CoChat

AI-Generated Content Disclosure and Prolonged Short-Video Engagement: A Heuristic-Systematic Risk-Trust Model Among Late-Adolescent and Emerging-Adult TikTok Users

Open access

The DOI matches its Crossref record; 86 of the 86 references it cites were found in a registry or an index; none of the 52 sources its claims cite is checked for support yet.

Abstract

Prolonged short-video engagement in the generative-AI era may be shaped by interface cues that encourage or interrupt repeated continuation decisions in algorithmic feeds. This study examines whether AI-generated content disclosure functions as interface-level digital friction for prolonged short-video engagement among late-adolescent and emerging-adult TikTok users. Prolonged watching intention is treated as a cognitive-behavioral proximal tendency relevant to problematic social media use (PSMU), rather than as a clinical diagnosis or an emotional-disturbance outcome. Drawing on the heuristic-systematic model, we tested a dual-pathway risk-trust model in which disclosure directly affects prolonged watching intention, while perceived risk and content trust operate as mediators and AI literacy operates as a person-level boundary condition. An online between-subjects experiment was conducted with 720 valid participants aged 18–24. Disclosure had a positive direct effect on prolonged watching intention, suggesting that AI labels can initially work as salient curiosity and novelty cues. At the same time, disclosure increased perceived risk and reduced content trust, generating negative indirect pathways that constrained prolonged watching intention. AI literacy strengthened both appraisal pathways. The findings reposition AI disclosure from a mere transparency notice to a behavioral cue that can simultaneously attract attention and activate protective appraisal. They contribute to developmental and media-psychological research on prolonged engagement and PSMU-relevant mechanisms without overstating clinical implications.

Overview

CoChat Summary

TL;DR

In an online between-subjects experiment with 720 valid TikTok users aged 18–24, AI-generated content disclosure had a positive direct effect on prolonged watching intention but also increased perceived risk and reduced content trust, producing negative indirect effects. AI literacy strengthened both appraisal pathways, so disclosure acted as a double-edged cue that attracted attention while activating protective appraisal.

Key findings

  • Disclosure directly increased prolonged watching intention (β = 0.474, p < 0.001), while increasing perceived risk (β = 0.304, p < 0.001) and decreasing content trust (β = −0.282, p < 0.001).
  • Perceived risk reduced prolonged watching intention (β = −0.413, p < 0.001) and content trust (β = −0.393, p < 0.001); content trust increased prolonged watching intention (β = 0.425, p < 0.001).
  • Mediation was competitive: indirect via perceived risk b = −0.235, 95% CI [−0.315, −0.160]; via content trust b = −0.224, 95% CI [−0.317, −0.152]; chain via risk then trust b = −0.095, 95% CI [−0.131, −0.063]; total indirect b = −0.553, 95% CI [−0.684, −0.436]; direct b = 0.883, 95% CI [0.719, 1.049].
  • AI literacy moderated both pathways: disclosure × AI literacy on perceived risk b = 0.192, SE = 0.072, t = 2.664, p = 0.008, ΔR² = 0.008; on content trust b = −0.270, SE = 0.081, t = −3.340, p = 0.001, ΔR² = 0.013. Simple slopes showed the risk effect increased from 0.267 to 0.677 and the absolute trust effect increased from 0.532 to 1.108 across low to high AI literacy; the H7 interaction was small and below the 80% sensitivity threshold (f² ≈ 0.008 vs ≈ 0.011), while H8 was above it (f² ≈ 0.013).

Methods

  • Single-factor, two-level between-subjects online experiment; independent variable was AI-generated content disclosure (label vs no label). Stimulus was a weather-report short-form video selected via pretest (81 valid responses; three videos did not differ in interest F = 1.46, p = 0.235, coherence F = 1.12, p = 0.329, or fluency F = 0.88, p = 0.417; Cochran’s Q = 11.58, p = 0.003; Video 2 retained with lowest AI recognition at 17.3%).
  • 720 valid participants aged 18–24 recruited through TikTok-related online spaces and Credamo; random assignment yielded 356 in the AI-disclosure group and 364 in the no-disclosure group; effective response rate was 67.8%. Measures used adapted scales: perceived risk 3 items (α = 0.806), content trust 3 items (α = 0.810), prolonged watching intention 3 items (α = 0.812), AI literacy 12 items (α = 0.940), all on 7-point Likert scales.
  • Analysis used structural equation modeling with maximum likelihood, bootstrap mediation with 2000 resamples, and regression-based moderation with mean-centered AI literacy and covariates (gender, education, occupation, daily short-form video exposure). Model fit was good: χ² = 36.852, df = 30, p = 0.182, χ²/df = 1.228, CFI = 0.997, RMSEA = 0.018, RMSEA 90% CI [0.000, 0.035].

Limitations

  • Stated: the stimulus was a synthetic weather report, a genre with strong truth, numerical-accuracy, public-safety, and civic-trust expectations, so findings cannot be seamlessly extrapolated to entertainment, comedy, lifestyle, aesthetic-remix, or other short-video streams. The selected clip also had low unlabeled AI recognition (17.3% identified it as AI-generated; 82.7% did not), making the control a high-realism, low-detection synthetic baseline rather than a neutral unlabeled baseline; the formal no-label group was not separately asked about perceived human versus AI authorship.
  • Stated: the sample was educationally concentrated (96.94% with some post-secondary education; 59.44% undergraduate, 33.19% junior/technical college, 4.31% graduate or above), raising restriction-of-range concerns; participants were legally adult users aged 18–24, limiting generalization to younger adolescents or minors; recruitment was voluntary through online TikTok-related spaces and a Credamo-hosted questionnaire without probability sampling, and geographic distribution could not be described.
  • Stated and evident from the text: the manipulation was binary (label vs no label), whereas real disclosures vary in wording, iconography, color, position, timing, persistence, and provenance links. The outcome was self-reported prolonged watching intention measured after one clip, not observed prolonged viewing or a validated PSMU diagnosis, so it may overrepresent reflective cognition and cannot capture infinite-scroll micro-decisions such as swipe latency, dwell time, autoplay, or time distortion.

Model-written summary. Check it against the paper before relying on it.

Integrity signals

Registry

Crossref · 10.3390/bs16071179

Cites

86 of 86 references verified0 citation errors · 0 not checkable

Statistics

Nothing to recompute.

Figures

Figure 1 Research hypothesis model for AI disclosure, protective appraisal, and prolonged watching intention. Note. AI = AI-generated content disclosure; PR = perceived risk; CT = content trust; PWI = prolonged watching intention; AL = AI literacy. Blue solid arrows indicate the hypothesized main effects through perceived risk and content trust (H2–H6); the orange solid arrow indicates the direct effect of AI disclosure on PWI (H1); orange dashed arrows indicate the moderating effects of AI literacy (H7–H8).
Figure 1 Research hypothesis model for AI disclosure, protective appraisal, and prolonged watching intention. Note. AI = AI-generated content disclosure; PR = perceived risk; CT = content trust; PWI = prolonged watching intention; AL = AI literacy. Blue solid arrows indicate the hypothesized main effects through perceived risk and content trust (H2–H6); the orange solid arrow indicates the direct effect of AI disclosure on PWI (H1); orange dashed arrows indicate the moderating effects of AI literacy (H7–H8).CC BY · PubMed Central
Figure 2 Experimental workflow and sample derivation. Note. Blue boxes indicate the procedural stages, whereas the orange box highlights the final analytic sample.
Figure 2 Experimental workflow and sample derivation. Note. Blue boxes indicate the procedural stages, whereas the orange box highlights the final analytic sample.CC BY · PubMed Central
Figure 3 Schematic comparison of AI-generated content disclosure and no-disclosure conditions.
Figure 3 Schematic comparison of AI-generated content disclosure and no-disclosure conditions.CC BY · PubMed Central
Figure 4 Standardized path coefficients of the structural equation model. Note. AI = AI-generated content disclosure; PR = perceived risk; CT = content trust; PWI = prolonged watching intention; *** p < 0.001.
Figure 4 Standardized path coefficients of the structural equation model. Note. AI = AI-generated content disclosure; PR = perceived risk; CT = content trust; PWI = prolonged watching intention; *** p < 0.001.CC BY · PubMed Central
Figure 5 Moderating effect of AI literacy on the relationship between AI-generated content disclosure and perceived risk. Note. The vertical axis uses the perceived-risk composite score.
Figure 5 Moderating effect of AI literacy on the relationship between AI-generated content disclosure and perceived risk. Note. The vertical axis uses the perceived-risk composite score.CC BY · PubMed Central
Figure 6 Moderating effect of AI literacy on the relationship between AI-generated content disclosure and content trust. Note. The vertical axis uses the content-trust composite score.
Figure 6 Moderating effect of AI literacy on the relationship between AI-generated content disclosure and content trust. Note. The vertical axis uses the content-trust composite score.CC BY · PubMed Central

Verification

Claims the paper makes

0 Valid · 0 Not valid · 36 Not checked · 87 Own results and interpretations

51 more claims

Create a free account to see all 123 claims and to verify the 43 references still open.

See all claims — free account
  • It may state two separate claims as one. 54
  • Probably not a substantial claim. 30
  • The quote only partly states this claim. 1

Create a free account to see these claims.

Citations

Cites

Show all 86 references

Metadata

Authors
Xiao, Y., Du, J., Ding, Y., Zhang, M., Jiang, Y., Yang, Y., & Liu, J.
Affiliations
  • Huazhong University of Science and Technology (CN)
  • Nanjing University (CN)
  • Shanghai Normal University (CN)
Corresponding author
  • Zhang, Minyang
  • Jiang, Yumei
Venue
Behavioral Sciences
Publisher
MDPI AG
Published
2026-07-13
Volume
16
Issue
7
Pages
1179
ISSN
2076-328X (ISSN-L)
OpenAlex
W7168187070
Cited by
2
References
86
Citations by year
2026: 2
FWCI
8.04
Type
journal article
Language
en
Open access
Locations
  • Behavioral Sciences (publishedVersion, cc-by)
  • PubMed Central (publishedVersion, cc-by)
  • PubMed (publishedVersion)
  • DOAJ (DOAJ: Directory of Open Access Journals) (submittedVersion)
Funders
  • 2025 Jiangsu Province Postgraduate Practice Innovation Program (SJCX25_0012)
  • Huazhong University of Science and Technology Party-Building Research Project (DJSZ202568)
  • General Program of the National Social Science Fund of China (22BXW056)
Topics
  • Social Sciences › Sociology and Political Science › Impact of Technology on Adolescents
  • Social Sciences › Communication › Digital Media and Philosophy
  • Social Sciences › Sociology and Political Science › Privacy, Security, and Data Protection
Keywords
TikTok, heuristic-systematic model, risk perception, AI literacy, social media addiction, late adolescence, emerging adults, algorithmic feeds

About these checks

Registry, retraction, Cites and statistics results are computed, not judged, with one exception: a printed reference that no index lists is read once by a model before it counts as a citation error. Claims are extracted by a model and kept only when their quote appears verbatim in the text; each cited source is then checked, with a model’s help, for whether it says what the claim relies on. Those verdicts can be wrong — read the claim’s quote and the cited source before relying on one.

A language model drew the claims from the paper on 2026-11-20, under version 3 of the claim definition (model deepseek-v4.1-flash, prompt extract-v3).

A second model labelled each claim on 2026-11-20 (model jev-1.13-20260917, definition revision gold-guideline-v0.2).

Statistics and arithmetic are recomputed, not judged.

Confidence bands are shown only for fields calibrated on our annotated test set. None is calibrated yet, so no band is shown.

What counts as a claim ›How we measure confidence ›How we check this ›

Are you an author, or do you see a problem with this page? Request a review

Data sources

Bibliographic data: Semantic Scholar, OpenAlex, Crossref.

Contains information from the Semantic Scholar Academic Graph Datasets, made available under the ODC Attribution License (ODC-BY 1.0).

Semantic Scholar’s data is described in: Kinney, R., et al. (2023). The Semantic Scholar Open Data Platform. arXiv:2301.10140.