Humans and Recommendation Algorithms: The Full Investigation

HUMANS AND RECOMMENDATION ALGORITHMS

The Full Investigation into Attention, Preference, Identity, Social Networks, Culture, and Algorithmic Feedback

Complete research record

“Do recommendation algorithms mostly learn what people already want, or do they gradually teach people what to want?”

Full investigation archive. Research, rival explanations, evidence testing, corrections, and synthesis are preserved in sequence.

Research Landscape: Humans and Recommendation Algorithms as a Co-Evolving System

The central question is deceptively simple:

Do recommendation algorithms mostly learn what people already want, or do they gradually teach people what to want?

The strongest current research suggests that the answer is both, but not in equal measure across all contexts, people, platforms, or outcomes. Users supply the behavioral data from which recommendation systems learn; platforms rank and amplify selected material back to those users; users then respond to the resulting environment, generating the next round of training signals. This creates a recurrent human–algorithm feedback loop.

There is strong evidence that recommendation algorithms can alter exposure, attention, engagement, following behavior, and patterns of interaction. There is credible but more variable evidence that they can influence attitudes, preferences, emotional states, and social identification. Evidence that they routinely manufacture deep ideological commitments or stable identities from scratch is considerably weaker.

That distinction will be essential if we later ask where this system could take human society by 2045–2046.

I. Overview

1. What exactly are we studying?

The object of investigation is not simply "social media algorithms."

It is better defined as a dynamic sociotechnical system involving:

people → behavior → data → recommendation model → ranked environment → new behavior → updated data → revised model

The process can be expressed conceptually as:

That last arrow—whether exposure actually alters future preferences—is the controversial one.

The system contains at least three different phenomena that are often mistakenly collapsed together.

Personalization

The system predicts what an individual is likely to engage with.

Example:

A person repeatedly watches woodworking videos, and YouTube recommends more woodworking.

This may be primarily preference discovery rather than preference creation.

Amplification

Some types of material receive disproportionately greater visibility because they score highly on whatever objectives the platform is optimizing.

For example, Twitter research has demonstrated measurable algorithmic differences in the amplification of political content. (DOI.org)

Preference formation

Repeated exposure changes what the user subsequently values, believes, seeks, or identifies with.

This is the most consequential mechanism—and the hardest to prove.

2. Why the distinction matters

Suppose someone watches increasingly extreme political videos.

Several causal stories are possible.

The algorithm might have radicalized the person.

But the person might already have been radicalizing and the algorithm merely detected the change.

Or both might be occurring simultaneously.

Researchers therefore face a classic endogeneity problem:

The algorithm chooses content partly because of a user's previous behavior, while the user's future behavior may partly reflect what the algorithm chose.

This makes causal inference unusually difficult.

Recommendation systems are therefore not merely machines acting upon humans.

Humans are part of the machine's input.

3. Necessary boundaries

This investigation should primarily concern algorithmically ranked feeds and recommendation systems, including platforms such as:

• TikTok
• Instagram
• Facebook
• YouTube
• X/Twitter
• Reddit
• emerging AI-mediated social environments

Search engines, Amazon recommendations, Spotify, Netflix, dating apps, and news aggregators are highly relevant comparative cases but should remain secondary unless a later branch specifically investigates them.

The deeper question is about social behavior and identity formation, not merely recommendation accuracy.

II. Core Elements

1. The user as trainer

A common misconception portrays the algorithm as watching passive users.

In reality, users continually train recommendation systems.

Signals may include:

• clicking
• pausing
• scrolling
• replaying
• liking
• disliking
• commenting
• sharing
• following
• subscribing
• searching
• abandoning content
• watching to completion
• returning later
• creating similar content

TikTok publicly describes interactions such as watches, likes, shares, follows, comments, and content creation as recommendation signals. It has also said that completing a longer video can be weighted more heavily than relatively weak contextual signals such as geographic coincidence. (TikTok Newsroom)

YouTube similarly identifies watch history, search history, subscriptions, likes, dislikes, explicit negative feedback, and satisfaction surveys among its recommendation signals. (Google Help)

Documented fact: user behavior materially shapes recommendations.

So the algorithmic feed is partly a behavioral reflection.

But it is not a mirror.

It is a selective prediction machine.

2. The algorithm as curator

The recommender must decide which possible item gets shown.

Facebook has described feed ranking in terms of:

inventory → signals → predictions → ranking score.

The system predicts such things as whether a user will comment, share, react, or otherwise engage. (About Facebook)

Thus the system doesn't simply ask:

What does Jeff like?

It effectively asks something closer to:

Among everything we could show Jeff right now, which item best satisfies the ranking objective?

Those are very different questions.

The objective function matters enormously.

An algorithm optimized for:

watch time

will produce a different media environment from one optimized for:

user satisfaction

which will differ again from one optimized for:

meaningful social interaction

or:

diversity of viewpoint.

Facebook explicitly altered News Feed ranking in 2018 to prioritize posts predicted to generate "meaningful interactions" and conversations. (About Facebook)

This demonstrates an important principle:

Platform values are partly encoded into the information environment through ranking objectives.

3. Revealed preference versus reflective preference

One of the most important emerging distinctions is between:

what people engage with

and

what people say they actually want.

These are not necessarily identical.

A 2025 preregistered audit of Twitter's engagement-based ranking found that the ranked feed amplified more partisan, emotionally negative, and out-group-hostile political material than a reverse-chronological baseline. Importantly, users did not necessarily report preferring the material that engagement-based ranking selected. (PubMed Central (PMC))

This creates an important conceptual problem.

Suppose someone angrily reads something for thirty seconds.

The algorithm may interpret that as:

valuable engagement.

The person may experience it as:

irritation I wish I hadn't seen.

So behavioral data may contain a systematic error:

attention ≠ preference.

This may be one of the most important mechanisms in the entire subject.

4. Feedback loops

The fundamental technical concern has been studied explicitly in recommender-system research.

Chaney, Stewart, and Engelhardt demonstrated in simulations that recommendation systems trained on behavior generated after previous recommendations can produce algorithmic confounding—where observed preferences become increasingly entangled with prior recommendations. Their simulations produced greater behavioral homogeneity without corresponding gains in user utility. (arXiv)

Jiang and colleagues similarly analyzed what they termed degenerate feedback loops, where recommender systems and changing user preferences interact dynamically over time. (arXiv)

The theoretical danger is:

1. an algorithm recommends A,
2. A becomes disproportionately consumed,
3. consumption of A becomes evidence that people prefer A,
4. the system recommends still more A.

The result can resemble preference formation even if the initial preference difference was tiny.

5. Popularity cascades

There is also a collective dimension.

Suppose millions of users independently show small preferences for something.

Recommendation systems may transform those microscopic differences into large visibility disparities.

More visibility produces:

more consumption → more engagement → more visibility.

This creates rich-get-richer effects.

Recent feedback-loop research continues to find that recommendation systems can redistribute collective demand and sometimes increase popularity concentration even while an individual's apparent menu of consumption may initially look diverse. (arXiv)

Thus algorithmic personalization can paradoxically produce:

individual customization

alongside

collective homogenization.

That possibility deserves substantial investigation.

III. User Agency

The user is not simply manipulated.

Users learn.

They experiment with algorithms.

They manipulate them.

They resist them.

They consciously train them.

TikTok culture contains explicit examples:

"I need to fix my FYP."

Users intentionally like, ignore, search, follow, block, or watch particular content because they believe doing so will reshape the system.

Karizat, Delmonaco, Eslami, and Andalibi's 2021 TikTok research describes what scholars call algorithmic folk theories—informal theories users develop about how algorithms work. Participants sometimes intentionally altered behavior in an effort to modify how the algorithm represented their interests or identities. (Icahn School of Medicine at Mount Sinai)

This creates an extraordinary recursive situation.

The algorithm models the user.

The user models the algorithm.

Then:

the algorithm modifies its behavior because of the user, while the user modifies behavior because of their beliefs about the algorithm.

That is genuine co-adaptation.

IV. Creator Adaptation

There is another participant in the system:

the creator.

Creators learn which behaviors algorithms reward.

They alter:

• thumbnails
• titles
• video duration
• opening seconds
• emotional intensity
• posting frequency
• language
• topics
• editing style
• controversy level

YouTube's own creator guidance explicitly encourages creators to examine viewer responses such as whether users choose a video, continue watching it, and express satisfaction, because those signals affect recommendation performance. (Google Help)

Thus creators evolve in response to algorithmic incentives.

The complete feedback system is therefore closer to:

users → algorithm → creators → content → users → algorithm

This creates cultural selection pressure.

Some scholars are beginning to model this explicitly as an evolutionary platform–creator dynamic. (Sage Journals)

V. Historical and Intellectual Background

1. Before algorithmic social media

Information selection has always existed.

Human communities relied upon:

• family
• clergy
• teachers
• publishers
• journalists
• editors
• broadcasters
• librarians
• political organizations
• peers

These institutions filtered information.

So mediated reality is not new.

What changes with algorithmic recommendation is the combination of:

personalization + scale + automation + behavioral measurement + rapid feedback.

2. Collaborative filtering

A major technical ancestor appeared in the early 1990s.

The GroupLens project demonstrated collaborative filtering of Usenet articles, using people's evaluations of material to predict what others with similar preferences might value. The foundational GroupLens work was published at ACM CSCW in 1994. (DOI.org)

The idea was elegant:

People like you liked this, so perhaps you will too.

This principle later became foundational across digital recommendation systems.

3. Social feeds become ranked

Early social platforms often presented information largely chronologically.

As networks expanded, users generated more material than anyone could realistically consume.

Platforms therefore moved toward ranking systems.

Facebook eventually formalized systems that predicted which stories would generate engagement and interpersonal interaction. (About Facebook)

The decisive shift was conceptual:

old model

User selects sources.

new model

Platform selects items from available sources.

TikTok pushed this further.

The For You feed made algorithmic discovery—not merely one's explicitly chosen social network—the central interface of the platform. TikTok described the For You feed in 2020 as individually tailored from user interactions, video attributes, and contextual information. (TikTok Newsroom)

This is historically significant.

The dominant organizing principle shifted from:

"people I follow"

toward:

"content predicted to hold my attention."

VI. Political and Ideological Effects

This is perhaps the most intensively studied—and most frequently overstated—part of the field.

Evidence for amplification

A major Twitter experiment examined millions of users comparing algorithmic and chronological feeds.

Researchers found systematic political amplification differences, including greater algorithmic amplification of mainstream-right political content in six of seven countries studied. Importantly, they did not find evidence that extreme political groups were universally favored over moderate ones. (DOI.org)

So:

Documented: algorithms can alter political exposure.

But exposure is not identical to persuasion.

VII. The Facebook Experiments: An Important Corrective

Large experiments associated with the 2020 U.S. election are especially valuable because they tested causal effects rather than simply observing correlations.

Users were randomly moved from Facebook and Instagram's algorithmically ranked feeds to reverse chronological ones.

The change substantially affected behavior:

• users spent less time on the platforms,
• activity declined,
• exposure patterns changed.

Yet researchers found no significant changes in several major political-attitude outcomes during the study period, including affective polarization and issue polarization. (PubMed)

Another experiment reduced exposure to politically like-minded sources by about one-third among more than 23,000 Facebook participants.

Exposure changed substantially.

Political attitudes did not measurably change across eight preregistered attitudinal outcomes. (Nature)

Another experiment removed reshared content.

Again:

information exposure changed substantially.

Political opinions did not detectably change. (PubMed)

These findings provide a powerful warning against simplistic claims such as:

"The algorithm shows partisan material, therefore it creates polarization."

That inference is not presently justified.

VIII. But the X Experiment Complicates the Picture

A particularly important 2026 Nature study provides counterevidence.

Researchers randomly assigned active U.S. X users to algorithmic or chronological feeds for seven weeks.

Switching users toward the algorithmic feed increased engagement and shifted some political attitudes in a more conservative direction. Researchers found changes concerning policy priorities, perceptions surrounding Donald Trump's legal investigations, and attitudes toward the war in Ukraine.

They also found that algorithmic exposure encouraged users to follow conservative political activist accounts, and those follows persisted even after exposure conditions changed. (PubMed)

This result is important for two reasons.

First:

algorithmic exposure can sometimes produce measurable political effects.

Second:

the effect may become partly embedded in a user's network.

That produces a fascinating mechanism:

In other words, an algorithm may influence behavior temporarily but leave behind a structural change that persists.

That may be one of the most consequential mechanisms discovered so far.

IX. YouTube and the "Rabbit Hole" Debate

YouTube became one of the most discussed examples of alleged algorithmic radicalization.

Early public narratives often proposed:

moderate video → stronger video → extremist video.

The evidence is considerably more complicated.

A 2023 PNAS study using approximately 100,000 automated accounts found that YouTube recommendations became ideologically congenial for partisan users and that problematic channels became somewhat more prevalent deeper in some recommendation trajectories.

However, the researchers did not find a general progression toward increasing ideological extremity. (PubMed Central (PMC))

A broader systematic review of 23 studies found:

• 14 studies implicated YouTube recommendations in pathways toward problematic content,
• 7 produced mixed evidence,
• 2 did not.

But the reviewers warned that researchers lacked full knowledge of YouTube's internal recommendation architecture, limiting causal certainty. (PubMed Central (PMC))

Therefore:

Probable interpretation: recommendation systems sometimes facilitate pathways toward problematic material.

Not established: YouTube routinely converts politically moderate users into extremists.

X. Identity Formation

The identity question may ultimately be more important than political persuasion.

Social media increasingly supplies users with categories through which they understand themselves.

Examples include communities organized around:

• fashion
• neurodiversity
• sexuality
• religion
• fitness
• mental health
• political identity
• gender
• relationship styles
• hobbies
• illness
• lifestyle
• aesthetic subcultures

A recommendation system might repeatedly expose someone to one category because of weak initial signals.

The person may then:

watch → recognize themselves → join a community → adopt terminology → alter self-description → produce matching content.

At that point, distinguishing discovery from construction becomes difficult.

Karizat and colleagues' TikTok research is important here because participants described attempts to bring their "algorithmic identity" into alignment with their self-understanding and expressed concern that algorithms might misrepresent or suppress aspects of identity. (Icahn School of Medicine at Mount Sinai)

This suggests a new social phenomenon:

People may increasingly encounter themselves partly through machine-generated representations of who they appear to be.

XI. The Algorithmic Mirror Hypothesis

One interpretation views recommendation systems primarily as mirrors.

The user already possesses:

interests → preferences → personality → beliefs.

Algorithms merely infer them.

Under this model:

This interpretation is supported by the enormous role behavioral signals play in recommendation.

YouTube even frames its system in terms of following audience preferences rather than dictating them. (Google Help)

But the mirror metaphor is incomplete.

XII. The Algorithmic Prism Hypothesis

A better metaphor may be a prism.

A prism receives incoming light but changes how that light is distributed.

Similarly:

A slight preference can become disproportionately visible.

This produces:

amplification without creation.

Someone mildly interested in bodybuilding may suddenly inhabit a media environment saturated with bodybuilding.

That environment could subsequently influence identity.

XIII. The Algorithmic Sculptor Hypothesis

A stronger interpretation proposes that repeated personalized exposure actually alters preferences.

Under this model:

This is plausible and supported in certain contexts, but its general magnitude remains uncertain.

Technical simulation research demonstrates that such preference feedback is possible. (Brandon Stewart)

Real-world causal evidence exists but is mixed.

The X experiment provides unusually strong evidence that feed selection can affect some attitudes and subsequent following behavior. (PubMed)

But the Meta experiments demonstrate that substantial exposure changes do not necessarily translate into measurable ideological change. (Nature)

XIV. Social Interaction

Recommendation systems influence not just what people know, but increasingly who finds whom.

This matters because relationships themselves change identity.

Algorithms can help users discover:

• political communities
• medical-support groups
• hobby communities
• fandoms
• conspiracy communities
• religious movements
• social movements
• professional networks

The algorithm may therefore influence identity indirectly:

This could be substantially more powerful than direct persuasion.

Humans are profoundly influenced by communities.

An algorithm may not need to convince someone of an idea.

It only needs to introduce them to people who eventually do.

XV. Emotional Selection

Another major branch concerns emotional content.

Engagement systems may unintentionally reward material that causes:

• anger
• fear
• outrage
• moral condemnation
• tribal solidarity
• excitement
• surprise

The 2025 Twitter audit found that engagement-based ranking amplified negative emotions and partisan hostility relative to chronological ranking. Anger was especially amplified in political material. (PubMed Central (PMC))

This does not prove platforms deliberately optimize for anger.

Rather:

reasonable inference: if anger systematically produces measurable engagement, optimization for engagement may indirectly select anger-producing material.

That distinction is important.

XVI. Mental Health and Wellbeing

Research connecting social media and mental health remains much more complicated than public discourse often implies.

A 2024 systematic review and meta-analysis found that general social-media use showed relatively weak associations with depression and anxiety, while problematic social-media use showed stronger associations with depression, anxiety, poorer wellbeing, and sleep problems. (ScienceDirect)

This cautions against statements such as:

"Social media causes depression."

The relationship is heterogeneous and may be bidirectional.

However, researchers have increasingly shifted toward examining specific design mechanisms rather than generic "screen time."

A 2026 grounded-theory study specifically investigated algorithmic recommendation as a social and commercial determinant of adolescent mental health. (Springer)

Similarly, recent work has examined how recommendation systems may repeatedly expose young users to misogynistic or toxic ideological material. (Frontiers)

The research frontier is moving from:

How many hours were you online?

toward:

What system selected what you saw, why did it select it, and what sequence followed?

That is a much more sophisticated question.

XVII. Major Interpretations

The research landscape can currently be divided into roughly six schools.

1. Preference-Revelation Model

Algorithms mostly discover existing preferences.

Users remain the primary causal actors.

Recommendations largely reflect human demand.

Strongest evidence

Behavioral personalization works because user signals are predictive.

Weakness

Observed behavior is partly produced by previous recommendations.

Therefore "preference" and "exposure" cannot always be disentangled.

2. Amplification Model

Algorithms do not necessarily create preferences but magnify some existing ones.

This interpretation currently has substantial empirical support.

Examples include:

• political amplification on Twitter,
• congenial recommendation on YouTube,
• increased engagement under algorithmic feeds. (DOI.org)

This may presently be the safest general model.

3. Preference-Shaping Model

Algorithms alter preferences through repeated exposure.

Evidence exists, especially for some short-term political effects and behavioral changes.

But general claims of large-scale ideological transformation are not yet warranted.

4. Sociotechnical Co-Evolution Model

Users and systems continually adapt to one another.

This is conceptually powerful because it avoids asking which party is "really" responsible.

Instead:

people influence algorithms while algorithms modify the environment in which people's preferences develop.

Feedback-loop research strongly supports the plausibility of this model. (arXiv)

5. Political-Economy Model

This interpretation emphasizes platform incentives.

Recommendation systems are not optimized in a social vacuum.

Platforms may value:

• engagement
• retention
• advertising opportunities
• subscription conversion
• growth
• creator participation

Therefore human behavior becomes partially organized around commercial objectives.

This perspective overlaps with arguments associated with scholars such as Shoshana Zuboff, José van Dijck, Tarleton Gillespie, Taina Bucher, Safiya Noble, and others studying platform power and algorithmic governance.

The strongest form of this argument—claiming platforms intentionally manipulate society in a unified direction—is generally harder to substantiate.

The weaker version is much better supported:

Commercial optimization criteria influence the architecture through which information and social interaction are distributed.

6. User-Agency Model

Users learn platform mechanics and strategically influence them.

Creators optimize content.

Users manipulate feeds.

Communities coordinate engagement.

Governments and campaigns attempt algorithmic influence.

Bots and coordinated networks may exploit ranking systems.

The ecosystem therefore contains many competing agents, not one all-powerful algorithm.

XVIII. Regulation Enters the Feedback Loop

Governments are increasingly treating recommender systems themselves—not merely individual posts—as sources of systemic risk.

The European Union's Digital Services Act has created obligations concerning recommendation transparency, systemic-risk assessment, researcher access, and, for very large platforms, options that are not based on profiling in certain circumstances. (Digital Strategy)

In October 2024 the European Commission specifically requested information from TikTok, YouTube, and Snapchat about recommender-system risks concerning mental wellbeing, elections, civic discourse, illegal content, and potential "rabbit holes." (Digital Strategy)

By 2026 the Commission had intensified attention to addictive design and personalized recommendation systems; in July 2026 it preliminarily found Meta in breach of the DSA over addictive-design concerns involving features including infinite scroll, autoplay, notifications, and highly personalized recommendations. This remains a regulatory finding rather than scientific proof of a particular causal mental-health mechanism, but it demonstrates how seriously governments now view the architecture itself. (Digital Strategy)

XIX. What Might the Next Twenty Years Look Like?

This section must be treated as scenario analysis, not prediction.

Current evidence does not justify claiming that any particular future is inevitable.

But several trajectories deserve investigation.

Scenario A: Hyperpersonalized individuality

Algorithms become extraordinarily effective at discovering individual preferences.

People inhabit increasingly customized media ecosystems.

Paradoxically this might produce greater diversity between individuals:

one person inhabits medieval history TikTok;

another synthetic biology;

another obscure jazz;

another woodworking.

Digital culture fragments into millions of microcultures.

Scenario B: Collective cultural convergence

Feedback loops might instead produce increasing concentration.

A small number of formats, behaviors, creators, and aesthetics dominate because algorithms repeatedly reward content already demonstrating strong performance.

Humans adapt themselves to machine-recognizable success patterns.

Culture becomes algorithmically optimized.

We may already see weak versions of this in:

• thumbnail conventions
• influencer speech patterns
• short-video pacing
• reaction formats
• viral audio
• title structures
• standardized aesthetics

The long-term possibility is striking:

Humans might increasingly shape creative expression around what machine-ranking systems can recognize and reward.

Scenario C: Algorithmic identity formation

Young people may increasingly encounter potential identities through recommendation systems.

Someone's developmental sequence might become:

Algorithms would not dictate identity.

But they might increasingly mediate the menu of identities encountered.

That distinction could become sociologically profound.

Scenario D: Behavioral self-optimization for algorithms

People may increasingly act with algorithmic audiences in mind.

We already see:

• creators optimizing for ranking,
• professionals optimizing LinkedIn presence,
• daters optimizing profiles,
• businesses optimizing review platforms,
• musicians optimizing streaming discovery.

Over twenty years, algorithmic legibility may become a major social skill.

People might routinely ask:

How does the system perceive me?

That could alter self-presentation itself.

Scenario E: AI agents join the loop

The next development may be substantially larger than social-media recommendation.

Generative AI agents may increasingly:

• select information,
• summarize news,
• recommend friends,
• choose entertainment,
• recommend purchases,
• help form opinions,
• mediate communication,
• generate social content.

The feedback system could become:

This would introduce a new actor between humans and information environments.

Its consequences remain largely unexplored.

XX. Open Questions

The most important unresolved questions are not merely technical.

They concern human nature.

1. Do recommendation systems discover preferences or manufacture them?

Probably both.

We do not yet know the relative magnitude.

2. How long must exposure persist before durable preference change occurs?

Many experimental studies last weeks or months.

Human identity develops over years.

This creates a major evidence gap.

3. Are adolescents uniquely susceptible?

This seems plausible because identity formation is developmentally active, but high-quality causal longitudinal evidence remains limited.

4. Which preferences are most malleable?

Possible differences may exist among:

politics, aesthetics, sexuality, consumption, hobbies, diet, religion, relationships, health behavior, and moral attitudes.

They should not be assumed equivalent.

5. Does algorithmic personalization increase or decrease human diversity?

Current theoretical work suggests both outcomes can occur depending on level of analysis.

Individual consumption may diversify while collective attention becomes concentrated.

6. Can algorithms create entirely new social categories?

Potentially.

Online communities already generate new identity labels, subcultures, political categories, and aesthetic movements.

The question is whether recommendation systems merely accelerate cultural evolution or fundamentally change its structure.

7. Is engagement optimization psychologically distorting?

The gap between revealed engagement and stated preference may be critical. (OUP Academic)

If people systematically engage with things they do not actually value, recommendation systems may be optimizing toward behavioral vulnerabilities rather than wellbeing.

8. Does algorithmic influence persist after the algorithm disappears?

The 2026 X study suggests one mechanism whereby algorithmic exposure alters whom people follow, potentially creating lasting network effects. (PubMed)

This deserves extensive investigation.

9. Will people become more algorithmically literate?

Users already develop folk theories and deliberately manipulate feeds. (Icahn School of Medicine at Mount Sinai)

Future generations may treat recommender systems as environments to be actively engineered rather than passively consumed.

10. Can algorithms be designed for human flourishing rather than engagement?

This may eventually become the central engineering question.

Possible optimization objectives might include:

• satisfaction
• learning
• viewpoint diversity
• wellbeing
• meaningful relationships
• novelty
• long-term regret minimization
• autonomy

But these raise difficult normative questions:

Who defines flourishing?

XXI. Initial Research Map

The present landscape points toward eight major branches that deserve deeper investigation.

Branch 1 — The causality problem

Determine how much recommendation systems actually alter underlying preferences versus revealing them.

This is foundational.

Branch 2 — Long-term identity formation

Examine whether algorithmic environments influence personality, political identity, religion, lifestyle, sexuality, aesthetics, and social belonging across years rather than weeks.

This may be the most consequential human question.

Branch 3 — Childhood and adolescence

Investigate whether developmental-stage exposure produces effects that differ fundamentally from adult exposure.

Branch 4 — Engagement versus true preference

Study whether what people click, watch, and share systematically differs from what they consciously value.

If so, much of modern recommendation may be optimizing a misleading target.

Branch 5 — Collective cultural evolution

Examine whether recommender systems produce cultural homogenization, fragmentation, polarization, or simultaneous combinations of all three.

Branch 6 — Algorithmic social-network formation

Investigate the indirect mechanism:

recommendation → follow → community → social influence → identity.

The X experiment makes this branch especially promising. (PubMed)

Branch 7 — Creator evolution

Examine how creators alter speech, aesthetics, emotional tone, political content, and cultural production to satisfy algorithmic selection pressures.

This is essentially the evolutionary ecology of algorithmic culture.

Branch 8 — AI-mediated society after recommender feeds

Investigate what changes when recommendation algorithms are joined by personalized generative agents capable not merely of selecting information but of interpreting, creating, persuading, conversing, and acting.

This is likely where the twenty-year horizon becomes most consequential.

Provisional Synthesis

The present evidence does not support either extreme version of the debate.

The claim that humans are passive subjects manipulated wholesale by algorithms is too strong.

The claim that algorithms merely reflect what people independently want is also increasingly untenable.

A more defensible model is:

The strongest evidence currently concerns exposure, engagement, amplification, network formation, and behavioral adaptation.

Evidence becomes progressively less certain as we move toward:

preference change → identity change → personality change → civilization-scale behavioral transformation.

Yet those weaker-evidence areas may ultimately be the most important.

The deeper possibility emerging from this research landscape is therefore not simply that recommendation algorithms "control people."

It is something subtler.

For most of human history, culture evolved primarily through interactions among people, institutions, environments, and inherited traditions.

We may now be entering a period in which adaptive machine systems become persistent participants in cultural evolution itself.

Whether those systems ultimately make humanity more diverse or more homogeneous, more autonomous or more predictable, more connected or more fragmented remains genuinely unresolved.

And that is where the investigation becomes most interesting.

Deeper Investigation: Human–Algorithm Co-Evolution on Social Media

A deeper pass changes the shape of the problem.

The most useful question is no longer simply:

Do users shape algorithms more, or do algorithms shape users more?

That formulation assumes two separable actors exerting influence on one another. In practice, the system is closer to a recursive social environment in which preferences, behaviors, ranking systems, creators, peer groups, business incentives, and institutional rules continually modify one another.

A more complete model is:

The arrows do not have equal evidentiary support.

We have relatively strong evidence for the early parts of this chain—behavior affects ranking; ranking affects exposure; exposure affects engagement. We have substantially weaker evidence for the distant parts—long-term personality, durable identity, moral values, or civilization-scale behavioral change.

That distinction becomes even more important after examining the literature at finer resolution.

I. The First Hidden Problem: What Counts as a "Preference"?

The previous investigation used the word preference repeatedly, but that term conceals several different phenomena.

A recommender may infer that you "prefer" something because you:

• watched it,
• paused on it,
• replayed it,
• clicked it,
• commented angrily,
• shared it to criticize it,
• searched for related material,
• followed its creator,
• rated it positively,
• explicitly said you wanted more of it,
• or continued returning to the platform after seeing it.

Those signals are psychologically different.

Yet they can all enter recommendation systems as evidence of interest.

This means that algorithms often do not observe preferences directly.

They observe behavioral traces.

That gives us at least four categories.

1. Revealed behavioral preference

"What did the person do?"

Examples:

• watch,
• click,
• linger,
• share.

2. Stated preference

"What does the person say they want?"

3. Reflective preference

"What would the person prefer after deliberation rather than instantaneous reaction?"

4. Welfare preference

"What actually improves the person's long-term wellbeing?"

Those four can diverge radically.

A person might:

click something,

dislike having clicked it,

continue watching anyway,

and later say:

I wish the platform wouldn't show me this.

That is not a hypothetical problem.

A 2025 preregistered audit of Twitter/X found that engagement-based ranking amplified more partisan, angry, and out-group-hostile political posts than a chronological baseline. Critically, the political content selected by engagement ranking was not the content users said they most wanted to see. (DOI.org)

A newer 2026 PNAS study adds another layer. Among 715 U.S. X users, posts from accounts people themselves chose to follow tended to reflect their explicitly stated values, but the platform's amplification of posts showed an overall negative relationship with those values. The authors found that users' engagement behavior itself could diverge from their stated values, creating a pathway by which the algorithm learns an apparently distorted representation of them. (DOI.org)

This leads to a more precise possibility than "the algorithm manipulates users":

Algorithms may sometimes faithfully optimize the wrong representation of the user.

That is a fundamentally different failure mode.

II. The Algorithm May Be Learning Your Reflexes Rather Than Your Values

A recommender trained on engagement encounters a measurement problem.

Consider two hypothetical users.

User A

Watches:

• gardening,
• history,
• architecture.

Likes all three.

User B

Watches:

• gardening,
• history,
• rage-inducing political clips.

User B watches the political material longest precisely because it is upsetting.

An engagement model may infer:

Political conflict is one of User B's strongest preferences.

But the better interpretation could be:

Political conflict is unusually effective at capturing User B's involuntary attention.

This distinction resembles the difference between wanting and liking studied in psychology and behavioral economics.

A recommender can become excellent at predicting:

"What will cause another interaction?"

without being equally good at answering:

"What does this person endorse?"

or:

"What will this person be glad they consumed tomorrow?"

That suggests one of the most consequential design questions of the next twenty years:

What exactly should an artificial system optimize when human preferences are internally inconsistent?

There is no technical answer alone.

It is partly philosophical.

III. Users Shape Algorithms at More Than One Level

The earlier analysis treated "user input" mainly as clicks and watches.

That is incomplete.

Users influence recommendation systems through at least five layers.

Layer 1 — Immediate interaction

The obvious signals:

clicking, scrolling, watching, liking.

Layer 2 — Network construction

Users decide whom to follow, mute, block, subscribe to, friend, or unfollow.

That creates the pool from which many algorithms operate.

This matters because a ranking system cannot be blamed for every partisan exposure if the user has independently built a highly partisan network.

The 2023 Meta experiments demonstrated precisely this complexity. Reducing like-minded Facebook content changed exposure substantially, but people became more likely to engage with congenial content when they did encounter it. (Nature)

User preference therefore continued operating even after the platform altered ranking.

Layer 3 — Search and exploration

Search is often more intentional than passive feed consumption.

A person who searches:

"Why is inflation happening?"

is producing a stronger expression of interest than someone who merely pauses on an inflation video.

Different platforms assign different weights to such signals.

Layer 4 — Deliberate feed training

Users increasingly understand that their actions teach systems.

Some intentionally:

• avoid liking certain material,
• deliberately watch desired topics,
• use "not interested",
• create alternate accounts,
• clear histories,
• reset recommendation profiles.

Survey research from YouTube and Instagram users suggests that greater understanding of recommendation systems is associated with greater intention to manipulate those systems intentionally. (ScienceDirect)

Layer 5 — Content creation

Users become creators.

Their posts then train the platform at population scale.

This is where the feedback loop becomes societal rather than personal.

IV. Algorithms Do Not Merely Select Content; They Define the Opportunity Set

A subtle distinction was underdeveloped previously.

Suppose someone chooses video B over video A.

Was that a free preference?

Partly.

But what if the platform selected A and B from ten million possible videos?

The user selected within a menu the platform created.

This is a form of choice architecture.

The platform does not necessarily dictate:

"Choose B."

It influences the prior question:

"What choices are available to you at all?"

That difference is fundamental.

Recommendation power is often less about persuasion than agenda formation.

An algorithm can influence behavior without changing opinions by changing:

• which issue becomes salient,
• which hobby is discovered,
• which personality is encountered,
• which aesthetic appears normal,
• which community becomes visible,
• which controversies seem ubiquitous.

This resembles older agenda-setting theories in mass communication, but personalization makes the process individualized.

Traditional broadcast media could make a society think about Topic X.

A personalized recommender can potentially make:

Person A think about X,

Person B obsess over Y,

and

Person C never encounter either.

That is qualitatively different.

V. Exposure, Persuasion, and Salience Must Be Separated

The earlier discussion risked collapsing these mechanisms.

They are distinct.

Exposure effect

The system changes what you see.

Strongly established.

Attention effect

The system changes what receives your limited cognitive resources.

Also strongly plausible and measurable.

Salience effect

Repeated exposure makes an issue feel more important or common.

Probable, but highly context-dependent.

Persuasion effect

The material changes your opinion.

Mixed evidence.

Identity effect

Repeated exposure contributes to changing who you understand yourself to be.

Plausible and supported qualitatively in specific settings, but weakly established causally.

Personality effect

Long-term exposure alters relatively stable psychological traits.

Currently much less established.

This hierarchy matters.

A system need not persuade someone ideologically to exert enormous social power.

If it determines what they think about, its influence can still be profound.

VI. The Meta Experiments Were Both Stronger and Weaker Than They First Appear

The 2023 Facebook and Instagram studies are among the strongest evidence available because they randomized actual users inside real platforms.

Participants' feeds were altered during the 2020 U.S. election.

Researchers tested interventions including:

• reducing content from like-minded sources,
• switching to chronological feeds,
• reducing reshared material.

The manipulations produced substantial changes to people's information environments.

For example, reducing like-minded Facebook content lowered exposure to such sources from roughly 54% to about 36%. It also reduced exposure to misinformation-associated sources and uncivil material. (Nature)

Yet political attitudes—including affective polarization and ideological extremity—did not measurably change. (Nature)

That result is important.

But several limitations must be understood.

Duration

The experiments lasted months, not years.

Population

Participants were adults who already had established political identities.

Historical context

They occurred during an extraordinarily polarized U.S. presidential election.

Outcomes measured

Political attitude scales capture only certain forms of influence.

General-equilibrium effects

Changing someone's ranking system temporarily does not reverse the larger information ecosystem that existed before the experiment.

This last limitation is critical.

Imagine that ten years of ranking shaped:

• whom someone followed,
• what media institutions grew,
• which influencers became famous,
• which political language entered public discourse.

Switching that user to chronological order for three months does not erase those accumulated structures.

Therefore the Meta results strongly argue against:

"Changing ranking immediately changes ideology."

They do not prove:

"Recommendation systems have no long-term political effect."

The latter conclusion would overreach the evidence.

VII. The 2026 X Experiment Reveals a Different Mechanism

The newer X experiment is particularly important because it found measurable political effects where the Meta experiments largely did not.

Active U.S. users were randomized to algorithmic versus chronological feeds for seven weeks.

Turning on X's algorithmic feed:

• increased engagement,
• increased conservative political exposure,
• reduced relative exposure to traditional news media,
• shifted some political views toward more conservative positions.

But it did not significantly change party identity or affective polarization. (Nature)

The most interesting finding may not be attitude change.

It is network change.

Users exposed to the algorithm followed conservative political activist accounts, and some of those follows persisted after the experimental condition ended. (Nature)

That suggests:

This is a form of path dependence.

Temporary algorithmic influence can potentially create lasting environmental change without permanently changing the underlying algorithm.

This deserves far more attention than the simplistic question:

"Did the user's attitude move?"

VIII. Network Effects May Be More Important Than Direct Persuasion

Suppose TikTok introduces a teenager to a community devoted to a niche lifestyle.

The algorithm itself may never persuade the teenager of anything.

Instead:

1. it recommends a creator,
2. the teenager follows them,
3. the teenager finds a community,
4. friendships form,
5. new vocabulary is learned,
6. group norms become meaningful,
7. offline behavior changes.

The causal structure is now:

The human group performs much of the persuasive work.

This means conventional studies measuring direct media persuasion might underestimate algorithmic effects because they ask the wrong causal question.

The relevant mediator may be:

whom the algorithm causes you to meet.

IX. Identity Formation Requires Better Definitions

The previous investigation referred broadly to "identity."

That needs subdivision.

Identity can mean:

Personal identity

"Who am I as an individual?"

Social identity

"Which groups am I one of?"

Narrative identity

"What story do I tell about my life?"

Performative identity

"How do I present myself to others?"

Aspirational identity

"Who do I want to become?"

Algorithmic identity

"Who does the system appear to think I am?"

These can diverge.

TikTok research by Karizat and colleagues used interviews with 15 U.S. users to examine this phenomenon. Participants described attempting to alter their behavior to make TikTok's inferred "algorithmic identity" better match their self-understanding. (ResearchGate)

This is meaningful but should not be overstated.

It is:

peer-reviewed qualitative evidence of user experience.

It is not causal proof that TikTok transforms identity.

That distinction was insufficiently emphasized earlier.

Still, the concept reveals something historically novel.

Humans may now encounter a machine-generated portrait of themselves indirectly through the content they receive.

The feed effectively says:

"People like you apparently care about these things."

That may itself become socially meaningful.

X. The Algorithmic Mirror Can Become a Self-Fulfilling Prophecy

Imagine a teenager displaying weak interest in five subjects:

• guitar,
• fitness,
• fashion,
• politics,
• gaming.

By chance, the first fitness video generates unusually long watch time.

The recommender increases fitness content.

The teenager responds.

Soon the feed becomes heavily fitness-oriented.

The teenager now receives:

• fitness influencers,
• diet content,
• gym culture,
• clothing,
• terminology,
• communities.

After six months, fitness may genuinely become central to the teenager's identity.

Did the algorithm discover the preference?

Or create it?

The correct answer might be:

It selected one of several possible selves and helped stabilize it.

That is a more sophisticated model than either "manipulation" or "reflection."

Call it algorithmic path selection.

The system may not invent a preference ex nihilo.

It may influence which weakly developed preference receives enough reinforcement to become enduring.

This is one of the most important new threads emerging from the deeper analysis.

XI. Adolescents Represent a Special Case—but Evidence Must Be Handled Carefully

The developmental argument is intuitively powerful.

Adolescence involves:

• identity exploration,
• heightened peer sensitivity,
• social-status learning,
• emotional development,
• experimentation.

Therefore algorithmic environments could plausibly exert greater effects.

But the causal evidence remains limited.

A 2026 BMC Public Health grounded-theory study interviewed young people and developed the concept of algorithmically structured exposure. Participants described repeated recommendation patterns, emotional feedback loops, and difficulty separating voluntary interest from content repeatedly supplied by algorithms. (Springer)

The study is valuable for mechanism discovery.

It does not establish population-level causal effects.

A 2026 CESifo working paper used the 2016 introduction of Instagram's algorithmic feed as a quasi-experimental event and reported negative adolescent mental-health effects, including social isolation and social comparison. But this remains a working paper, not settled peer-reviewed consensus, and its difference-in-differences design depends on assumptions about comparison groups and parallel trends. (ifo Institut)

So the current evidence balance is:

plausible concern + suggestive evidence + inadequate long-term causal research.

That is considerably weaker than many public claims imply.

XII. Emotional Feedback Loops Are More Complicated Than "Algorithms Promote Anger"

The earlier analysis noted anger amplification.

That deserves refinement.

The 2025 Twitter audit found engagement ranking strongly amplified negative emotionality, particularly anger in political posts. (DOI.org)

But several mechanisms could generate that result.

Mechanism A — Human negativity bias

Humans naturally attend more to threat and conflict.

Mechanism B — Moral-emotional transmission

Emotionally moralized language can spread effectively through social networks.

Mechanism C — Platform optimization

Engagement models learn that emotional material performs well.

Mechanism D — Creator adaptation

Creators learn that outrage works and produce more outrage.

Mechanism E — audience sorting

People disproportionately follow emotionally intense political creators.

The final information environment may therefore emerge from all five simultaneously.

The correct causal story is not necessarily:

Platform chooses outrage.

It may instead be:

Human psychology rewards outrage → models detect that reward → creators adapt → supply increases → users encounter more outrage → engagement reinforces the model.

That is genuine evolutionary feedback.

XIII. The Creator Side Changes the Entire Problem

Most public discussion treats recommendation systems as choosing among a fixed set of content.

But the content pool is not fixed.

Creators adapt to ranking incentives.

This creates what economists would call an endogenous supply response.

If short, emotionally intense videos receive greater reach, creators produce more short, emotionally intense videos.

Now the recommender influences culture even if no individual recommendation persuades anyone.

Recent theoretical research explicitly models this.

A 2026 paper on platform–creator dynamics found that recommendation incentives can generate different ecosystem equilibria, including a low-quality outcome driven by short-term optimization. Its results are game-theoretic rather than direct observational proof, but they highlight possible tipping and path-dependence mechanisms. (Sage Journals)

Another 2026 theoretical study found that common online-learning approaches can generate incentives for creators to reduce effort under certain model assumptions and proposed alternative algorithms designed around producer incentives. (DROPS)

Again:

these are models.

They demonstrate possible dynamics, not evidence that TikTok or YouTube currently operates exactly that way.

But they reveal an overlooked question:

What does a recommendation system cause creators to manufacture?

That may matter as much as what it recommends.

XIV. Cultural Evolution Can Occur Through Selection Rather Than Persuasion

This opens a powerful analogy.

Darwinian evolution does not require organisms to consciously choose their traits.

Differential reproduction changes the population.

Similarly, algorithmic culture can evolve through differential visibility.

Content possessing certain traits may receive more distribution.

Creators imitate successful formats.

Unsuccessful formats disappear.

Over time:

Possible selected traits include:

• duration,
• facial expression,
• emotional intensity,
• editing rhythm,
• controversy,
• novelty,
• narrative structure,
• thumbnail style,
• ideological framing.

This mechanism could reshape culture even if individual users retain substantial autonomy.

The algorithm does not need to "brainwash" anyone.

It merely needs to alter which cultural variants reproduce most successfully.

That is a much more plausible and analytically powerful long-term mechanism.

XV. There Is No Single Social-Media Algorithm

A major conceptual error in popular discourse is treating all recommenders as interchangeable.

They are not.

Facebook

Historically grounded heavily in social-graph relationships.

The algorithm ranks content within an environment built substantially from people and pages the user already knows or follows.

YouTube

A hybrid discovery and sequential-consumption system.

Recommendations can determine what users watch next after an initial video.

TikTok

Extremely discovery-oriented.

A user can encounter creators without deliberately constructing the social graph first.

X

Its "For You" feed mixes material from followed and unfollowed accounts, making recommendation capable of altering both ranking and discovery. The 2026 X experiment exploited exactly this distinction. (Nature)

These architectures imply different causal power.

A TikTok-style discovery system may exert greater influence over what communities a user encounters.

A Facebook-style ranking system may exert more influence over which existing relationships become salient.

A YouTube-style watch-next system may matter more for sequential content trajectories.

Generalizing one platform's effect to another is therefore methodologically dangerous.

XVI. "Echo Chamber" Is an Overused and Ambiguous Term

At least three phenomena are routinely conflated.

Homophily

People choose relationships with similar people.

This long predates social media.

Selective exposure

People preferentially consume information consistent with existing beliefs.

Again, not uniquely algorithmic.

Algorithmic filtering

A ranking system preferentially displays congenial information.

These can interact, but they should not be treated as equivalent.

The Meta experiment demonstrated that algorithmic filtering matters for exposure but also showed that human preference remains powerful after algorithmic intervention. (Nature)

A 2026 study of 825 young Indian voters reported relationships among network homophily, algorithmic recommendation, perceived echo-chamber strength, and polarization. But because the study relied on survey measures and structural equation modeling rather than randomized exposure, causal interpretation requires caution. (DOI.org)

That example also reveals a broader problem:

Most high-quality causal studies come disproportionately from the United States and Western Europe.

The global evidence base remains uneven.

XVII. Regional Differences May Be Much Larger Than We Assume

Algorithms operate inside different societies.

The same technical recommender can interact differently with:

• multilingual populations,
• ethnic divisions,
• authoritarian states,
• fragmented media systems,
• weak journalistic institutions,
• strong public broadcasting,
• high or low digital literacy,
• different political party structures.

Consider a recommender in:

Finland

versus

India

versus

Brazil

versus

Myanmar.

The surrounding information ecology is entirely different.

Even preferences toward personalization differ internationally.

A preregistered six-country survey covering Brazil, Germany, Japan, South Korea, the United Kingdom, and the United States found that roughly one-fifth of respondents indicated they would opt out of personalized recommender systems, with substantial variation related to awareness, privacy attitudes, and nationality; German respondents showed particularly high willingness to contest personalization. (DOI.org)

Thus human–algorithm co-evolution may not produce one global trajectory.

It could create different algorithmic cultures.

XVIII. The Chronological Feed Is Not a Neutral Baseline

Another subtle methodological issue:

Researchers frequently compare algorithmic ranking with a chronological feed.

But chronological ranking is also an algorithmic rule:

More importantly, chronological feeds inherit:

• whom the user follows,
• how frequently those people post,
• what creators learned from previous algorithms,
• what content the platform allows,
• which accounts became popular historically.

So "algorithmic versus nonalgorithmic" is usually misleading.

The true comparison is:

one ranking policy versus another ranking policy.

This matters because chronological feeds can sometimes produce undesirable outcomes too.

Nature's summary of the 2023 Meta studies notes that chronological ranking increased exposure to some untrustworthy sources under the tested conditions. (Nature Media)

Therefore:

algorithmic ≠ harmful

and

chronological ≠ neutral.

XIX. Algorithms Are Moving Targets

Research suffers from another fundamental obstacle.

Platforms change constantly.

A study might analyze:

YouTube recommendations in 2019.

But by 2022:

• objectives may have changed,
• moderation may have changed,
• candidate generation may have changed,
• satisfaction signals may have changed,
• creator incentives may have changed.

A scientific finding can therefore be historically true but operationally obsolete.

Researchers face a system whose treatment itself evolves.

This is uncommon in traditional social science.

The "instrument" under study may be redesigned before the paper is published.

XX. Researchers Usually Cannot See the Full System

External auditing has serious limitations.

Modern recommendation architectures can involve:

1. candidate generation,
2. retrieval,
3. multiple ranking models,
4. safety filtering,
5. diversity constraints,
6. business rules,
7. user controls,
8. advertising,
9. experimentation systems.

Researchers often see only:

input → displayed feed.

They do not see the internal decision process.

A 2024 analysis of sociotechnical transparency identifies limited platform data access, difficulty isolating variables, and complex interactions among algorithm, users, and environment as central obstacles to understanding recommender systems. (Springer)

This means external research frequently studies a black box.

Internal platform studies have the opposite problem:

they may have extraordinary data access but raise questions regarding independence and replicability.

Neither approach is ideal.

XXI. Simulation Research Is Useful—but Frequently Overinterpreted

The influential Chaney, Stewart, and Engelhardt work showed through simulation that recommender feedback can homogenize user behavior and reduce utility. (arXiv)

That result is intellectually important.

But simulations establish:

"This dynamic can occur under these assumptions."

They do not establish:

"This is what Facebook actually does."

The same applies to newer "algorithmic drift" frameworks designed to quantify how recommendations could change preferences over repeated interactions. (ScienceDirect)

Simulations are especially useful for studying twenty-year trajectories that cannot yet be observed.

But they should remain mechanism-generating tools, not empirical substitutes.

XXII. One of the Biggest Research Problems Is Time Scale

Current studies operate at several very different temporal scales:

seconds — click behavior minutes — session consumption days — feed adaptation weeks — experiments months — political campaigns years — identity formation decades — cultural evolution

Most rigorous causal evidence exists in the first five levels.

Our question concerns the last two.

That creates an unavoidable epistemic gap.

We are trying to infer twenty-year human consequences using systems that:

• are less than twenty years old,
• continually change,
• have only recently become heavily personalized,
• now face another disruption from generative AI.

Any confident 2046 forecast should therefore be treated with suspicion.

Scenario analysis is more defensible than prediction.

XXIII. Social Norms May Be Another Missing Mechanism

Algorithms need not persuade people directly.

They can alter what appears popular.

Humans infer norms from visibility.

A feed filled with a behavior can create the impression:

"Everyone is doing this."

That distinction involves:

Descriptive norms

What people appear to do.

Injunctive norms

What people appear to approve of.

However, a 2026 experiment involving 1,021 social-media users found that simply labeling material as socially or algorithmically recommended did not significantly change perceived norms in a single-exposure setting. Perceived norms themselves did predict engagement intentions, but recommendation labels alone did not manufacture those norms. (Springer)

That is a valuable corrective.

It suggests that:

Repeated environmental exposure may matter more than merely telling users "the algorithm recommends this."

XXIV. Social Proof and Algorithmic Proof May Eventually Merge

A potentially important future mechanism becomes visible here.

Traditional social proof says:

"Millions of people liked this."

Algorithmic personalization introduces another signal:

"This was selected specifically for you."

Those are psychologically different.

The first indicates popularity.

The second implies machine knowledge of self.

As AI systems become more trusted, a recommendation may eventually carry informational authority:

"The system knows me better than I know myself."

Whether humans develop that level of algorithmic deference is uncertain.

But it could fundamentally change recommendation power.

XXV. The Most Important Feedback Loop May Operate at Population Level

The original framing emphasized:

The deeper model is:

Consider:

1. people reward emotional content,
2. algorithms amplify it,
3. creators imitate it,
4. journalism adapts to competition,
5. political campaigns adapt,
6. people encounter a more emotional environment,
7. new engagement data reinforce the pattern.

No single participant intended the final result.

That is an emergent property.

This distinction is crucial because it removes the need for a mastermind.

Societal outcomes can emerge from incentive-compatible individual behavior.

XXVI. A New Distinction: First-Order vs Second-Order Algorithmic Effects

This may help organize future research.

First-order effects

Direct consequences of ranking.

Examples:

• which post you see,
• whether you click,
• whether you watch.

These are relatively easy to measure.

Second-order effects

Consequences of people adapting to ranking.

Examples:

• creators altering content,
• users changing how they present themselves,
• politicians changing rhetoric,
• journalists changing headlines.

Harder to measure.

Third-order effects

Institutional and cultural adaptations.

Examples:

• new professions,
• influencer economies,
• political campaigning styles,
• social norms about attractiveness,
• attention conventions.

These are much harder.

Fourth-order effects

Developmental and evolutionary changes across generations.

Examples:

• children developing expectations around personalized information,
• changes in self-conception,
• changes in how relationships form.

The twenty-year question primarily concerns third- and fourth-order effects, where evidence is currently weakest.

XXVII. Identity May Become Increasingly Performative Toward Machines

Humans historically performed identity for:

• family,
• peers,
• employers,
• communities,
• romantic partners.

We may increasingly perform identity for ranking systems.

Creators already think:

"Will the algorithm understand this?"

Businesses think:

"Will search rank this?"

Professionals think:

"Will LinkedIn surface me?"

Dating-app users think:

"Will the system match this profile?"

That could lead to a new social competency:

algorithmic legibility

The ability to make oneself interpretable and attractive to automated selection systems.

Over two decades, this could shape speech, aesthetics, resumes, dating behavior, and creative expression.

XXVIII. But Human Resistance Will Also Evolve

A deterministic future is unlikely because humans adapt defensively.

Users already:

• create private group chats,
• use chronological modes,
• delete apps,
• manipulate algorithms,
• use multiple accounts,
• reset recommendations.

The six-country contestability research found meaningful willingness to opt out of personalized recommendation. (DOI.org)

As algorithmic literacy grows, users may develop cultural norms analogous to earlier media literacy:

Don't believe everything television tells you.

could become:

Don't let your feed decide what you care about.

Therefore increasing algorithmic sophistication could generate increasing algorithmic resistance at the same time.

XXIX. Personalization Could Produce Both Fragmentation and Homogenization

This apparent paradox deserves deeper attention.

At the individual level, recommendations can generate highly specialized worlds.

Person A gets:

woodworking.

Person B:

ancient history.

Person C:

Korean cooking.

This increases diversity between individual information diets.

Yet at the population level, the same systems may concentrate attention around relatively few:

• creators,
• narratives,
• formats,
• aesthetics.

Chaney et al.'s simulation work demonstrates how recommendation feedback can produce behavioral homogenization under certain conditions. (OAR Princeton)

So both claims may be true:

People become more specialized.

and

Culture becomes more standardized.

This is not contradictory once different levels of analysis are separated.

XXX. The Long-Term Human Question May Be Less About Beliefs Than Attention

One of the most important conclusions from this deeper pass is that ideological persuasion may be the wrong focal point.

Human life is constrained by limited attention.

Every minute spent consuming Topic X is a minute not spent on Topics Y and Z.

Therefore recommendations shape:

which shapes:

which shapes:

which shapes:

which shapes:

An algorithm need never persuade someone politically to substantially influence their life.

It can merely repeatedly answer:

"What should you pay attention to next?"

That may be the deepest form of influence.

XXXI. New Thread: Opportunity Costs of Algorithmic Exposure

Recommendation research rarely measures what did not happen.

Suppose a teenager spends 1,000 hours over three years watching short-form entertainment.

The relevant comparison is not:

Did those videos make them happier or sadder?

It may be:

What activities disappeared from those 1,000 hours?

Possible counterfactuals include:

• reading,
• sleep,
• sports,
• face-to-face interaction,
• creative practice,
• boredom,
• exploration.

This is extremely difficult to study.

But twenty-year societal effects may depend heavily on those displaced behaviors.

XXXII. New Thread: The Loss of Boredom and Unstructured Discovery

Historically, humans often encountered information through:

• wandering,
• browsing,
• boredom,
• accidental discovery.

Highly optimized feeds reduce friction.

Every moment can contain predicted stimulation.

That raises an underexplored question:

What happens to human curiosity when informational environments become increasingly predictive?

Two opposite futures are plausible.

Expansion hypothesis

Recommendation systems expose people to niches they would never discover.

Compression hypothesis

Systems continually exploit known interests and reduce truly random exploration.

Which dominates probably depends on recommender design.

This should become a distinct research branch.

XXXIII. New Thread: Preference Ossification

A recommender may become so good at predicting past behavior that it stabilizes previous versions of a person.

Suppose someone liked heavy metal at age 18.

At age 23 they might naturally drift toward jazz.

But personalized recommendation continually feeds heavy metal because that preference has extensive historical data.

The system might therefore produce preference inertia.

Human development often involves abandoning identities.

Recommendation histories may resist forgetting.

This creates an important design problem:

Should algorithms contain intentional mechanisms for forgetting who users used to be?

This may become increasingly important.

XXXIV. New Thread: Exploration as a Human Right or Design Objective

Machine-learning recommenders already face an exploration–exploitation dilemma.

Exploit: recommend what seems likely to work.

Explore: recommend something uncertain to learn more.

Normally this is framed as an optimization problem.

But socially it becomes:

How much unfamiliarity should humans encounter?

Too much exploitation could trap users in increasingly precise preference loops.

Too much exploration could make feeds irrelevant.

The optimal answer may not be purely commercial.

It may involve developmental and democratic values.

XXXV. New Thread: Algorithmic Memory and Identity Continuity

Future AI systems may retain years of behavioral data.

A personal assistant might know:

• what you watched at 16,
• what you bought at 20,
• whom you dated at 25,
• what political ideas interested you at 30.

That creates unprecedented continuity in machine representations of individuals.

Humans forget.

Friends forget.

Communities change.

Machines may not.

The social consequences of persistent algorithmic memory deserve serious research.

XXXVI. New Thread: From Recommendation to Delegation

The next twenty years may not be dominated by feeds at all.

Recommendation asks:

"Here are five restaurants you might like."

Agentic AI may say:

"I made the reservation."

Recommendation asks:

"Here are articles about this issue."

Agentic AI may say:

"I read them and summarized what matters."

Recommendation asks:

"These people might interest you."

Agentic AI might someday say:

"I think you should meet this person."

This represents a fundamental shift from:

attention mediation

to

decision mediation.

The co-evolutionary loop becomes more consequential because algorithms may increasingly influence actions rather than merely information.

XXXVII. Revised Evidence Ladder

After the deeper review, the evidence can be ranked more precisely.

Very strong

User behavior influences recommendation.

Recommendation algorithms materially change exposure.

Ranking design changes engagement and platform use.

Strong

Some engagement-optimized systems amplify emotional, partisan, or hostile material relative to alternative ranking policies.

Creators and users strategically adapt behavior to algorithmic incentives.

Moderate

Algorithmic exposure can sometimes alter following behavior, short-term attitudes, and social-network structure.

Recommendation systems can reinforce existing interests and produce path dependence.

Suggestive

Long-term algorithmic environments contribute to identity development.

Adolescents may be especially sensitive.

Feedback loops may alter population-level cultural diversity and content quality.

Weak / unresolved

Recommendation algorithms systematically reshape stable personality.

They routinely create extreme ideologies in previously neutral individuals.

They will predictably homogenize humanity over twenty years.

They will predictably fragment humanity into isolated individualized realities.

Those last claims remain scenarios, not established findings.

XXXVIII. A More Sophisticated Co-Evolutionary Model

The original two-way loop can now be replaced.

Simultaneously:

And around both sits:

This is no longer merely a recommender system.

It is an adaptive cultural ecosystem.

XXXIX. Where the Research Landscape Now Points

The deeper investigation substantially changes the priority of the next questions.

The most important unknown may not be:

Can algorithms persuade people?

That question is too narrow.

The higher-value questions are now:

1. Can algorithms select among possible future versions of a person's identity by disproportionately reinforcing initially weak interests?
2. Do temporary recommendations leave permanent effects through network formation, creator discovery, and altered social environments?
3. How much culture changes because creators adapt to machine-selection pressures rather than because audiences consciously change preferences?
4. Are engagement signals systematically measuring attention capture rather than reflective human preference?
5. Does lifelong personalization reduce exploration, preference change, boredom, and accidental discovery?
6. What happens when recommender systems evolve into AI agents that increasingly act rather than merely recommend?

Those questions move the investigation from a debate about "social media manipulation" into something considerably larger:

Whether adaptive computational systems are becoming an active selection mechanism in human cultural evolution.

That proposition is not yet an established fact.

But unlike stronger claims of mass algorithmic mind control, it does not require speculative assumptions about direct persuasion. It follows from mechanisms we already know exist: behavioral measurement, differential amplification, creator adaptation, network formation, and recursive learning.

The unresolved question is not whether those mechanisms exist.

They do.

The unresolved question is how large their cumulative effects become when the loop operates continuously across a generation.

High-Value Research Questions: Human–Algorithm Co-Evolution

The investigation has reached the point where the central problem can be narrowed substantially.

The most important uncertainty is no longer whether recommendation algorithms influence people at all. They plainly influence exposure, attention, engagement, and sometimes subsequent network formation. The harder questions concern where in the causal chain durable human change occurs, how large it is, and whether that change accumulates across years into altered identity, culture, and social organization.

The research agenda should therefore prioritize questions that can distinguish among four rival broad models:

1. Reflection: algorithms mainly discover pre-existing preferences.
2. Amplification: algorithms disproportionately reinforce preferences that already exist.
3. Path selection: algorithms help one of several weak or latent interests become disproportionately important.
4. Preference construction: repeated algorithmic exposure materially creates durable preferences, identities, or beliefs that would probably not otherwise have developed.

The highest-value research questions are those capable of telling these apart.

I. Decisive Questions

1. If two otherwise comparable people receive systematically different algorithmic environments for several years, do their enduring preferences, identities, relationships, and behavior measurably diverge?

This is the single most decisive question.

Almost every larger claim about algorithmic co-evolution ultimately depends on it.

Short-term experiments establish that feeds alter exposure and sometimes attitudes. The 2026 randomized X study, for example, found that seven weeks of algorithmic-feed exposure shifted some political attitudes and caused users to follow more conservative activist accounts. Crucially, those following changes persisted after the algorithm was removed. (Nature)

But seven weeks is tiny compared with:

• adolescence,
• university years,
• political development,
• relationship formation,
• career development,
• identity maturation.

What would change our understanding?

If multi-year exposure produces durable divergence: The stronger co-evolutionary thesis becomes substantially more credible. Recommendation systems would have to be understood as developmental environments rather than merely media filters.

If exposure effects repeatedly disappear when the algorithm changes: The reflection/amplification interpretation becomes stronger.

If attitudes remain stable but networks, habits, and interests diverge: The emerging thesis would need reframing: algorithmic influence may operate primarily through life-path architecture rather than persuasion.

II. Foundational Questions

Before making larger claims, several underlying questions must be resolved.

2. What exactly is a "preference," and which observable signals reliably measure it?

This is more foundational than it initially appears.

A recommendation system may treat:

• watch time,
• click-through,
• comments,
• sharing,
• replaying,
• likes,
• subscriptions,

as evidence of preference.

But these can measure different psychological states.

A person can pay attention to something because they:

• enjoy it,
• hate it,
• fear it,
• find it shocking,
• disagree with it,
• cannot stop watching it.

Recent evidence on X strengthens this concern. A 2026 PNAS study found tensions between users' explicitly stated values, their engagement behavior, and what the platform ultimately amplified. (DOI)

The foundational research question is therefore:

How well do engagement signals correlate with reflective preferences, stated values, long-term satisfaction, and welfare?

Necessary experiment

For the same individuals, researchers should compare:

immediate engagement

against:

immediate stated preference

against:

24-hour retrospective satisfaction

against:

long-term desire for similar material.

Without this distinction, claims that "the algorithm learns what users want" remain partly circular.

III. The Path-Selection Question

3. Do recommendation systems primarily strengthen already dominant preferences, or can they determine which weak interests become dominant?

This may ultimately be more important than persuasion.

Imagine someone has mild initial interests in:

• fitness,
• politics,
• history,
• music,
• cooking.

If stochastic early engagement causes the recommender to disproportionately develop the fitness branch, does that person become meaningfully more fitness-oriented several years later?

The competing explanations are:

Reflection model

Fitness would have become dominant anyway.

Reinforcement model

The algorithm accelerated an existing trajectory.

Path-selection model

Several futures were plausible; the recommender helped stabilize one.

Construction model

The resulting identity largely exists because of the algorithmic environment.

Distinguishing these requires repeated measurements before substantial personalization occurs.

This makes new-account studies, adolescent cohorts, and natural experiments especially valuable.

IV. Persistence Versus Reversibility

4. Which algorithmically induced changes persist after exposure ends?

This question can discriminate temporary media effects from genuine co-evolution.

Possible outcomes should be measured separately:

• opinion,
• habit,
• following network,
• friendship,
• community membership,
• consumption behavior,
• self-identification,
• creator preference.

The X experiment is especially important because it found asymmetric effects: switching the algorithm on changed some political attitudes and following behavior, whereas switching it off did not simply reverse those effects. (Nature)

That suggests possible hysteresis:

does not imply

when the original stimulus is removed.

Decisive test

After six months of personalized exposure, remove personalization and follow participants for another year.

If effects rapidly disappear, influence is mostly transient.

If network structure persists, influence may be infrastructural.

If preferences persist independently of network structure, stronger preference-formation mechanisms become plausible.

V. Direct Persuasion Versus Network Mediation

5. Does the algorithm change people primarily through content, or through the humans and communities it helps them discover?

This could radically reframe the subject.

The typical model assumes:

But the more important pathway may be:

The 2026 X experiment provides unusually strong evidence for part of this mechanism: algorithmic exposure changed whom users followed, altering even the content subsequently available in their chronological feeds. (DOI)

Discriminating experiment

Randomly assign:

Group A: repeated content from Creator X without ability to follow.

Group B: limited exposure but ability to follow Creator X.

Group C: exposure plus community interaction.

Then measure longer-term change.

If Group C changes most, social-network effects may dominate direct algorithmic persuasion.

VI. The Creator-Evolution Question

6. How much does recommendation ranking reshape the supply of culture itself?

This is crucial because most research focuses on consumers.

Yet recommender systems also determine which creators succeed.

Creators then adapt.

The resulting loop is:

Recent formal work increasingly treats recommender systems as multi-agent ecosystems in which creator incentives interact with platform learning, although much of this literature remains theoretical rather than direct evidence about specific platforms. (Proceedings of Machine Learning Research)

High-value research questions

Do recommendation incentives measurably increase:

• emotional intensity?
• ideological extremity?
• visual standardization?
• video brevity?
• repetitive formats?
• clickbait?
• novelty?
• outrage?
• sensationalism?

Most importantly:

Would creators produce materially different culture under different ranking objectives?

That can be experimentally tested.

VII. The Engagement-Objective Question

7. What cultural ecosystem emerges when algorithms optimize different objectives?

This could reveal whether apparent social harms result from recommendation itself or from particular optimization criteria.

Imagine identical platforms optimizing separately for:

watch time

immediate engagement

user-rated satisfaction

24-hour satisfaction

long-term retention

learning

social connection

viewpoint diversity

novelty

reflective preference

The resulting environments could differ profoundly.

This is one of the most valuable experiments the field could conduct.

It would distinguish:

"Personalized algorithms inevitably create these outcomes"

from:

"These outcomes result from specific objective functions."

That distinction has enormous policy implications.

VIII. The Attention Question

8. Is the most important algorithmic effect not persuasion but cumulative allocation of human attention?

This deserves elevation to a major research question.

The relevant outcome may not be:

"Did your political belief change?"

It may instead be:

"What did you spend 5,000 hours thinking about?"

The counterfactual problem becomes:

If someone spends years consuming personalized media, researchers need to know what disappears:

• books,
• sleep,
• conversation,
• sports,
• creative work,
• outdoor activity,
• boredom,
• unstructured thinking.

Why decisive?

Even if algorithmic content has zero persuasive effect, altered time allocation could still transform:

education,

skills,

friendships,

health,

careers,

and culture.

That would substantially reframe the whole investigation.

IX. The Exploration Question

9. Does personalization expand discovery or gradually narrow the range of possible interests?

Two plausible mechanisms compete.

Exploration hypothesis

Algorithms expose users to material they would never have discovered otherwise.

Exploitation hypothesis

Algorithms increasingly optimize around known interests and reduce genuine novelty.

The outcome may follow an inverted-U:

but

A strong longitudinal study would measure not simply content diversity but distance from prior preference space.

Seeing ten new football channels is not genuine exploration if the person's existing interest is football.

Discovering archaeology after years of sports consumption represents something different.

X. Preference Ossification

10. Do recommender systems make people's preferences more stable than they otherwise would be?

This is distinct from homogenization.

Humans naturally change.

Algorithms retain behavioral histories.

That creates the possibility that systems repeatedly present users with their own past.

The key question is:

Does personalization slow natural preference drift?

This could be tested by comparing:

• heavily personalized users,
• users receiving deliberate exploration,
• chronological users,
• users whose recommendation histories are periodically reset.

If preference change occurs faster in reset/exploration groups, personalization may be creating identity inertia.

That would have major developmental implications.

XI. Adolescence and Sensitive Periods

11. Are there developmental windows during which algorithmic influence has unusually durable effects?

The hypothesis is plausible but insufficiently established.

The crucial question is not simply whether adolescents spend more time online.

It is:

Does identical algorithmic exposure produce stronger or more persistent effects at age 14 than at age 24 or 44?

Research should measure:

• identity exploration,
• peer-network formation,
• body image,
• political attitudes,
• self-concept,
• aspirations,
• risk behavior,
• social comparison.

The strongest design would follow cohorts longitudinally rather than relying primarily on retrospective surveys.

XII. Individual Susceptibility

12. Which people are most influenceable, and under what conditions?

Average treatment effects may conceal enormous heterogeneity.

Possible moderators include:

• age,
• prior belief strength,
• loneliness,
• social-network density,
• personality,
• political sophistication,
• algorithmic literacy,
• novelty seeking,
• social conformity,
• uncertainty about identity.

The 2026 X findings already suggest that treatment effects may depend on predispositions and the composition of available content. (Nature)

A strong future theory must predict not merely:

"Algorithms influence people."

but:

which algorithm influences which person on which dimension under which circumstances.

XIII. Platform Architecture

13. Which recommendation architectures possess the greatest capacity for behavioral shaping?

This question is necessary because "social media algorithm" is not one treatment.

Compare:

Social-graph ranking

Primarily ranks content from people already followed.

Discovery ranking

Introduces unknown creators and communities.

Sequential recommendation

Determines what comes immediately after current consumption.

Search-driven recommendation

Responds primarily to active intentional queries.

Agentic recommendation

Interprets goals and potentially takes action.

The hypothesis worth testing is:

Discovery systems may have greater identity-selection power because they influence whom and what users encounter before users deliberately choose them.

If confirmed, TikTok-like architectures would need to be studied separately from Facebook-style ranking.

XIV. Cultural Homogenization Versus Fragmentation

14. At what level does personalization increase diversity, and at what level does it decrease it?

This question could resolve an apparent contradiction.

Measure separately:

Within-person diversity

How broad is one person's information diet?

Between-person diversity

How different are two people's diets?

Producer concentration

How many creators dominate attention?

Format diversity

How varied are presentation styles?

Idea diversity

How varied are represented concepts?

Possible outcome:

while simultaneously:

and

That would mean personalization creates niche subjects delivered through increasingly standardized cultural forms.

XV. Cross-Cultural Generalizability

15. Are current findings properties of algorithms or properties of algorithms interacting with particular societies?

Most high-quality causal research remains concentrated disproportionately in Western platforms and populations.

Yet recommender effects could vary according to:

• political institutions,
• language,
• media systems,
• education,
• religious structure,
• collectivism/individualism,
• ethnic fragmentation,
• platform penetration,
• state censorship.

The decisive design would deploy the same intervention across multiple countries.

If effects vary dramatically, algorithmic influence is better understood as:

rather than a universal platform effect.

XVI. Discriminating Questions

Several questions can directly distinguish competing theories.

16. Does algorithmic exposure matter after controlling for prior preference with unusually high precision?

If no:

reflection gains support.

If yes, but only slightly:

amplification gains support.

If small initial preferences produce increasingly divergent outcomes:

path selection gains support.

If entirely novel preferences arise systematically:

construction gains support.

17. Does random early exposure create long-term divergence?

This is an especially elegant test.

Give new users slightly different random early recommendations before enough behavioral history exists for meaningful personalization.

Then allow normal recommendation.

Years later ask:

Did those tiny initial differences create different interests, networks, or identities?

If yes, recommendation ecosystems may exhibit strong path dependence.

18. Does removing recommendation reverse the effect?

Already discussed, but this is particularly discriminating.

Rapid reversal → temporary amplification.

Persistent behavioral structure → network/path dependence.

Persistent preference after environment disappears → stronger internal preference change.

19. Does exposure without social interaction produce the same effect as exposure with community participation?

If yes:

content persuasion is more important.

If no:

social-network mediation becomes central.

20. Do users given explicit control over recommendation objectives behave differently?

Give users settings such as:

• show me what I usually like,
• surprise me,
• challenge me,
• teach me,
• maximize wellbeing,
• minimize outrage.

If users voluntarily choose environments different from engagement-optimized feeds, this would strengthen the claim that revealed engagement frequently fails to represent reflective preference.

XVII. Missing-Evidence Questions

The greatest weaknesses in the current literature arise from evidence we simply do not possess.

21. Where are the decade-long longitudinal cohorts?

Ideally researchers would have:

10–15 years of:

• recommendation exposures,
• interaction histories,
• social networks,
• psychological assessments,
• identity measures,
• offline behavior,
• relationships,
• education,
• employment,
• political attitudes.

Almost nothing equivalent exists with sufficient granularity.

Without this, twenty-year forecasts remain largely extrapolative.

XVIII. The Missing Counterfactual

22. What would the same person's life have looked like without personalized recommendation?

This is fundamentally unknowable at the individual level.

But approximations could come from:

• randomized longitudinal studies,
• natural experiments,
• platform rollouts,
• geographic differences,
• age-threshold policies,
• historical cohorts.

The important outcome is not just digital behavior.

Researchers should measure:

offline life trajectories.

XIX. Missing Internal Platform Data

23. What ranking objectives and weighting systems are actually operating?

External researchers rarely know enough about:

• candidate generation,
• feature weights,
• ranking stages,
• safety filters,
• experiment assignments,
• commercial overrides,
• exploration rates,
• model updates.

Without those records, researchers often study:

without knowing the mechanism.

Ideal evidence would include archived platform versions so researchers could reconstruct:

exactly which model was operating for which user at which time.

XX. Missing Creator Data

24. How do creators actually change production after algorithmic incentives change?

Researchers would ideally have:

• creator drafts,
• upload histories,
• analytics dashboards,
• recommendation exposure,
• monetization data,
• editing decisions,
• creator interviews.

The decisive natural experiment would be a major ranking change where researchers observe creator behavior before and after.

XXI. Missing Developmental Evidence

25. What happens to children exposed to personalized recommendation from early childhood?

The first generation experiencing highly personalized short-form recommendation during substantial portions of childhood is only now reaching adulthood.

We therefore do not yet possess full life-course evidence.

This is a genuine historical limitation, not merely a research failure.

XXII. Missing Evidence About AI Agents

26. Does moving from recommendation to delegation produce qualitatively different human adaptation?

This evidence barely exists because the technology is emerging.

We need experiments comparing:

information-selection AI

"Here are five choices."

against

decision-support AI

"I think option B is best."

against

delegated agents

"I selected B for you."

The central outcome would be whether users gradually:

• reduce independent search,
• defer more frequently,
• lose skills,
• become more efficient,
• develop greater trust,
• become more autonomous in other domains.

This may become one of the most important research questions of the 2030s.

XXIII. Ranked Research Agenda

Based on expected informational value, I would currently rank the agenda as follows.

XXIV. Why the Top Three Are Especially Decisive

The first three questions could materially reorganize the entire investigation.

If Question 1 produces weak effects

Then much of the twenty-year concern would need to be scaled back. Algorithms might be extraordinarily powerful at controlling attention while relatively weak at constructing durable human preferences.

That would still matter enormously, but it would be a different thesis.

If Question 1 produces strong durable effects

Then recommendation systems should probably be conceptualized alongside other developmental institutions:

family, school, peers, religion, media, neighborhood—and algorithms.

That would be a major revision to our understanding of socialization.

If Question 2 finds strong path dependence

This would support perhaps the most interesting hypothesis uncovered so far:

Recommendation systems may not manufacture people wholesale; they may influence which one of several plausible versions of a person becomes realized.

That is subtler than manipulation and potentially more consequential.

If Question 3 shows network mediation dominates

Then much current research is looking in the wrong place.

Instead of asking:

"Did this video persuade you?"

the central question becomes:

"Whom did the system introduce you to, and what happened afterward?"

That would move research toward sociology, network science, developmental psychology, and anthropology rather than treating recommendation mainly as a media-effects problem.

XXV. The Single Highest-Value Experiment

If one unusually ambitious study could be conducted, I would design it around new users before substantial personalization begins.

Recruit a large, diverse cohort—ideally including adolescents and adults—and randomly vary the degree and kind of recommendation exploration while leaving users otherwise free to behave normally.

Track for several years:

• every exposure,
• engagement,
• searches,
• follows,
• unfollows,
• social relationships,
• interests,
• values,
• political views,
• self-described identity,
• hobbies,
• aspirations,
• offline behavior,
• wellbeing.

Periodically alter or remove personalization.

Most importantly, measure participants before recommendation begins.

That would allow researchers to estimate something existing studies struggle to observe:

That question sits at the heart of this entire investigation.

Research Agenda Synthesis

The evidence so far increasingly suggests that the weakest way to investigate this subject is to ask whether "algorithms brainwash people."

That framing sets up an unnecessarily extreme standard.

The scientifically richer issue is whether computational recommendation becomes one of the selection pressures through which human interests, relationships, identities, and culture develop.

The decisive research program therefore needs to establish four things in sequence:

↓

↓

↓

We have reasonably strong answers to the first part of that sequence, increasingly strong evidence concerning the second, intriguing evidence concerning the third, and very limited evidence concerning the fourth.

That fourth level—cumulative developmental and cultural path dependence—is where the investigation should now concentrate.

It is also where the most extraordinary possibilities lie, and where our current confidence should remain lowest.

Evidentiary Audit: Human–Algorithm Co-Evolution on Social Media

This audit materially narrows several claims made earlier. The investigation has accumulated many plausible mechanisms, but it has accumulated far less direct evidence about long-term human transformation than the breadth of the discussion might imply.

The strongest empirical foundation concerns a relatively short causal chain:

Evidence becomes progressively thinner as we extend the chain:

That gradient in evidentiary strength should govern everything that follows.

I. Major Claims and Present Evidentiary Status

The table is the most important result of this audit: the empirical core is real, but the long-range interpretation remains far more provisional.

II. Strongest Claim: Users Really Do Train Recommendation Systems

Classification: Established / strongly supported

This is one of the least controversial claims in the investigation.

YouTube explicitly identifies watch history, search history, subscriptions, likes, dislikes, "not interested" feedback, channel exclusions and satisfaction surveys as recommendation signals. It also states that different surfaces use those signals differently—for example, current-video context is especially important for "Up Next," while watch history is central to homepage recommendations. (Google Help)

TikTok similarly states that its For You system uses user interactions—including likes, shares, follows, comments and content creation—along with video information and contextual signals. TikTok also says some behavioral signals, such as completing a longer video, are weighted more heavily than weaker contextual variables. (TikTok Newsroom)

Source quality

These are primary corporate descriptions, which are excellent evidence for what the companies say their systems consider.

They are less useful for establishing:

• exact weights,
• undisclosed signals,
• causal social consequences,
• proprietary model architecture.

Source criticism

Platform documentation has obvious incentives toward presenting recommendation as user-serving. YouTube describes its goals in terms of helping viewers find content they want and maximizing long-term satisfaction. That claim should be treated as a statement of declared design objectives, not independent proof that the system always achieves them. (Google Help)

Still, the narrow factual point that user behavior feeds future recommendation is exceptionally well supported.

III. Algorithms Really Do Change Exposure

Classification: Established / strongly supported

This is supported by randomized experiments rather than mere correlations.

The 2023 Facebook experiment involving 23,377 consenting users reduced exposure to politically like-minded sources by about one-third. The intervention materially altered what users saw, reducing uncivil material and exposure to sources that repeatedly posted misinformation while increasing cross-cutting or politically neutral exposure. (Nature)

Another part of the 2020 Facebook/Instagram Election Study randomized users to reverse-chronological rather than default algorithmic feeds and documented substantial changes in platform behavior and information exposure. (PubMed)

The 2026 X experiment likewise directly randomized actual users between algorithmic and chronological feeds. (Nature)

There is little serious evidentiary reason to doubt the basic proposition:

ranking policy causally changes the information environment.

The harder question begins with what happens afterward.

IV. Algorithms Change Engagement More Reliably Than Beliefs

Classification: Established / strongly supported

The X experiment found that enabling the algorithmic feed increased engagement relative to chronological use. (Nature)

The Meta election experiments similarly found substantial behavioral consequences when default ranking was replaced by reverse chronology, even while measured political attitudes remained largely stable. (PubMed)

This is one of the strongest conclusions of the entire research sequence:

has considerably firmer evidence than:

That difference should remain explicit going forward.

V. Engagement Is Not the Same Thing as Preference

Classification: Established in an important context; probable more generally

Our earlier discussion was directionally sound, but it sometimes generalized too quickly.

A 2025 preregistered PNAS Nexus audit compared Twitter's engagement-based ranking with a reverse-chronological baseline. Engagement ranking amplified emotionally charged and out-group-hostile political content, and users did not report preferring the political tweets selected by the engagement algorithm. (OUP Academic)

That is high-value evidence because it directly tests the relationship between:

revealed behavior

and

stated preference.

However, it does not establish that all engagement signals on TikTok, Instagram, YouTube and every other platform systematically misrepresent user preferences.

Indeed, YouTube explicitly says it supplements engagement with satisfaction surveys, dislikes, "not interested" feedback and other signals designed to capture something beyond raw watch time. (Google Help)

Corrected formulation

The defensible claim is:

Engagement can diverge from reflective preference, and ranking systems built heavily around engagement can consequently optimize toward content users do not explicitly say they prefer. How severe this problem is varies by platform, objective function and signal mix.

That is stronger than saying "attention ≠ preference" as an absolute.

VI. Anger and Divisive Content

Classification: Established for one audited system; probable but not universal

The 2025 Twitter audit provides direct evidence that the engagement-based ranking system amplified angry, partisan and out-group-hostile political material relative to chronological ranking. (OUP Academic)

What it does not establish is that:

all recommendation algorithms inherently favor outrage.

Several mechanisms remain possible:

• human negativity bias,
• selective audience behavior,
• engagement optimization,
• creator adaptation,
• subject matter,
• specific platform architecture.

Therefore our broader model—

—is probable as a mechanism, not universally established.

A particularly important falsifier would be platforms optimizing explicitly for satisfaction or wellbeing that do not exhibit equivalent emotional amplification.

VII. Political Persuasion: The Evidence Is Genuinely Mixed

This requires perhaps the greatest discipline.

Claim: algorithmic feeds can affect political attitudes

Classification: Established in at least one important context

The 2026 Nature experiment on X randomly assigned active U.S. users to algorithmic or chronological feeds for seven weeks. Turning the algorithmic feed on shifted several measured outcomes toward more conservative positions, including policy priorities and attitudes surrounding Trump's investigations and the war in Ukraine. Effect sizes were measurable but modest; for example, some composite effects were around 0.08–0.12 standard deviations. (Nature)

That is strong causal evidence.

Claim: algorithmic ranking generally polarizes people

Classification: Disputed / not established

The 2023 Facebook study is a major counterweight.

Reducing like-minded exposure by roughly one-third among 23,377 users produced no measurable effect on eight preregistered political-attitude measures, and the study reports sufficient precision to rule out effects larger than roughly ±0.12 standard deviations for those outcomes during the intervention. (Nature)

The study removing reshared Facebook material similarly altered information exposure but found no detectable effect on beliefs or opinions during the three-month experiment. (PubMed)

Evidence balance

Therefore:

Algorithms can affect some political attitudes under some conditions.

is supported.

But:

Algorithmic feeds are a general major cause of political polarization.

is not established by the evidence reviewed here.

This is a significant boundary.

VIII. Source Criticism of the Meta Experiments

These studies are unusually strong but not dispositive.

Strengths

They were:

• randomized field experiments,
• conducted on real users,
• large,
• preregistered,
• based on actual platform exposure data rather than self-report alone.

The Facebook like-minded-source paper explicitly states that academic investigators retained control over the analysis plan and manuscript and Meta could not suppress publication. Data and replication materials were archived under controlled access. (Nature)

Limitations

Participants differed from Facebook's general user population: they were disproportionately white, female, college educated and Democratic-leaning, though weighting was used. (Nature)

More importantly, the experiment lasted about three months.

It therefore speaks strongly to:

short-term effects of altering an already mature user's feed during the 2020 election.

It speaks much less strongly to:

childhood development, ten-year exposure, identity formation or cumulative cultural effects.

This distinction is crucial.

A null three-month attitude effect cannot falsify a twenty-year developmental effect.

But neither can twenty-year effects simply be assumed because the experiment was short.

IX. Source Criticism of the X Experiment

The 2026 X study is presently one of our strongest pieces of affirmative causal evidence.

Strengths

It was:

• randomized,
• conducted with real active X users,
• seven weeks long,
• able to observe content, political attitudes and following behavior.

The algorithmic feed both reordered followed content and introduced material from accounts the users did not follow, making it a stronger discovery intervention than pure ranking within an existing network. (Nature)

Important limitation

It investigated X in summer 2023, during a particular period of the platform under Elon Musk.

The authors found the algorithm promoted more conservative material and demoted traditional-media content under those conditions. (Nature)

That does not imply:

• every X algorithm always behaves similarly,
• TikTok behaves similarly,
• Facebook behaves similarly,
• recommendation inherently shifts people rightward.

The direction of political movement was partly a property of the content environment and ranking system actually present in that platform-period.

This is a major generalizability constraint.

X. Persistent Network Effects

Classification: Established in one strong experiment; plausible more broadly

The X experiment found that algorithmic-feed exposure induced users to follow conservative political activist accounts and that they continued following many of those accounts even after the algorithmic condition ended. The authors use this network change to help explain why turning the algorithm off did not simply reverse the political effects. (Nature)

This is highly important evidence.

Our larger interpretation—

—is therefore not merely speculative.

But only the first demonstration is established.

The stronger claim that this is the dominant long-term mechanism of algorithmic social influence remains plausible, not demonstrated.

We need replication across:

• TikTok,
• YouTube,
• Instagram,
• nonpolitical communities,
• adolescents,
• multi-year timescales.

XI. The YouTube "Rabbit Hole" Claim

Claim: YouTube can lead users toward problematic material

Classification: Probable

A systematic review screened 1,187 candidate studies and retained 23 examining YouTube recommendation pathways. Fourteen implicated recommendation in facilitating problematic-content pathways, seven produced mixed findings, and two did not. (PubMed Central (PMC))

A 2023 PNAS audit using approximately 100,000 automated accounts found that YouTube recommendations became ideologically congenial for partisan users and that deeper recommendation trails contained a growing proportion of content from extremist, conspiratorial or otherwise problematic channels, especially for right-leaning profiles. (PubMed Central (PMC))

That is meaningful evidence for content pathways.

Claim: YouTube recommendations progressively make users more ideologically extreme

Classification: Unsupported as a general proposition

The same 2023 PNAS audit found no meaningful general increase in ideological extremity of recommendations, despite increased prevalence of problematic channels deeper in some pathways. (PubMed Central (PMC))

The systematic review itself explicitly cautions that recommending problematic content is not equivalent to demonstrating that users themselves become more radical. (PubMed Central (PMC))

Correction

Our earlier treatment was appropriately cautious here, and that caution should be strengthened.

Exposure pathway ≠ psychological radicalization.

XII. Algorithmic Feedback and Homogenization

Classification: Mechanistically established in models; real-world societal magnitude unverified

Chaney, Stewart and Engelhardt demonstrated through simulation that recommender systems trained on behavior already influenced by prior recommendations can create algorithmic confounding, increasing behavioral homogeneity without improving simulated user utility. (arXiv)

This is important theoretical work.

But it is simulation evidence.

It proves:

such a feedback dynamic can arise under specified assumptions.

It does not prove:

TikTok, YouTube or society as a whole does exhibit that magnitude of homogenization.

Our earlier language occasionally approached the latter implication too closely.

Audit classification

Feedback-loop possibility: strongly supported theoretically.

Large real-world homogenization effect: plausible but presently unverified.

XIII. User Resistance and "Algorithmic Folk Theories"

Classification: Strong qualitative evidence; limited population inference

Karizat and colleagues conducted semi-structured interviews with 15 U.S.-based TikTok users and documented user theories about how TikTok's For You algorithm represented identity and how some users intentionally altered behavior to resist or retrain it. (Icahn School of Medicine at Mount Sinai)

This is excellent evidence that the phenomenon exists.

It is not strong evidence for claims such as:

"most users deliberately train their algorithms."

The study's sample was 15 people and qualitative by design.

Therefore:

Existence of algorithmic folk theories and resistance: established.

Prevalence, frequency and societal importance: probable but incompletely quantified.

Cross-national survey research does show meaningful interest in contesting personalization: a preregistered six-country study involving 6,217 respondents found roughly 20% said they would choose a nonpersonalized recommender if offered one, with significant country variation. (DOI)

That supports user agency, though an expressed intention to opt out is not identical to actual behavioral resistance.

XIV. Algorithmic Identity

Classification: Plausible, with good qualitative evidence but weak causal evidence

The TikTok interview study provides genuine evidence that users think about platform recommendations in relation to personal and social identity and sometimes attempt to alter their behavior to change how the system appears to categorize them. (Icahn School of Medicine at Mount Sinai)

What it does not demonstrate is:

The direction could equally be:

or, most likely:

Our concept of algorithmic identity remains useful analytically, but causal claims require upgrading evidence.

XV. "Algorithmic Path Selection"

Classification: Speculative synthesis

This was one of our most interesting original hypotheses:

A recommendation system may not invent someone's interests but may disproportionately reinforce one of several weak possibilities until it becomes a major component of identity.

For example:

No study reviewed here directly establishes this longitudinal mechanism.

The X experiment provides a small analogue—temporary recommendation changed network-following patterns—but not identity development. (Nature)

Therefore this hypothesis should be marked explicitly:

reasonable, testable and theoretically important—but currently speculative.

A multi-year randomized or natural experiment beginning before strong personalization would dramatically upgrade it.

XVI. Adolescents and Mental Health

This portion needs substantial evidentiary downgrading.

Claim: adolescents experience algorithmically structured exposure

Classification: Strongly supported qualitatively

A 2026 BMC Public Health study conducted photo-elicitation interviews with 27 UK adolescents and young adults aged 14–19. Participants showed researchers their actual TikTok For You and Instagram Explore pages. Researchers developed the grounded-theory concept of "algorithmically structured exposure," including reinforcement, emotional feedback loops and perceived difficulty controlling recommendations. (Springer)

That is rich evidence about lived experience and perceived mechanism.

It is not causal population evidence.

Claim: recommendation algorithms harm adolescent mental health

Classification: Disputed / suggestive

A 2026 CESifo working paper exploited Instagram's 2016 algorithmic-feed introduction using a difference-in-differences design and Dutch longitudinal data. It reported negative teenage mental-health effects and proposed social isolation and social comparison as mechanisms. (ifo Institut)

This is materially stronger than a cross-sectional correlation but remains:

• a working paper,
• quasi-experimental rather than randomized,
• dependent on identifying assumptions,
• focused on one historical rollout.

The BMC study provides supportive qualitative mechanisms but cannot estimate causal effect sizes. (Springer)

Therefore the evidence does not yet justify:

"Recommendation algorithms cause adolescent mental-health deterioration."

The safer formulation is:

There is credible emerging evidence that particular recommendation environments may contribute to adolescent mental-health outcomes, but causal magnitude, heterogeneity and platform specificity remain unresolved.

XVII. Adolescents as a Uniquely Sensitive Population

Classification: Plausible, presently underverified

Developmental psychology gives good theoretical reasons to suspect higher sensitivity during adolescence.

But our investigation has not established a clean causal comparison such as:

The BMC study examines only young people. (Springer)

The CESifo paper reports effects among teenagers but does not establish a universal developmental sensitivity mechanism. (ifo Institut)

So the claim remains plausible rather than established.

XVIII. Creator Adaptation

Classification: Probable, but less empirically demonstrated than our earlier prose suggested

YouTube explicitly advises creators to examine whether viewers choose to watch, continue watching and report satisfaction, because those signals affect content performance and recommendation. (Google Help)

This establishes that:

creators are explicitly given information about behaviors relevant to recommendation.

It does not by itself prove broad cultural adaptation.

The inference that creators strategically adapt thumbnails, pacing, subjects and formats is highly plausible and widely observable, but our investigation has not yet assembled strong longitudinal causal creator data demonstrating exactly how much cultural production changes because of ranking.

Therefore:

creator adaptation exists: probable to strongly supported.

algorithmic incentives systematically make culture shorter, angrier, more standardized, etc.: plausible and outcome-specific, not established.

XIX. Recommendation Systems as "Cultural Natural Selection"

Classification: Our synthesis; plausible but not empirical fact

We proposed:

The analogy is analytically useful.

The constituent mechanisms are individually plausible:

• content varies,
• platforms distribute it unequally,
• creators observe performance,
• successful formats can be imitated.

But calling algorithms a selection pressure in cultural evolution remains a theoretical synthesis.

There is not yet evidence sufficient to quantify:

• selection coefficients,
• cultural diversity effects,
• long-term equilibrium,
• whether algorithmic selection outweighs ordinary audience taste.

This concept deserves deeper study, not presentation as established fact.

XX. Personalization and Cultural Homogenization

Classification: Plausible but unresolved

Simulation evidence demonstrates that recommendation feedback can produce homogenization under some conditions. (arXiv)

However, real platforms can simultaneously produce:

• niche discovery,
• personalized subcultures,
• concentrated creator attention,
• shared viral formats.

We have not yet established whether the net effect is:

homogenization,

fragmentation,

or a combination operating at different levels.

The claim that individuals could become more specialized while formats or attention become more concentrated is a reasonable multilevel hypothesis, not yet an empirical conclusion.

XXI. Chronological Feeds as a "Neutral" Baseline

Classification: Strongly supported criticism

Our deeper investigation correctly noted that chronological ranking is not equivalent to an unmediated information environment.

A chronological feed still depends on:

• whom the user previously followed,
• historical platform dynamics,
• what creators produced,
• moderation,
• posting frequency.

The Meta study illustrates this particularly well: users' information diets remained strongly shaped by their chosen networks even when exposure was experimentally altered. (Nature)

Therefore experimental comparisons should be described as:

algorithmic policy A versus ranking policy B

rather than:

algorithm versus no algorithm.

This correction is conceptually important.

XXII. Platform Generalization

Classification: Earlier discussion occasionally generalized too freely

This deserves explicit correction.

Evidence on:

• 2020 Facebook,
• 2023 X,
• YouTube,
• TikTok,

cannot automatically be pooled as though those systems implement one treatment.

The X algorithm studied in 2023 introduced posts from accounts users did not follow in addition to reordering content. (Nature)

YouTube distinguishes homepage, Up Next, Shorts and other recommendation surfaces and states that different signals matter differently across them. (Google Help)

TikTok's For You feed is designed around discovery from a ranked pool that extends far beyond explicit follows. (TikTok Newsroom)

Thus:

"social-media algorithms cause X"

is usually too broad.

We should instead specify:

which system, which version, which objective, which user population, which outcome and which time period.

XXIII. Geographic Generalization

Classification: Present evidence is Western-heavy and incomplete

Our earlier discussion correctly raised this issue.

The six-country recommender-contestation survey found meaningful national variation, with German respondents particularly likely to favor contesting personalization and Brazilian respondents less likely. (DOI)

But much of the strongest causal evidence in this investigation comes from:

• the United States,
• U.S. election settings,
• Western platform ecosystems.

That means claims about:

• India,
• Africa,
• Latin America,
• authoritarian regimes,
• multilingual environments,

remain much less grounded.

Human–algorithm co-evolution should not yet be assumed to follow a single universal trajectory.

XXIV. Transparency and Black-Box Limitations

Classification: Strongly supported methodological problem

Researchers usually cannot observe the complete operational recommendation pipeline.

A 2024/2025 AI and Ethics paper summarizes recurring problems including limited platform-data access, difficulty grounding models, inability to isolate variables in coupled sociotechnical systems and the limits of simulation. (Springer)

The YouTube systematic review reaches essentially the same caution: researchers do not know enough about the production recommender architecture to ensure their reconstructed models perfectly represent actual recommendation pathways. (PubMed Central (PMC))

This substantially lowers confidence in claims based purely on external audits or simulations.

XXV. One Earlier Claim Should Be Explicitly Retracted Pending Verification

In the previous deepening stage I referred to a "2026 PNAS study of 715 U.S. X users" purporting to show a divergence among stated values, engagement and platform amplification.

In this audit I could not recover an authoritative source matching that description.

The verified 2025 PNAS Nexus study does establish a related but different finding: engagement-based Twitter ranking selected more divisive political material than users explicitly preferred. (OUP Academic)

Therefore the separate claimed 2026/715-user result should presently be classified as:

Unsupported or presently unverified

It should not be carried forward as evidence unless the actual publication can be independently identified and checked.

This is exactly the kind of accumulated-conversation drift the research protocol is designed to catch.

XXVI. Claims That Currently Have the Thinnest Foundation

Several ideas became prominent because they are intellectually interesting. They should be visibly tagged as hypotheses.

Preference ossification

Status: Speculative.

We have not established that personalized recommendation prevents people naturally changing interests.

Loss of boredom

Status: Speculative.

The psychological importance of boredom has a broader literature, but we have not established a causal recommender-specific effect in this investigation.

Reduced serendipity

Status: Plausible but unresolved.

Personalization could decrease accidental discovery—or increase discovery of previously inaccessible niches.

Both remain viable.

"Algorithmic legibility" reshaping human self-presentation

Status: Plausible synthesis.

There are obvious creator and platform examples, but broad longitudinal effects on identity remain unproven.

Algorithms selecting among "possible selves"

Status: Speculative but high-value.

No decisive longitudinal test has been presented.

Cultural homogenization

Status: Plausible.

Simulation supports possible mechanisms; civilization-scale empirical evidence does not yet establish it.

Long-term personality transformation

Status: Unsupported.

This should not presently be asserted.

Twenty-year transformation of humanity

Status: Scenario analysis only.

Any claim about 2046 remains conditional.

XXVII. Evidence We Are Missing

The biggest gaps are structural rather than merely numerical.

We lack long-duration randomized evidence. The strongest experiments last weeks or months; identity and development unfold over years.

We lack pre-personalization baselines. Most studies observe people after they already possess mature digital histories.

We lack full exposure histories linked to offline outcomes. A real developmental study would need to know not only what users clicked but their relationships, education, hobbies, political development and life trajectories.

We lack clean creator-side causal evidence tracing ranking-policy changes into specific alterations in creative production.

We lack cross-platform replication of network-persistence effects like those observed on X.

We lack cross-cultural causal replication.

And crucially, we lack the counterfactual:

What would the same person have become under a different algorithmic environment?

That remains the hardest question in the entire project.

XXVIII. Upgrade and Downgrade Conditions

Claim: algorithms cause durable preference change

Current: Plausible.

Upgrade if

Large randomized or quasi-randomized multi-year studies show persistent divergence after controlling carefully for baseline preference.

Downgrade if

Repeated long-term interventions show exposure changes but preferences consistently revert once ranking changes.

Claim: algorithmic path selection affects identity

Current: Speculative.

Upgrade if

Random early exposure among otherwise comparable new users predicts durable differences years later in hobbies, networks or self-identification.

Downgrade if

Initial exposure effects disappear once baseline dispositions and self-selection are modeled adequately.

Claim: network mediation is a major long-term pathway

Current: Plausible, with one strong causal demonstration.

Upgrade if

X's following effect replicates across platforms and social domains and explains substantial downstream behavioral change.

Downgrade if

Following differences prove short-lived or have little downstream effect.

Claim: engagement optimization distorts preference

Current: Probable.

Upgrade if

Across multiple platforms, randomized comparisons repeatedly show engagement-ranked content performing worse on reflective satisfaction than alternative objectives.

Downgrade if

Platforms incorporating satisfaction and explicit feedback produce little divergence between engagement and reflective preference.

Claim: algorithms harm adolescent mental health

Current: Disputed / emerging.

Upgrade if

Multiple independent longitudinal natural experiments or randomized design changes converge on significant effects with credible mechanisms.

Downgrade if

Better-controlled studies show apparent associations are mostly explained by prior mental-health state, family environment or self-selection.

Claim: creator adaptation changes culture

Current: Probable.

Upgrade if

Ranking-policy shocks predict systematic, measurable changes in creator formats, emotionality, topics or production effort across large datasets.

Downgrade if

Creators continue producing largely the same material despite large ranking changes.

Claim: recommender systems homogenize culture

Current: Plausible.

Upgrade if

Longitudinal platform-wide studies demonstrate declining producer, stylistic or conceptual diversity causally attributable to ranking.

Downgrade if

personalization consistently expands cultural variety after accounting for concentration effects.

XXIX. Revised Evidence Pyramid

After auditing the entire investigation, I would now organize our knowledge like this:

Tier 1 — Very high confidence

Tier 2 — High-to-moderate confidence

Tier 3 — Plausible but incomplete

Tier 4 — Active hypotheses

Tier 5 — Future scenarios

These may ultimately prove important. They are not yet evidence-backed predictions.

Final Audit Judgment

The central thesis survives the audit, but in a narrower and more defensible form.

It is well supported that humans and recommender systems form a recursive system in which:

That is empirically grounded.

There is also strong evidence that this loop sometimes changes more than mere content ordering: under particular conditions it can alter engagement, political attitudes and social-network formation. (Nature)

But we have not yet established that the same feedback loop routinely restructures durable identity, stable personality or civilization-scale culture.

The most important correction is therefore conceptual:

We have good evidence for human–algorithm behavioral co-adaptation. We have suggestive but incomplete evidence for human–algorithm psychological co-evolution, and presently weak evidence for long-term developmental or civilization-scale co-evolution.

That distinction should govern the next stage.

The strongest emerging research target is consequently not proving that algorithms "shape humanity." It is determining whether repeated short-term behavioral effects accumulate, persist, propagate through social networks and creator ecosystems, and eventually become developmental effects.

That is the missing bridge between what the evidence already demonstrates and the much larger possibility this investigation has uncovered.

Chronological Reconstruction: Human–Algorithm Co-Evolution on Social Media

Chronology clarifies this topic because several very different technological and intellectual stages are often compressed into the single phrase "the social-media algorithm." They should not be.

The history is better understood as a sequence of transformations:

Only near the end of that sequence do researchers begin seriously asking the question at the center of this investigation:

What happens when humans do not merely use algorithms, but continually adapt to systems that are themselves continually adapting to humans?

I. Before Social Media: The Intellectual Prehistory

Before the 1990s — Human filtering precedes algorithmic filtering

The underlying problem predates computers.

Human societies have always confronted information scarcity and information overload through:

• editors,
• publishers,
• librarians,
• journalists,
• teachers,
• clergy,
• professional gatekeepers,
• peer groups,
• friendship networks.

Likewise, several psychological and communication concepts relevant to modern recommender systems existed long before social media:

selective exposure — people preferentially seek or attend to information congruent with existing interests or beliefs;

homophily — people disproportionately associate with similar people;

social proof — perceived popularity influences behavior;

agenda setting — media can influence which subjects become salient even without determining what people believe about them.

These older mechanisms matter because they create a major chronological correction.

Algorithmic systems did not invent:

• ideological sorting,
• sensational media,
• peer influence,
• status competition,
• selective exposure,
• commercial competition for attention.

The historical question is therefore not whether those phenomena began with Facebook or TikTok.

It is whether computerized personalization changed their speed, granularity, scale, feedback frequency, or capacity for individual targeting.

That distinction should govern the chronology.

II. 1994: Collaborative Filtering Becomes Explicit

One useful technical starting point is the GroupLens project.

In October 1994, Paul Resnick, Neophytos Iacovou, Mitesh Suchak, Peter Bergstrom, and John Riedl published GroupLens: An Open Architecture for Collaborative Filtering of Netnews at ACM CSCW.

The system helped people navigate Usenet articles using ratings supplied by other users. Its core assumption was straightforward:

people who agreed in the past are more likely to agree again.

Users rated articles; those ratings were aggregated; predicted ratings were then generated for other users. (DOI)

This is an important early layer because the original model was largely:

The system sought to infer what users already liked.

It was not yet designed around continuous behavioral surveillance or an infinitely refreshing attention stream.

Historical significance

The user's role was relatively transparent.

The person explicitly supplied the preference signal.

Later recommendation systems increasingly replaced:

"Tell me what you like"

with:

"I'll infer what you like from what you do."

That transition becomes extremely important.

III. Late 1990s–Early 2000s: Recommendation Becomes Commercial Infrastructure

During the following decade, personalized recommendation migrated from academic prototypes into large consumer services.

Online stores, music services and media platforms increasingly used:

• collaborative filtering,
• similarity models,
• purchase histories,
• click histories,
• item characteristics.

The technical problem remained mainly:

Given too many available items, which few are most relevant to this particular user?

This stage should be distinguished from later social-media recommendation.

The primary goal was often choice assistance:

book → another book song → another song product → another product.

The later social feed introduces something more powerful:

continuously determining which fragments of the social world receive a person's attention.

IV. 2005–2006: The Social Graph Becomes a Feed

Facebook launched News Feed in September 2006.

Its initial significance was not simply algorithmic ranking.

News Feed transformed Facebook from something users largely navigated deliberately—visiting profiles, groups and pages—into a continuously assembled stream of social updates.

The launch itself generated substantial privacy backlash, leading Facebook almost immediately to add additional privacy controls. Facebook's September 8, 2006 announcement explicitly referred to intense user reaction to the newly launched News Feed and Mini-Feed. (About Facebook)

This represents a crucial historical transition:

Earlier

Feed architecture

Even before sophisticated personalization, the feed changed informational agency.

The platform increasingly determined the opportunity set.

V. 2006–2010s: Information Overload Makes Ranking Necessary

As social networks grew, simple chronological presentation became increasingly impractical.

Facebook later described the problem explicitly: by the mid-2010s, users' friends, pages and other connections generated far more material than any individual could consume, so News Feed stories had to be ranked. (About Facebook)

This is an important causal link:

Ranking was therefore not initially imposed merely as a manipulation technology.

There was a genuine information-management problem.

But once ranking existed, another possibility followed:

deciding what information should outrank what other information.

That converted a practical filtering mechanism into a potentially powerful behavioral architecture.

VI. YouTube Before 2012: Click Prediction

YouTube provides one of the clearest documented cases of recommendation objectives changing over time.

By 2012, YouTube stated that its Related and Recommended systems had already operated for years and were attempting to predict what viewers wanted to watch next. (YouTube Blog)

Initially, clicks were a major signal.

That produced an obvious incentive:

But YouTube concluded that clicks were an imperfect indicator.

A compelling thumbnail could generate a click even when the resulting video disappointed the viewer.

This is historically important because platforms were already confronting the exact problem central to our investigation:

Observed behavior may not equal underlying satisfaction.

VII. 2012: YouTube Moves from Clicks Toward Watch Time

In March 2012 YouTube publicly announced a significant change.

It would increasingly prioritize time watched in Related and Recommended videos rather than emphasizing clicks. (YouTube Blog)

By August, YouTube explicitly described its broader goal as encouraging people to spend more time watching, interacting and sharing on the service. (YouTube Blog)

In October, watch-time optimization was extended further into search ranking. (YouTube Blog)

This is one of the most important chronological turning points.

The objective moves approximately from:

toward:

And that changes creator incentives.

YouTube explicitly told creators that videos keeping viewers engaged would receive greater recommendation visibility. (YouTube Blog)

Proposed causal mechanism

This is an early documented example of the platform not merely learning from creators but giving creators incentives to learn the platform.

The co-evolutionary loop is beginning to become visible.

VIII. 2012: ByteDance and the Recommendation-First Model

A parallel development occurred in China.

ByteDance was founded in March 2012 and launched Toutiao later that year. ByteDance's own corporate history identifies Toutiao as one of its earliest products. (ByteDance)

Toutiao's significance in the broader history is that ByteDance developed around algorithmic content recommendation rather than primarily around a pre-existing social graph.

That distinction later becomes central to TikTok.

In a traditional social network:

In a recommendation-first architecture:

The latter potentially gives the recommender greater power over discovery.

IX. 2015: Evidence Begins Separating Human Choice from Algorithmic Filtering

A landmark Facebook study by Eytan Bakshy, Solomon Messing and Lada Adamic examined 10.1 million U.S. Facebook users and their exposure to ideologically diverse news.

The authors distinguished three stages:

1. what people's friends shared,
2. what Facebook's News Feed exposed users to,
3. what users actually clicked.

They found that Facebook ranking reduced exposure to some ideologically cross-cutting material, but users' own choices played an even larger role in limiting what they ultimately consumed. (PubMed)

This study becomes historically important because it begins empirically dismantling a simplistic story:

algorithm → echo chamber.

Instead:

This is an early strong expression of the co-production model.

Source limitation

The authors were employed and funded by Facebook, though the paper states that Facebook did not restrict design or publication beyond its data and ethics requirements. (DOI)

More importantly, it was observational rather than a randomized test of attitude change.

It could show differences in exposure.

It could not establish that ranking produced downstream ideological transformation.

X. 2016: Deep Learning Changes Recommendation Capacity

A major technical shift becomes explicit in YouTube's 2016 paper Deep Neural Networks for YouTube Recommendations.

Paul Covington, Jay Adams and Emre Sargin described a two-stage architecture:

1. candidate generation, selecting promising videos from an enormous corpus;
2. ranking, determining which candidates should receive greatest priority.

They reported substantial performance gains from deep learning. (Google Research)

This matters beyond YouTube.

The old recommender problem—

"people like you liked these items"

—is increasingly superseded by highly nonlinear models processing enormous behavioral datasets.

The user's algorithmic representation becomes less transparent.

The system can combine many signals into predictions that humans cannot easily reconstruct intuitively.

XI. 2016: Instagram Abandons Pure Chronology

In March 2016 Instagram announced that it would reorganize feeds based partly on predicted interest rather than maintaining strict reverse chronology.

The company said users were missing roughly 70% of feed content and proposed ranking according to likelihood of interest, relationship and recency. (Yahoo)

Twitter had introduced a related optional ranked timeline shortly beforehand.

Thus by roughly 2015–2016, ranked feeds were becoming the industry norm rather than a Facebook-specific feature.

The causal sequence was broadly:

But another effect followed:

Creators and publishers increasingly had reason to produce content that recommendation systems would reward.

XII. 2016–2018: TikTok Changes the Importance of the Social Graph

ByteDance launched Douyin in China in September 2016 and launched TikTok outside mainland China in May 2017. ByteDance acquired Musical.ly in November 2017 and merged Musical.ly into TikTok in August 2018. (ByteDance)

TikTok's architecture represented a significant departure from the traditional follower-first model.

The For You feed became the central experience.

Users did not need an established friend or follower network before receiving highly personalized material.

TikTok later described the feed as ranking videos based on:

• user interactions,
• followed accounts,
• comments,
• created content,
• captions,
• sounds,
• hashtags,
• contextual information.

It also stated that stronger behavioral signals, such as completing a longer video, could outweigh relatively weak contextual information. (TikTok Newsroom)

Why this is chronologically significant

Earlier social recommendation was often:

TikTok moved closer to:

This potentially strengthens the path-selection mechanism we identified earlier because the algorithm can introduce entirely new creators and communities instead of merely reprioritizing familiar ones.

That stronger psychological consequence, however, remains a hypothesis rather than an established historical fact.

XIII. 2017–2018: Public Debate Shifts from Filtering to Social Consequences

Following the 2016 U.S. election, concerns about:

• misinformation,
• political polarization,
• computational propaganda,
• algorithmic amplification,
• addictive design,

became far more prominent.

Facebook itself changed ranking objectives in January 2018, announcing that News Feed would prioritize posts expected to generate "meaningful interactions," including back-and-forth discussion among friends. (About Facebook)

Facebook subsequently reduced the priority of some public content and introduced separate quality signals for news. (About Facebook)

This is significant because ranking objectives were no longer framed purely as:

"predict relevance."

They incorporated normative objectives such as:

• meaningful interaction,
• trustworthiness,
• relationship strength,
• perceived value.

This demonstrates that recommender systems are never purely neutral predictors.

Their architects make choices about what type of response counts as success.

XIV. 2017–2019: Researchers Formalize Feedback-Loop Risks

At approximately the same time, recommender-system researchers began explicitly modeling the possibility that systems trained on user responses could alter the behavior from which they subsequently learn.

Chaney, Stewart and Engelhardt developed the concept of algorithmic confounding: recommendation affects consumption, consumption becomes training data, and future recommendations are then based partly upon preferences that the system itself helped produce.

Jiang and colleagues' 2019 Degenerate Feedback Loops in Recommender Systems went further, explicitly analyzing systems in which recommendations can affect user beliefs or preferences, which then affect subsequent feedback. (arXiv)

The intellectual progression becomes:

Early recommender assumption

Feedback-loop model

That second formulation is essentially the theoretical foundation for our current investigation.

But the chronological distinction matters:

theoretical possibility preceded strong real-world causal evidence.

Researchers could model feedback loops before they could reliably measure how powerful they were in real populations.

XV. 2018–2020: TikTok Makes Algorithmic Discovery Culturally Central

TikTok's rapid international growth made personalized discovery an ordinary mass-cultural experience.

When TikTok and Musical.ly merged in August 2018, the company explicitly described the new service as offering a personalized viewing experience and enabling creators to reach audiences without already possessing celebrity status or a large follower base. (TikTok Newsroom)

By 2020 TikTok described For You as central to its user experience and explicitly acknowledged a recognized risk:

personalization could produce an increasingly homogeneous stream.

TikTok said it therefore intentionally inserted diverse recommendations and occasionally surfaced content apparently outside established interests. (TikTok Newsroom)

That is historically important for two reasons.

First, platforms themselves were now publicly acknowledging the exploration–exploitation problem.

Second, the system's designers were intentionally intervening in what might otherwise emerge from pure preference optimization.

In other words:

Platform designers deliberately introduce:

• diversity,
• safety constraints,
• exploration,
• content eligibility rules.

So the feed is never simply a behavioral mirror.

XVI. 2019–2021: Platforms Begin Supplementing Engagement With Satisfaction

Another historical layer complicates simplistic descriptions of "engagement optimization."

Facebook introduced "Worth Your Time" surveys in 2019 and later incorporated direct user feedback more broadly into News Feed ranking. (About Facebook)

YouTube likewise describes incorporating:

• likes,
• dislikes,
• watch history,
• explicit "not interested" feedback,
• satisfaction surveys,

rather than relying only on raw watch time.

Thus the history is not:

in one linear direction.

It is partly:

Whether those corrections are adequate is a separate empirical question.

XVII. 2021: Large-Scale Evidence of Algorithmic Political Amplification

The political consequences of ranking became measurable at much larger scale.

Huszár and colleagues analyzed a long-running Twitter experiment involving a reverse-chronological control group of nearly two million daily active accounts, legislators in seven countries, and millions of U.S. news articles.

They found that mainstream political-right content received greater algorithmic amplification than mainstream-left content in six of seven countries studied.

Importantly, they did not find evidence that the algorithm systematically amplified extreme political parties over moderates. (DOI)

This creates an important chronological correction.

By 2021 there was strong evidence that ranking could alter political visibility.

There was still far less evidence that this visibility difference caused political persuasion.

Those two questions were often conflated in public debate.

XVIII. 2021–2022: "Rabbit Holes" Become an Empirical Research Question

Concerns that YouTube recommendation might progressively draw users toward extremist material produced a growing audit literature.

A systematic review ultimately examined 23 qualifying studies. Fourteen implicated YouTube recommendations in pathways to problematic material, seven produced mixed findings and two did not. (PubMed Central (PMC))

Yet the review emphasized a major methodological barrier:

external researchers did not have full visibility into YouTube's production recommendation system.

Thus by this stage the field had evidence for:

but still lacked strong evidence for the much stronger proposition:

This distinction remains unresolved today.

XIX. 2022: Recommendation Becomes an Explicit Regulatory Object

The European Union adopted the Digital Services Act (Regulation 2022/2065) in October 2022.

Its significance for our chronology is substantial.

Recommender systems ceased being treated only as private product architecture and became an explicit subject of law, transparency obligations and systemic-risk regulation.

The regulation entered into force in 2022 and became broadly applicable from 17 February 2024, with certain provisions applying earlier to designated very large platforms. (EUR-Lex)

This marks a transition:

Earlier

recommendation = product engineering

Increasingly

recommendation = social infrastructure subject to governance

That institutionalization followed years of political controversy, academic research and platform dependence.

The causal mechanism was not simply scholarly influence.

It involved:

• public controversies,
• election integrity debates,
• child-safety concerns,
• misinformation,
• platform market power,
• regulatory pressure,
• emerging empirical research.

No single research finding "caused" the DSA.

It arose from a broader political ecosystem.

XX. 2023: The Meta Experiments Challenge Simplistic Causal Narratives

The 2020 Facebook and Instagram Election Study was conducted during the 2020 U.S. election, but its major findings were published in 2023.

This gap between event date and publication date is itself chronologically important.

Researchers experimentally:

• reduced like-minded content,
• removed reshared material,
• replaced algorithmic ranking with chronological feeds.

The interventions caused substantial changes to users' information environments.

But they produced little detectable change in major political-attitude outcomes during the experimental period. (Nature)

This is perhaps the most important empirical correction in the chronology.

By 2023:

was strongly demonstrated.

But:

did not reliably follow.

The historical narrative therefore had to become more complex.

XXI. An Earlier-Later Layer Problem: The Experiments Tested Mature Users

The Meta experiments contain an important chronological limitation.

Participants in 2020 were not blank slates.

They had already spent years experiencing:

• social media,
• political communication,
• ranked feeds,
• established social networks.

Temporarily changing ranking in 2020 could not undo:

Therefore a three-month null result cannot tell us what the same adults would have become had they spent the previous decade in a different media environment.

But the reverse mistake is equally invalid:

the absence of long-term evidence does not permit us simply to assume such effects occurred.

This remains one of the central missing chronological layers.

XXII. 2024: Regulation Moves From Principles to Enforcement

On 17 February 2024, the DSA became fully applicable to regulated online intermediaries in the EU. (European Commission)

In October 2024 the European Commission specifically requested information from:

• YouTube,
• Snapchat,
• TikTok

about recommender-system design and risks, including:

• civic discourse,
• electoral processes,
• mental wellbeing,
• addictive behavior,
• "rabbit holes",
• protection of minors. (Digital Strategy)

This is historically significant because concepts that originated partly in academic and public debate had now been translated into regulatory categories.

But that institutionalization must not be mistaken for scientific proof.

A regulator asking whether recommender systems produce addictive behavior is evidence that the concern has become institutionally important.

It is not evidence that every proposed causal mechanism has been scientifically established.

XXIII. 2023 Experiment → 2026 Publication: The X Turning Point

One of the most consequential experiments in the present investigation was conducted in 2023 but published in Nature in 2026.

Researchers randomized active U.S. X users between algorithmic and chronological feed conditions for seven weeks.

The algorithmic feed:

• increased engagement,
• increased exposure to conservative political material,
• reduced relative exposure to traditional media,
• shifted some measured political attitudes in a conservative direction.

Crucially, the algorithm also caused users to follow more right-leaning political accounts, and many of those follows persisted. (Nature)

This supplies an empirical bridge missing from much earlier work:

The user is no longer merely consuming algorithmically selected content.

The algorithm helps modify the user's future information architecture.

XXIV. Why the 2026 X Result Does Not Contradict the 2023 Meta Results

Chronologically, it might appear that science changed its mind:

2023

"Algorithms don't change political attitudes."

2026

"Algorithms do change political attitudes."

That would be incorrect.

Different experiments examined:

• different platforms,
• different historical moments,
• different recommendation systems,
• different content distributions,
• different populations,
• different interventions.

Facebook's intervention largely altered ranking within an existing mature social environment.

X's For You system also introduced material from accounts users did not follow.

That means the treatment itself differed.

The most defensible historical interpretation is therefore:

By the mid-2020s, evidence showed that recommender effects are conditional rather than universal. Some ranking interventions substantially alter exposure without changing measured attitudes; others alter exposure, some attitudes, and even network structure.

That is a major maturation of the research field.

XXV. Earlier and Later Layers of Interpretation

The intellectual history can now be separated into distinct layers.

Layer 1 — Prediction, 1990s

Question: What does the user probably like?

Dominant concept:

Layer 2 — Ranking, mid-2000s

Question: Which items deserve scarce attention?

Dominant problem:

Layer 3 — Engagement, early 2010s

Question: Which recommendation produces the strongest behavioral response?

Dominant signals increasingly include:

• watch time,
• clicks,
• retention,
• interaction.

Layer 4 — Deep personalization, mid-2010s

Question: Can large machine-learning systems infer individual interest at enormous scale?

Deep learning and large behavioral datasets dramatically increase prediction capacity.

Layer 5 — Discovery, late 2010s

TikTok-style systems increasingly determine not merely which followed content appears first but what unfamiliar creators and interests users encounter.

Layer 6 — Feedback-loop concern, late 2010s–early 2020s

Researchers increasingly ask:

What happens when recommendation changes the behavior used to train subsequent recommendations?

Layer 7 — Causal social-effects research, 2020s

Field experiments begin separating:

• exposure,
• engagement,
• attitudes,
• political polarization,
• network formation.

Layer 8 — Institutional governance, 2020s

Recommendation systems become explicit subjects of regulation, transparency demands and systemic-risk policy.

Layer 9 — Emerging agentic era, mid-2020s onward

This final stage is only beginning and must remain provisional.

Recommendation systems increasingly coexist with generative AI capable of:

• interpreting information,
• summarizing it,
• conversing about it,
• recommending actions,
• potentially executing decisions.

The likely transition is from:

toward:

But the downstream behavioral consequences remain speculative.

XXVI. Transmission and Intermediaries

There was no single technological lineage in which GroupLens simply became Facebook, which became YouTube, which became TikTok.

The transmission was more complex.

Academic → commercial transmission

Collaborative filtering and machine-learning techniques moved through:

• computer-science research,
• industrial research laboratories,
• recommendation conferences,
• engineering practice,
• rapidly growing consumer platforms.

The ACM Recommender Systems community became an important intellectual intermediary.

Cross-platform competitive transmission

Platforms watch one another.

When ranked and discovery-based feeds prove commercially successful, competitors adopt related product structures.

TikTok itself claimed in 2020 that its For You approach had changed how other technology companies thought about recommendation and discovery. That is a corporate claim rather than independent proof, but the broader shift toward short-form discovery feeds across the industry is historically visible. (TikTok Newsroom)

User → platform transmission

Users supply behavioral data.

Platforms convert those signals into model updates.

Platform → creator transmission

Analytics and distribution outcomes teach creators which forms perform well.

YouTube's 2012 communications provide direct evidence of this route: creators were explicitly told that watch-time performance would affect recommendation visibility and were given analytics to examine it. (YouTube Blog)

Creator → user transmission

Creators adapt content.

That altered content then becomes the environment experienced by users.

User → user transmission

Recommendation introduces users to:

• creators,
• communities,
• ideological groups,
• lifestyles.

Human social influence then takes over.

The 2026 X experiment gives direct evidence that algorithmic exposure can alter whom users follow, demonstrating at least one such transition from algorithmic selection to persistent human network structure. (Nature)

XXVII. The Full Causal Transmission Chain

The mature system can therefore be reconstructed as:

↓

behavior

↓

machine learning

↓

differential visibility

↓

attention/engagement

↓

following/community formation

↓

creator adaptation

↓

new training data

This is the strongest sense in which the system is genuinely co-evolutionary.

The individual arrows are not equally established.

But unlike the strongest speculative claims about identity transformation, the recursive architecture itself is no longer hypothetical.

XXVIII. Chronological Problems

Several problems prevent a clean historical reconstruction.

1. Algorithms have no stable "version"

Unlike a printed text published on a fixed date, recommendation systems change continuously.

YouTube said as early as May 2012 that it had already made more than 100 changes to suggested and recommended-video algorithms during its first seven years. (YouTube Blog)

Thus:

"the YouTube algorithm in 2018"

may itself conceal dozens or hundreds of iterations.

This makes historical causation unusually difficult.

XXIX. Research Is Always Looking Backward

Major findings often appear years after the relevant platform state.

Examples:

2020 Meta intervention → 2023 publication

2023 X intervention → 2026 publication

By the time researchers characterize an algorithm scientifically, that production system may already have changed.

This generates a persistent problem:

The scientific literature is partly an archaeology of recently obsolete systems.

XXX. We Lack the First Decade of Fine-Grained Longitudinal Evidence

Facebook's News Feed dates to 2006.

YouTube was optimizing recommendation extensively by the early 2010s.

TikTok's international recommendation-first model emerged in the late 2010s.

Yet we do not possess comprehensive datasets tracking individuals continuously through:

while simultaneously recording precise recommendation exposure.

Thus the period most relevant to long-term developmental effects is partly undocumented.

XXXI. Historical Platform Changes Confound Human Development

Suppose someone born in 2005 used:

• Facebook in 2016,
• Instagram in 2018,
• TikTok in 2020,
• YouTube throughout,
• generative AI after 2022.

Their behavior cannot be attributed to a stable "algorithmic environment."

The environment itself evolved during their development.

This means:

while both sides change over time.

That is precisely why conventional one-direction causal models struggle.

XXXII. The chronology does not support a simple "algorithms became more manipulative" story

The historical evidence instead shows repeated objective changes.

Platforms moved among:

• clicks,
• watch time,
• engagement,
• satisfaction,
• meaningful interaction,
• diversity,
• trust,
• safety constraints.

TikTok explicitly says it intentionally injects diversity into recommendations partly to avoid repetitive or homogeneous feeds. (TikTok Newsroom)

Facebook incorporated direct "worth your time" feedback. (About Facebook)

So the chronology is not linear technological escalation toward maximal behavioral control.

It is an adaptive arms race among competing objectives:

XXXIII. Revised Chronological Narrative

The most defensible historical account is therefore this.

1990s

Recommendation begins primarily as information filtering.

Systems such as GroupLens attempt to predict preferences explicitly from ratings and similarity. (DOI)

Early–mid 2000s

Large online services confront exploding information abundance.

The feed becomes an important organizational structure; Facebook launches News Feed in 2006. (About Facebook)

Late 2000s–early 2010s

Ranking becomes necessary at scale.

Platforms increasingly infer preferences from behavior rather than explicit ratings.

YouTube's documented 2012 pivot toward watch time demonstrates the move from simple clicks toward deeper engagement measures. (YouTube Blog)

Mid-2010s

Machine learning becomes dramatically more powerful.

YouTube documents deep neural recommendation architecture in 2016. Instagram moves away from strict chronology. (Google Research)

At the same time, early empirical work shows that algorithmic filtering and user self-selection interact rather than operating independently. (DOI)

2016–2018

ByteDance develops a recommendation-first short-video ecosystem; TikTok expands internationally and absorbs Musical.ly. (ByteDance)

Recommendation increasingly determines discovery, not merely ordering.

Late 2010s

Researchers begin explicitly modeling algorithm–user feedback loops.

Platforms also alter objectives in response to concerns about wellbeing, interaction quality, misinformation and user satisfaction.

The conceptual object shifts from:

recommendation system

toward:

sociotechnical feedback system.

Early 2020s

Large-scale audits confirm that ranking measurably alters political exposure.

But systematic research increasingly shows that exposure cannot automatically be equated with persuasion. (DOI)

2022–2024

Recommendation systems become regulatory objects under the EU Digital Services Act.

Society begins treating ranking architecture itself—not merely individual posts—as potentially consequential infrastructure. (EUR-Lex)

2023 research published in 2026

The randomized X experiment provides unusually important evidence that algorithmic exposure can cause not only greater engagement and some political attitude change but persistent modifications to users' following networks. (Nature)

This becomes one of the first strong empirical examples of the transition:

XXXIV. What the Chronology Does—and Does Not—Establish

The chronology strongly supports one conclusion:

Recommendation systems evolved from relatively simple tools for predicting preferences into adaptive social infrastructures that both learn from human behavior and alter the informational environments within which subsequent human behavior occurs.

That development is documented.

The chronology also establishes that creators, users, companies and regulators have progressively adapted to recommendation systems, rather than users being passive recipients.

What chronology does not establish is the stronger proposition:

algorithms have already caused durable transformation of human identity, personality or civilization.

We simply lack the longitudinal evidence necessary to reconstruct that causal sequence.

This creates an interesting asymmetry.

We can document:

with considerable confidence.

But the corresponding human developmental chronology—

has not yet been observed with comparable rigor.

That missing chronology may now be the central evidentiary problem in the investigation.

The most defensible revised narrative

The story is therefore not that algorithms suddenly appeared and began controlling human behavior.

It is that, over roughly three decades, filtering systems gradually acquired four characteristics that did not originally coexist:

continuous behavioral observation,

individualized prediction at enormous scale,

control over discovery as well as ranking,

and

rapid recursive feedback.

Humans simultaneously adapted by learning to:

click for the system, create for the system, perform for the system, resist the system, regulate the system, and increasingly understand themselves through environments partially organized by the system.

The first half of that co-evolution is now historically visible.

The second half—whether repeated exposure accumulates into durable changes in identity, preference formation, social structure and cultural evolution—is only beginning to become measurable.

That is the point at which the chronology ends and the genuinely unresolved investigation begins.

Meaningful Connections Not Yet Fully Explored

The deepest new connections in this investigation are not really between isolated “algorithm effects.” They are between recommendation systems and older theories of choice architecture, performative prediction, cultural evolution, social learning, identity performance, and institutional adaptation.

Several of these connections are already evidential. Others are conceptual bridges that could become major research branches if tested properly.

Direct connections

1. Recommendation systems and choice architecture

This connection is now explicit in the literature.

Recommender systems do not merely predict what a user may want. They also determine the set, ordering, prominence, and sequence of available choices. That puts them squarely within the tradition of “choice architecture” and digital nudging. A 2021 systematic treatment of digital nudging and recommender systems argues that automated recommendation influences decision-making precisely by structuring what information is readily accessible and how it is presented. More recent work on social-media design likewise treats recommender systems as part of a broader choice architecture alongside interface design, affordances, and platform rules. (ScienceDirect)

This adds an important mechanism to our existing model:

rather than only:

That distinction matters because it offers a way for algorithms to shape behavior without ever directly “convincing” anyone.

Alternative explanations: the observed behavior could still reflect pre-existing preferences, because people choose among the options offered. The key causal question is how much the construction of the option set matters relative to the user's underlying disposition.

2. Recommender feedback loops and “performative prediction”

There is a surprisingly close connection to the machine-learning concept of performative prediction.

Ordinary prediction assumes:

while the world being predicted remains independent.

Performative prediction instead recognizes:

That is almost exactly the problem identified by Chaney et al. and Jiang et al. in recommender systems: deployed recommendations affect the user behavior that becomes future training data. (arXiv)

The conceptual connection is now stronger because 2026 work explicitly studies polarization and fairness under performative prediction, where the deployed model changes the data distribution on which future model behavior depends. (AAAI Publications)

This may give us a better technical vocabulary for the whole investigation.

Instead of saying only:

“the algorithm shapes the user,”

we can say:

the prediction target is endogenous to the predictor.

That is a much more precise claim.

3. Engagement ranking and proxy failure

Another direct connection is to a broader optimization problem: the system cannot optimize “what is good for the user” directly, so it uses proxies.

Examples include:

• clicks,
• watch time,
• likes,
• replies,
• shares,
• retention.

The 2025 PNAS Nexus audit demonstrated that Twitter's engagement ranking could increase time/engagement while selecting political tweets users did not actually say they preferred, and could amplify anger and out-group hostility relative to chronological ranking. (OUP Academic)

This links recommender systems to a general principle:

Once the proxy becomes the optimization target, the system may become increasingly good at producing the proxy rather than the underlying human objective.

This resembles Goodhart-type problems, but we should be careful: the investigation has not established a direct historical lineage from Goodhart's Law into recommender design. The connection is structural, not genealogical.

The significant point is that the system may not be manipulating a correctly modeled preference.

It may be amplifying measurement error.

4. Recommendation and social-network construction

The 2026 X experiment makes this connection unusually concrete.

Algorithmic exposure did not merely change which posts users saw. It changed which political accounts they followed, and those follows partly persisted after the experimental condition ended. (PubMed)

That creates a bridge between recommender systems and network formation:

This mechanism is more important than it first appears because it means a temporary algorithm can leave behind a durable human-created structure.

Once a user follows an account, even a chronological feed can continue delivering that account.

So one of the least explored connections is:

algorithmic influence can become embedded in ostensibly nonalgorithmic social structure.

That helps explain why removing an algorithm may fail to reverse its effects.

Structural parallels

5. The algorithmic loop resembles cultural selection

There is now a genuine research bridge between recommender systems and cumulative cultural evolution.

A 2025 peer-reviewed theoretical paper modeled intelligent algorithms as intermediaries in human social learning and found that algorithmic mediation can alter cultural accumulation, with effects depending on the structure of the social network. The authors explicitly situate recommender systems alongside search engines and LLMs as new mediators of cultural transmission. (PubMed)

This strengthens—but does not prove—our earlier cultural-selection analogy:

The structural parallel with biological selection is:

• variation = different content,
• selection = differential visibility/engagement,
• inheritance = imitation/reposting/format copying,
• mutation = modification/remixing,
• environment = platform ranking system.

The analogy becomes especially interesting because the “environment” is itself adaptive.

Thus:

This is more complex than standard cultural evolution.

Alternative explanations

Algorithms may merely accelerate ordinary cultural selection already generated by audience demand.

Or they may introduce a genuinely new selection pressure.

The relevant discriminating test is whether creator populations evolve differently after ranking-objective changes while controlling for audience composition.

6. Creator adaptation parallels evolutionary niche construction

A less obvious connection follows from the previous one.

In evolutionary theory, organisms can alter the environments that subsequently exert selection upon them—niche construction.

Creators do something structurally similar.

Creators:

1. observe which content gets distributed,
2. alter content to improve distribution,
3. collectively change the content environment,
4. generate new user responses,
5. thereby change future training data.

So creators are not simply organisms being selected by the platform.

They partly reconstruct the environment that will select them next.

This gives us:

This is currently a conceptual parallel, not an established application of niche-construction theory to social-media recommenders. But it looks highly productive.

7. Identity performance and the “looking-glass self”

The TikTok identity literature suggests a striking connection to older sociological theories of identity.

Karizat et al. found that users developed folk theories about how TikTok “saw” them and sometimes deliberately altered behavior so that their algorithmic identity better matched their own sense of self. The study explicitly documents changed user performances directed toward the system. (ResearchGate)

Structurally, this resembles Charles Horton Cooley's classic looking-glass self:

people partly form their self-conception through how they imagine others perceive them.

The novel element is that the perceived observer may now be a machine.

The loop becomes:

That is potentially more important than “algorithmic identity” as a standalone label.

The machine becomes a kind of social mirror, even though it has no human consciousness.

Evidentiary status

The behavioral component is supported qualitatively.

The stronger idea that machine-mediated reflected appraisal materially changes long-term identity is still speculative.

8. The recommender as both mirror and environment

Earlier we treated the algorithm as either:

• mirror,
• prism,
• sculptor.

A deeper structural connection suggests those metaphors may be incomplete.

The system can simultaneously be:

a mirror, because it learns from past behavior;

a map, because it classifies the user into predictive categories;

a gatekeeper, because it controls exposure;

an environment, because users live inside repeated outputs;

a teacher, because repeated exposure conveys implicit norms;

a market maker, because it reallocates attention among creators.

This matters because competing interpretations may each be correct at different stages.

For example:

then

then

then

Thus arguments about whether algorithms “reflect” or “shape” humans may be falsely dichotomous because reflection can itself become a mechanism of shaping.

Linguistic and conceptual connections

9. “Preference” has quietly shifted meaning across disciplines

One of the biggest conceptual problems is linguistic rather than empirical.

In economics, a “revealed preference” traditionally refers to preference inferred from choice behavior.

In recommender engineering, however, behaviors such as clicks, watch time, lingering, sharing, and replies may be treated as signals of relevance or preference.

In psychology, those same behaviors may reflect:

• arousal,
• compulsion,
• anger,
• novelty,
• habit,
• attention capture.

The 2025 Twitter audit demonstrates that engagement-defined “revealed preference” can diverge from stated preference. (OUP Academic)

So the same word—preference—may refer to:

This semantic slippage is not trivial.

A large amount of disagreement about whether algorithms “give people what they want” may be partly definitional.

A major research priority should therefore be a formal taxonomy separating:

• behavioral preference,
• reflective preference,
• affective response,
• stated value,
• long-term welfare.

10. “Personalization” may conceal two very different operations

The term personalization often combines:

Retrieval personalization

Find more of what I already appear to like.

and

Discovery personalization

Decide what new things I should encounter.

These are not equivalent.

Retrieval personalization primarily reflects existing preference.

Discovery personalization has much greater potential for path selection.

The X study is relevant because its algorithm introduced content from unfollowed accounts, and that exposure subsequently altered following behavior. (PubMed)

TikTok's identity research also matters because the For You Page is inherently discovery-oriented. (Icahn School of Medicine at Mount Sinai)

This distinction deserves much greater emphasis.

“Personalization” is not one causal treatment.

11. “User control” may itself be a misleading concept

There is an unexpected conceptual connection between regulation and platform design.

Recent analysis of the EU Digital Services Act describes a double choice architecture: regulation gives users controls over recommendation settings, but platforms still design the interface through which those controls are exercised. (Taylor & Francis Online)

So even:

“Let the user choose”

does not eliminate platform mediation.

The platform can still determine:

• which settings exist,
• default values,
• visibility,
• wording,
• complexity,
• reversibility.