Humans and Recommendation Algorithms: A Standalone Research Synthesis

HUMANS AND RECOMMENDATION ALGORITHMS

A Standalone Research Synthesis on Attention, Preference, Identity, Social Networks, and Algorithmically Mediated Opportunity

Synthesis of the full research investigation

“Our behavior increasingly helps construct the informational environment that will solicit our next behavior.”

Prepared as a standalone exhibit. Evidence claims, interpretation, and speculation are explicitly separated.

Executive Thesis

Social-media recommendation systems are adaptive components of larger sociotechnical ecosystems. They learn from human behavior and causally redistribute attention, exposure, visibility, and opportunities for connection. Their effects on immediate behavior are well established. Effects on social-network choices and some political attitudes have been demonstrated in particular settings. Effects on durable preferences, identity, personality, and long-run cultural development remain substantially less certain.

The strongest evidence does not support either popular extreme. It is too strong to say that users are merely passive subjects whose beliefs are manufactured wholesale by algorithms. It is also too strong to say that algorithms simply reflect independent human wants. Recommendation systems operate between those poles: they infer from prior behavior, construct an opportunity set, amplify some signals, introduce novelty, alter salience, and sometimes trigger decisions - such as follows - that persist after the immediate recommendation disappears.

The central unresolved question is therefore longitudinal: when do repeated changes in attention, discovery, network structure, and content supply accumulate into persistent changes in preferences, identity, relationships, or institutions? Existing research is much better at measuring minutes, weeks, and months than years or decades.

A second conclusion emerged only after the investigation widened beyond persuasion: the largest societal effects may occur even if average effects on beliefs are modest. Recommendation systems can alter who is discovered, which creators receive attention, which communities become reachable, which topics become salient, and which cultural variants receive visibility. In that sense, the problem may be less about direct mind-changing than about the distribution of human opportunity.

The appropriate confidence gradient is:

• Very high confidence: user behavior affects recommendation; ranking changes exposure; ranking changes engagement and platform behavior.

• High confidence: engagement is not a transparent measure of reflective preference; some systems asymmetrically amplify emotional or partisan material; large exposure changes do not necessarily produce large attitude changes.

• Moderate-to-high confidence: some recommendation systems can alter some political attitudes and following behavior in specific settings.

• Moderate confidence: prior disposition and algorithmic amplification interact; network mediation and creator adaptation are important mechanisms but incompletely mapped.

• Low-to-moderate confidence: recommendation contributes to durable preference formation, differential developmental effects, and broad cultural evolution.

• Low confidence: routine durable identity reshaping, stable personality transformation, generalized radicalization of moderates, or a predictable twenty-year psychological trajectory.

Background and Context

Information has always been filtered. Families, teachers, clergy, editors, publishers, broadcasters, librarians, political organizations, advertisers, and peer groups all shape what people encounter. Algorithmic recommendation did not invent mediation, selective exposure, homophily, persuasion, or cultural gatekeeping.

What changed is the combination of personalization, scale, automation, behavioral measurement, and rapid feedback. Earlier gatekeepers typically selected one newspaper front page, one broadcast schedule, or one curriculum for many people. A recommender can construct a different information environment for each person and update it after individual interactions.

A major technical ancestor was collaborative filtering. The GroupLens project, published at ACM CSCW in 1994, used evaluations of Usenet material to predict what similar users might value. The foundational idea was simple: people like you liked this, so perhaps you will too. Later systems increasingly used implicit behavior rather than explicit ratings.

As social platforms grew, chronological display became less practical because users generated more material than anyone could consume. Ranking systems increasingly predicted which items would generate interaction or satisfaction. Facebook developed ranked News Feed systems; YouTube shifted from click-based signals toward watch time and later satisfaction-oriented signals; TikTok's For You feed made recommendation-driven discovery rather than the explicit social graph central to the user experience.

The historical shift is therefore not from 'unmediated' to 'mediated' reality. It is from relatively static and mass-mediated selection toward individually adaptive environments that learn from the user's own behavior.

The Core System: Behavior, Inference, Ranking, and Feedback

The investigation initially asked whether users shape algorithms or algorithms shape users. That binary proved misleading because the two causal directions are entangled.

The early arrows in this loop are strongly supported. Users supply signals such as watches, pauses, skips, searches, likes, dislikes, comments, shares, follows, subscriptions, completion, and explicit feedback. Platforms use those signals to rank candidate content. The resulting exposure then changes what users can attend to and how they behave.

The difficult arrow is the transition from exposure and behavior to durable preference. A system can cause someone to see or click something without changing what they endorse. It can increase platform time without altering ideology. It can introduce a creator whom a user later follows without changing a deeply held belief.

The full sociotechnical system is therefore broader than a user and a model. A more complete representation is:

This distinction matters because different research traditions often measure different stages of the chain. Machine learning measures prediction. Psychology measures attention and motivation. Political experiments measure attitudes. Network science studies ties. Sociology studies communities and identity performance. Economics studies incentives and welfare. None of these alone can establish the whole causal sequence.

The Preference Problem

One of the most consequential conceptual findings is that the word 'preference' hides several different objects. Recommender systems usually observe behavior, yet researchers and public debate often interpret that behavior as if it directly revealed what the user truly wants.

At least four distinct notions should be separated:

• Behavioral preference: what the person clicked, watched, searched, shared, replayed, or returned to.

• Stated preference: what the person explicitly says they want.

• Reflective preference: what the person would choose after deliberation rather than immediate reaction.

• Welfare preference: what contributes to the person's longer-term wellbeing or valued goals.

These can diverge. A person may watch upsetting political content longer precisely because it is upsetting. An engagement model can correctly predict that the content will capture attention while misrepresenting whether the person endorses or values it.

The 2025 preregistered audit by Milli and colleagues is especially important. Engagement-based Twitter ranking amplified more partisan, negative, and out-group-hostile political material relative to reverse chronology, and in the political domain users did not necessarily report preferring what the engagement-ranked system selected.

The lesson is narrower than 'attention is never preference.' Engagement is a useful behavioral signal, but its psychological meaning depends on context. The system may sometimes be learning reflexes, curiosity, anger, habit, or social obligation rather than reflective endorsement.

This has a practical design consequence: there is no purely technical answer to what a recommender ought to optimize when human impulses, stated desires, and long-term goals conflict. Any attempt to optimize 'wellbeing' or 'flourishing' requires normative choices about what counts as welfare and who has authority to define it.

Mechanisms of Influence

The investigation ultimately separated several mechanisms that public debate often collapses together.

Personalization is prediction: the system estimates what a particular user is likely to engage with. Amplification is distribution: some items receive disproportionately greater visibility under a ranking objective. Preference formation is psychological change: repeated exposure changes what a person later values, believes, seeks, or identifies with. These mechanisms can coexist, but they are not equivalent.

A more complete causal hierarchy is:

The opportunity-set mechanism emerged as particularly important. A recommender need not persuade the user to choose B. It can matter by selecting A and B from millions of possibilities. That is a form of computational choice architecture. Influence can occur before persuasion, through deciding which topics, creators, communities, and options become salient enough to be considered.

The network mechanism may be more persistent than the content mechanism. A temporary recommendation can lead to a follow; the follow can persist; the new relationship can generate future exposure even after the original treatment disappears. The machine may therefore act as a matchmaker that changes the user's future human social environment.

Creator adaptation adds a population-level mechanism. Creators watch which formats and topics receive distribution and adapt thumbnails, length, pacing, language, emotional intensity, and posting strategy. The evidence that creators respond to platform incentives is strong descriptively, while the causal magnitude of ranking-induced changes in cultural production remains undermeasured.

Core Evidence

1. User behavior materially shapes recommendation

Platform documentation and recommender-system research converge on a basic fact: interaction data are central inputs. TikTok has described watches, likes, shares, follows, comments, content characteristics, and contextual signals as components of recommendation. YouTube has described watch and search history, subscriptions, likes, dislikes, explicit negative feedback, and satisfaction surveys. The exact weights change over time and are not fully transparent, but the role of behavioral data is not seriously disputed.

2. Ranking changes political exposure

Huszár and colleagues' large-scale randomized Twitter study, published in PNAS in 2022, compared algorithmic and reverse-chronological timelines across several countries. It found systematic differences in political amplification, including greater amplification of mainstream-right political content than left content in six of seven countries. Importantly, the study did not establish that ideological extremes were universally favored over moderates. The result demonstrates asymmetric exposure, not a universal direction of persuasion.

3. The Meta experiments: exposure can change while attitudes remain stable

The 2023 Meta election studies are among the strongest causal evidence in the field because they randomized real users inside active platforms during the 2020 U.S. election. Interventions included reverse-chronological feeds, reducing like-minded-source exposure, and removing reshared content.

The interventions substantially altered behavior and information exposure. Users in chronological-feed conditions spent less time on the platforms and behaved differently. In the like-minded-source experiment, exposure to congenial sources fell from roughly 54 percent to roughly 36 percent among more than 23,000 participants. Yet researchers detected no measurable effects across eight preregistered political-attitude outcomes during the intervention. Removing reshares similarly changed exposure without detectable political-opinion effects.

These null findings are not evidence that algorithms never matter. They are strong negative evidence against a simple model in which changing partisan exposure quickly changes established political attitudes. They also show that algorithmic ranking and user selection are separable: when congenial material became scarcer, users became more likely to engage with it when it did appear.

The limitations are equally important. Participants were mature adults with established political histories. The interventions lasted months, not years. Temporarily changing ranking could not erase previously formed networks, creators, or political habits. A null three-month effect cannot falsify a twenty-year developmental effect - but the absence of long-term evidence cannot be used to assume such an effect either.

4. The 2026 X experiment: some attitudes and networks can change

The 2026 Nature study of X, based on a seven-week randomized field experiment with active U.S. users, produced one of the investigation's most important affirmative findings. Exposure to the algorithmic feed increased engagement, increased conservative political exposure, reduced relative exposure to traditional media, and shifted several measured political opinions in a conservative direction. The study did not find significant effects on affective polarization or self-reported partisanship.

The most consequential result may have been network formation. Users exposed to the algorithmic feed followed more conservative political activist accounts, and those follows persisted beyond the immediate treatment. This creates a plausible mechanism of path dependence: temporary ranking changes can leave behind a changed human information network.

The Meta and X results should not be treated as science reversing itself. They examined different platforms, historical moments, user populations, content supplies, and interventions. Facebook's interventions mainly changed ranking within established social environments. X's For You feed also introduced material from unfollowed accounts. The broader lesson is that recommender effects are conditional rather than universal.

5. YouTube and the rabbit-hole question

Haroon and colleagues' 2023 PNAS audit used approximately 100,000 sock-puppet accounts to examine ideological and problematic recommendations on YouTube. Partisan profiles received more ideologically congenial recommendations, and problematic channels became somewhat more prevalent along some deeper trajectories, particularly for right-leaning profiles. But the study did not show a general progression toward increasing ideological extremity.

A broader systematic review of 23 studies found mixed evidence: many studies implicated recommendations in pathways to problematic content, several were mixed, and a small number did not. The major limitation is that external researchers cannot observe the full production recommendation architecture.

The defensible conclusion is therefore that recommendation can facilitate access to problematic material. The stronger claim that YouTube routinely turns politically moderate people into extremists is not supported by the evidence reviewed.

6. Feedback-loop theory and simulation

Chaney, Stewart, and Engelhardt showed in simulation that recommendation systems trained on behavior generated after prior recommendations can create algorithmic confounding and increasing homogeneity without corresponding gains in user utility. Jiang and colleagues likewise modeled degenerate feedback loops in which changing user preferences and recommendation interact over time.

These studies establish that self-reinforcing dynamics are possible under specified assumptions. They do not by themselves establish the magnitude of those dynamics on TikTok, YouTube, Facebook, or X. The investigation repeatedly corrected itself on this point: theoretical possibility must not be reported as observed population effect.

7. User agency and algorithmic folk theories

Karizat, Delmonaco, Eslami, and Andalibi's 2021 TikTok study provides qualitative evidence that users develop 'algorithmic folk theories' about how recommendation works and sometimes deliberately change behavior to influence the system's representation of them. Participants described efforts to align the feed with their self-understanding and experiences of algorithmic representation.

This is valuable evidence that users are not simply passive recipients. But the small qualitative sample cannot establish the prevalence or population-level effectiveness of these strategies. It supports co-adaptation, not a claim that TikTok causally transforms durable identity.

Competing Interpretations

No single interpretation explains the entire literature. The strongest explanation is conditional interaction, but several rivals remain serious.

The reflection model is the strongest skeptical rival because it requires few additional mechanisms. It predicts that long-term randomized differences should largely converge once genuine preferences are accurately learned. The amplification model adds a causal role for the algorithm without requiring preference creation; it predicts the largest effects among users with weak pre-existing tendencies.

The attention model deserves more emphasis than it often receives. A system can substantially alter a person's life without changing beliefs if it repeatedly changes what receives time, practice, emotion, or consideration. The network model makes a similar move: the recommendation may matter because it changes whom the person encounters rather than because the content directly persuades.

The most discriminating future studies would randomize early exposure among new users, separate passive content exposure from opportunities to follow or join communities, remove personalization for long periods after it has matured, and vary exploration versus exploitation. These designs could force reflection, amplification, path-dependence, and preference-construction explanations apart.

Anomalies, Contradictions, and Uncertainty

The single most important contradiction is the coexistence of large environmental effects and often-small measured attitude effects. Meta shows that changing the information stream can leave political attitudes stable. X shows that other recommendation environments can shift some attitudes and construct different follow networks. YouTube shows that problematic-content pathways can increase without a general monotonic rise in ideological extremity.

These findings suggest several possibilities that remain unresolved:

• Attitudes may be relatively stable while attention, habits, and networks remain highly malleable.

• Recommendation may affect different platforms through different mechanisms because social-graph, discovery, and sequential-consumption systems are not equivalent.

• Short-term treatments may underestimate accumulated effects because mature users already carry years of prior network and platform history.

• Average effects may conceal strong effects for a susceptible minority or near-zero effects for most users.

• Adding algorithmic exposure may have different consequences from removing it; effects may be asymmetric and partially irreversible when follows, habits, or communities persist.

• A 'chronological' feed is not a neutral no-algorithm baseline; it inherits prior follows, popularity, moderation, creator adaptation, and platform history.

The field also faces a moving-target problem. Platforms change objectives, models, moderation rules, interface design, and creator incentives faster than academic publication cycles. A finding can be scientifically valid for the studied system and operationally obsolete by publication.

External researchers usually observe inputs and displayed outputs but not the full system of candidate generation, safety filtering, ranking, diversity constraints, ads, experiments, and business rules. Internal platform research has superior data access but raises independence and replicability concerns.

The evidence base is also geographically uneven. Strong causal work is concentrated in the United States and other wealthy democracies. Recommendation may interact differently with multilingual populations, authoritarian politics, weak or strong journalistic institutions, differing family structures, and different levels of digital literacy.

Several populations are underrepresented, including adolescents in long-term causal comparisons, older adults, disabled and neurodivergent users, low-resource-language communities, and nonusers who may nonetheless experience platform spillovers through culture, journalism, politics, and commerce.

A deeper blind spot is that the unit of analysis may not be a single platform at all. People inhabit multiple recommender systems simultaneously - social media, video, music, shopping, search, streaming, and increasingly AI assistants. One system may introduce an interest and another reinforce it. Cross-platform longitudinal evidence remains scarce.

Cross-Disciplinary Interpretation

Recommender-systems engineering shows that observed behavior is affected by exposure and that prediction, ranking, and causal estimation are distinct tasks. Causal inference emphasizes the missing counterfactual: what would the same person have done under a different ranking policy? Network science warns that similarity among connected users can reflect homophily, peer influence, or both.

Psychology adds mechanisms of attention, reinforcement, salience, habit, emotion, and motivated reasoning, while also warning against assuming unlimited malleability. Sociology distinguishes internal identity from social and performative identity and emphasizes communities, status, institutions, and group norms. Human-computer interaction shows that users build models of the algorithm, sometimes modifying their behavior because of what they believe the system rewards.

Media studies provides historical continuity: agenda setting and gatekeeping existed before machine learning. The distinctive feature of algorithmic recommendation is individualized, continuous, data-driven gatekeeping. Economics adds platform incentives, multi-sided markets, search costs, and the distinction between engagement and welfare. Law reveals growing institutional concern but cannot convert regulatory concern into scientific causal proof. Philosophy clarifies that preference, autonomy, welfare, manipulation, and attention are partly normative concepts, not merely measurable quantities.

Across these disciplines, four agreements are particularly robust:

1. Observed behavior is not identical to underlying or reflective preference.

2. Context mediates effects; there is no universal coefficient for 'the algorithm.'

3. Attention allocation is a real outcome even when persuasion is weak.

4. Social networks complicate direct media-effects models because recommendation can change who becomes part of the user's future informational environment.

The integrated object is therefore not 'an algorithm acting on a user.' It is a sociotechnical ecosystem in which humans, models, creators, markets, networks, interfaces, moderators, advertisers, regulators, and cultural institutions interact.

Interpretive Dimensions

The empirical investigation asks what recommendation systems do. Interpretive frameworks ask what the phenomenon means or resembles. These should not be confused.

Symbolically, the strongest image is the mirror that answers back. The feed reflects a portrait assembled from behavioral traces, but then uses that portrait to alter what the person sees next. The image is not evidence that the system has discovered the user's 'true self.' It is a warning against confusing a partial behavioral model with the whole person.

The labyrinth offers another useful symbol. A traditional maze is fixed; a personalized feed is adaptive. The traveler partly constructs the maze by walking through it. The oracle captures the subjective experience of a system predicting an interest before the user consciously articulates it, but statistical prediction is not supernatural knowledge.

Jungian concepts such as persona, projection, and shadow can illuminate the tension among public self-presentation, private behavior, and machine-inferred identity. These are interpretive frameworks within analytical psychology, not causal evidence that recommendation systems transform identity.

Contemplative traditions concerned with conditioning, attention, craving, aversion, and self-observation can provide serious comparative perspectives. Buddhist dependent-origination frameworks, for example, analyze recursive conditioning of contact, feeling, craving, and action. A recommendation loop can be structurally compared with repeated stimulus-response conditioning, but there is no historical claim that Buddhist traditions anticipated machine-learning systems.

The metaphysical questions are genuine even when no mystical claims are made. If a predictive model helps construct the environment from which the predicted person's next action emerges, prediction becomes entangled with causation. If a person's past behavior continually shapes the opportunities shown to the present person, questions arise about temporal identity and how much authority one's past should have over one's future.

Nothing in the investigation supports claims that recommendation algorithms are conscious, possess mystical agency, reveal hidden spiritual essence, or fulfill ancient prophecy. Symbolic and contemplative readings remain interpretive lenses.

Contemporary Relevance and Broader Social Consequences

The topic matters because recommendation is increasingly infrastructure for discovery. The key social question is no longer only who can publish, but who gets found. Computational ranking allocates visibility among creators, communities, news sources, products, ideas, and people.

This reframes several practical issues. A personalized feed should never be mistaken for a representative sample of society. 'My feed shows everyone thinks this' is a category error. For individuals, a more useful question is whether the content repeatedly served corresponds to what they deliberately want more of in their lives. For education, algorithmic literacy should include the principle that a feed is a constructed environment, not the world itself.

The blind-spot review widened the frame further. Recommendation can affect the people being recommended by distributing followers, professional opportunities, dating visibility, social prestige, and creator income. It can affect nonusers when platform-selected content escapes into journalism, elections, commerce, and everyday conversation.

Language and moderation are also upstream causal layers. If a system poorly understands a low-resource language, its ranking and moderation may be structurally different before personalization even begins. Advertising and paid promotion create additional ranking systems layered over organic recommendation. Bots, coordinated campaigns, and synthetic engagement can contaminate the behavioral signals platforms treat as evidence of demand.

A serious minority interpretation deserves more attention: recommendation can be emancipatory infrastructure. It can reduce search costs for rare-interest communities, minority-language material, educational resources, niche creators, and support networks. The problem may not be recommendation itself but centralized control over its objectives. User-governed, plural, or federated recommendation is therefore a legitimate research direction rather than a fringe idea.

Findings by Evidentiary Status

The investigation's final claims can be organized by strength rather than narrative appeal.

Established strongly

• User behavior is a material input into modern recommendation systems.

• Ranking changes exposure and the distribution of attention.

• Ranking changes engagement and platform behavior.

• Exposure is not equivalent to persuasion.

• Changing exposure substantially can leave deeper political attitudes unchanged over periods of months.

• Engagement is not a transparent synonym for reflective preference or welfare.

Established in important but bounded contexts

• Algorithmic systems can amplify political content asymmetrically.

• Engagement-based ranking can amplify negative emotionality and out-group hostility relative to alternative ranking policies.

• Some recommendation systems can causally change some political attitudes.

• Recommendation can causally change following behavior and thereby alter the user's future information environment.

Probable

• Recommendation usually interacts with prior dispositions rather than simply replacing them.

• Amplification is a more general explanation than wholesale preference creation.

• Network formation is an important candidate mechanism for persistent effects.

• Users and creators adapt strategically to platform incentives.

• Attention allocation may be a larger societal mechanism than direct persuasion.

• Recommendation contributes to differential cultural visibility and selection.

Plausible but uncertain

• Repeated recommendation contributes to durable preference formation in some domains.

• Early discovery can produce path-dependent interests or social trajectories.

• Adolescents may show different long-term susceptibility from adults.

• Creator adaptation may reshape cultural supply at population scale.

• Algorithmic personalization may produce simultaneous individual specialization and collective concentration.

Speculative

• Recommendation systems routinely reshape durable identity at population scale.

• They alter stable personality traits.

• They routinely radicalize politically moderate users.

• They will produce a predictable twenty-year trajectory of human psychology or culture.

• Current algorithmic systems are conscious or metaphysically agentic.

Highest-Value Unresolved Research

The project's most important unanswered question is not whether feedback exists; it is whether repeated feedback accumulates into durable divergence. Several research programs would materially change confidence:

1. Multi-year new-user cohorts with strong pre-personalization baselines, randomized to meaningfully different recommendation environments.

2. Direct randomization of exploration: low, medium, and high exposure to novel but plausible interests, followed by long-term observation of whether interests persist.

3. Experiments separating passive content exposure from the ability to follow, join communities, or form social ties.

4. Long-duration removal or reset studies to test reversibility and distinguish reflection from path dependence.

5. Cross-platform longitudinal studies that identify which system first introduced an interest and which systems merely reinforced it.

6. Creator-side natural experiments around major ranking changes, tracking topic, format, effort, emotional tone, reach, income, and career decisions.

7. Cross-cultural and low-resource-language studies that test whether the same behavioral profile produces different recommendation environments across languages and institutions.

8. Household, friendship-network, and nonuser spillover studies to capture social effects missed by individual-level experiments.

9. Welfare-oriented measures asking whether users were glad they received a recommendation, whether it advanced valued goals, and what it displaced.

10. Research on the transition from recommendation to AI delegation, where systems may increasingly select, summarize, generate, transact, and act on users' behalf.

Principal Sources and Frameworks

Foundational recommender-systems and feedback-loop research

Resnick, Paul, Neophytos Iacovou, Mitesh Suchak, Peter Bergstrom, and John Riedl. “GroupLens: An Open Architecture for Collaborative Filtering of Netnews.” ACM CSCW, 1994. DOI: 10.1145/192844.192905.

Chaney, Allison J. B., Brandon M. Stewart, and Barbara E. Engelhardt. “How Algorithmic Confounding in Recommendation Systems Increases Homogeneity and Decreases Utility.” RecSys, 2018. DOI: 10.1145/3240323.3240370.

Jiang et al. “Degenerate Feedback Loops in Recommender Systems.” 2019. DOI: 10.1145/3306618.3314288.

Political and causal field experiments

Huszár et al. “Algorithmic Amplification of Politics on Twitter.” Proceedings of the National Academy of Sciences, 2022. DOI: 10.1073/pnas.2025334119.

Guess et al. “How Do Social Media Feed Algorithms Affect Attitudes and Behavior in an Election Campaign?” Science, 2023. DOI: 10.1126/science.abp9364.

Nyhan et al. “Like-Minded Sources on Facebook Are Prevalent but Not Polarizing.” Nature, 2023. DOI: 10.1038/s41586-023-06297-w.

Guess et al. “Reshares on Social Media Amplify Political News but Do Not Detectably Affect Beliefs or Opinions.” Science, 2023. DOI: 10.1126/science.add8424.

Gauthier et al. “The Political Effects of X’s Feed Algorithm.” Nature, 2026. DOI: 10.1038/s41586-026-10098-2.

Engagement, preference, and problematic-content research

Milli et al. “Engagement, User Satisfaction, and the Amplification of Divisive Content on Social Media.” PNAS Nexus, 2025. DOI: 10.1093/pnasnexus/pgaf062.

Haroon et al. “Auditing YouTube’s Recommendation System for Ideologically Congenial, Extreme, and Problematic Recommendations.” Proceedings of the National Academy of Sciences, 2023. DOI: 10.1073/pnas.2213020120.

Hosseinmardi et al. “Examining the Consumption of Radical Content on YouTube.” Proceedings of the National Academy of Sciences, 2021. DOI: 10.1073/pnas.2101967118.

Hosseinmardi et al. “Causally Estimating the Effect of YouTube’s Recommender System Using Counterfactual Bots.” Proceedings of the National Academy of Sciences, 2024. DOI: 10.1073/pnas.2313377121.

User agency, identity, and human-computer interaction

Karizat, Nadia, Alejandro Delmonaco, Motahhare Eslami, and Nazanin Andalibi. “Algorithmic Folk Theories and Identity: How TikTok Users Co-Produce Knowledge of Identity and Engage in Algorithmic Resistance.” Proceedings of the ACM on Human-Computer Interaction, 2021. DOI: 10.1145/3476046.

Platform descriptions and institutional material

TikTok Newsroom. “How TikTok Recommends Videos #ForYou.” June 18, 2020.

YouTube / Google Help materials on how YouTube recommendations work and recommendation signals.

Meta / Facebook materials on News Feed ranking and the 2018 shift toward “meaningful interactions.”

European Union Digital Services Act materials on recommender transparency, systemic risk, researcher access, and non-profiling recommendation options for very large platforms and search engines.

European Commission information requests and investigations concerning recommender-system risks, including 2024 requests to YouTube, Snapchat, and TikTok and later enforcement developments discussed in the investigation.

Cross-disciplinary and interpretive frameworks

Network science and social-network economics: homophily, peer influence, diffusion, and endogenous network formation.

Psychology: reinforcement learning, attention, salience, habit, emotional response, social reward, and motivated reasoning.

Sociology and HCI: identity performance, norms, communities, algorithmic folk theories, and strategic user adaptation.

Media studies: agenda setting, gatekeeping, selective exposure, and the historical continuity of mediated information.

Economics and political economy: search costs, multi-sided platforms, creator incentives, advertising, market power, and welfare.

Philosophy of attention, autonomy, identity, and welfare; the investigation specifically discussed contemporary work treating recommender systems as allocators of scarce attention.

Carl Jung’s analytical-psychology concepts of persona, projection, shadow, and symbolic self-representation, used interpretively rather than as empirical evidence.

Buddhist dependent-origination and contemplative frameworks concerning contact, feeling, craving, conditioning, and awareness, used as comparative interpretation rather than historical prediction of algorithmic systems.

Norbert Wiener’s cybernetic framework of feedback and control, relevant to recursive systems.

Conclusion

The investigation began with a question that seemed to require a verdict: do recommendation algorithms learn what people want, or teach them what to want? The evidence does not permit a clean victory for either side because the premise is too simple.

Humans enter platforms with prior interests, values, habits, identities, social networks, moods, and vulnerabilities. Their behavior materially trains recommendation systems. Those systems then alter the opportunities presented back to them: which content is seen, which creator is discovered, which topic becomes salient, which community becomes reachable. Users respond, resist, follow, search, leave, or create. The resulting behavior becomes the next generation of data.

That reciprocal behavioral feedback is real. Its deeper psychological consequences are conditional and uneven. Strong experiments show that large changes in exposure can leave established attitudes stable. Other experiments show that algorithmic feeds can move some opinions and construct different follow networks. YouTube research shows that problematic pathways can increase without a general march toward extremity. These are not contradictions to be resolved by choosing one favorite study; they are evidence that platform architecture, context, prior disposition, and mechanism matter.

The investigation's most important shift is therefore away from a narrow psychology of persuasion. Recommendation systems may exercise their largest influence by allocating attention, visibility, discovery, and connection. A system can matter enormously without 'brainwashing' anyone if it helps determine which people meet, which creators survive, which ideas become visible, and which options enter the field of consideration.

At the same time, the research does not justify technological fatalism. Users actively shape and sometimes resist feeds. Prior preferences remain powerful. Recommendation can expose people to knowledge, rare-interest communities, support networks, and opportunities they might otherwise never encounter. The central normative issue is not whether personalization is inherently good or bad, but who controls the objectives, what proxies stand in for human preference, how much meaningful user choice exists, and how effects are measured over time.

The deepest empirical unknown is cumulative persistence. The field knows much more about what algorithms do to exposure and engagement than about what years of repeated adaptation do to lives. To resolve that question will require longitudinal, cross-platform, cross-cultural research that measures not only attitudes but attention, relationships, skills, careers, communities, and offline behavior.

Recommendation systems do not simply reflect human preference, and current evidence does not show that they routinely manufacture the human self. They are adaptive systems that help structure the environment in which preferences, choices, relationships, and culture continue to develop. The unresolved question is how often those environmental interventions become persistent enough to redirect human trajectories.

That is a narrower conclusion than the original hypothesis. It is also more important, because it directs attention toward the mechanisms the evidence can actually test - and toward the forms of influence that may matter even when beliefs remain unchanged.

Appendix: Compact Evidence Map