PaperlayerView exact revision on arXiv

Source-linked AI summary

Fear Doesn't Go Viral, It Lingers: Spread, Affect, and Discourse of Food-Fear Content on TikTok

Hunjun Shin, Hongpeng Fu, Divya Chaudhary, Mohit Singhal

arXiv:2610.33898v1cs.SIcs.CY

TL;DR

Prior work has not examined how everyday food-fear content spreads or how audiences respond to it. This study analyzes TikTok food and nutrition videos, comments, and replies, finding that fear videos reach fewer people but are shared more per view and discussed longer. Their threads are emotionally negative but contested, and fear frames broaden over time.

  • Problem

    Prior work has focused on political and crisis misinformation and health-video accuracy, leaving the spread and audience responses to everyday food-fear content unexamined.

    Source: §1 · pp. 1–2

  • Method

    The study labels 97,234 TikTok food and nutrition videos with a zero-shot multimodal classifier validated against human annotation, then analyzes comments, replies, and discourse.

    Source: §3 · p. 3, §6 · p. 7

  • Results

    Fear videos get 62% fewer views but 1.7–2.1× more shares per view and 1.6× longer discussions, while their threads are more negative but do not reinforce negativity more than non-fear threads.

    Source: §1 · p. 2, §8 · p. 8

  • Takeaways & Limitations

    Persistence-based signals may better identify food-fear content than signals that wait for viral reach, and its contested threads offer a setting for soft moderation and community-based correction.

    Source: §8 · p. 9

    The corpus is keyword-defined rather than a random sample of TikTok food content, comments are incompletely collected, and recommendation-feed exposure is unobserved.

    Source: §7 · p. 8

Abstract

from arXiv · show

TikTok has become a primary channel for nutrition information, and alongside evidence-based content, it hosts a growing stream of \emph{fear-based} wellness messaging that frames ordinary foods as ``toxic,'' ``poison,'' or things to ``never eat.'' Prior work has extensively studied political and public-health misinformation and audited the accuracy of health videos, but to the best of our knowledge, no work has examined how everyday food-fear content spreads and how audiences respond to it. To investigate this, we collected 97,234 TikTok food and nutrition videos, labeled them \textit{Fear} or \textit{Non-Fear} with a zero-shot multimodal classifier built on Gemini~3 Flash that is validated against human consensus (Cohen's $\kappa=0.81$), and analyzed more than 1.1M comments and 272K parent-reply pairs. Contrary to the intuition that fear ``goes viral,'' fear videos reach fewer people than non-fear videos. Their distinctive signature lies in dynamics: they are shared more per view (1.7--2.1$\times$ after controls) and their discussion persists longer (comment half-life 1.62$\times$ after controls). These shifts are modest in size but hold after controlling for reach. Comments on fear videos are more negative, angrier, and more disgusted, yet negativity is \emph{not} more self-reinforcing: negative comments draw negative replies at the same rate as in non-fear threads, while positive comments are far less likely to stay positive. Fear threads are contested rather than one-sided, with both more corrective and more alarm-amplifying replies. Fear discourse concentrates in frames about industrial ingredients, contaminants, and institutional distrust, and its repertoire of frames widens over 2020--2026. Our results argue for persistence-based rather than virality-based signals when platforms seek harmful health content.

1 Introduction

Food-fear messaging has received less attention than event-driven misinformation, despite appearing in everyday food choices and TikTok nutrition content. The paper asks how it spreads, shapes discussion, and changes over time, finding that persistence and conversation—not reach—distinguish it.

1 Introduction

  • Everyday food-fear messaging receives less attention than election and crisis misinformation, although it targets ordinary foods and ingredients.TikTok is a primary nutrition-information source, and less accurate nutrition posts draw more engagement.Source: §1 · p. 1
  • 62% of fear videos use a threat-and-secret-knowledge hook, compared with 37% of non-fear videos; cancer or “killing you” claims are 20 times more common.“Banned in other countries” claims are 45 times more common, and hidden-truth appeals are 10 times more common.Source: §1 · pp. 1–2
  • Fear videos reach fewer people, but are shared more per view and sustain discussion longer, motivating persistence-based rather than virality-based signals.The paper frames its questions around spread, audience affect and conversation, and changing discourse.Source: §1 · p. 2
  • The paper characterizes food-fear content as a category that is rarely viral and often misleading rather than clearly false.Its introduction argues that tools aimed at content selected for falsity, virality, or news events may not fit this setting.Source: §1 · p. 2
  • Fear threads are angrier and less able to sustain positivity, yet negativity is reinforced at the same rate as in non-fear threads.The introduction also reports more corrective replies and more alarm-amplifying voices in fear threads.Source: §1 · p. 2
  • Fear discourse targets industrial ingredients, contaminants, and institutions, while its repertoire of frames widens over time.The paper presents its contribution as mapping what food fear is about and how its frames multiply.Source: §1 · p. 2

2 Related Work

Prior work audits health-video accuracy and studies online attention, emotional contagion, and correction, but the paper shifts focus to the dynamics and discourse of an organic food-fear category.

Health misinformation on social and video platforms

  • Prior health-misinformation work spans platforms and topics, while TikTok diet content is often low quality and less accurate posts attract more engagement.Most cited work audits accuracy about particular conditions; this paper characterizes a broader fear-oriented category.Source: §2 · pp. 2–3

Attention dynamics and diffusion on the Web

  • Earlier diffusion studies emphasize how far and how fast content travels; this paper adds the shape of attention over time as an outcome.Prior work covers attention decay, activity around peaks, cascade structure, and emotional or false content diffusion.Source: §2 · p. 3

Affect, contagion, and echo chambers in comment threads

  • At the parent–reply level, negativity is reinforced at the same rate in fear and non-fear threads, while positivity fails to persist.This contrasts with prior accounts predicting negativity amplification in fear threads.Source: §2 · p. 3

Correction and counter-speech

  • Fear threads contain more corrective replies and more alarm-amplifying replies than other food threads.Prior work examines correction and counter-speech elsewhere, but their presence in organic fear content had not been measured.Source: §2 · p. 3

Fear appeals

  • The paper studies fear appeals as they occur organically in a social feed and interprets audience emotions through the EPPM lens.The cited theory holds that fear appeals paired with efficacy can change behavior, while those without efficacy can trigger defensive fear control.Source: §2 · p. 3

3 Methodology

The study builds a keyword-based TikTok sample, labels videos using a multimodal classifier and a defined codebook, and checks labels against human annotation.

Video collection

  • The study sampled 134,365 videos stratified by keyword and upload year, capping each creator at two videos, then collected metadata and comments.The corpus was drawn from 17 fear and 22 control keywords with at least 1,000 videos and 48 months of coverage.Source: §3 · p. 3

Definition

  • A video is labeled Fear when at least two codebook indicators are present, or when one indicator accompanies an alarming overall tone.Indicators include alarmist terms, urgency, absolutist claims, hidden-truth framing, warning imagery, and rejection imagery; balanced information, responsible alerts, satire, and debunking are Non-Fear.Source: §3 · p. 3

Human annotation

  • Two annotators double-labeled 986 videos, reaching 90.5% agreement and Cohen’s κ=0.58.Fear was the minority class, which the authors say makes κ much lower than raw agreement.Source: §3 · p. 3

Video collection

  • Table 1 reports the share of videos using common fear hooks, with differences Holm-adjusted at p<10^-70.“Toxic” and “stop eating” overlap with search keywords and also appear in many Non-Fear videos.Source: Table 1 · p. 4

LLM classification

  • The zero-shot Gemini 3 Flash classifier matched agreed human labels 96.3% of the time (κ=0.81), with 0.99 Fear recall and 0.72 precision.Against adjudicated labels on all double-labeled videos, Fear precision was 0.64; the authors note that this error biases group differences toward zero.Source: §3 · p. 4

4 Analysis Methods

The analysis measures sharing and comment timing, labels affect and reply stance, and uses statistical comparisons and regression controls to assess group differences.

Spread metrics

  • Discussion half-life t50 is the median days since upload by which half of comments arrive, measured for videos with at least five comments in a 0–365 day window.The analysis also fits per-video truncated power laws to comment arrival times; a smaller exponent α means slower decay.Source: §4 · p. 4

Affect

  • Comment and reply sentiment is labeled positive, neutral, or negative, with a 208-comment check showing 68% accuracy, macro-F1 0.69, and κ=0.53.The model catches 91% of comments judged negative, but often classifies neutral comments as positive or negative.Source: §4 · p. 4
  • 24.4% of comments under fear videos were negative versus 12.9% under non-fear videos after reweighting, a difference of 11.5 percentage points.The 95% bootstrap confidence interval for the difference is 1.9–19.6 percentage points; emotion labels were also applied to a stratified sample of 150,000 English comments.Source: §4 · p. 4

Reply direction and stance

  • Reply sentiment models test whether the Fear effect varies with parent sentiment, while stance labels distinguish correction from alarm amplification.Replies were modeled with GEE logistic models and robust standard errors clustered by video; 100 manually annotated replies checked stance labels.Source: §4 · p. 4

Statistics

  • Group comparisons use nonparametric tests and regressions controlling for reach, duration, upload year, and keyword, with creator fixed effects in a stricter within-creator specification.The authors report effect sizes alongside significance tests and correct 16 headline tests using Holm–Bonferroni.Source: §4 · p. 4

5 Results

Fear videos reach fewer people but are shared more per view and sustain discussion longer, even after controls. Their threads are more negative and contested, while fear discourse centers on food-related threats and institutional distrust and broadens over time.

5.1 Fear Reaches Fewer People but Lingers Longer (RQ1)

  • Fear videos had lower median views (1,033 vs. 2,741) but higher per-view sharing (2.32×) and longer discussion (1.82 vs. 1.03 days); controls retained 1.71×–2.05× sharing and a 1.62× half-life ratio.The controlled differences are modest, with substantial overlap between distributions.Source: §5.1 · pp. 4–5, Table 2 · p. 5

5.2 Fear Threads Are Angrier but Not Echo Chambers (RQ2)

  • Negative comments were 32.5% vs. 20.9%, with more disgust (26.0% vs. 16.9%) and anger (21.9% vs. 14.0%) under fear videos.Source: §5.2 · p. 5, Table 3 · p. 5
  • Fear threads dampened positivity rather than reinforcing negativity: negative-to-negative replies matched (37.8% vs. 37.5%), while positive continuity fell (39.6% vs. 57.8%).Source: §5.2 · p. 5, Figure 2 · p. 5, Table 4 · p. 6
  • Replies under fear videos were more often corrective (13.9% vs. 4.4%) and alarm-amplifying (9.1% vs. 1.0%), indicating contested threads.Source: §5.2 · pp. 5–6

5.3 Fear Discourse Targets Industrial Food and Institutions (RQ3)

  • Fear discourse overrepresented GMO and bioengineered ingredients (84%), seed oils (76%), and institutional distrust, alongside chemicals, additives, microplastics, and detox topics.Source: §5.3 · p. 6, Table 5 · p. 6
  • Fear-overrepresented transcript topics grew from 5–7 per year in 2020–21 to the low twenties from 2022 onward; the corpus-share rise to 15.8% in 2025 is suggestive only.The authors qualify the corpus-share trend because keyword composition and API coverage vary across years.Source: §5.3 · p. 6, Figure 3 · p. 6, Figure 4 · p. 7

5.4 Robustness Checks

  • Creator-matched estimates shrink (share-rate ratio 1.22×; half-life and negativity nonsignificant), while age, comment-coverage, and confident-label checks preserve the main patterns.Human-label results reproduce the share-rate and negativity gaps, but the half-life comparison is inconclusive.Source: §5.4 · pp. 6–7

6 Discussion

Fear content’s influence is better characterized by persistent, per-view engagement than broad reach, though within-creator comparisons weaken the pattern. Its threads are contested, and the discussion motivates persistence-aware platform responses.

6 Discussion

  • Fear videos reached fewer people but were shared more per view and discussed longer; these modest shifts weakened in strict within-creator comparisons.The authors note that part of the pattern may reflect which creators produce fear content.Source: §6 · p. 7
  • The authors cannot distinguish viewers returning to fear threads from TikTok resurfacing videos to new viewers without platform-held exposure data.Frequent sharing could prompt feed recommendations, bringing new commenters weeks after upload.Source: §6 · p. 7
  • Fear’s higher per-view sharing aligns with prior links between high-arousal emotion and sharing, but its videos accumulated fewer views and likes.The authors describe this as higher per-capita engagement among a smaller, more activated audience rather than broad virality.Source: §6 · pp. 7–8
  • Anger and disgust, rather than fear itself, dominate audience responses, a pattern consistent with fear control rather than constructive behavior change.The authors note that these videos rarely offer actionable efficacy and that institutional distrust is a persistent audience theme.Source: §6 · p. 8
  • Fear threads are contested: negativity is no more self-reinforcing, positivity persists less, and corrective and alarm-amplifying replies both appear more often.Corrective replies receive as many likes as corrective replies in non-fear threads, so the authors argue that interventions should make existing counter-speech more visible.Source: §6 · p. 8
  • Platforms could detect persistence instead of waiting for reach spikes, add friction at sharing, apply claim-based labels consistently, and prebunk recurring manipulation frames.The authors suggest that recognizing techniques such as hidden-truth promises may generalize better than chasing each new claim.Source: §6 · p. 8
  • Fear threads contain corrective replies three times as often as non-fear threads, and their days-to-weeks duration leaves time for crowd-sourced context to matter.The authors connect this longer discussion window to community notes, which have struggled when content spreads too fast for them to arrive in time.Source: §6 · p. 8
  • Because first replies shape thread tone and positive comments in fear threads rarely stay positive, surfacing well-received corrective or expert replies could counter that pattern.The authors cite prior evidence that expert corrections are the most effective at reducing health misperceptions.Source: §6 · p. 8

7 Limitations and Future Work

The study’s scope and measurements constrain how its findings should be interpreted. The authors propose testing interventions and extending the analysis to other platforms and languages.

7 Limitations and Future Work

  • The keyword-defined corpus is not a random sample of TikTok food content, and its composition varies across years.The authors also note moderate classifier precision, possible biases in automated sentiment, emotion, and stance labels, and limited human validation.Source: §7 · p. 8
  • The study lacks creator follower counts and platform exposure data, while incomplete, capped comment collection limits observation of discussion.The within-creator comparison is underpowered, and recommendation exposure cannot be separated from viewers returning to videos.Source: §7 · p. 8
  • Future work will test context labels and highlighted expert replies, track creators and audiences over time, and extend analysis to other platforms and languages.The authors report using public content in aggregate without profiling users or reporting personally identifying information, and plan to release video IDs, labels, the codebook, prompt, and analysis code rather than raw comment text.Source: §7 · p. 8

8 Conclusion

Fear videos reach fewer people but are shared more per view and sustain discussion longer. Their comments are more negative, yet threads remain contested and corrective replies are present.

8 Conclusion

  • Fear videos receive 62% fewer views, but are shared 1.7–2.1× more per view and keep discussion going 1.6× longer.The paper characterizes this pattern as lingering rather than spiking.Source: §8 · p. 8
  • Fear-video comments are angrier and more disgusted, while negative comments do not attract negative replies more often than in non-fear threads.Positive comments are less likely to stay positive (39.6% vs. 57.8%), and corrective replies are three times as common.Source: §8 · p. 8
  • Fear discourse centers on industrial ingredients, contaminants, and distrust of food-regulating institutions, while its distinct frames grow from a handful in 2020 to about twenty a year.The authors argue that persistence-based signals may help platforms find harmful health content that does not trend.Source: §8 · pp. 8–9

A Codebook

The codebook defines Fear by unnecessary fear-inducing food or nutrition framing and applies a structured set of verbal and visual indicators. It distinguishes such content from responsible warnings and other non-fear material.

A Codebook

  • Fear content uses language, urgency, alarmist claims, or sensationalized framing to create unnecessary fear about food or nutrition.The codebook asks whether a reasonable person would feel unnecessarily fearful after watching.Source: A Codebook · pp. 10–11
  • Indicators span alarmist terms, urgency, absolutist and hidden-truth claims, conspiracy undertones, alarming text, fear imagery, dramatic editing, shock visuals, and rejection imagery.Examples include “toxic,” “stop eating NOW,” and products crossed out or thrown away.Source: A Codebook · pp. 10–11
  • Two or more indicators mean Fear; one qualifies only when the overall tone is alarming, while no indicators means Non-Fear.Non-Fear includes responsible recalls and allergen alerts, educational or balanced content, recipes, and satire or debunking that does not reinforce fear.Source: A Codebook · p. 11

B Analysis Samples

The appendices identify dataset and analysis sample sizes, list fear and control search keywords, and provide the multimodal classification prompt and its decision criteria.

B Analysis Samples

  • Table 7 presents dataset and analysis sample sizes.The supplied caption does not state the table's values.Source: Table 7 · p. 11

C Keywords

  • The search-term sets comprise 17 fear keywords and 22 control keywords; the cited passage enumerates all terms in each set.Source: C Keywords · p. 11

D Classification Prompt

  • The classification prompt directs the model to assess both what is heard and seen in each TikTok video.It returns a Fear or Non-Fear label, confidence, indicators, and brief reasoning in JSON format.Source: D Classification Prompt · p. 11
  • The prompt defines Fear as unnecessary food-related fear and specifies ten verbal and visual indicators, with responsible warnings classified as Non-Fear.It sets two or more indicators as the usual Fear threshold and treats one indicator as Fear only when the overall tone is alarming.Source: D Classification Prompt · p. 11

E Full Regression Results

The full regression table reports Fear coefficients with reach, duration, year, and keyword-group controls. Standard errors are clustered by creator, with a within-creator specification for creators who posted both video types.

E Full Regression Results

  • Table 8 reports Fear coefficients controlling for log views, duration bin, upload-year, and keyword-group fixed effects.The supplied table caption and prose do not provide the coefficient values.Source: Table 8 · p. 11, E Full Regression Results · p. 11
  • Standard errors are clustered by creator; the within-creator model adds creator fixed effects for the 1,620 creators with both Fear and Non-Fear videos.The table caption specifies additional controls for t50 and negativity: log comment count.Source: Table 8 · p. 11

F Comment Topic Categories

Table 9 groups comment topics into categories based on sampled-comment review and reports each category’s comment count and share from fear videos against the 18.3% corpus baseline.

F Comment Topic Categories

Related papers

Loading 2610.33898v1…