Investigative Report Technical Report · 2025-01-15

Reconstructing the Hidden Objective Functions of Modern Personalized Feeds

InTelluric / Alnitak Group — Los Angeles, CA

Abstract

The public record does not expose the exact production reward functions used by modern personalized feeds, but it exposes enough architecture, metrics, and organizational behavior to reconstruct the class of objective functions that plausibly drove them. Across official disclosures by YouTube, Meta Platforms, and TikTok, the recurring pattern is a multi-stage system: retrieve candidates from very large corpora, score them with many signals, optimize several predicted actions, and rerank under latency, diversity, and policy constraints.

The strongest reconstruction is that the hidden objective function was not a single scalar like click-through rate. It was a composite continuity objective: maximize the probability that a user stays in the platform-controlled behavioral loop, returns soon, stays longer, generates more predictable feedback, and remains monetizable. The YouTube record is especially clear. Its 2012 post states directly that discovery features shifted away from driving views and toward increasing time spent watching — "not only on the next view, but also successive views thereafter" — with more watching also opening more revenue opportunities.

That continuity-maximization reconstruction matters because it matches the observed harm cluster better than the older "screen time" framing. The clearest public-health pattern is not all internet use versus no internet use. It is high-frequency, passive or semi-passive, socially evaluative, personalized, recommendation-driven exposure at adolescent developmental stages. CDC data show the suicide rate for ages 10–24 rose 62% from 2007 to 2021. The U.S. Surgeon General's advisory states that children and adolescents using social media more than three hours per day face double the risk of poor mental-health outcomes including depression and anxiety symptoms.

The behavioral-science lineage strengthens rather than weakens the reconstruction. A system that reliably identifies and sequences prompts tied to curiosity, peer validation, self-comparison, uncertainty, outrage, and relief can shape time allocation, sleep timing, affective state, and repeated re-entry while remaining fully compatible with users experiencing the behavior as self-directed. Adolescent susceptibility is documented: brain regions linked to attention, feedback, and reinforcement from peers become increasingly sensitive around age 10.

The most useful investigative conclusion is narrower and stronger than generic blame. The hidden objective function can be reconstructed as a continuity-maximizing, monetization-compatible, risk-managed control system acting on individualized behavioral forecasts. The report outputs a prioritized FOIA request plan, litigation discovery target list, and falsification tests for the reconstruction.

Key Terms

hidden objective function production reward surface watch time optimization session continuity successive views continuity maximization engagement maximization behavioral entrainment variable ratio reinforcement personalized recommendation algorithms recommendation system audit algorithmic amplification DARPA Narrative Networks DARPA SMISC IARPA OSI YouTube 2012 watch time shift adolescent mental health youth suicide rates post-2012 CDC YRBS Surgeon General advisory Murthy v. Missouri persuasive technology multi-objective ranking two-tower retrieval For You page ranking Instagram Explore funnel News Feed objective function long-term value optimization ad total value auction sequence learning ads

Key Findings

—The production reward surface across YouTube, Meta, and TikTok is reconstructible as a continuity-maximizing objective — maximize expected watch time, session depth, return frequency, and monetizable attention — with safety constraints layered on later as regulatory defense, not primary reward design.
—YouTube's 2012 shift from view-maximization to watch-time-and-successive-views optimization is a documented, public statement of long-horizon session objective. It precedes the measurable inflection in adolescent mental-health indicators by approximately one to three years.
—The DARPA Narrative Networks and SMISC programs, documented 2011–2012, describe capabilities — narrative influence on cognition, sentiment analysis, social-media-scale information-flow modeling — that are functionally isomorphic with feed recommendation objectives. Direct causal linkage to production reward functions is not established by open sources; institutional overlap and method convergence are.
—The mechanism does not require a label called "depression." A reward mix that concentrates exposure on emotionally sticky, socially evaluative, self-referential, or difficult-to-disengage stimuli is sufficient to shift population-scale mental-health indicators without the system having explicit harm targets.
—The most accessible investigative path to the hidden reward function is not source code. It is the internal metric and experiment vocabulary — session_length, long_watch, next_view_rate, return_24h, time_well_spent — which constitutes the operational definition of "better" and is recoverable through FOIA, discovery, and procurement records.

Reconstructed Production Score

Derived from YouTube 2010/2012/2016 disclosures, Meta News Feed and Instagram Explore documentation, TikTok factor description, and Meta ads-auction publication. Not proprietary code — the minimal objective class consistent with all public disclosures simultaneously.

Score(u, i, t) =
  Σ_k w_k · P(action_k | u, i, t)    // multi-action engagement
  + α · E[watch_time | u, i, t]      // session depth
  + β · P(next_view | u, i, t)       // successive views
  + γ · P(return_within_horizon | u)  // return frequency
  + δ · MonetizationValue(u, i, t)   // ad yield
  - λ · PolicyRisk(u, i, t)         // regulatory/brand defense
  - μ · Redundancy(i, slate_t)       // diversity constraint
  + ν · Freshness(i, t)             // recency signal

Lineage Convergence Timeline

1997Stanford Behavior Design Lab formalizes persuasive technology and ethics frameworks. Identifies motivation, ability, and prompts as the three necessary conditions for behavior change.
2005YouTube launches.
2010YouTube publishes recommendation system paper. Co-visitation graph, seed expansion, multi-hop candidate generation, A/B metrics including session length, long CTR, time until first long watch.
2011IARPA OSI program introduced: continuous real-time monitoring of social media, web search, news feeds, internet traffic, and Wikipedia edits to anticipate significant societal events.
2011DARPA SMISC BAA published. Objectives: misinformation detection, linguistic cues, sentiment and opinion analysis, information-flow modeling, narrative influence on cognition and behavior.
2011DARPA Narrative Networks BAA published. Objectives: how narratives influence cognition and behavior in security contexts including radicalization and social mobilization.
2012YouTube shifts discovery objective from view-maximization to watch time and successive views. Public statement: "not only on the next view, but also successive views thereafter." Monetization relevance explicitly linked.
2016YouTube deep-learning paper formalizes deep candidate generation and deep ranking. Presented "at a high level" only. Production reward mix not disclosed.
2021Meta publicly frames News Feed ranking as objective-function optimization toward "long-term value." Multiple ML predictions aggregated into single score per user.
2023Instagram Explore four-stage retrieval/ranking funnel published. Two-Tower multi-objective retrieval, tunable source weights, real-time and pre-generated candidates.
2024CDC YRBS analysis published: frequent social-media use associated with bullying victimization, persistent sadness or hopelessness, seriously considering suicide, making a suicide plan.
2024European Commission requests recommender-system parameters from YouTube, Snapchat, and TikTok under DSA, specifically addressing contribution to addictive behavior and mental well-being risk.

Investigative Outputs

The report includes a prioritized FOIA request plan, litigation discovery target list, procurement records methodology, and documentary leads by lineage. Primary targets:

—HHS, Surgeon General, CDC, FBI, CISA: platform-meeting calendars, slide decks, "insights" reports, recommendation-algorithm discussions. Date range 2020–2023 per Murthy record.
—DARPA and IARPA: SMISC, Narrative Networks, OSI solicitations, performer lists, technical reports, transition documents. Procurement records for social-media analytics and narrative analysis contractors.
—Litigation discovery: experiment registries, ranking dashboards, model-card repositories, youth-wellbeing review decks, feature-store schemas containing session_length, long_watch, next_view_rate, return_24h, time_well_spent, healthy_session.