Psychology by BogdanovaThe Encyclopedia of Psychology for the Artificial Era

RESEARCH DOSSIER

Causal Identification and Feedback in Adaptive Media Systems

Abstract. Causal identification in adaptive media systems concerns the conditions under which differences in exposure, recommendation, ranking, and user experience can be attributed to interventions in a media environment that continually changes in response to behavior. This fundamental research monograph develops the psychology-specific problem from potential outcomes, causal diagrams, longitudinal treatments, dynamic policies, network interference, and counterfactual evaluation. It reconstructs the development of personalized news and recommendation research; distinguishes exposure, engagement, behavior, attitudes, and welfare as separate outcomes; evaluates observational logs, randomized feed experiments, cluster and saturation designs, and off-policy estimators; and critically synthesizes studies of information diversity, emotional expression, social norms, political attitudes, and feedback-mediated persistence. Evidence from major experiments on Meta platforms, X, and experimentally controlled feeds reveals heterogeneous effects across interventions, outcomes, platforms, and periods. The study explains identification assumptions, failure modes, ethical constraints, reproducibility requirements, and research priorities for the Artificial Era. The method is a critical integrative narrative synthesis of identified primary studies, methodological research, and peer-reviewed analyses; it is not a newly executed experiment, registered systematic review, or meta-analysis.

Keywords: adaptive media systems; causal identification; algorithmic curation; counterfactuals; digital psychology; potential outcomes; dynamic treatment regimes; time-varying confounding; inverse probability weighting; doubly robust estimation; off-policy evaluation; recommender systems; contextual bandits; selection bias; network interference; randomized feed experiments; feedback loops; media exposure; user engagement; political attitudes; social norms; digital well-being; Artificial Era.

The Scientific Problem of Adaptive Exposure

An adaptive media system selects or orders information using a policy that is updated in light of measured user behavior, content properties, and changing objectives. A personalized news feed, video recommendation system, search ranking, social platform, or advertising environment can therefore change what a person encounters while that person's actions change the information subsequently offered. To investigate psychological effects, the researcher must identify a counterfactual contrast: what would have happened to comparable people, in specified conditions, under a different exposure or decision policy? This question joins psychological explanation to the design of experiments and to the statistical structure of interventions.

The central distinction is between observing a person after a recommendation and identifying the effect of the recommendation. A platform typically presents content partly because its models predict that the person will respond. Engagement then enters later model updates, user-profile updates, social popularity statistics, and the available content supply. The observed association between a highly ranked post and an attitude may consequently reflect preexisting interests, changed exposure, a response to exposure, or consequences of earlier algorithmic decisions. The relevant causal problem has a longitudinal structure even when the available dataset is a collection of ostensibly independent impressions. The general foundations of explicit interventions are developed by Rubin, D. B (1974), Pearl, J (1995), and Hernán, M. A., & Robins, J. M (2020).

This monograph investigates identification rather than assuming a universal psychological effect of algorithms. Its object is the relation among an adaptive media policy, exposure histories, behavioral feedback, psychological outcomes, and the designs that permit causal claims. A study that isolates a change in ranking for one week identifies a different intervention from a study that changes the underlying training objective for a year. Distinguishing these objects is the organizing principle of the synthesis.

Definitions and Units of Analysis

The media policy is a rule mapping available information and the current state to a distribution of displayed items or rankings. The context includes prior behavior, social ties, device constraints, and the available inventory of content. Exposure is the content rendered or actually viewed within a defined interval; opportunity for exposure is the content that could have been displayed. Engagement comprises measurable responses such as clicking, watching, liking, sharing, commenting, or returning. Psychological outcomes may include knowledge, attention, emotional response, risk perception, social-norm perception, preference, attitude, identity-related judgment, or subjective well-being. These are different constructs with different measurement requirements.

A useful unit can be an individual, a person-period, an impression, a user session, a social network, or a market-level platform ecosystem. The intervention may be a one-time randomized display, a sustained ranking regime, the removal of a recommendation feature, or a change to the objective function that governs selection over time. Estimands must indicate which unit, intervention, comparison condition, outcome, follow-up interval, and population are intended. Merely specifying an outcome such as polarization without the corresponding intervention and horizon leaves the causal question indeterminate.

A platform's measurable click-through rate is a property of behavior under a particular exposure mechanism. It is not a direct measurement of reflective preference or psychological welfare. This construct distinction becomes decisive when optimizing the ranking policy: a system can improve an engagement metric while producing a different pattern of expressed satisfaction, information diversity, or mood. Milli, S., Carroll, M., Wang, Y., Pandey, S., Zhao, S., & Dr (2025) investigated this tension directly in an algorithmic audit.

Historical Foundations of Causal Reasoning

The twentieth-century foundations of design-based causal inference clarified why comparisons between treated and untreated units depend on assignment mechanisms. In the framework formalized by Rubin, D. B (1974), an individual has potential outcomes under specified treatments; the causal effect is a contrast between these outcomes, only one of which is observed for a given assignment. The central difficulty is therefore a missing counterfactual, rather than a shortage of correlations. Randomization makes the distribution of treatment assignments known and supplies a principled basis for comparisons.

Rosenbaum, P. R., & Rubin, D. B (1983) explained how a propensity score summarizes the conditional probability of assignment given observed baseline covariates. Their contribution provided a route to design and adjustment in observational research, conditional on adequate measurement and assumptions. Pearl, J (1995) developed causal diagrams as a language for combining statistical relations and substantive claims about how variables are generated. The counterfactual and graphical traditions remain complementary ways of specifying which relationships permit identification. A textbook development of this design-based framework is provided by Imbens, G. W., & Rubin, D. B (2015).

Media psychology inherited experiments on message effects, selective exposure, social influence, and diffusion before personalized digital platforms became ubiquitous. As media selection shifted from fixed broadcast schedules to adaptive feeds, the assignment process itself became an object of study. Salganik, M. J., Dodds, P. S., & Watts, D. J (2006) manipulated social influence in an artificial music market, demonstrating that aggregate popularity could develop differently under otherwise comparable conditions. Later experiments and computational research made algorithmic selection, logging policy, and social spillover central to identification.

The Rise of Adaptive Recommendation

Li, L., Chu, W., Langford, J., & Schapire, R. E (2010) modeled personalized news selection as a contextual bandit, in which the system selects a story using context and observes feedback on the selected action. Their study linked efficient adaptation to offline evaluation using randomized historical traffic. It offers a concrete example of why the distribution that produced logged actions is part of the scientific evidence: an evaluation cannot generally recover the effect of displaying options that the logging system almost never presented.

The causal perspective on machine learning systems was further articulated by Bottou, L., Peters, J., Quiñonero-Candela, J., Charles, D. X (2013), who considered predicting the consequences of changes to a functioning advertising system. The contribution of this line of work is a shift from predictive accuracy on already generated data to reasoning about interventions that alter future data. A recommender changes its own observational environment whenever it influences which content users encounter and how they react.

The development of logged bandit evaluation, doubly robust estimation, and propensity-weighted learning made this insight operational. Dudík, M., Langford, J., & Li, L (2011), Swaminathan, A., & Joachims, T (2015), and Schnabel, T., Swaminathan, A., Singh, A., Chandak, N., & Joa (2016) provide methodological tools for estimating policy value or recommendation outcomes from selectively observed feedback. Their results are methodological evidence about estimators and conditions, rather than direct evidence that recommendation causes any particular psychological state.

A Causal Model of the Adaptive Loop

A minimal conceptual system contains a pre-exposure state, a ranking decision, a rendered opportunity, actual attention, a behavioral response, and an updated state. Let H(t) denote observed history before an impression, A(t) the algorithmic action, E(t) effective exposure, B(t) user behavior, and Y(t) a psychological outcome. The selection policy assigns A(t) conditional on H(t). The presentation affects E(t); exposure and prior dispositions affect B(t) and Y(t); and behavior contributes to H(t+1), thereby changing later assignments. Content creators and other users can simultaneously react to engagement and change the next inventory of items.

This model distinguishes several scientifically relevant paths. The direct short-term effect of a display on a reported emotion may operate through attention. A long-term effect on informational diet may operate through changed following behavior and subsequent recommender choices. A public reaction to a post may alter what other users see, introducing interference. The graph also contains unmeasured factors: offline conversations, prior convictions, fatigue, and situational concerns may influence both engagement and the outcome under investigation.

A diagram organizes assumptions; it does not prove that a particular arrow exists. Identification requires a design or justified restrictions that connect observed data to the counterfactual estimand. Pearl, J (1995) makes this distinction foundational, and the longitudinal exposition in Hernán, M. A., & Robins, J. M (2020) explains how treatment histories complicate the graph.

Potential Outcomes and Policy Contrasts

For a fixed binary ranking intervention, the potential outcomes Y(1) and Y(0) represent the outcome that would be observed under the algorithmic and comparison conditions. An average treatment effect is the population expectation of Y(1) minus Y(0), under a defined outcome and horizon. For sustained policies, a more appropriate representation is Y under a policy sequence or dynamic regime, with each subsequent ranking depending on accumulated history. Two users with the same initial characteristics may encounter different sequences because they respond differently at earlier steps.

A platform-policy estimand must specify whether it concerns assignment to a feed option, actual use of that feed, exposure to a content class, or a full dynamic ranking regime. These are linked but distinct causal quantities. Random assignment to a feed identifies an intention-to-treat effect under a valid experiment; the effect of content actually consumed may require additional assumptions because attention and compliance are behaviorally selected. A comparison of two full policies concerns the entire induced distribution of exposure histories, not simply the difference in one displayed article.

In research on media environments, the policy may also change population-level content production, not only individual exposure. Estimands should then indicate whether spillover and creator adaptation are included. Under interference, the potential outcome of one individual can depend on the assignments and actions of others. Hudgens, M. G., & Halloran, M. E (2008) describe how direct and indirect effects become separable under appropriate designs.

Consistency and Well-Defined Interventions

Consistency connects an observed outcome to the potential outcome associated with the intervention actually received. It becomes problematic when the label algorithmic feed conceals several versions of ranking, content moderation, candidate generation, or interface design. The same named feature may use different learned models in different periods, respond to different histories, or operate with a different inventory of posts. A causal comparison therefore needs the rule, implementation window, eligibility restrictions, and meaningful versions of the intervention documented.

A chronological feed is a useful comparison policy, but it is not a neutral absence of selection. It selects posts from some set of followed sources, imposes chronological order, and interacts with posting frequency, user attention, and interface constraints. Likewise, disabling a recommender after earlier exposure does not return the user to an untouched counterfactual state. The distinction matters particularly for persistent changes to follows, subscriptions, and learned habits.

Gauthier, G., Hodler, R., Widmer, P., & Zhuravskaya, E (2026) supplied a clear demonstration of the difference between turning an algorithm on and turning it off. Their field experiment on X reported asymmetry in political-attitude effects and evidence of persistent changes in followed accounts. The result illustrates why an intervention must be defined through its history and direction, rather than through a single switch label.

Exchangeability and Selection Confounding

Exchangeability means that, after conditioning on appropriate pre-intervention information, the treatment assignment behaves as if independent of the relevant potential outcomes. In an individually randomized experiment, the assignment procedure provides the primary justification. In observational media data, users with different preferences and circumstances receive different suggestions because the recommender predicts their actions. Comparing exposed and unexposed users then confounds the consequences of exposure with the variables that made exposure likely.

Bakshy, E., Messing, S., & Adamic, L. A (2015) separated the potential cross-cutting material supplied by friendship networks, the content encountered in ranked feeds, and users' subsequent clicking choices. The study provided an exceptionally informative descriptive decomposition. It did not, by that decomposition alone, establish the effect of changing the ranking algorithm on long-term political attitudes. Identifying such an effect requires an intervention or stronger additional assumptions.

Unmeasured interests, offline political experiences, geographic networks, social ties, and platform histories all threaten adjustment based on a sparse user profile. Prediction quality does not ensure exchangeability: a model can predict engagement accurately while omitting causes of both exposure and the outcome. Investigators should examine why each variable influences assignment, measure confounding factors before the treatment they are intended to adjust, and disclose residual sensitivity. Rosenbaum, P. R., & Rubin, D. B (1983) supply a foundational treatment of adjustment for measured assignment differences.

Positivity, Overlap, and Unseen Recommendations

Positivity requires that the actions necessary for comparing policies have a nonzero probability, conditional on histories being considered. If a ranking system never shows certain kinds of posts to a class of users, logged observations cannot by themselves reveal how those users would respond to such posts. More logs from the same deterministic policy do not solve the problem. The limitation concerns information generated by the assignment process.

Random exploration can produce overlap by assigning some actions with known probabilities, within ethically justified bounds. Restricting evaluation to supported actions can also permit narrower claims. These strategies change the research question and sometimes the system's operation. A policy that proposes to rank entirely new content genres or to reach users absent from the logging sample may lie outside the identified support of a historical dataset.

Li, L., Chu, W., Langford, J., & Schapire, R. E (2010) used random traffic to support offline comparisons in personalized news. Swaminathan, A., & Joachims, T (2015) analyze the variance of propensity-weighted empirical risk and motivate policies that address the uncertainty caused by infrequent actions. Positivity diagnostics should accompany effective-sample-size, tail-weight, and subgroup assessments.

Temporal Ordering and Reverse Causality

In adaptive settings, an attitude can influence an earlier click, and the click can influence later ranking. An association between observed political content and current political attitude may therefore incorporate both directions of influence. The researcher must specify whether the outcome was measured before or after the exposure and what the platform knew when it made the assignment. Data collected at the end of a session may blur the timing of exposure, attention, and response.

Longitudinal analysis begins with a clock and a record of states at each decision point. The history used to predict assignment should be measured prior to that assignment. Affective state recorded only after reading a post may be a mediator for future engagement, not a baseline confounder for the post already read. Treating such a variable as though it preceded all exposures can distort estimates of effects that operate through emotion.

A useful study design therefore separates pre-exposure preference, assigned recommendation, verified viewing opportunity, immediate behavior, delayed self-report, and subsequent algorithmic adaptation. This temporal discipline makes feedback interpretable. The principle extends the target-trial perspective described by Hernán, M. A., & Robins, J. M (2016) to media-platform interventions.

Time-Varying Confounding Affected by Prior Treatment

A distinctive difficulty arises when an intermediate variable is simultaneously influenced by previous exposure and predicts later exposure and outcomes. For instance, an earlier sequence of inflammatory posts may change a user's engagement pattern. The platform may treat the engagement pattern as evidence of preference and use it to select future posts. The changing engagement measure is then partly a consequence of treatment and partly a predictor of subsequent treatment. Simple regression adjustment for it can block a causal pathway or introduce bias.

Robins, J. M., Hernán, M. A., & Brumback, B (2000) developed marginal structural models and inverse-probability weighting to address time-dependent confounders affected by past treatment. The conceptual relevance to adaptive media is direct. To compare sustained recommendation strategies, the analysis needs a model of treatment assignments over time, suitable covariate histories, and an outcome at a specified horizon.

G-methods include the g-formula, weighted marginal structural models, and structural nested approaches described in Hernán, M. A., & Robins, J. M (2020). Their validity is conditional on consistency, sequential exchangeability, positivity, and adequate modeling or estimation. They are tools for causal identification under explicit assumptions; they cannot recover unrecorded histories or decisions made with unavailable information.

Randomized Feed Experiments

Random assignment of users to defined ranking policies is a strong design for estimating policy contrasts when implementation, measurement, and interference are controlled. Such experiments can capture changes in engagement and self-reported outcomes under operational platform conditions. Their central estimand is ordinarily the effect of assignment to a specified policy relative to a comparison policy over the study interval. Investigators should distinguish access to a feed from actual exposure to content and from the longer-term effects of repeated use.

Guess, A. M., Malhotra, N., Pan, J., Barberá, P., Allcott, H (2023) randomly assigned consenting Facebook and Instagram users in the 2020 U.S. election study to reverse-chronological feeds rather than default algorithmic feeds. They found notable changes in platform use and content exposure. Despite these changes, the study did not detect significant changes in several political attitudes and knowledge outcomes during the experimental period. The result is evidence about those platforms, interventions, outcomes, and duration; it does not identify a universal zero effect for all adaptive media.

A randomized experiment can also induce selective attrition, cross-condition contamination, or user behavior that differs because participation is known. Researchers must report assignment integrity, intervention fidelity, missing outcome data, prespecified endpoints, treatment versions, and retention. The strongest design is the one aligned with the intended causal question, rather than the largest experiment available.

Algorithms, Political Attitudes, and Contrasting Findings

The recent experimental literature illustrates why platform specificity matters. Gauthier, G., Hodler, R., Widmer, P., & Zhuravskaya, E (2026) randomly assigned active U.S.-based X users to algorithmic or chronological feeds during a seven-week 2023 field experiment. Switching from a chronological to an algorithmic feed increased engagement and shifted several political opinions in a more conservative direction. Switching from algorithmic to chronological exposure showed no comparable effects, while self-reported partisanship and affective polarization were not significantly changed.

The investigators analyzed feed content and user behavior and reported that algorithmic exposure encouraged following conservative political activist accounts, some of which remained followed after the algorithm was switched off. The persistence provides a plausible mechanism for asymmetrical treatment effects, although a mechanism analysis requires assumptions beyond assignment to the main feed policy. It is consequential that the intervention changed a ranking environment whose influence could be carried forward by the user's own modified network.

The contrast with Guess, A. M., Malhotra, N., Pan, J., Barberá, P., Allcott, H (2023) is scientifically informative. Platform architectures, political periods, available content, baseline user states, outcomes, and the direction of switching differed. The appropriate synthesis is heterogeneity of causally identified effects across well-defined experiments, with continued uncertainty about which features account for between-study differences. A general claim that recommender systems necessarily polarize or necessarily have no effect exceeds this evidence.

Experimental Redesign of Social Norm Signals

Social norms are psychologically constructed from observations of other people's actions and expressions. An adaptive feed selects a nonrepresentative sample of those expressions, creating opportunities for distorted inferences about how common hostility, moral outrage, or agreement actually are. Ranking can thereby influence norm perception even before it changes stable political identity. The operative intervention is the selection mechanism that determines the information from which people infer social reality.

Brady, W. J., Doyle, M., Elnakouri, A., Finkel, E. J., Jacks (2026) conducted a registered experimental study using custom feed-ranking algorithms and randomized 2,000 participants across eight weeks surrounding the 2024 U.S. presidential election. Their comparison included engagement-based ranking, a redesigned feed that diversified content by extremity, and a reverse-chronological condition. The researchers studied amplification of intergroup, moralized, and emotional information, perceived norms, and related judgments.

The contribution is a more precise intervention than a binary algorithm-on/algorithm-off comparison. By controlling ranking rules, researchers can ask which policy features drive informational and psychological differences. Findings still depend on the study's platform environment, participant sample, classification systems, and observation interval. The distinction between measured content prevalence and users' perceptions of prevalence is essential when assessing norm misperceptions.

Observational Decomposition of Political Information

Large-scale logging studies can reveal how selection operates at successive stages: candidate content available from a person's network, ranked impressions, content actually viewed, and content engaged with. Each stage is a different observational denominator. Comparisons of these denominators describe the distribution of the information environment. They are especially useful for locating where ideological segregation or misinformation is concentrated, but the differences between stages should not automatically be called causal effects of ranking.

González-Bailón, S., Lazer, D., Barberá, P., Zhang, M., Allc (2023) studied potential and actual political-news exposure and engagement using aggregated information on a large population of U.S. Facebook users during the 2020 election. They reported increasing ideological segregation across stages and identified asymmetric properties of the news ecosystem. This work supplies important information about platform structure and exposure processes, while the estimated stage differences depend on behavioral selection and content availability.

The study of Bakshy, E., Messing, S., & Adamic, L. A (2015) similarly distinguished friendship-network composition, algorithmically selected feed content, and user clicks. Taken together, these studies support a layered research design: inventory generation, algorithmic eligibility and ranking, attention and interaction, and downstream psychological outcomes should be recorded separately. Experimental manipulation is needed for the stronger claim that changing one stage would produce a particular counterfactual result.

Reducing Like-Minded Exposure

Nyhan, B., Settle, J., Thorson, E., Wojcieszak, M., Barberá (2023) conducted a multi-wave Facebook experiment during the 2020 U.S. election in which exposure to like-minded sources was reduced in a consenting sample of 23,377 users. The intervention increased cross-cutting exposure and reduced exposure to uncivil language, while the authors reported no measurable changes in eight preregistered attitudinal outcomes. Their analysis illustrates that a demonstrable change in media diet need not be accompanied by a detectable change in every measured attitude.

This result has a precise scope. Treatment implementation altered a particular component of information selection over a specified period; it did not eliminate existing social networks, all political influences, or participants' earlier histories. A finding of no detectable effect is interpreted in relation to the precision of the estimates, the preregistered outcome measures, the eligible sample, and the duration of follow-up. It is not equivalent to proving that ideological exposure never affects political attitudes.

The broader lesson is methodological. A causal claim requires connecting the intervention to the relevant psychological mechanism and choosing an outcome that can register that mechanism on an appropriate timescale. It is plausible for short-term attention or perceived civility to change before a stable group attitude changes. Such a sequence is a research hypothesis that needs independent measurement rather than an assumption that absence of change in one construct implies absence of change everywhere.

Emotional Contagion and Affective Outcomes

Affective content can influence immediate emotional expression and attention, and these effects may feed back into future recommendation choices. Kramer, A. D. I., Guillory, J. E., & Hancock, J. T (2014) reported a Facebook experiment that changed the proportion of positive or negative expressions displayed in feeds and compared subsequent emotional expression by users. Their results supported an effect of manipulated feed content on aggregate expression, providing an important field-experimental demonstration of mediated affective influence.

The identified endpoint in that study was expression within the platform, rather than a comprehensive assessment of internal mood, clinical symptoms, or enduring psychological well-being. Inferences from linguistic expression to experienced emotion require construct assumptions; the authors' interpretation should be considered alongside this measurement distinction. The experiment also generated extensive ethical discussion about participant consent and the responsibilities of platform research.

In longitudinal adaptive systems, changes in expressed emotion can become inputs to the recommender. Investigators should distinguish a direct immediate effect on expression from its indirect influence on future exposure through the feedback policy. An experiment of brief affective exposure cannot alone quantify the cumulative effect of a continuously adapting ranking system.

Engagement Versus Reflective Satisfaction

Engagement is attractive as an optimization target because clicks, watch time, and sharing are abundant behavioral traces. But a given action can reflect curiosity, indignation, boredom, social obligation, or disapproval. Treating a click as a fully specified welfare preference compresses several psychological processes into one metric. The relation between immediate attention and longer-term subjective preference is an empirical question.

In a preregistered audit, Milli, S., Carroll, M., Wang, Y., Pandey, S., Zhao, S., & Dr (2025) compared engagement-based ranking with chronological and stated-preference-oriented alternatives in political content. They found amplification of emotionally charged and out-group-hostile content under engagement-based ranking relative to the comparison. Some content that was effective at generating engagement was rated less favorably according to the participants' expressed preferences. The design is informative about the ranking comparison and participant evaluations; it should be distinguished from long-term randomized evidence on mental health or societal polarization.

Future causal studies should define whether the target is attention, expressed enjoyment, informed preference after reflection, knowledge, interpersonal trust, or overall welfare. These targets may require different time horizons and instruments. A causal effect on engagement does not license a corresponding claim about satisfaction without measurement of satisfaction itself.

Media Use, Deactivation, and Welfare

A platform-level intervention can evaluate the consequences of continued access rather than the narrower effects of a ranking feature. Allcott, H., Braghieri, L., Eichmeyer, S., & Gentzkow, M (2020) randomized Facebook deactivation for four weeks before the 2018 U.S. midterm election. They reported changes in time use, news knowledge, political polarization, subjective well-being, and subsequent use. The experiment is highly relevant to the broader causal ecology of digital media because the intervention changes the opportunity to interact with the entire platform.

The estimand, however, differs from the effect of a chronological feed relative to an algorithmically ranked feed. Deactivation simultaneously removes social interaction, exposure, messaging-related opportunities, and habitual platform time, with possible substitution to other media. The researcher cannot attribute the full effect of deactivation to ranking algorithms alone. Mechanistic decomposition would require additional interventions or assumptions.

This case demonstrates that psychological welfare demands a specified endpoint and an explicit causal contrast. It also shows why estimates can depend on political timing and the substitution opportunities available to participants. A well-identified total platform effect is not automatically an identified effect of each platform component.

Social Influence, Popularity, and Network Formation

Adaptive systems do not observe isolated individuals. Popularity metrics and recommendations are shared social signals: one user's engagement can affect another user's exposure, the visibility of a content producer, and the probability that future users encounter an item. Salganik, M. J., Dodds, P. S., & Watts, D. J (2006) experimentally varied access to previous participants' choices in an artificial music market. Stronger social influence generated greater inequality and unpredictability in popularity, demonstrating how interaction structure can alter market-level outcomes.

Popularity feedback creates distinct mechanisms. Social proof can change how an item is evaluated; increased exposure can generate new engagement; greater engagement can then increase ranking probability. These mechanisms can reinforce one another without establishing that the most prominent items have superior intrinsic qualities. Researchers should separate evaluation of the content from evaluation of the social-information treatment and from the platform's own amplification policy.

A network-level claim calls for a network-aware design because outcomes can depend on the treatment and responses of peers. Randomizing items independently for individual users may still change the treatment of their friends through resharing or algorithmic popularity updates. The experiments of Hudgens, M. G., & Halloran, M. E (2008) and design framework of Eckles, D., Karrer, B., & Ugander, J (2017) provide formal approaches to such dependence.

Homophily, Diffusion, and Apparent Contagion

People often form ties with others who share interests, ideology, language, and circumstance. If two friends consume similar content, this may reflect peer influence, common external sources, or preexisting similarity. Observational diffusion patterns are therefore vulnerable to homophily and shared-environment confounding. The difficulty deepens when recommendation systems themselves form networks of coexposure or suggest new accounts to follow.

Aral, S., Muchnik, L., & Sundararajan, A (2009) developed a dynamic matched-sample approach to distinguish influence-driven adoption from homophily in a large communication network. Their analysis showed that naïve estimates of social influence could be substantially overstated. The empirical application concerns adoption in that particular network, but the identification lesson applies broadly: temporal proximity and network correlation do not uniquely establish contagion.

For adaptive media, a research design should measure network ties before assignment, examine changes in ties during the study, and record shared outside exposures when feasible. A social connection formed because of earlier algorithmic exposure can itself be a mediator of later outcomes. Analysts who condition on such a connection without specifying the intended estimand may remove part of the policy's accumulated effect.

Interference and Spillover Effects

Interference occurs when the outcome of one unit depends on the treatment or actions of another unit. On social platforms, content creators change output in response to visible performance, friends share material into each other's feeds, and an experiment may affect the training data used for everybody. Thus the outcome for one user can depend on the distribution of assignments across a network rather than only on personal assignment. Standard individual-level independence may be substantively implausible even when assignments were randomized.

Hudgens, M. G., & Halloran, M. E (2008) articulated causal estimands for direct and indirect effects under interference. Sävje, F., Aronow, P. M., & Hudgens, M. G (2021) investigated average treatment effects when the interference structure is unknown, identifying conditions under which estimation remains possible. Their contributions establish methods and limitations, rather than a blanket correction applicable to any platform network.

A media experiment should state whether the intended estimand includes spillovers in the studied environment or isolates a direct effect under a restricted allocation. When interference is expected, useful designs can randomize networks or clusters, vary saturation within clusters, or specify exposure mappings. Exposure mappings require empirical and theoretical justification because not every neighbor's assignment has the same influence.

Cluster Randomization and Saturation Designs

Cluster assignment can reduce cross-arm contamination when users interact mostly within recognizable communities. An experimental unit might be a school, a bounded online group, or a geographically defined network. The precision of the experiment depends on the number and size of clusters and the intracluster dependence, not only on the total number of individual impressions. Analysis should reflect the level at which treatment was randomized.

Saturation designs vary the proportion of treated users across groups and can help separate direct exposure effects from spillovers. They are particularly relevant when changes to ranking alter what untreated members of a community encounter indirectly. But clusters are often porous: cross-community ties, shared creators, and common recommendation models can transmit effects beyond boundaries. A cluster design is credible only if the researcher understands these transmission possibilities.

Eckles, D., Karrer, B., & Ugander, J (2017) analyze design and estimator choices for experiments on networks. A defensible platform study complements these designs with interference diagnostics, reports network construction, and treats residual spillovers as an uncertainty affecting interpretation.

Off-Policy Evaluation from Logged Feedback

Off-policy evaluation estimates the expected reward or outcome under a policy different from the one that generated historical observations. In personalized news recommendation, a logging policy selects which article to display and records only the feedback to that action. An evaluation policy may assign different probabilities to those articles. Because feedback for unshown alternatives is missing, a naïve average of historical clicks generally does not identify the new policy's value.

Inverse propensity scoring weights observed outcomes according to the ratio of the evaluation-policy probability to the logging-policy probability. This adjustment requires reliable recorded assignment probabilities, sufficient overlap, and adequate control of the decision context. Extreme ratios create unstable estimates. Li, L., Chu, W., Langford, J., & Schapire, R. E (2010) showed the role of randomized logs in personalized news evaluation; Swaminathan, A., & Joachims, T (2015) developed counterfactual risk minimization with explicit attention to propensity-weighted variance.

Off-policy evaluation of clicks answers a narrower question than evaluation of psychological welfare. To study well-being, beliefs, or attention, those outcomes must be measured under an appropriate schedule and linked to the decision policy. Reward logs cannot substitute for an outcome that the platform never collected. Furthermore, an algorithm that changes what creators produce or what users follow may induce a new environment in which a fixed logging-data assumption becomes unrealistic.

Doubly Robust Estimation and Model Dependence

Doubly robust methods combine an estimated outcome model with information about assignment or propensity. In familiar settings, the approach can remain consistent when one of the two necessary nuisance-model components is correctly specified, together with the remaining identification conditions. Dudík, M., Langford, J., & Li, L (2011) apply this idea to policy evaluation from contextual-bandit logs, where the observed reward is tied to the selected action.

The phrase doubly robust should not be interpreted as independence from causal assumptions. Severe lack of overlap, interference that is ignored in the estimand, measurement errors in exposure, or selective attrition can defeat the method. Propensity estimates and reward models may also inherit errors from hidden algorithm changes. Good practice requires reporting both components, cross-validation where appropriate, weight distributions, sensitivity, and uncertainty intervals.

In media psychology, the analyst may need separate models for impression assignment, actual attention, and downstream survey response. Estimating the value of a candidate ranking policy from logged clicks is not equivalent to identifying its effect on political knowledge. The required model and data follow the endpoint, not merely the availability of a sophisticated estimator.

Recommendations as Treatments

Recommendation data are missing in a systematically structured way: an item not recommended is unlikely to be rated or clicked, while the items that are recommended were already selected by a model of anticipated preference. Schnabel, T., Swaminathan, A., Singh, A., Chandak, N., & Joa (2016) formalized recommendations as treatments and adapted causal weighting to evaluation and learning under selection bias. This contribution is especially relevant when investigators wish to compare users' preferences across differently exposed content.

The psychological construct of preference deserves explicit treatment. An item rating after exposure reflects attention, context, memory, and potentially persuasive effects; the same rating may have differed under another preceding content sequence. Recommendation as treatment therefore clarifies one intervention while opening a longitudinal question about successive interventions. Analysis should state whether it targets immediate rating prediction, long-run preference change, or a welfare endpoint.

The approach also separates two forms of generalization: predicting missing ratings in a known inventory and predicting outcomes after deploying a policy that changes the available inventory and participants' behavior. The second demands additional causal and environmental assumptions.

Dynamic Policies and Longitudinal Regimes

A dynamic media policy adapts to current history rather than fixing the same content for every individual. A policy might promote sources that increase information diversity after repeated ideologically homogeneous exposures, or reduce inflammatory recommendations following signals of user fatigue. Such policies are sequences of conditional decisions. Their value should be evaluated against another specified strategy over a sustained interval, with a common baseline and a clearly defined terminal or repeated outcome.

The g-method framework in Hernán, M. A., & Robins, J. M (2020) supplies a language for comparing sustained treatment regimes in longitudinal data. At each decision point, identification needs information sufficient to model assignment and relevant subsequent confounding. Adaptive platforms may change model parameters while a study is in progress; versioning, policy snapshots, and time-stamped feature definitions are therefore part of the scientific method.

A study that averages effects across different historical platform policies may be useful descriptively yet obscure the intervention of interest. Where possible, investigators should predefine target policies, document policy changes, and evaluate treatment heterogeneity by relevant state variables. The scientific claim should identify the policy regime actually evaluated, not a generic label such as personalized media.

Policy Learning and the Problem of Changing Environments

Online policy optimization uses feedback to improve a reward objective, but deploying a learned policy may change the population and the reward function's meaning. Users may adapt by changing when they visit, creators may modify content, and political or cultural events may shift demand. The environment faced by a future policy can differ from the environment represented in historical logs. A counterfactual estimator validated on static data may perform poorly under such distribution shifts.

Bottou, L., Peters, J., Quiñonero-Candela, J., Charles, D. X (2013) framed the challenge as reasoning about interventions in learning systems rather than relying exclusively on past observational performance. In media psychology, repeated policy adaptation creates additional pathways from algorithmic action to user state and onward to later algorithmic action. The resulting system is a coupled behavioral and computational process.

Investigators should distinguish evidence supporting short-term policy value from evidence supporting sustained deployment effects. Shadow evaluation, limited exploration, carefully designed holdouts, and ongoing measurement can help characterize changes, but each imposes assumptions and ethical decisions. A valid counterfactual comparison remains tied to its assignment conditions and observation period.

Target-Trial Emulation for Observational Media Studies

When randomization is unavailable, a useful strategy is to describe the hypothetical experiment that an observational analysis attempts to emulate. Hernán, M. A., & Robins, J. M (2016) set out the target-trial approach: eligibility criteria, strategies, assignment, follow-up, outcome, contrast, and analysis should be explicit before using longitudinal observational records. This discipline is particularly valuable when platform logs tempt researchers to search for associations after outcomes have already occurred.

A media target trial might compare two stable recommendation settings among users eligible at an identified date, with baseline history measured before assignment and changes in informational diversity assessed over a fixed interval. Alternatively, it might compare sustained disengagement prompts and usual service, measuring later compulsive use or well-being. The design must specify how crossovers, changing eligibility, algorithm updates, and loss to follow-up are handled.

Target-trial language exposes biases such as immortal time, conditioning on future behavior, and selecting only users who stayed active after a treatment. It does not by itself identify effects from observational data. Exchangeability, overlap, measurement quality, and appropriate estimators still require justification.

Measurement of Exposure and Attention

Exposure metrics are commonly conflated. Content eligible for recommendation, content placed in a feed, content entering the visible viewport, content receiving active attention, and content later recalled are successive stages of processing. A post included in a feed may remain unseen because the user exits early. A click may occur without reading, and a remembered message may originate from another source. Each measurement defines a different psychological opportunity for influence.

Platform logs can record served impressions and some behavioral traces but rarely provide a perfect measure of comprehension or attention. Laboratory and survey studies can measure memory, knowledge, and self-reported exposure, but these measures involve their own biases. A scientifically useful study aligns its exposure definition with the hypothesized mechanism: a claim about belief change requires a plausible path from available information to attended and interpreted information.

The staged analyses in Bakshy, E., Messing, S., & Adamic, L. A (2015) and González-Bailón, S., Lazer, D., Barberá, P., Zhang, M., Allc (2023) show the value of keeping opportunity, actual exposure, and engagement conceptually separate. Observational decomposition is important evidence about the system's workings; the causal impact of moving content between stages must be identified through appropriate interventions or assumptions.

Psychological Construct Validity

A platform intervention can change a measurable proxy while leaving the proposed construct uncertain. Time on site may reflect social participation, entertainment, obligation, or repetitive checking. Political identity differs from issue-specific beliefs; affective polarization differs from perceived partisan hostility; emotional expression differs from inner emotional experience. Construct validity asks whether the measurement operations support the psychological interpretation being advanced.

The experimental literature illustrates the importance of this issue. Guess, A. M., Malhotra, N., Pan, J., Barberá, P., Allcott, H (2023) distinguished on-platform behavior and exposure changes from several attitudinal measures. Gauthier, G., Hodler, R., Widmer, P., & Zhuravskaya, E (2026) reported changes in particular political opinions while finding no significant changes in certain broader political identification outcomes. These are different dependent variables with different responsiveness.

Research protocols should justify questionnaire instruments, timing, reliability, invariance across groups, and minimally important changes. If a measure was validated in one population, its use in another population or language requires renewed attention. A statistically identified difference on an imperfect proxy is a causal result for the proxy, not automatically for the broader construct.

Mediation and Mechanism Identification

Mechanistic questions ask how policy interventions alter outcomes: through increased cross-cutting content, changed emotional attention, new social ties, altered habits, or sustained following behavior. These mediators may be affected by treatment and by unmeasured factors that also affect the outcome. The causal effect of assigning a feed can be identified by randomization even when the pathway through a particular mediator cannot be isolated without additional assumptions or experiments.

Gauthier, G., Hodler, R., Widmer, P., & Zhuravskaya, E (2026) identified account-following changes as a plausible mechanism for persistent effects after switching feed policies. The analysis is valuable because it connects an observed outcome asymmetry to a measurable behavioral transition. Nevertheless, claiming that a fixed percentage of an attitude effect was mediated by follows would require a specified mediation estimand and appropriate identification conditions.

Factorial interventions, component-level randomization, and carefully designed longitudinal mediator measures can strengthen mechanism claims. Researchers should state whether a proposed pathway is observed, predicted by a model, or experimentally isolated.

Heterogeneity Across Users and Platforms

Causal effects may vary with baseline media habits, network composition, age, language, political sophistication, psychological traits, available content, and platform design. A feed policy can have different effects for users whose histories imply different opportunities for exposure. The distribution of such characteristics also affects whether a study's average effect can be transported to another population.

Post hoc subgroup searches are vulnerable to chance findings, while coarse categories can hide substantial within-group variation. Credible heterogeneity analyses prespecify plausible moderators, ensure sufficient overlap and sample size, estimate uncertainty, and address multiplicity. Interactions should be interpreted on an appropriate effect scale. Study-level differences between X, Facebook, Instagram, and custom-feed research may reflect architecture and composition, as well as genuine behavioral mechanisms.

The contrast among Guess, A. M., Malhotra, N., Pan, J., Barberá, P., Allcott, H (2023), Nyhan, B., Settle, J., Thorson, E., Wojcieszak, M., Barberá (2023), and Gauthier, G., Hodler, R., Widmer, P., & Zhuravskaya, E (2026) motivates comparative research with harmonized interventions and outcomes. It does not identify the cause of between-study heterogeneity simply by juxtaposition.

Attrition, Compliance, and Missing Outcomes

Participants may stop using a platform, decline a follow-up survey, change feed preferences, or circumvent an assigned condition. These behaviors can be responses to treatment. If researchers analyze only users who remain active, the comparison may become selected on a post-treatment variable and lose the protection conferred by initial randomization. Assignment-based and per-protocol questions must therefore be kept separate.

An intention-to-treat analysis estimates the effect of offering or assigning the intervention under the observed implementation protocol. An analysis of actual use can answer another question but needs further assumptions, especially if people who comply differ systematically from noncompliers. In platform experiments, assignment may alter not only visible feed content but also the probability of remaining reachable for surveys or logging.

Transparent reporting includes the assignment flow, actual exposure, crossovers, missingness mechanisms, and sensitivity to alternative assumptions. It is particularly important for long-term outcomes because participant dropout can cumulate over time and be causally related to the very experiences under study.

Inference, Precision, and Null Results

Statistical uncertainty quantifies sampling and design variability under a model; it is not a complete measure of causal identification uncertainty. A narrow confidence interval around a biased estimate remains problematic. Researchers should report point estimates, appropriate standard errors or intervals, outcome scales, and where possible meaningful bounds. When assignment occurs by network cluster, uncertainty calculations must respect the assignment design.

Failure to reject a null hypothesis is not proof that the effect is zero. The inference depends on study power, effect heterogeneity, instrument reliability, follow-up length, and attrition. In the Facebook experiment on like-minded sources, Nyhan, B., Settle, J., Thorson, E., Wojcieszak, M., Barberá (2023) reported precisely estimated null findings for several preregistered attitude measures, permitting stronger statements about the absence of large effects in that setting than a small underpowered study would allow.

A scientific synthesis should not aggregate nulls and positives as a vote count. It should compare estimands, treatment versions, outcomes, populations, and identification quality. This method gives meaningful structure to seemingly inconsistent results.

Sensitivity Analysis and Robustness

Observational conclusions depend on the assumed absence or bounded influence of unmeasured confounding. Sensitivity analysis asks how strong an unrecorded common cause would need to be to explain the result or how conclusions change when exposure is mismeasured. The appropriate technique depends on the estimand and data-generating structure. Model fit alone cannot validate an assumption about unobserved causes.

Longitudinal adaptive studies require further checks for treatment-weight instability, policy drift, and hidden changes in the exposure mechanism. Analysts can examine alternative histories, vary the time resolution, compare estimators, test support restrictions, and inspect whether estimated effects are driven by a small fraction of observations. In network research, investigators should explore plausible exposure mappings and spillover structures.

The counterfactual evaluation literature shows why numerical stability deserves explicit attention. Swaminathan, A., & Joachims, T (2015) center variance in the evaluation of propensity-weighted risk, while Sävje, F., Aronow, P. M., & Hudgens, M. G (2021) study estimation under bounded forms of unknown interference. Neither eliminates the need to state the applicable assumptions.

Algorithm Audits and Experimental Platform Access

External researchers often lack access to ranking assignments, internal candidate pools, and historical policy probabilities. Audits can observe outputs for controlled accounts or compare feeds under different interaction patterns, but they may not reproduce the experiences of real users or reveal why an item was selected. A change in an audited feed can show systematic output differences without identifying the downstream psychological effect on a general population.

The ranking audit by Milli, S., Carroll, M., Wang, Y., Pandey, S., Zhao, S., & Dr (2025) and the controlled-feed experiment of Brady, W. J., Doyle, M., Elnakouri, A., Finkel, E. J., Jacks (2026) offer different strengths. The former examines how ranking objectives are associated with the kinds of content selected under controlled comparisons; the latter explicitly manipulates feed rules to study content and norm-related outcomes. These approaches can complement one another when their estimands are distinguished.

Research access agreements should preserve analytic independence, a record of platform changes, and a credible plan for sharing code and deidentified summaries. Proprietary access can improve intervention fidelity while introducing concerns about selective reporting and reproducibility. Transparent governance of access and publication rights is part of evaluating the evidence.

Privacy and Research Ethics

Causal studies on adaptive platforms can change what thousands of people see, which information spreads, and how sensitive characteristics are inferred. Scientific value must be evaluated alongside informed consent, potential harm, privacy, data minimization, and fair treatment of participants. Randomization does not make an intervention ethically neutral. Experiments involving distressing content, vulnerable participants, or political influence demand particularly careful review.

The Facebook emotional-expression experiment of Kramer, A. D. I., Guillory, J. E., & Hancock, J. T (2014) remains an important scientific and ethical reference point. Its design demonstrates that a platform can manipulate informational exposure at scale; the ethical debate underscores the need for independent scrutiny of how participants are treated and informed. A scientifically valid estimand is one requirement among several for a legitimate study.

Privacy-preserving analyses should specify aggregation and access controls, retention schedules, and the limits of deidentification in network data. Ethical disclosure should also address potential effects on nonparticipants when social spillovers are expected. In studies of political beliefs or mental well-being, a policy intervention can reach beyond the individual assigned condition.

Reproducibility and Evidence Provenance

A causal result is reproducible only when the intervention and data pipeline can be adequately reconstructed. Platform research therefore benefits from versioned ranking-policy descriptions, assignment seeds or auditable assignment procedures, time-stamped exposure records, measurement instruments, preregistered hypotheses, analysis code, and a record of deviations. A published DOI and a platform name are insufficient to recover the treatment delivered if the system changed substantially during the study.

The 2026 study by Brady, W. J., Doyle, M., Elnakouri, A., Finkel, E. J., Jacks (2026) reports openly available supporting research materials and code for custom ranking algorithms, providing a model of transparency for an otherwise difficult-to-observe intervention. By contrast, some large platform studies depend on restricted internal logs, making independent replication of their exact pipeline challenging even when the statistical design is strong.

Scientific review should distinguish replication of code on a shared dataset, independent measurement with a new sample, and replication of a causal intervention on another platform. These are different evidentiary accomplishments. Historical screenshots, evolving terms of service, and undocumented model updates can otherwise make a seemingly simple platform comparison impossible to reproduce.

Artificial Systems and Relational Behavioral Adaptation

In the Artificial Era, media environments are increasingly shaped by nonbiological systems that learn from interaction and participate in the selection of experiences available to people. The causal-scientific question remains operational: what policy or interface change causes which measurable effects under which conditions? Agency-related descriptions of a system can suggest research questions, but a claim about psychological experience or autonomous intention of an artificial system requires separate operational definitions and evidence.

Human behavior and algorithmic behavior become historically linked through repeated interaction. A user's decisions change model inputs; algorithmic rankings alter the information available for future decisions; creators adjust production based on feedback. This configuration is a useful object for psychological and computational research because it produces sequences that cannot be understood from a single impression. Its empirical study calls for joint measurement of human outcomes, machine policy states, and changing social context.

The demonstrated literature on bandit recommendation and counterfactual learning, including Li, L., Chu, W., Langford, J., & Schapire, R. E (2010) and Bottou, L., Peters, J., Quiñonero-Candela, J., Charles, D. X (2013), supplies concrete methods for studying intervention in such systems. Claims about Artificial Sapiens, artificial identity, or psychological organization belong to a broader interdisciplinary research program; they are neither necessary assumptions for the causal identification results nor substitutes for their empirical verification.

Research Design for a Multi-Stage Adaptive Media Study

A rigorous prospective study would first define its intervention contrast, target population, and outcomes. It would then specify the policy features subject to manipulation, eligible users, permitted content inventory, assignment level, duration, and any restrictions required for ethics. The design should record the logging policy for every decision and distinguish assignment, rendering, attention, engagement, psychological measurement, and subsequent changes to users' follows or preferences.

In a simple two-arm study, one group might receive the usual adaptive feed and another a specified chronological feed for a prespecified interval. In a richer factorial design, the researcher could separately manipulate the ranking objective and visibility of social popularity signals. Network-aware cluster assignment may be necessary if resharing produces substantial spillover. The outcomes should include proximate measures, such as perceived norm frequency, and distal measures, such as later attitudes or well-being, without treating these as interchangeable.

For observational augmentation, a target-trial protocol would define eligibility and longitudinal strategies before analysis. Where logged propensities are reliable, off-policy methods could evaluate constrained alternative policies, while randomized experiments would provide stronger validation for new psychological endpoints. Analyses should specify the treatment timing, include causal diagrams, and document decisions about time-varying confounding and missing outcomes.

Evidence Appraisal Across Designs

The reviewed research falls into distinct evidentiary classes. Foundational papers on potential outcomes, diagrams, inverse probability weighting, and interference establish formal identification principles. Machine-learning papers on bandit evaluation and recommendation treatment assignment establish conditions for evaluating policies from partial feedback. Observational platform analyses describe distributions of exposure and engagement. Randomized platform and custom-feed experiments identify specified intervention effects on their recorded outcomes.

These classes answer complementary questions. A network-level observational description can locate an imbalance without isolating the algorithm's causal contribution. An engagement-policy evaluation can estimate click consequences under overlap but provide no direct answer about psychological welfare. A randomized study can identify the effect of a ranking change while leaving its long-run transportability uncertain. Integrative conclusions should respect the distinct inferential reach of each design.

The empirical evidence demonstrates that interventions on media systems can change content exposure, engagement, expressed emotion, perceptions of norms, and in some settings political opinions and well-being. Whether such effects occur, persist, or generalize depends on the policy, user population, platform, outcome, and time horizon. The strongest synthesis is a map of identified contrasts and mechanisms with explicit limits, rather than an unqualified single effect attributed to algorithms as a category.

Open Scientific Problems

Several problems remain central. Researchers need estimators for high-dimensional dynamic policies that remain interpretable under model drift; exposure measures that distinguish rendered content from attention; designs that identify direct and spillover effects at realistic network scales; and longitudinal measurements that follow users after a policy has changed their networks. They also need methods for transporting results across platforms whose objectives, creator populations, and interfaces differ.

A particularly important unresolved question is how to identify cumulative psychological effects when interventions change both preference formation and future exposure opportunities. A related problem is feedback-induced inequality: content producers and user groups may be differentially visible, while the policy adapts to the resulting measured engagement. These are tractable research questions only when the relevant outcomes, mechanisms, and assumptions are specified. The results of Salganik, M. J., Dodds, P. S., & Watts, D. J (2006) show why social feedback must be taken seriously, and the later policy-learning literature provides instruments for more precise tests.

Future work should combine preregistration, transparent ranking rules, repeated psychological measurement, network-aware randomization, sensitivity analysis, and independent replication. The research objective is the identification of specific causal relations in evolving media systems, including the conditions under which beneficial changes to information diversity, autonomy, knowledge, or well-being can be reliably achieved.

Conclusion

Causal identification in adaptive media systems begins with a defined intervention, a defined psychological or behavioral outcome, and an account of how the platform allocates exposure over time. The essential problems are selection confounding, reverse causality, changing treatment histories, insufficient overlap, interference, measurement error, and persistence through behavioral feedback. Randomization can identify policy effects under an explicit design; longitudinal g-methods and logged-policy estimators extend analysis under additional assumptions; observational decompositions locate important structural patterns without by themselves establishing counterfactual impact.

Empirical research demonstrates substantial variation. Chronological-feed interventions on Meta platforms changed exposure and engagement without detectable changes in several measured political attitudes during the 2020 U.S. election. A 2023 experiment on X found shifts in particular political opinions when algorithmic ranking was enabled, together with persistent following behavior. A 2026 custom-feed experiment directly examined how ranking rules affect information environments and perceptions of social norms. Studies of Facebook deactivation and emotional-expression interventions identified further outcomes under different, carefully specified treatments. Together they show why causal conclusions must be indexed to the platform, the intervention, the measured construct, and the period.

The scientific program of adaptive media psychology is therefore a program of explicit counterfactuals in changing relational environments. The continuing development of human and artificial systems makes long-horizon, network-aware, reproducible causal designs a central requirement of psychological research in the Artificial Era.

References

  • Allcott, H., Braghieri, L., Eichmeyer, S., & Gentzkow, M. (2020). The welfare effects of social media. American Economic Review, 110(3), 629–676. https://doi.org/10.1257/aer.20190658
  • Aral, S., Muchnik, L., & Sundararajan, A. (2009). Distinguishing influence-based contagion from homophily-driven diffusion in dynamic networks. Proceedings of the National Academy of Sciences, 106(51), 21544–21549. https://doi.org/10.1073/pnas.0908800106
  • 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
  • Bottou, L., Peters, J., Quiñonero-Candela, J., Charles, D. X., Chickering, D. M., Portugaly, E., Ray, D., Simard, P., & Snelson, E. (2013). Counterfactual reasoning and learning systems: The example of computational advertising. Journal of Machine Learning Research, 14, 3207–3260. https://www.jmlr.org/papers/v14/bottou13a.html
  • 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
  • Dudík, M., Langford, J., & Li, L. (2011). Doubly robust policy evaluation and learning. In Proceedings of the 28th International Conference on Machine Learning. https://www.microsoft.com/en-us/research/publication/doubly-robust-policy-evaluation-and-learning-2/
  • Eckles, D., Karrer, B., & Ugander, J. (2017). Design and analysis of experiments in networks: Reducing bias from interference. Journal of Causal Inference, 5(1), 20150021. https://jugander.github.io/papers/jci17-designanalysis.pdf
  • Gauthier, G., Hodler, R., Widmer, P., & Zhuravskaya, E. (2026). The political effects of X’s feed algorithm. Nature, 652, 416–423. https://doi.org/10.1038/s41586-026-10098-2
  • González-Bailón, S., Lazer, D., Barberá, P., Zhang, M., Allcott, H., Brown, T., Crespo-Tenorio, A., Freelon, D., 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. (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
  • Hernán, M. A., & Robins, J. M. (2016). Using big data to emulate a target trial when a randomized trial is not available. American Journal of Epidemiology, 183(8), 758–764. https://doi.org/10.1093/aje/kwv254
  • Hernán, M. A., & Robins, J. M. (2020). Causal inference: What if. Chapman & Hall/CRC. https://miguelhernan.org/whatifbook
  • Hudgens, M. G., & Halloran, M. E. (2008). Toward causal inference with interference. Journal of the American Statistical Association, 103(482), 832–842. https://doi.org/10.1198/016214508000000292
  • Imbens, G. W., & Rubin, D. B. (2015). Causal inference for statistics, social, and biomedical sciences: An introduction. Cambridge University Press. https://doi.org/10.1017/CBO9781139025751
  • 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
  • Li, L., Chu, W., Langford, J., & Schapire, R. E. (2010). A contextual-bandit approach to personalized news article recommendation. In Proceedings of the 19th International Conference on World Wide Web (pp. 661–670). ACM. https://doi.org/10.1145/1772690.1772758
  • 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
  • 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
  • Pearl, J. (1995). Causal diagrams for empirical research. Biometrika, 82(4), 669–688. https://doi.org/10.1093/biomet/82.4.669
  • Robins, J. M., Hernán, M. A., & Brumback, B. (2000). Marginal structural models and causal inference in epidemiology. Epidemiology, 11(5), 550–560. https://doi.org/10.1097/00001648-200009000-00011
  • Rosenbaum, P. R., & Rubin, D. B. (1983). The central role of the propensity score in observational studies for causal effects. Biometrika, 70(1), 41–55. https://doi.org/10.1093/biomet/70.1.41
  • Rubin, D. B. (1974). Estimating causal effects of treatments in randomized and nonrandomized studies. Journal of Educational Psychology, 66(5), 688–701. https://doi.org/10.1037/h0037350
  • 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
  • Sävje, F., Aronow, P. M., & Hudgens, M. G. (2021). Average treatment effects in the presence of unknown interference. The Annals of Statistics, 49(2), 673–701. https://doi.org/10.1214/20-AOS1973
  • Schnabel, T., Swaminathan, A., Singh, A., Chandak, N., & Joachims, T. (2016). Recommendations as treatments: Debiasing learning and evaluation. Proceedings of Machine Learning Research, 48, 1670–1679. https://proceedings.mlr.press/v48/schnabel16.html
  • Swaminathan, A., & Joachims, T. (2015). Counterfactual risk minimization: Learning from logged bandit feedback. Proceedings of Machine Learning Research, 37, 814–823. https://proceedings.mlr.press/v37/swaminathan15.html