus. At one point in a sequence, apparent popularity reflects an interaction between item characteristics and previous participants' contingent decisions. At a later point, people respond to that history as though it were independent information. Even if no participant deliberately deceives anyone, the aggregate can become path dependent. A study of consensus should therefore distinguish the total number of supporters from the trajectory through which they arrived, including early random differences in visibility and feedback. Observational comparisons among existing feeds cannot reconstruct the unobserved alternative history unless additional design assumptions hold Salganik et al. (2006) Bikhchandani et al. (1992).

An informative extension would repeat the same initial content pool across independently randomized digital groups while keeping interface features constant, then examine the variance of final rankings, agreement estimates, and private evaluations. Independent repetition is necessary to estimate unpredictability across possible histories, rather than confusing uncertainty in a single sample with inherent instability in a dynamic system. By measuring both ratings and confidence, the investigator could also assess whether increasing alignment creates increased subjective certainty. The cultural-market paradigm is therefore a bridge between psychological social proof and formal theories of cascades and feedback Salganik et al. (2006) Acemoglu & Ozdaglar (2011).

Randomized Feedback and Herding in Online Ratings

The randomized experiment reported by Muchnik, Aral, and Taylor is a central piece of direct causal evidence. On a social-news aggregation site, the researchers experimentally manipulated initial ratings and followed subsequent responses. Positive manipulation increased the probability of later positive ratings, and the average final score was also shifted upward. Negative manipulation produced different patterns, including corrective behavior. The experiment separates a created social signal from an independently chosen item quality, allowing causal inference about the effect of that initial intervention on later visible responses under the site's specific design Muchnik et al. (2013).

The study's importance lies as much in the decomposition of mechanisms as in the headline result. Later participants may have re-evaluated the quality of a comment after seeing approval, become more inclined to participate, or adjusted their response to match or contest other users. These processes do not all indicate the same mental state. Researchers who use a final popularity score as evidence of changed private attitudes risk overreading a behavioral metric. The experiment also reveals contextual heterogeneity: responses depended partly on topical content and social group relationships, an observation that complicates any single parameter for digital conformity Muchnik et al. (2013).

A robust research program would add independent judgment measures before exposure, randomly vary whether feedback is disclosed, and then collect choices on a later occasion without the manipulated counter. Such a sequence would help distinguish transient imitation, informational updating, and repeated social reinforcement. Calibration against the unmanipulated distribution would indicate how far apparent agreement diverges from initial opinion. Because exposed ratings become stimuli for subsequent users, variance estimation should reflect the nested or sequential structure rather than treating each later rating as an isolated independent trial. This is the experimental logic of algorithmically mediated public opinion, even when no machine-learning recommender is involved Muchnik et al. (2013) Salganik et al. (2006).

Large-Scale Platform Experiments and Network Spillovers

Bond and colleagues randomized social mobilization messages shown to approximately 61 million Facebook users during the 2010 United States congressional election. Their results showed effects on online expression, information seeking, and measured voting behavior, with social transmission especially associated with close ties. The experiment demonstrates that digitally delivered social cues can have measurable consequences beyond an on-platform reaction. Its generalizability, however, concerns the specific mobilization message, political event, platform population, and intervention window. It should not be interpreted as evidence that all forms of algorithmic ranking influence voting or that a visible majority cue necessarily changes underlying political preferences Bond et al. (2012).

Network interference is central to interpreting this work. If a treated user shares information that reaches an untreated friend, the treatment may have indirect effects and a standard individual-randomization analysis can conceal relevant exposure pathways. Social influence research accordingly requires clear estimands: direct effects among treated people, spillover effects among connected others, and total effects in the network. For algorithmic consensus, a ranking intervention might affect what one user believes and what that user subsequently broadcasts, altering another person's perceived social environment. Such cascades need explicitly identified links and cannot be inferred from cross-sectional association between similar users Bond et al. (2012) Centola (2010).

Centola's online social-network experiment showed that network structure can alter the spread of health-related behavior, complementing theoretical accounts of complex contagion that require reinforcement from multiple contacts. The findings indicate why digital information exposure and actual behavioral adoption deserve separate analysis. Social proof from one prominent source may be sufficient to make a post visible, whereas a more consequential behavior can require validation across several trusted relations. These principles inform design but do not automatically transfer to all norms or all contemporary recommendation systems. The present dossier uses network experiments to explain a mechanism of social reinforcement while leaving the independent topology of diffusion to its neighboring dossier Centola (2010) Centola & Macy (2007).

Emotional Exposure, Communication, and Research Ethics

Kramer, Guillory, and Hancock manipulated the emotional composition of Facebook news feeds in an experiment involving 689,003 users. Reducing exposure to positive or negative expressions altered the affective language of participants' subsequent posts in the expected directions. The large sample and direct intervention provide evidence that changes in displayed social information can affect expression without face-to-face interaction. Yet the dependent measure was textual expression, which is not equivalent to a clinical assessment of mood or long-term psychological well-being. Interpretation must preserve the difference between what users wrote after exposure and the broader claim that their emotional states had been transformed Kramer et al. (2014).

This study also remains a major case in platform-research ethics. PNAS published an editorial expression of concern about the ethical context of the experiment, and the publication record contains subsequent correction material. An academically complete discussion should retain these notices alongside the substantive result. Platform access can make experiments unusually powerful, but a high-capacity delivery mechanism does not itself resolve questions about participant consent, transparency, risk, and research governance. Reproducing an intervention scientifically requires consideration of whether the same question can be answered through consented recruitment, lower-risk simulations, or controlled observational designs Kramer et al. (2014).

For algorithmic consensus, emotion is a possible amplifier of perceived group beliefs. A highly moralized post may be both more likely to attract attention and more likely to be interpreted as evidence of widespread hostility. Research must separate the actual frequency of strong emotions, the probability that emotionally intense posts are displayed, and the perceptions formed by readers. Experiments that merely show changed emotional expression cannot establish accuracy of norm perception, and studies of norm distortion cannot alone establish altered affective states. The empirical bridge is supplied by designs that jointly observe authored emotion, displayed sampling, observer judgments, and later behavior Brady et al. (2023) Brady et al. (2026).

News Endorsements and Selective Exposure

Messing and Westwood experimentally examined how social endorsements influenced people's choices among online news stories. Stronger endorsements increased the chance that articles were selected, and the social recommendation context reduced partisan source-based selectivity in their test setting. The result shows that visible social approval can redirect attention independently of content accuracy or persuasion. It also indicates a potentially integrative function for endorsements: recommendations associated with a heterogeneous set of people may bring unfamiliar sources into the choice set. This is an important counterexample to a universal theory in which social metrics always strengthen existing group boundaries Messing & Westwood (2014).

Bakshy, Messing, and Adamic studied exposure to ideologically diverse material in Facebook's United States user population. Their analysis separated the ideological composition of friendships, feed ranking, and users' own click selections. Self-selection remained substantial, and the study's observational identification does not permit general causal attribution of opinion change to an algorithm. Nevertheless, the work introduced a valuable exposure-accounting principle: the available inventory, the algorithmically delivered material, and the consumed material must be measured separately. Any claim about consensus distortion should identify precisely which stage supplies the evidence and how that evidence might differ from a representative view of others Bakshy et al. (2015).

Evidence about selection can coexist with divergent consequences for knowledge, polarization, and interpersonal perception. Increased exposure to cross-cutting material need not produce persuasion; the observer can reject it or interpret it through existing commitments. Bail and colleagues' randomized social-media intervention found that exposure to opposing political accounts could increase polarization among some participants. These findings invite explicit measurement of defensive reactions, identity cues, perceived representativeness, and the duration of treatment. Strong research does not assume that diversity of exposure and convergence of attitudes have the same causal direction Bail et al. (2018) Messing & Westwood (2014).

Echo Chambers, Visibility, and the Limits of Descriptive Evidence

The expression echo chamber can refer to homophilous connections, repeated exposure to agreeable messages, segregation of information, or a psychological reduction in openness to disagreement. These are separate objects of measurement. Flaxman, Goel, and Rao examined patterns of online news consumption and the relation of search and social-network referral to ideological segregation. Cinelli and colleagues compared content and interaction patterns across several platforms, reporting variation in how communities clustered and circulated contested information. These observational studies document distributions and associations; they cannot by themselves isolate the effects of a particular ranking feature on subsequent private beliefs Flaxman et al. (2016) Cinelli et al. (2021).

A useful decomposition compares potential exposure, actual feed impressions, engagement, recall, and later beliefs. González-Bailón and collaborators analyzed Facebook news exposure during the 2020 United States election, comparing what could enter feeds with what was actually shown and consumed. Ideological segregation increased across these stages, demonstrating that an inventory-level diversity statistic may poorly describe an individual's experienced information environment. This observation supports studying mediated exposure but still requires experimental interventions to identify how a particular component changes judgments. Population-wide platform records provide scale yet often depend on access arrangements and proprietary measurement choices González-Bailón et al. (2023).

The measurement of ideological concentration must also be kept separate from normative inference. Two users with similar political news inventories may infer different group norms based on source trust, comment tone, and personal expectations. Conversely, a user can encounter a diverse mix yet believe most other people support one position if a small number of extreme posts are disproportionately prominent. A comprehensive account of algorithmic consensus therefore adds independent measures of perceived agreement and perceived hostility to descriptive exposure metrics. The relative scarcity of combined measures explains why the literature on echo chambers cannot be treated as a complete theory of psychologically experienced consensus Brady et al. (2023) Nyhan et al. (2023).

Randomized Changes to Meta Feed Algorithms

The 2020 United States Facebook and Instagram Election Study produced unusually informative experiments on algorithmic feed design. Guess and colleagues randomly assigned consenting users to reverse-chronological feeds or the platforms' default ranking systems during the election period. Chronological ordering substantially changed engagement and exposure composition, including changes to political content and other categories. Yet over the three-month study, the authors did not detect corresponding significant changes in the central political attitude and polarization outcomes they measured. This contrast establishes that substantial treatment effects on information exposure do not entail detectable effects on every downstream psychological outcome over a given horizon Guess et al. (2023).

Nyhan and colleagues used platform data and an intervention that reduced exposure to politically like-minded sources. They found that like-minded exposure was common and could be changed through a controlled feed intervention, but the measured political attitudes did not move detectably within the study period. The design provided important bounds for specified outcomes, with interpretive limits tied to sample, timing, adherence, and what was actually changed. It would be invalid to conclude that all ranking systems are harmless or that social influence has ceased to operate. The strongest supported conclusion concerns the particular intervention and attitude measures examined on Facebook in 2020 Nyhan et al. (2023).

These findings also challenge simplified causal stories about a single decisive algorithmic ingredient. A feed intervention can increase exposure to some types of material while decreasing others; users may adjust time spent online, seek content elsewhere, or react selectively to what remains. These responses can offset or reshape effects of the assigned condition. Furthermore, changing exposure composition does not automatically equal correcting false perceptions of social norms. A design targeted specifically at apparent consensus should measure respondents' estimates of attitudes and approved behavior in relevant reference groups. The Meta experiments establish important counterevidence against general claims of automatic persuasion while leaving narrower norm-learning hypotheses scientifically open Guess et al. (2023) Nyhan et al. (2023).

Platform Differences and the 2026 X Feed Experiment

Gauthier, Hodler, Widmer, and Zhuravskaya published a field experiment in 2026 involving United States users of X, randomly assigning algorithmic and chronological feed conditions during a seven-week period in 2023. Switching users from a chronological feed to the platform's algorithmic feed increased engagement and shifted selected issue attitudes in the direction of content emphasized by that algorithm. Switching established algorithmic-feed users to chronology did not yield comparable reversing effects. Neither transition significantly changed the study's affective-polarization or self-reported partisanship measures. These outcomes describe asymmetry across intervention direction and outcome variable, not a single generic measure of influence Gauthier et al. (2026).

The researchers examined content delivery and account-following behavior, observing that the algorithm prioritized particular political content and that exposure led some users to follow accounts that continued influencing their subsequent feed. This offers a mechanism for path dependence: algorithmic onboarding may alter the composition of later self-selected information even if the original ranking is switched off. It also demonstrates why studies that only switch a ranking feature off may understate the effects of earlier exposure. Such mechanisms should be tested in other systems before generalizing to all platforms or time periods Gauthier et al. (2026) Bikhchandani et al. (1992).

The contrast with the Meta findings is scientifically productive. Platform inventories, algorithm objectives, user populations, periods of political discussion, treatment uptake, and outcome batteries differ. The X study found detectable movement in some issue attitudes, while the Meta study did not find shifts in its central political outcomes across its intervention window. These results can coexist without logical contradiction because they are estimates of different interventions in different settings. The adequate synthesis therefore identifies moderators and causal pathways instead of declaring one platform experiment universally decisive for algorithmic consensus Guess et al. (2023) Gauthier et al. (2026).

Algorithmic Distortion and Norm-Correction Interventions in 2026

Brady and colleagues' 2026 registered report directly tested the connection between ranking objectives and perceptions of social norms. They built custom feed-ranking algorithms and randomly assigned 2,000 participants to use them for eight weeks around the 2024 United States presidential election. Engagement-based ranking amplified intergroup, moralized, emotionally expressive, and toxic content relative to reverse chronology, with particularly prominent increases in outrage-related and political material. The study also found changes in prescriptive norm perception accuracy and perceived partisan animosity, although it did not detect significant changes in users' own engagement behavior. Its design makes norm inference a measured psychological outcome rather than an assumption inferred from feed composition Brady et al. (2026).

A second condition in the registered report used a diversified-extremity algorithm that reduced the influence of extreme contributors. The alternative ranking reduced exposure to the measured contentious content categories and improved prescriptive norm accuracy while maintaining comparable reported platform enjoyment. The result indicates that the population visible in a feed can be made more representative without necessarily requiring a large sacrifice in perceived user experience. Its significance is bounded by the custom feed environment, eight-week duration, registered outcomes, and particular political context. It supports a causal intervention on measured norm perceptions; it does not establish a permanent universal solution for every form of algorithmic distortion Brady et al. (2026).

The registered-report structure and public research materials improve scrutiny of the tested predictions, even where the journal's article text is subscription-restricted. Independent experiments remain important because performance can change when accounts, topics, platforms, and norms vary. Geber and Stahel's 2026 survey experiment, for example, found no average effect of recommendation type on perceived social norms in its digital-immortality issue context. These studies support a conditional conclusion: algorithmic architecture can distort or correct particular norm perceptions, while the mere presence of an algorithmic recommendation does not guarantee influence. Manipulation strength and the informational meaning of cues are part of the treatment Geber & Stahel (2026) Brady et al. (2026).

Misinformation, Outrage, and Reflective Preference

False and emotionally striking content supplies a further test of feedback mechanisms. Vosoughi, Roy, and Aral investigated the diffusion of true and false news stories on Twitter and found differences in the spread of the categories they analyzed. Because their design examined observed diffusion rather than randomly assigning a ranking intervention, the study establishes patterns in the historical dataset rather than a platform-independent causal effect of recommendations. Likewise, an apparent consensus built around widely shared misinformation might arise from novelty, user incentives, network concentration, and repeated exposure. Its psychological effect must be measured separately from its circulation Vosoughi et al. (2018).

Pennycook and collaborators tested whether briefly drawing attention to accuracy improved the discernment of subsequent social-media sharing intentions. Their experimental results suggest that sharing behavior can be partly redirected by decision context, which complicates interpretations of shares as endorsements of accuracy. A count of shares may therefore reflect attention, identity, amusement, perceived relevance, or low scrutiny rather than belief that a statement is true. Such findings justify using direct belief assessments and credibility judgments alongside behavioral trace data, especially where collective counts are used by observers as social evidence Pennycook et al. (2020).

The preregistered audit of Milli and collaborators found that an engagement-ranked feed amplified divisive emotional content compared with chronological and stated-preference alternatives, and that what generated engagement did not always match what users said they wanted to see. This separates revealed platform interactions from reflective evaluation. A system trained to predict reactions can elevate content that users dislike in hindsight. The resulting popularity cues can appear as evidence of approval even when the underlying interaction was driven by anger or curiosity. For algorithmic consensus research, this is a serious construct-validity problem: the optimization target, observed engagement, experienced emotion, perceived approval, and considered preference must each be measured on their own terms Milli et al. (2025) Brady et al. (2023).

Measurement, Estimands, and Experimental Design

Scientific claims about algorithmic consensus require explicit estimands. A causal effect of displaying a high like count is not identical to a causal effect of choosing an engagement-optimized ranking; both differ from the effect of increasing the actual proportion of supporters in a social group. Studies should report the treatment assignment, the delivered intervention, the exposure actually received, and the relevant outcome at a specified horizon. Depending on the question, designs include randomized count manipulations, within-person ordering experiments, cluster-randomized feed interventions, repeated independent social worlds, and population surveys measuring an external benchmark. Outcomes should capture perceived descriptive norms, perceived injunctive norms, private attitudes, public expressions, and actions separately Muchnik et al. (2013) Brady et al. (2026).

Attention to treatment fidelity is especially important in adaptive feeds. Researchers should log not merely the assigned ranking condition but the actual impressions, their order, their exposure duration, and when possible the material the participant noticed. Individuals may circumvent an intervention by searching or switching apps; assignment-based estimates remain informative while exposure-conditioned analyses require additional assumptions. Interference can occur when users influence one another, and sequential popularity feedback makes treatment effects history dependent. A preanalysis plan must specify how clustered observations, repeated measures, spillovers, missing surveys, and attrition will be handled. The large experiments by Bond and colleagues and Guess and colleagues illustrate different scales and resulting measurement problems Bond et al. (2012) Guess et al. (2023).

Accuracy of perceived consensus requires a defensible criterion. A random sample of all citizens, all platform users, a person's close contacts, and the readers who encountered a specific post are different populations. A norm question that neglects the reference group cannot identify bias. Researchers can experimentally manipulate the displayed proportion of endorsers while independently measuring the actual prevalence of the attitude in the defined group. The resulting calibration curve relates expressed confidence and estimated prevalence to the benchmark. It can be compared across interface variants and repeated after a correction to determine whether inferred consensus is stable, learnable, or resistant to contradictory evidence Cialdini et al. (1990) Brady et al. (2023) Geber & Stahel (2026).

Validity, Replication, and Generalizability

Evidence strength is a function of design, measurement, independence, and the interpretation being drawn. A controlled experiment with a few hundred recruited respondents can isolate a specific cue yet have limited ecological realism. A platform experiment with millions of accounts has tremendous power for its delivered treatment and operational outcomes but may depend on opaque internal choices and a narrow participating population. Observational digital traces reveal naturalistic behaviors at scale but cannot generally resolve unmeasured selection into network ties, topics, or exposure. A serious synthesis should preserve these trade-offs rather than rank studies by participant count alone Salganik et al. (2006) Bakshy et al. (2015) Bond et al. (2012).

Replicability questions arise at several levels: whether the original analysis can be reconstructed, whether the effect reappears with new participants and materials, whether it generalizes to another platform, and whether the same mechanism operates over longer periods. The contrast between Guess and colleagues' randomized Meta interventions and Gauthier and colleagues' randomized X intervention shows why outcome comparability and intervention direction matter. Different results about polarization and selected issue attitudes should first be mapped to their distinct measurements. A null estimate should be interpreted with its precision and specified equivalence bounds; a statistically significant effect should be interpreted with its size, uncertainty, and actual consequence Guess et al. (2023) Nyhan et al. (2023) Gauthier et al. (2026).

Population generalization remains a major open problem. Platform age distributions, language, local norms, access conditions, political periods, and conventions for likes and shares differ. A finding from one United States election should not be automatically extended to everyday cultural coordination or to another country's communication environment. Cultural context can influence whether public agreement reflects private conviction, politeness, group signaling, or fear of sanctions. Future multi-site studies should use comparable instruments, establish measurement invariance where appropriate, and report heterogeneous effects with appropriate multiplicity control. Where a study only supplies an abstract or publicly available summary, this review retains conclusions within that verified source's scope Geber & Stahel (2026) Brady et al. (2026).

Transparency, Intervention Ethics, and Public Interpretation

Digital experiments combine a scientifically valuable capacity for controlled exposure with strong responsibilities toward participants. An intervention on a feed can alter what people see without their actively recognizing that an experiment is occurring. Research governance therefore needs proportionate assessment of consent, platform collaboration, deception, data protection, foreseeable psychological or social harms, and the risk of influencing public decisions. The ethical controversy surrounding the Facebook emotional-content intervention is part of the discipline's historical record and should inform future protocol review. An institution's access to a large population is a practical resource rather than an independent ethical justification Kramer et al. (2014).

Interventions designed to correct misperceived consensus must also avoid substituting one engineered appearance of social agreement for another. A scientifically defensible approach would disclose the relevant benchmark and its sampling uncertainty, improve representativeness of visible expressions, and evaluate whether perceived norms become more accurate. Ranking toward a desired substantive opinion is a different objective from ranking toward representative exposure, and the two should be evaluated separately. Brady and colleagues' diversified-extremity condition offers evidence that ranking choices can improve the measured accuracy of prescriptive norm perception under its conditions, providing an empirical basis for testing transparent norm-calibration interventions Brady et al. (2026).

Public reports should distinguish content-level, user-level, and collective outcomes. A platform may amplify sensational material without causing widespread attitude change; conversely, a seemingly small change in following behavior may reshape a user's future information environment. Communication of null results should specify the tested outcomes rather than imply universal absence of effects. Communication of positive results should state the intervention, relevant comparison, sample, and time horizon. Reliable scholarship also acknowledges independent research access and the potential influence of platform-controlled data availability on the questions investigators can ask. These norms improve interpretability and allow readers to assess empirical claims independently Guess et al. (2023) Gauthier et al. (2026) Milli et al. (2025).

Human–Artificial Social Relations and Future Research

The Artificial Era extends the analysis of digital social influence to environments where recommendation systems interact with generated text, synthetic participants, and artificial conversational agents. A person can treat a synthetic comment as one vote in a social discussion, a conversational recommendation as an expert cue, or an automated cluster of reactions as evidence of community approval. The primary research task is to measure what observers believe about authorship, source independence, reliability, and group representativeness. Artificially generated social signals can be studied through the same psychological variables as human-originating cues, while the causal implementation of generating systems requires additional auditing of provenance and coordination Eslami et al. (2015) Geber & Stahel (2026).

A productive design factorially manipulates stated authorship, number of visible endorsements, source independence, and ranking explanation. It then measures private beliefs, perceived group norms, confidence, willingness to interact, and memory at several follow-up intervals. The scientific aim is to determine whether people interpret many artificial responses as many independent social witnesses, as one coordinated source, or as a feature of platform presentation. Such a study can distinguish numerical agreement from epistemic independence. Its outcomes should never be inferred merely from the apparent fluency or person-like character of an artificial system; they require behavioral evidence from the interacting population Bikhchandani et al. (1992) Brady et al. (2023).

Further research should compare short-lived reactions with persistent changes in follows and peer interaction; distinguish issue-level opinions from group affect; identify when users reject apparently popular misinformation; and examine how calibrated summaries of representative opinion compete with engagement-weighted cues. Longitudinal studies can track whether early exposure generates self-sustaining social feedback and whether intervention timing changes efficacy. The broader research horizon concerns relations between Homo and Artificial within technologically mediated systems of social knowledge. In empirical psychology, this horizon becomes scientifically tractable through explicit source identities, manipulated exposures, independently measured outcomes, and claims proportionate to observed evidence Gauthier et al. (2026) Brady et al. (2026) Pennycook et al. (2020).

Conclusion

Digital social influence is an empirically established family of effects through which the presentation of social information changes what people notice, select, express, infer, and sometimes believe or do. Algorithmic consensus designates the appearance of majority opinion or normative agreement generated when selective exposure and visible interactions are treated as representative evidence. Classical research on informational influence, descriptive and injunctive norms, and informational cascades explains the structure of the problem. Randomized cultural-market and online-rating studies show that early social signals can produce path dependence and herding, while large network experiments show that digitally delivered cues can influence behavior beyond the immediate interface Deutsch & Gerard (1955) Salganik et al. (2006) Muchnik et al. (2013) Bond et al. (2012).

The contemporary evidence establishes substantial heterogeneity. Feed ranking can alter content visibility, emotional tone, engagement, and perceived hostility; these changes do not always produce detectable movement in the same downstream attitudes. The 2023 Meta election experiments reported important exposure changes alongside null effects on the political outcomes examined. The 2026 X experiment identified shifts in selected issue attitudes and asymmetric responses to activating versus deactivating algorithmic feeds. The 2026 registered report on redesigned algorithms provided direct evidence concerning the accuracy of norm perception, while another recommendation experiment reported no average norm effect in its distinct context. These findings make the research program more precise by identifying the treatment, mechanism, outcome, and boundary conditions Guess et al. (2023) Nyhan et al. (2023) Gauthier et al. (2026) Brady et al. (2026) Geber & Stahel (2026).

A lasting scientific account treats the feed as a selected sample, engagement as an imperfect expression of approval, and perceived consensus as a measurable inference about a defined population. It requires independent benchmarks for norms, reliable measures of actual exposure, deliberate attention to network interference, and replications across platforms and cultural settings. These principles establish the next stage of investigation into human and artificial participation in digital social environments. The essential contribution of the present synthesis is an integrated evidence map linking historical psychological theory to causally tested ranking interventions, subjective norm accuracy, and the dynamics of publicly visible agreement Acemoglu & Ozdaglar (2011) Brady et al. (2023) Milli et al. (2025) Brady et al. (2026).

References

  • Acemoglu, D., & Ozdaglar, A. (2011). Opinion dynamics and learning in social networks. Dynamic Games and Applications, 1(1), 3–49. https://doi.org/10.1007/s13235-010-0004-1
  • 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
  • Bikhchandani, S., Hirshleifer, D., & Welch, I. (1992). A theory of fads, fashion, custom, and cultural change as informational cascades. Journal of Political Economy, 100(5), 992–1026. https://doi.org/10.1086/261849
  • 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., McLoughlin, K. L., Torres, M. P., Luo, K. F., Gendron, M., & Crockett, M. J. (2023). Overperception of moral outrage in online social networks inflates beliefs about intergroup hostility. Nature Human Behaviour, 7, 917–927. https://doi.org/10.1038/s41562-023-01582-0
  • 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
  • Centola, D. (2010). The spread of behavior in an online social network experiment. Science, 329(5996), 1194–1197. https://doi.org/10.1126/science.1185231
  • Centola, D., & Macy, M. (2007). Complex contagions and the weakness of long ties. American Journal of Sociology, 113(3), 702–734. https://doi.org/10.1086/521848
  • Cialdini, R. B., Reno, R. R., & Kallgren, C. A. (1990). A focus theory of normative conduct: Recycling the concept of norms to reduce littering in public places. Journal of Personality and Social Psychology, 58(6), 1015–1026. https://doi.org/10.1037/0022-3514.58.6.1015
  • 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
  • Deutsch, M., & Gerard, H. B. (1955). A study of normative and informational social influences upon individual judgment. Journal of Abnormal and Social Psychology, 51(3), 629–636. https://doi.org/10.1037/h0046408
  • Epstein, R., & Robertson, R. E. (2015). The search engine manipulation effect (SEME) and its possible impact on the outcomes of elections. Proceedings of the National Academy of Sciences, 112(33), E4512–E4521. https://doi.org/10.1073/pnas.1419828112
  • Eslami, M., Rickman, A., Vaccaro, K., Aleyasen, A., Vuong, A., Karahalios, K., Hamilton, K., & Sandvig, C. (2015). “I always assumed that I wasn’t really that close to [her]”: Reasoning about invisible algorithms in news feeds. Proceedings of the 33rd Annual ACM Conference on Human Factors in Computing Systems, 153–162. https://doi.org/10.1145/2702123.2702556
  • 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., 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. (2023). 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
  • Kallgren, C. A., Reno, R. R., & Cialdini, R. B. (2000). A focus theory of normative conduct: When norms do and do not affect behavior. Personality and Social Psychology Bulletin, 26(8), 1002–1012. https://doi.org/10.1177/01461672002610009
  • Kramer, A. D. I., Guillory, J. E., & Hancock, J. T. (2014). Experimental evidence of massive-scale emotional contagion through social networks. Proceedings of the National Academy of Sciences, 111(24), 8788–8790. https://doi.org/10.1073/pnas.1320040111
  • Messing, S., & Westwood, S. J. (2014). Selective exposure in the age of social media: Endorsements trump partisan source affiliation when selecting news online. Communication Research, 41(8), 1042–1063. https://doi.org/10.1177/0093650212466406
  • Milli, S., Carroll, M., Wang, Y., Pandey, S., Zhao, S., & Dragan, A. D. (2025). Engagement, user satisfaction, and the amplification of divisive content on social media. PNAS Nexus, 4(3), pgaf062. https://doi.org/10.1093/pnasnexus/pgaf062
  • Muchnik, L., Aral, S., & Taylor, S. J. (2013). Social influence bias: A randomized experiment. Science, 341(6146), 647–651. https://doi.org/10.1126/science.1240466
  • 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
  • Pennycook, G., McPhetres, J., Zhang, Y., Lu, J. G., & Rand, D. G. (2020). Fighting COVID-19 misinformation on social media: Experimental evidence for a scalable accuracy-nudge intervention. Psychological Science, 31(7), 770–780. https://doi.org/10.1177/0956797620939054
  • Salganik, M. J., Dodds, P. S., & Watts, D. J. (2006). Experimental study of inequality and unpredictability in an artificial cultural market. Science, 311(5762), 854–856. https://doi.org/10.1126/science.1121066
  • 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