pith. machine review for the scientific record. sign in

arxiv: 2604.14832 · v1 · submitted 2026-04-16 · 💻 cs.SI · cs.CY

Recognition: unknown

Seeking Help, Facing Harm: Auditing TikTok's Mental Health Recommendations

Amir Ghasemian, Homa Hosseinmardi, Pooriya Jamie

Authors on Pith no claims yet

Pith reviewed 2026-05-10 08:42 UTC · model grok-4.3

classification 💻 cs.SI cs.CY
keywords TikTokrecommender systemsmental health recommendationsalgorithmic audituser intentcontent exposuresocial media platforms
0
0 comments X

The pith

TikTok's recommendation system saturates feeds with mental health content based on user engagement patterns rather than search intent.

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The paper runs a seven-day controlled audit of TikTok's For You page with thirty new accounts and simulated agents that begin with either distress-focused or help-seeking searches. It then applies three different interaction styles: active engagement, avoidance, or passive viewing. The results show that engagement behavior shapes exposure outcomes far more than the initial search type. Active users receive mental health videos in roughly 45 percent of daily recommendations, while avoidance and passive approaches cut this to 11-20 percent but leave some exposure intact. Help-seeking searches tilt the mix toward more supportive content, yet low levels of potentially harmful material, including suicide and self-harm videos, still appear across all conditions.

Core claim

The controlled seven-day experiment with thirty fresh accounts shows that interaction behavior is the dominant factor in exposure: active engagement leads to approximately 45 percent of daily recommendations being mental health content, whereas avoidance and passive viewing lower this share to between 11 and 20 percent without removing it entirely. Framing the initial search as help-seeking increases the share of potentially supportive videos relative to distress-focused searches, but videos from potentially harmful categories, including Suicide/Self-Harm, continue to appear at low but non-zero rates in all conditions.

What carries the argument

The experimental audit that systematically varies initial search framing between distress and help-seeking queries alongside three interaction strategies across multiple new accounts to measure resulting changes in the composition of the For You page.

If this is right

  • Active engagement with mental health content quickly increases its prevalence in future recommendations.
  • Avoidance behaviors reduce but do not stop exposure to such content.
  • Help-oriented searches lead to more supportive content but do not eliminate potentially harmful videos.
  • Recommendation systems show weak response to explicit user intent signals for mental health topics.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • These patterns may indicate that engagement metrics override other signals in algorithmic decisions for sensitive content.
  • Similar audits on other social media platforms could reveal whether this is a general issue with recommendation engines.
  • Platform designers might consider adding user-set preferences for mental health content sensitivity to improve outcomes.

Load-bearing premise

That the simulated agents accurately copy real human search choices, timing, and interpretations, and that manual and automated labeling of videos into supportive versus harmful categories is accurate and free of bias.

What would settle it

A direct comparison of exposure rates between the simulated accounts and data from real TikTok users who document their own search framing and daily interaction habits with mental health topics.

Figures

Figures reproduced from arXiv: 2604.14832 by Amir Ghasemian, Homa Hosseinmardi, Pooriya Jamie.

Figure 1
Figure 1. Figure 1: Cross-sectional exposure patterns across personas. (A) MH saturation (proportion of MH-relevant videos among 40 daily recommendations; dashed line is the global mean). (B) Average daily counts of potentially harmful vs. potentially supportive content. (C) Harm subtype proportions within potentially harmful content. ing a random sample of 100 videos (approximately 1% of the corpus), independently annotated … view at source ↗
Figure 2
Figure 2. Figure 2: Feed evolution (Days 1–7). (A) MH ratio over time by persona. MH-Engaged saturates while MH-Avoidant and Passive-Observer drop after Day 1 but remain non-zero. (B) Potentially supportive and (C) potentially harmful ratios, with harmful exposure persisting across conditions. Error bars show 95% CIs. Across all personas, Toxic Positivity and Self-Diagnosis ac￾count for the largest shares of potentially harmf… view at source ↗
read the original abstract

Recommender systems on social media increasingly mediate how users encounter mental health content, yet it remains unclear whether they distinguish help-seeking from distress expression. We conduct a controlled 7-day audit of TikTok's "For You" page using 30 fresh accounts and LLM-guided agents that vary initial search framing (distress- vs. help-initiated) and interaction strategy (engaged, avoidant, passive). Across 8,727 recommended videos, interaction behavior dominates exposure outcomes: engagement rapidly saturates feeds with mental health content (~45% of daily recommendations), while avoidance and passive viewing reduce but do not eliminate exposure (~11-20%). Search framing mainly shifts composition rather than volume--help-initiated searches yield more potentially supportive material, yet potentially harmful content persists at low but non-zero levels, including content in the Suicide/Self-Harm category. These findings suggest limited sensitivity to user intent signals in TikTok's recommendations and motivate context-aware safeguards for sensitive topics.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Referee Report

3 major / 2 minor

Summary. The manuscript reports results from a 7-day controlled audit of TikTok's For You page using 30 fresh accounts and LLM-guided agents that vary initial search framing (distress- vs. help-initiated) and interaction strategy (engaged, avoidant, passive). Across 8,727 recommended videos, the central claim is that interaction behavior dominates exposure: engagement saturates feeds with mental health content (~45% of daily recommendations), avoidance and passive viewing reduce but do not eliminate it (~11-20%), while search framing mainly shifts composition rather than volume, with potentially harmful content (including Suicide/Self-Harm) persisting at low but non-zero levels.

Significance. If the measurements hold, the work supplies direct empirical evidence that TikTok's recommender shows limited sensitivity to user intent signals on mental health topics. The controlled multi-account, multi-strategy design is a strength, enabling reproducible comparisons of exposure under varied conditions and motivating platform safeguards. The findings contribute to the literature on algorithmic auditing of sensitive content recommendation.

major comments (3)
  1. [Methods] Methods (audit design and video labeling): The abstract and methods description provide no details on video classification methods, inter-rater reliability, or human validation for labeling videos into mental health, supportive, harmful, or Suicide/Self-Harm categories. These percentages (e.g., ~45%, 11-20%) are load-bearing for the dominance-of-interaction claim; without them the quantitative results cannot be assessed for bias or accuracy.
  2. [Methods] Methods (LLM agent simulation): No validation, robustness tests, or comparison to real-user behavior is reported for the LLM-guided agents' reproduction of search framing, interaction timing, dwell time, or avoidance actions. This assumption is central to the headline result that interaction strategy dominates search framing; failure here would collapse the reported exposure differences.
  3. [Results] Results (statistical controls): The abstract reports no controls or checks for account age, regional effects, or other confounders in the 30-account design. This is load-bearing for cross-strategy comparisons, as unaccounted variation could explain the observed saturation differences rather than interaction behavior alone.
minor comments (2)
  1. [Abstract] Abstract: The total video count (8,727) is stated but could be paired with per-condition breakdowns for immediate context on sample sizes underlying the ~45% and 11-20% figures.
  2. The manuscript would benefit from an explicit limitations subsection addressing potential artifacts from automated agents and labeling, even if brief.

Simulated Author's Rebuttal

3 responses · 0 unresolved

We thank the referee for their constructive and detailed review. The comments highlight important areas for clarification in the methods and results. We address each point below and have revised the manuscript to strengthen the presentation of our audit design and findings.

read point-by-point responses
  1. Referee: [Methods] Methods (audit design and video labeling): The abstract and methods description provide no details on video classification methods, inter-rater reliability, or human validation for labeling videos into mental health, supportive, harmful, or Suicide/Self-Harm categories. These percentages (e.g., ~45%, 11-20%) are load-bearing for the dominance-of-interaction claim; without them the quantitative results cannot be assessed for bias or accuracy.

    Authors: We agree that the original submission omitted key details on the labeling pipeline. The revised manuscript now includes an expanded Methods subsection describing the hybrid classification approach: an initial LLM-based classifier (fine-tuned on a manually annotated seed set of 500 videos) followed by human review of a 10% random sample per category. We report inter-annotator agreement (Cohen's kappa = 0.81 across the four primary categories) and provide the annotation codebook in the appendix. These additions allow readers to evaluate potential bias in the reported exposure percentages. revision: yes

  2. Referee: [Methods] Methods (LLM agent simulation): No validation, robustness tests, or comparison to real-user behavior is reported for the LLM-guided agents' reproduction of search framing, interaction timing, dwell time, or avoidance actions. This assumption is central to the headline result that interaction strategy dominates search framing; failure here would collapse the reported exposure differences.

    Authors: The paper describes the agent parameters (search queries, dwell-time distributions, and action probabilities) as calibrated to publicly available TikTok usage statistics and platform documentation. Direct head-to-head validation against real-user traces is not possible without access to proprietary platform logs. In the revision we have added a dedicated limitations paragraph and a sensitivity analysis varying dwell times and avoidance thresholds by ±20%, confirming that the relative ordering of conditions (engaged vs. avoidant/passive) remains stable. We also note that the controlled, within-platform design still supports causal comparisons across the tested strategies even if absolute behavior deviates from any single real user. revision: partial

  3. Referee: [Results] Results (statistical controls): The abstract reports no controls or checks for account age, regional effects, or other confounders in the 30-account design. This is load-bearing for cross-strategy comparisons, as unaccounted variation could explain the observed saturation differences rather than interaction behavior alone.

    Authors: All 30 accounts were created on the same day, assigned identical device and location settings, and aged uniformly before data collection began. The revised Results section now explicitly states these controls and includes a supplementary table confirming no statistically significant baseline differences in initial feed composition across the six experimental cells (Kruskal-Wallis p > 0.4). We further added a mixed-effects regression controlling for day-of-experiment and account ID as random effects; the interaction-strategy coefficients remain significant and larger in magnitude than the search-framing coefficients. These additions address the concern about unmeasured confounders. revision: yes

Circularity Check

0 steps flagged

No circularity: direct empirical counts from controlled audit

full rationale

The paper reports observational results from a 7-day TikTok audit using 30 accounts and LLM-guided agents, yielding 8,727 videos whose mental-health content shares are tallied directly (e.g., ~45% under engagement, 11-20% under avoidance). No equations, fitted parameters, or derived quantities appear; the central claims are raw proportions and composition shifts obtained from the collected recommendations. No self-citations to prior author work are invoked as load-bearing premises, and no uniqueness theorems or ansatzes are smuggled in. The derivation chain is therefore the data-collection and counting procedure itself, which is self-contained and does not reduce to its own outputs by construction.

Axiom & Free-Parameter Ledger

0 free parameters · 2 axioms · 0 invented entities

The central claims rest on the untested fidelity of LLM agents as user proxies and on the validity of content categorization; no numerical free parameters or new physical entities are introduced.

axioms (2)
  • domain assumption LLM-guided agents can accurately simulate human search framing and interaction behaviors on TikTok
    Invoked to create the distress- versus help-initiated and engaged/avoidant/passive conditions.
  • domain assumption Videos can be reliably labeled into supportive, harmful, and Suicide/Self-Harm categories
    Required to report the ~45%, 11-20%, and non-zero harmful percentages.

pith-pipeline@v0.9.0 · 5468 in / 1447 out tokens · 59507 ms · 2026-05-10T08:42:38.118806+00:00 · methodology

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

6 extracted references · 3 canonical work pages

  1. [1]

    arXiv:2503.20231

    Dynamics of Algorithmic Content Amplification on TikTok. arXiv:2503.20231. Boeker, M.; and Urman, A

  2. [2]

    InProceedings of the ACM Web Conference 2022, WWW ’22, 2298–2309

    An Empirical Investiga- tion of Personalization Factors on TikTok. InProceedings of the ACM Web Conference 2022, WWW ’22, 2298–2309. New York, NY , USA: Association for Computing Machin- ery. Coppersmith, G.; Ngo, K.; Leary, R.; and Wood, A

  3. [3]

    InPro- ceedings of the 2016 CHI conference on human factors in computing systems, 2098–2110

    Discovering shifts to suicidal ideation from mental health content in social media. InPro- ceedings of the 2016 CHI conference on human factors in computing systems, 2098–2110. Eltaher, F.; Gajula, R. K.; Miralles-Pechu ´an, L.; Crotty, P.; Mart´ınez-Otero, J.; Thorpe, C.; and McKeever, S

  4. [4]

    arXiv:2505.11160

    Pro- tecting Young Users on Social Media: Evaluating the Ef- fectiveness of Content Moderation and Legal Safeguards on Video Sharing Platforms. arXiv:2505.11160. Franklin, J. C.; Ribeiro, J. D.; Fox, K. R.; Bentley, K. H.; Kleiman, E. M.; Huang, X.; Musacchio, K. M.; Jaroszewski, A. C.; Chang, B. P.; and Nock, M. K

  5. [5]

    I See Me Here

    “I See Me Here”: Mental Health Content, Commu- nity, and Algorithmic Curation on TikTok. InProceedings of the 2023 CHI Conference on Human Factors in Comput- ing Systems, CHI ’23. New York, NY , USA: Association for Computing Machinery. Mosnar, M.; Skurla, A.; Pecher, B.; Tibensky, M.; Jakub- cik, J.; Bindas, A.; Sakalik, P.; and Srba, I

  6. [6]

    arXiv:2403.12410

    TikTok and the Art of Personal- ization: Investigating Exploration and Exploitation on Social Media Feeds. arXiv:2403.12410. Xue, L.; Corso, F.; Fontana, N.; Liu, G.; Ceri, S.; and Pierri, F