Next Research Direction: Humans and Recommendation Algorithms

NEXT RESEARCH DIRECTIONS

Five High-Value Investigations to Test, Deepen, or Overturn the Thesis

Research agenda following the full human-algorithm feedback investigation

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

Prepared as a companion research agenda to the standalone synthesis and full investigation.

The investigation itself already identifies the decisive gap: we have strong evidence for reciprocal behavioral feedback, but weak evidence about what accumulates over years. The missing counterfactual remains, essentially, “What would the same person have become under a different algorithmic environment?”

Highest-Value Research Branches

1. Algorithmic Path Dependence: Can Recommendation Change Which Future Person Emerges?

Why this matters

This is the highest-value branch because it directly tests the strongest unresolved implication of the entire investigation.

behavior → recommendation → exposure → new behavior

different exposure → different interests → different relationships → different opportunities → different future person

The existing investigation specifically identifies multi-year durable divergence as its highest-ranked unanswered question.

Reflection: algorithms eventually discover who the person already was.

Amplification: algorithms strengthen tendencies already present.

Path selection: small early differences cause one of several plausible developmental trajectories to become more likely.

Preference construction: exposure actually produces durable interests or preferences for which little meaningful antecedent existed.

The investigation already developed an unusually clean discriminator: randomly different early recommendations among genuinely new users. Reflection predicts convergence toward baseline preferences; amplification predicts divergence primarily where weak prior tendencies existed; path dependence predicts persistent effects from initially random differences.

Key questions

• Can initially random recommendations create durable interests?
• Do initially similar people become measurably different after years in different recommendation environments?
• Do early algorithmic differences compound or eventually wash out?
• Can genuinely novel interests emerge without detectable baseline preference?
• Are effects strongest during particular developmental periods?
• Does personalization slow natural preference change by repeatedly returning people to their historical behavior?
• Do effects survive removal or resetting of personalization?
• Which outcomes are most path-dependent: hobbies, politics, aesthetics, religion, relationships, career interests, aspirations, or identity?
• Are algorithms actually constructing preferences, or merely discovering latent preferences sooner?
• At what point, if any, does behavioral divergence become meaningful life-course divergence?

Evidence to seek

This branch should deliberately move beyond another literature review of ordinary social-media effects.

• longitudinal cohorts with pre-platform or pre-personalization baselines;
• experiments involving genuinely new platform users;
• platform rollout natural experiments;
• randomized recommendation/exploration experiments;
• recommendation-history reset experiments;
• chronological versus personalized-feed longitudinal studies;
• age-threshold policy natural experiments;
• historical cohorts exposed to different recommendation architectures;
• datasets linking digital exposure to education, employment, hobbies, relationships, political behavior, and offline activities;
• research on preference formation and path dependence outside social media;
• developmental psychology on sensitive periods;
• cultural-evolution research;
• network-formation studies;
• causal-inference research specifically addressing time-varying treatments.

The ideal study identified in the investigation would recruit new users before meaningful personalization begins, establish extensive baseline measurements, randomly vary early recommendation environments, follow participants for years, and periodically alter or remove personalization.

Possible outcomes

Strong convergence: initially different feeds eventually lead similar people toward similar interests.

That would substantially strengthen the reflection model and weaken the strongest version of human-algorithm co-evolution.

Baseline-dependent divergence: recommendation changes trajectories primarily where weak prior interests already existed.

That would favor amplification/reinforcement.

Persistent random divergence: initially arbitrary exposure differences predict later hobbies, communities, relationships, or self-identification.

That would provide major evidence for algorithmic path dependence.

Novel durable preferences: randomized exposure creates persistent interests with little detectable baseline antecedent.

That would be the most consequential finding. Recommendation would have to be treated as a genuine mechanism of preference formation rather than principally preference discovery.

It could justify placing algorithmic environments alongside traditional developmental institutions such as family, peers, school, neighborhood, religion, and mass media. The existing investigation explicitly identifies that as the implication of strong durable effects.

2. The Network Hypothesis: Does the Algorithm Matter Most Because of Whom It Introduces Us To?

Why this matters

One of the investigation's most important discoveries was that persuasion may be the wrong unit of analysis.

The 2026 X experiment supplied an important causal bridge:

recommendation → discovery → follow → persistent network change

The follow can remain after the original recommendation disappears.

That raises a much larger possibility.

Perhaps algorithms do not primarily transform people by repeatedly showing persuasive content.

Perhaps they transform people's environments by introducing them to other humans.

The long-run mechanism might instead be:

Algorithm → creator/community discovery → follow → repeated social contact → belonging → norms → identity/behavior

If so, much algorithm research is concentrating on the wrong outcome.

Key questions

• How often do algorithmically induced follows persist?
• Do those follows subsequently change behavior or beliefs?
• Does passive repeated exposure have weaker long-run effects than joining a community?
• Which platforms produce the greatest network restructuring?
• Are effects stronger for socially isolated users?
• Do algorithmically introduced relationships migrate offline?
• Can recommendation alter friendship, romantic, professional, religious, political, or hobby networks?
• Does removing the algorithm reverse effects after the social network has already changed?
• Do communities mediate apparent "algorithm effects"?
• Are algorithms better conceptualized as social matchmakers than persuaders?

Evidence to seek

• replications of the X following effect;
• longitudinal follow/unfollow datasets;
• randomized experiments separating exposure from follow/community capability;
• Discord, Reddit, Facebook Groups, TikTok, YouTube, Instagram, X, and other community-discovery systems;
• network-science studies of homophily versus influence;
• socialization and peer-effects research;
• studies linking online introductions to durable offline relationships;
• community membership histories;
• migration between platforms;
• qualitative studies of how people first discovered important communities.

A particularly discriminating experiment already emerged from the investigation: randomize users between repeated exposure without social-connection options, exposure with following capability, and exposure with active community connection. If persistent effects concentrate in the latter conditions, the network hypothesis becomes much stronger.

Possible outcomes

If network changes explain little downstream variance, direct attention and content mechanisms regain importance.

If algorithm-induced network differences persist but attitudes do not, recommendation may primarily restructure social opportunity rather than psychology.

If network changes mediate substantial long-term changes in beliefs, interests, behavior, or identity, the investigation would undergo a major conceptual shift:

The most important question would cease to be “What did the algorithm show you?” and become “Whom did the algorithm cause you to meet?”

That possibility was already identified as potentially requiring a shift from conventional media-effects research toward sociology, network science, developmental psychology, and anthropology.

3. The Creator-Side Loop: Are Algorithms Changing People Indirectly by Changing Culture?

Why this matters

Most research asks what recommendation does to consumers.

But the algorithm simultaneously acts upon producers.

Creators learn what receives distribution and adapt accordingly. The investigation therefore reconstructed a larger chain involving differential visibility, attention, following, creator adaptation, and new training data.

This creates a powerful alternative mechanism:

Algorithm → creator incentives → changed content supply → changed cultural environment → users

The implication is striking.

Individual users might be psychologically resistant to recommendation while nevertheless inhabiting a cultural environment increasingly shaped by recommendation incentives.

This could explain how algorithms produce large societal consequences despite modest average persuasion effects.

Key questions

• What changes after major ranking-policy changes?
• Do creators alter topics, emotionality, duration, aesthetics, vocabulary, thumbnails, pacing, controversy, or posting frequency?
• Do creators consciously imitate algorithmically successful competitors?
• Does content become more homogeneous in form even while topics become more diverse?
• Do ranking incentives increase emotional extremity?
• Which creator adaptations survive when algorithms change again?
• How quickly do cultural production norms diffuse across platforms?
• Does creator adaptation precede changes in audience demand?
• Are creators responding to genuine audience preference or platform-generated incentives?
• Can recommendation systems change culture even among users who disable personalization?

Evidence to seek

The current investigation explicitly identifies the missing evidence: creator drafts, upload histories, analytics dashboards, recommendation exposure, monetization data, editing decisions, and interviews surrounding major ranking changes.

Seek natural experiments involving major changes such as:

• YouTube's historical optimization changes;
• Instagram feed/recommendation changes;
• TikTok distribution changes;
• X algorithm changes;
• demonetization or monetization-policy shocks;
• introduction/removal of short-form video;
• platform migrations;
• creator analytics before and after ranking shocks.

The critical design is difference-in-differences around ranking changes, comparing heavily affected creators with less-affected controls.

Possible outcomes

If creators barely change production after major ranking shocks, the cultural-supply hypothesis weakens substantially.

If formats change but underlying ideas do not, algorithms may primarily standardize presentation.

If topics, emotionality, ideology, or creative investment change systematically, recommendation becomes a much more consequential cultural institution.

The strongest result would show:

ranking change → creator adaptation → changed content supply → later audience change

That would demonstrate a general-equilibrium mechanism almost entirely missed by user-level feed experiments.

4. What Does the Algorithm Actually Learn: Preference, Attention, or Reflex?

Why this matters

The entire recommendation ecosystem rests upon a deceptively simple assumption:

Behavior contains information about what the user wants.

But the investigation found that the algorithm does not observe "preference" directly. It observes behavior generated partly by its own previous interventions.

A user can watch something because they love it, hate it, fear it, are confused by it, cannot look away, or simply because the system repeatedly placed it before them.

This branch asks whether the fundamental data-generating process behind personalization has been conceptually misunderstood.

Key questions

• How closely does engagement correlate with stated preference?
• What about reflective preference after deliberation?
• What about later satisfaction?
• What about wellbeing?
• Which signals best distinguish curiosity from endorsement?
• Are outrage, fear, sexual interest, novelty, and social comparison particularly easy to misinterpret?
• Does completion mean satisfaction or compulsion?
• What happens when users choose the objective themselves?
• Would people select "surprise me," "teach me," or "challenge me" instead of "maximize engagement"?
• Can recommender systems optimize simultaneously for immediate enjoyment and long-term goals?

Evidence to seek

Seek randomized comparisons among:

engagement optimization / stated-preference optimization / satisfaction optimization / diversity-exploration optimization / user-selected objectives / wellbeing-oriented recommendation

Also prioritize research involving delayed satisfaction surveys, regret, content hiding, repeated choice, voluntary disabling, counterfactual recommendations, and behavioral versus reflective preference.

The existing investigation specifically proposed allowing users to choose objectives such as "show me what I usually like," "surprise me," "challenge me," "teach me," or "minimize outrage." If users systematically choose environments different from engagement optimization, that would be important evidence that observed engagement is an imperfect representation of reflective preference.

Possible outcomes

If engagement reliably predicts later satisfaction across domains, concerns about proxy misalignment would narrow.

If divergence occurs only in politics or outrage-heavy domains, the problem is contextual rather than intrinsic.

If substantial divergence appears broadly, the interpretation of personalization changes profoundly.

The system may not be learning:

“What do I want?”

but rather:

“What reliably produces a measurable response from me?”

That would sharpen the distinction between a predictive model of behavior and a model of the person.

5. From Recommendation to Delegation: What Happens When the Algorithm Begins Acting for Us?

Why this matters

This branch moves beyond the historical investigation into its most consequential emerging extension.

Today's recommendation loop largely says:

“Here is something you may want.”

Agentic systems increasingly introduce:

“Here is what I think you should choose.”

and eventually:

“I chose and acted for you.”

The existing investigation already identified this as a potentially major research question of the 2030s and emphasized that the evidence is presently sparse because the technology is emerging.

Recommendation allocates opportunities for choice.

Delegation can allocate choices themselves.

That is a qualitative change.

Key questions

• Does repeated AI delegation reduce independent information seeking?
• Does it improve decision quality?
• Does it produce skill atrophy?
• Does trust generalize from low-stakes to high-stakes domains?
• Does personalization become stronger when an AI possesses memory across domains?
• Does the system gradually become a mediator between the individual and the world?
• What happens when recommendation incorporates explicit goals rather than inferred engagement?
• Can delegation paradoxically increase autonomy by freeing cognitive resources?
• How do people calibrate trust after correct and incorrect AI decisions?
• Does an agent eventually influence preferences by controlling which alternatives are ever considered?

Evidence to seek

information selection — “Here are five options.”

decision support — “I recommend option B.”

delegation — “I selected B for you.”

Measure independent search, decision quality, confidence, learning, skill retention, trust calibration, regret, autonomy, time saved, preference change, and willingness to delegate increasingly consequential decisions.

This branch should eventually include AI assistants, shopping agents, travel agents, financial assistants, educational tutors, health-navigation systems, workplace agents, and personalized media systems.

Possible outcomes

Delegation might produce substantial efficiency while leaving autonomy intact.

It might increase autonomy by removing low-value cognitive work.

Or repeated delegation might progressively reduce independent search and skill while making the agent increasingly influential over the opportunity set itself.

The last possibility would transform the original feedback loop:

Human behavior → algorithm → recommendation → human choice

into:

Human goals + behavioral history → agent → action → changed environment → future goals

That would represent a substantially deeper form of human-machine coupling than the systems investigated so far.

Priority Ranking

There is an important reason I would not make adolescent mental health, polarization, radicalization, filter bubbles, or cultural homogenization one of the five primary branches. They are valuable questions, but each investigates a particular outcome. The five above investigate the underlying causal architecture. If we understand that architecture, we can return to those outcomes with much greater precision.

Branch #1 stands clearly above the rest because it attacks the missing counterfactual identified throughout the investigation: whether different algorithmic environments create genuinely different future people. The source investigation itself says that long-duration randomized evidence, pre-personalization baselines, full exposure histories tied to offline outcomes, and cross-platform persistence evidence are precisely what is missing.