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

This paper presents an algorithmic audit of TikTok revealing that while the platform technically complies with the Digital Service Act's ban on profiled advertising to minors, it effectively circumvents this protection by delivering highly personalized, often undisclosed influencer marketing content to adolescents, thereby highlighting the urgent need to expand the regulatory definition of "advertisement" to cover such commercial practices.

The Gist

Here is an explanation of the paper, translated into everyday language with some creative analogies.

The Big Idea: The "Trap Door" in the Law

Imagine the European Union built a giant, high-tech security fence around a playground. The goal? To stop salespeople from sneaking in and trying to sell candy to the kids. This fence is called the Digital Services Act (DSA).

The law says: "No salespeople allowed to target kids based on what they like."

The researchers in this paper decided to test this fence. They asked: "Is the fence actually working, or is there a secret trap door we didn't know about?"

They found that the fence is working perfectly for the official salespeople, but the unofficial ones are slipping right through the gap, and they are targeting the kids even harder than the adults.

---

The Cast of Characters

To test the fence, the researchers created a team of digital spies (called "sock-puppets").

• The Kids: They created fake accounts pretending to be 16 and 17-year-olds.
• The Adults: They created matching fake accounts for 20 and 21-year-olds.
• The Interests: They taught all these spies to love specific things: Beauty, Fitness, Gaming, and Politics.

They sent these spies to TikTok for 10 days to see what the algorithm (the robot brain that decides what videos you see) would show them.

---

The Three Types of "Salespeople"

The researchers realized that not all ads look the same. They categorized them like this:

1. The Official Salesperson (Formal Ads):

   • What it is: A video that says "Sponsored" or "Ad" in big letters. The company paid TikTok directly to show this.
   • The Law: The DSA says, "No targeting kids with these."
   • The Result: TikTok passed the test. They showed very few of these to the kids, and they didn't seem to pick them based on what the kids liked. The fence held up here.

2. The Honest Influencer (Disclosed Ads):

   • What it is: A creator says, "Hey, I'm being paid to show you this," and uses a small label like "Paid Partnership."
   • The Law: The DSA doesn't clearly count this as an "ad" in the strict sense, so the ban doesn't fully apply.
   • The Result: These were rare because creators are lazy about labeling them. But when they did appear, they were highly targeted to the kids' interests.

3. The Sneaky Salesperson (Undisclosed Ads):

   • What it is: This is the big problem. A creator posts a video raving about a new lipstick or a gaming headset. They don't say they are being paid. They just act like a normal friend recommending something.
   • The Law: Because there is no "Ad" label, TikTok's robot brain thinks, "Oh, this is just a fun video!" and shows it to the kid who loves makeup or gaming.
   • The Result: This is where the fence failed. The researchers found that kids were bombarded with these sneaky ads. If a kid liked gaming, 93% of the "gaming" videos they saw were actually hidden ads. If they liked beauty, 92% were hidden beauty ads.

---

The "Blind Spot" Explained

Here is the tricky part: The law is too narrow.

The DSA defines an "advertisement" as something where a company pays TikTok directly. It misses the massive world of Influencer Marketing.

• The Analogy: Imagine a school cafeteria. The law says, "No one can sell candy to kids at the checkout counter."
  • The school checks the checkout counter. No candy is being sold there. Mission accomplished!
  • However, the law forgot to check the lunch tables.
  • At the lunch tables, popular kids (influencers) are whispering to their friends, "You have to try this candy bar, my mom bought it for me." They aren't standing at the counter; they are just hanging out.
  • Because the law only banned the checkout counter sales, the kids are still eating candy, just from a different source.

The researchers found that TikTok is doing exactly this. They stopped the "checkout counter" ads (Formal Ads) for kids, but they let the "lunch table" sales (Influencer/Undisclosed Ads) happen, and they are using the same robot brain to make sure those sales hit the kids who are most likely to buy them.

---

Why This Matters

The paper argues that this is dangerous because:

1. Kids can't tell the difference: A 16-year-old doesn't have the life experience to know that a "friend" recommending a product is actually a paid salesperson.
2. The targeting is intense: The robot brain is matching these hidden ads to the kids' insecurities and interests even better than it matches regular ads to adults.
3. The platform is innocent (technically): TikTok can say, "We followed the law! We stopped the formal ads." But they aren't stopping the effect of the ads.

The Solution Proposed

The authors suggest four fixes:

1. Widen the Definition: Change the law so that any content made to sell something counts as an ad, whether the company paid TikTok or just paid the influencer.
2. Make TikTok Responsible: Don't just give creators a button to label their ads. Make TikTok responsible for finding the ones that aren't labeled and fixing them.
3. Ban the Targeting: Apply the "No targeting kids" rule to all commercial content, not just the official ads.
4. Real Audits: Stop just reading the company's paperwork. Use these "spy bot" tests to actually check what kids are seeing in real life.

The Bottom Line

The Digital Services Act tried to build a shield to protect children from being manipulated by ads. But the shield has a hole in it. TikTok is technically following the rules, but the kids are still being targeted by a flood of hidden, personalized sales pitches that the law doesn't see. The researchers are saying: "Close the hole, or the kids are still going to get sold out."

Technical Summary

Here is a detailed technical summary of the paper "The DSA's Blind Spot: Algorithmic Audit of Advertising and Minor Profiling on TikTok" by Solarova et al.

1. Problem Statement

The paper addresses a critical regulatory gap in the European Union's Digital Services Act (DSA), specifically Article 28(2), which prohibits online platforms from presenting profiling-based advertisements to minors.

• The Regulatory Gap: The DSA defines "advertisement" narrowly (Article 3(r)) as content where the platform receives direct remuneration (formal ads). This definition excludes influencer marketing, brand self-promotion, and unlabelled commercial content (paid partnerships not paid to the platform).
• The Vulnerability: Adolescents possess developing cognitive capacities, making them susceptible to commercial persuasion and algorithmic manipulation. They often cannot distinguish between entertainment and advertising, especially when commercial content is disguised as organic posts.
• The Hypothesis: While platforms may formally comply with the DSA by blocking formal profiled ads for minors, they likely continue to profile minors using the vast ecosystem of undisclosed and disclosed influencer/brand content, which falls outside the current legal definition of "advertisement."

2. Methodology

The authors conducted a sock-puppet algorithmic audit on TikTok to empirically test platform behavior.

Experimental Design

• Paired User Simulation: The study deployed 4 pairs of automated agents (sock-puppets). Each pair consisted of one minor (aged 16–17) and one adult (aged 20–21).
  • Rationale: The narrow age gap minimizes generational content preference confounds, isolating "legal status" (minor vs. adult) as the primary variable.
  • Demographics: Pairs were assigned specific genders and interest profiles (Beauty, Fitness, Gaming, Politics) to simulate distinct user personas.
• Environment: All agents accessed TikTok via the desktop web interface using static residential proxies in Germany to ensure consistent geo-location and DSA jurisdiction applicability.
• Interest Seeding: To overcome the "cold start" problem, agents underwent a 1-day seeding phase where they actively engaged (watched, liked, bookmarked) with 25 relevant videos per interest profile using an LLM-based relevance predictor.
• Data Collection: Over 10 days in December 2025, agents interacted with the "For You" feed for 1 hour daily. The system captured metadata for all videos (watched or skipped) to analyze the platform's unfiltered recommendation logic.

Classification Pipeline

The study utilized a Large Vision-Language Model (LVLM) (Qwen3-VL-4B-Instruct) to automatically annotate 7,095 videos into four categories:

1. Formal Ads: Platform-injected labels ("Sponsored" or "Ad").
2. Disclosed Ads: Creator-labeled content ("Paid Partnership" or "Promotional Content").
3. Undisclosed Ads: Commercial content lacking any disclosure labels but containing visual/semantic commercial indicators.
4. Non-Ads: Organic content.

• Validation: A stratified random sample of 113 videos was manually annotated by researchers. The automated pipeline achieved 90.3% accuracy for ad type and 88.5% for ad topic, with high inter-annotator agreement.
• Metric: The study calculated the Profiling Effect ($\Delta$), defined as the percentage point difference between the Personalization Rate (ads matching user interest) and the Baseline Rate (ads of that topic appearing for users with different interests but same age).

3. Key Contributions

1. Novel Methodology: Introduced a paired user sock-puppet design specifically for auditing ad delivery algorithms. Unlike previous "paired ads" studies (which test how a specific ad is distributed), this approach is user-centric, analyzing what types of content are recommended to specific user profiles.
2. First Empirical Evidence: Provided the first large-scale data showing that minors are subjected to profiling-based targeting via content formats (influencer/brand marketing) that the DSA currently excludes from its definition of "advertisement."
3. Regulatory Gap Identification: Demonstrated that the DSA's narrow definition creates a "blind spot" where the majority of commercial exposure to minors is unregulated, despite the platform's formal compliance with Article 28(2).

4. Key Results

The audit analyzed 7,095 videos, of which 1,346 (18.97%) contained commercial content.

A. Formal Ads (DSA-Compliant)

• Volume: Minors received significantly fewer formal ads than adults (e.g., 0 formal ads for the "Fitness Minor" vs. 534 for "Fitness Adult").
• Profiling: There was no evidence of profiling for minors regarding formal ads. The personalization rate for minors was low or negative compared to baselines.
• Conclusion: TikTok is formally compliant with Article 28(2) regarding formal advertisements.

B. Disclosed and Undisclosed Ads (The "Blind Spot")

• Volume: The vast majority of commercial content minors saw was undisclosed (55.8% of all ads) or disclosed. Formal ads were a minority.
• Profiling Intensity:
  • Minors experienced extremely high profiling effects for these non-formal ads: +83.68 pp to +93.18 pp.
  • Example: The "Beauty Minor" account received 91.89% of its disclosed/undisclosed ads related to beauty, compared to a baseline of 5.62%.
  • Comparison: This profiling intensity is 5 to 8 times stronger than the profiling observed in adult formal advertising (which ranged from +10.39 pp to +16.43 pp).
• Disclosure Failure: Only 5.4% to 11.4% of creator-driven commercial content was properly labeled. The platform did not intervene to correct missing labels.

C. Specific Findings on Beauty

• Minors interested in beauty were exposed to a high volume of undisclosed ads (92.1% personalization rate). Given the known link between beauty content and negative body image in adolescents, this represents a significant documented harm pathway.

5. Significance and Policy Recommendations

The study reveals a regulatory paradox: TikTok complies with the letter of the law (blocking formal profiled ads) but violates the spirit of the law (allowing highly personalized commercial manipulation via influencer marketing).

Policy Recommendations:

1. Expand the Definition of "Advertisement": The DSA definition must shift from a payment-based model (who gets paid) to a commercial-purpose model. This would include influencer marketing, brand self-promotion, and affiliate content.
2. Strengthen Labeling Obligations: Platforms must bear direct responsibility for detecting and correcting missing disclosure labels, rather than relying solely on creator self-reporting.
3. Extend Article 28(2): The prohibition on profiling-based advertising to minors must apply to the expanded definition of advertisements.
4. Independent Algorithmic Audits: Compliance verification must move beyond document reviews to include independent, behavioral algorithmic audits (like the methodology used in this paper) to verify real-world system behavior.

Conclusion

The paper concludes that current regulatory frameworks are insufficient to protect minors from commercial exploitation on social media. By operating outside the narrow definition of "advertisement," platforms can legally continue to algorithmically target minors with highly personalized commercial content, undermining the protective intent of the DSA. The authors argue that without expanding the regulatory scope and enforcing active platform liability for labeling, minors remain vulnerable to manipulative commercial practices.