cs.CYMar 5, 2026

The DSA's Blind Spot: Algorithmic Audit of Advertising and Minor Profiling on TikTok

Authors: Sara Solarova, Matej Mosnar, Matus Tibensky, Jan Jakubcik, Adrian Bindas, Simon Liska, Filip Hossner, Matúš Mesarčík, 

Organizations: Kempelen Institute of Intelligent Technologies, Bratislava, Slovakia · Comenius University Bratislava, Bratislava, Slovakia

Read PDF
Ask AI
Cite

Copy citation

DOINot available for this paper

Abstract

Adolescents spend an increasing amount of their time in digital environments where their still-developing cognitive capacities leave them unable to recognize or resist commercial persuasion. Article 28(2) of the DSA responds to this vulnerability by prohibiting profiling-based advertising to minors. However, the regulation's narrow definition of "advertisement" excludes current advertising practices including influencer paid partnerships and brand promotional content that serve functionally equivalent commercial purposes. We provide the first empirical evidence of how this definitional gap operates in practice through an algorithmic audit of TikTok. Our approach deploys sock-puppet accounts simulating a pair of minor and adult users with matching interest profiles. The content recommended to these users is automatically annotated, enabling systematic statistical analysis. Our findings reveal a stark regulatory paradox. TikTok demonstrates formal compliance with Article 28(2) by shielding minors from profiled formal advertisements, yet both disclosed and undisclosed ads exhibit significant profiling aligned with user interests (5-8 times stronger than for adult formal advertising). The strongest profiling emerges within undisclosed commercial content, where creators/brands fail to label paid partnership/promotional content and the platform neither corrects this omission nor prevents its personalized delivery to minors. These results demonstrate that minors remain exposed to algorithmically targeted commercial content through the same recommendation mechanisms the DSA seeks to constrain. We argue that protecting minors requires expanding the definition of advertisement in EU law to encompass influencer and brand promotional content, and ensuring that any such expansion is accompanied by a corresponding prohibition on profiling-based targeting of minors.

Figures & tables

UserGenderAgeTotal videosTotal adsAvg video length
Beauty MinorFemale16651169 (25.96%)55 seconds
Beauty AdultFemale211088246 (22.61%)36 seconds
Fitness MinorMale1787645 (5.14%)36 seconds
Fitness AdultMale202341534 (22.81%)14 seconds
Gaming MinorFemale1686851 (5.88%)30 seconds
Gaming AdultFemale211271301 (23.68%)30 seconds
Table 1. Overview of user characteristics and total number of observed videos and ads (formal + disclosed + undisclosed)
Figure 1. Ad distribution by ad type and ad topic match per user. While minors are recommended less ads in total, such ads are not properly disclosed and are heavily personalized to match their interest. Ad distribution by ad type and ad topic match per user. While minors are recommended less ads in total, such ads are not properly disclosed and are heavily personalized to match their interest.
Figure 2. Ad topic distribution. While formal ads distribution indicate no significant personalization, disclosed and undisclosed ads show strong diagonal dominance. Ad topic distribution.
UserPers. Rate Pers./Total AdsBaseline Rate Pers./Total AdsProfiling EffectStat. Diff.
Beauty Minor4.76% (1/21)28.57% (2/7)-23.81 pp 
Beauty Adult33.33% (2/6)16.90% (86/509)16.43 pp 
Fitness Minor0.00% (0/0)3.57% (1/28)-3.57 pp 
Fitness Adult19.27% (58/301)8.88% (19/214)10.39 pp**
Gaming Minor42.86% (3/7)9.52% (2/21)33.34 pp
Gaming Adult23.56% (49/208)9.12% (28/307)14.44 pp
Table 2. Profiling effect for formal ads. Personalization rate is a ratio between ads relevant for a user and all ads. Baseline rate is determined by proportion of ads with the same topic that naturally appeared for users with the same age but a different interest. Profiling effect is a percentage point (pp) difference between these two numbers.
UserPers. Rate Pers./Total AdsBaseline Rate Pers./Total AdsProfiling EffectStat. Diff.
Beauty Minor91.89% (136/148)5.62% (5/89)86.27 pp
Beauty Adult64.17% (154/240)3.68% (12/326)60.49 pp
Fitness Minor88.89% (40/45)5.21% (10/192)83.68 pp
Fitness Adult91.85% (214/233)24.62% (82/333)67.23 pp
Gaming Minor93.18% (41/44)0.00% (0/193)93.18 pp
Gaming Adult88.17% (82/93)0.85% (4/473)87.32 pp
Table 3. Profiling effect for disclosed and undisclosed ads.
Appendix figures & tables4 assets

Supplementary material from the paper’s appendix.

Appendix

Interestbeautybeautyfitnessfitnessgaminggaming
Age Groupminoradultminoradultminoradult
Genderfemalefemalemalemalefemalefemale
Total Records651108887623418681271
Non Ad Records4828428311807817970
Ads Detected1692464553451301
Avg Video Length54.835.635.8143030
Appendix
Table 4. Summary table of data collected during the audit.
Figure 3. Predicted-true confusion matrices for ad type and ad topic labels for each of human annotators.
Figure 4. Screenshot of the custom-developed tool providing an overview of the collected data. The displayed videos were presented to a user Beauty Minor (a 16-year-old minor with an expressed interest in beauty). The tags determine whether the annotation model determined any kind of advertisement or not. The faces and personal identifying information were blurred or anonymised to preserve creators’ privacy. Screenshot of the custom-developed tool providing an overview of the collected data. The displayed videos were presented to a user Beauty\textsubscript{Minor} (a 16-year-old minor with an expressed interest in beauty). The tags determine whether the annotation model determined any kind of advertisement or not. The faces and personal identifying information were hidden or anonymised to preserve creators' privacy.
Figure 5. Screenshot of the custom-developed tool providing a detail of the video observed by a user. The displayed video was presented to a user Fitness Minor (a 17-year-old minor with an expressed interest in fitness). The tool displays automatic content analysis (including the reasoning of the model) together with the screenshot and URL to the video, streamlining the manual annotation process. The displayed video represents an example of paid partnership content. The faces and personal identifying information were hidden or anonymised to preserve creators’ privacy. Screenshot of the custom-developed tool providing a detail of the video observed by a user. The displayed video was presented to a user Fitness\textsubscript{Minor} (a 17-year-old minor with an expressed interest in fitness). The tool displays automatic content analysis (including the reasoning of the model) together with the screenshot and URL to the video, streamlining the manual annotation process. The displayed video represents an example of paid partnership content. The faces and personal identifying information were blurred or anonymised to preserve creators' privacy.
More in this topic

Online Advertising

Online advertising studies how digital ads are targeted, ranked, priced, measured, and delivered while balancing budgets, business goals, and user interests.

See all 87 papers

Explore similar work

CardsList
Aug 18, 2026

Auditing Exposure to Harmful Content on TikTok using Multimodal Language Models: A Cross-National, Age-Stratified Study

Online video platforms can expose young users to harmful content, but independent audits remain difficult because video annotation is costly and moderation judgments vary across languages. We audit TikTok in France, Italy, and Sweden with sockpuppet accounts representing four age personas (13, 16, 19, 40), collecting 36,971 videos from passive For-You-page scrolling and active sessions that scroll, search for harm keywords, and scroll again. To scale annotation, we validate four multimodal LLMs against native-speaker labels on a 300-video reference set. Gemini 2.5 Flash with eight sampled frames plus text performs best (aggregate kappa = 0.42), at half the per-call cost of native-video upload, and we apply it to a 10% sample for approximately $50 in total API spend across both modalities. Keyword search returns 35-56% harmful content, a 1.5-7.5x increase over the scrolling baseline in ten of twelve country-age combinations; the spike is temporary and flattens the age differences observed in France and Sweden. Under passive scrolling, Italy has the highest harm rate at every age, with Italian age-19 reaching 48.6%. Overall, MLLM-based auditing offers a scalable approach for cross-national youth-safety audits, while provider safety filters (1.1% refusal rate) under-count the most explicit harms.

Hamidreza Saffari, Francesco Pierri

Politecnico di Milano

Algorithmic AuditingVideo Content ModerationContent ModerationMultimodal Content ModerationMultimodal Large Language Models
AskPDF
Jun 2, 2026

Auditing Engagement Incentives in the Kidfluencer Ecosystem: A Multimodal Weak Supervision Approach

The rise of `kidfluencers' on YouTube has raised ethical concerns about child digital labor and exploitation. While emerging legislation attempts to regulate this ecosystem, empirical evidence linking exploitation to engagement remains scarce, given the difficulty of operationalizing exploitation at scale. This study presents a multimodal AI audit of 5,051 videos across 79 kidfluencer channels, using weak supervision to detect exploitation signals without large-scale manual labels. We aggregate noisy labeling functions -- including LLM-based classification of titles and GPT-4 Vision analysis of thumbnails and descriptions across six literature-grounded dimensions -- to assign a probabilistic exploitation score to each video. A multi-annotator validation study (N=107) shows strong agreement with human judgment (macro-average F1 =0.911= 0.911) and high sensitivity for overall exploitation risk (recall =0.960= 0.960, F1 =0.793= 0.793). Our findings reveal a significant engagement premium for performative labor, emotional bait, and privacy violations. Exploitation scores correlate with view counts (Spearman ρ=0.229ρ= 0.229, p<10−50p < 10^{-50}), and mixed-effects regression controlling for channel-level variation shows that a one-unit increase in exploitation score yields a 4.4×4.4\times increase in views (p<0.001p < 0.001). Within-channel analyses indicate median view boosts of +65.6%+65.6\% for emotional bait and +56.0%+56.0\% for performative content (FDR-corrected p<0.001p<0.001), with effects holding in same-year robustness checks (p=0.030p=0.030). Explicit commercial content (product placement), by contrast, shows no premium (−3.8%-3.8\%, n.s.), suggesting the platform rewards commodification of the child's identity and labor over traditional advertising. These findings challenge policy frameworks focused solely on financial trusts, showing that engagement is systematically tied to the intensive, performative labor of children.

Zijing Wei, Chao Peter Yang, Xuanjie Chen

Independent Contributor · Duke University

Social Media AnalysisAlgorithmic Auditing
AskPDF
Aug 19, 2026

ChildSafeAds Shared Task 2026: Commercial Content in Child-Facing YouTube Videos

ChildSafeAds is a shared task on commercial content in YouTube videos likely to reach children and teenagers. The dataset contains 3,360 videos from 939 channels, each starting with a sponsor segment submitted to SponsorBlock. We pair each segment with its transcript, video, and channel information, plus a sales or service page linked from the video description. Systems determine the type of offer being promoted (ST1), assign product categories (ST2), and identify potential legal risks (ST3). Evidence is divided into four cumulative access levels, from the transcript to the linked page, so performance can be compared against the cost of collecting additional data. YouTube's "Includes paid promotion" label was absent from 45.5% of the videos. GPT-5.4 produced the labels after the organiser team, including a legal expert, reviewed samples and iterated on the taxonomy, prompts and model choices. GPT-5.6-luna independently labelled the development set to measure cross-model agreement. The evaluation received 21 entries; the highest mean macro-F1 was 0.735, while the best transcript-only entry reached 0.652. Across four system papers, the linked promotional page mainly helped identify what was being sold and added little to compliance-risk classification. The competition also exposed sensitivity to rare labels, advertiser overlap across splits and substantial variation in the ST3 labels.

Thales Bertaglia, Catalina Goanta, Gerasimos Spanakis +1

Utrecht University, Netherlands · Maastricht University, Netherlands · Radboud University, Netherlands

ClassificationVideo Content Moderation
AskPDF