Predicting Teen Porn Use: Parental Control Barely Helped

The paragraph beginning “This fragment is already well-translated and matches the Korean source…” should be removed entirely, as it is an internal editorial note that leaked into the body.

---

Around the same time in Korea, the question of protecting minors on AI character chat apps like Zeta, Crack, and Baby Chat came up during the National Assembly’s annual audit. Let’s read these two stories side by side.

What a Survey of 33,543 Spanish Teens Found

A research team at Spain’s Universitat Rovira i Virgili (URV) published a paper in the journal Computation in August 2026. The question was simple: what traits show up alongside teens who say they’ve viewed or used pornography in the past 30 days?

The data came from the 2023 edition of ESTUDES¹, a drug-use survey that the Spanish government runs every two years among secondary school students. The original sample had 42,208 respondents, and 33,543 who answered every required question made it into the analysis. Most were 15 to 18 years old. Of these, 42.6% (14,287 students) said they had viewed pornography in the past 30 days.

The team built four models to predict this response: one logistic regression², a method long used in statistics, plus three gradient-boosting³ models—XGBoost, LightGBM, and CatBoost. They then used an explanation technique called SHAP⁴ to work out which variables moved the prediction, and by how much. The variables fed into the models included gender and age, relationships with father and mother, family conflict, parental monitoring both inside and outside the home, tobacco/alcohol/cannabis use, some sexual experience, psychological distress such as anxiety attacks, and problematic internet use.

Let me flag up front what this study doesn’t do. Because it’s a cross-sectional survey⁵ asking questions at a single point in time, it can’t establish what causes what. And since the outcome variable is just a binary—viewed in the past 30 days or not—it can’t tell us how often someone views pornography or whether that use has become problematic. The research team themselves were explicit that this model shouldn’t be used to diagnose or stigmatize individual students, but rather to inform education and prevention policy.

After Gender, the Heaviest Variable Is Trouble Disconnecting From the Internet

By far the biggest variable in the prediction was gender. By logistic regression, the odds⁶ that a female student had viewed pornography in the past 30 days were 92.8% lower than for male students. Age was also related, but its SHAP contribution was far smaller than that of the leading variables.

Second after gender was problematic internet use. The research team built this index from 14 items — things like difficulty disconnecting, cutting sleep because of internet use, putting off tasks that need doing, and going online to escape worry or bad feelings. The team specifically noted that this index is not the same thing as a clinical diagnosis of internet addiction. In the logistic regression, moving this index from its lowest to its highest value multiplied the odds of pornography use by about 5.3x.

Next came alcohol, cannabis, and family conflict. For alcohol, the prediction shifted most sharply between students who had never drunk and those who had drunk even once over a weekend. For cannabis, the biggest divide was simply whether a student had ever tried it or not. Sexual risk behaviors — condomless sex, sex regretted the next day — also added to the prediction, but their effect size was small.

On the other end, some variables explained almost nothing: parental control at home and parental control outside the home. When included in the model alongside other variables, the SHAP values for both were close to 0 for most students. The research team acknowledged that prior studies link parental supervision to lower exposure, but concluded that once individual, behavioral, and family factors are considered together, parental control is hard to treat as a major predictor.

The ordering shifts a bit when you split by sex. Among the remaining variables excluding gender itself, alcohol was the top predictor for male students, while problematic internet use topped the list for female students. Family conflict, anxiety attacks, and the relationship with parents all showed noticeably larger prediction contributions among female students. Differences across age groups were small.

Problematic internet use didn’t operate alone, either. It appeared in six of the model’s ten strongest interactions, and its effect size shifted depending on its entanglement with age, cannabis, alcohol, and the relationship with the father. Gender was the single strongest variable but never once appeared among the top interactions. By the research team’s reading, gender is a variable that pushes hard on its own, while problematic internet use is one that grows larger when it overlaps with other conditions.

Even the complex model barely won—and that’s a finding in itself

The three machine learning models named in the paper’s title beat logistic regression by only a hair. Across ten repeated rounds of cross-validation within the training data, AUC⁷—which scores classification performance on a scale from 0.5 to 1—came out to 0.817 for logistic regression and around 0.822 for the three boosting models. On the held-out test data, the gap was similarly narrow, about 0.005–0.006. The difference was statistically significant, but the research team itself called the effect size small, and there was no meaningful difference among the boosting models.

Which model is “better” turned out to depend on which kind of mistake you’re trying to avoid. Using 0.5 as the cutoff, logistic regression caught more actual users (sensitivity⁸ of about 0.76) but misclassified roughly a quarter of non-users as users. The boosting models cut that false-positive rate to around a fifth, but missed more actual users in the process. The team wrote that logistic regression makes more sense if the goal is to catch as many users as possible, while a boosting model may be the better fit when misidentified students would be routed into interviews, counseling, or other limited resources. Either way, they reiterated the same caveat: don’t use this to single out individual students.

The researchers were upfront about the limitations, too. The study only covered students who were enrolled in school, and all responses were self-reported. Dropping respondents with missing data may have slightly underrepresented students engaged in higher-risk behavior. Because sample weighting and survey design weren’t factored in, the findings should be read cautiously even when generalizing to Spanish adolescents as a whole—and even more cautiously when applied to other countries.

Meanwhile in Korea, AI character chat hit the National Assembly audit

This paper’s subject was the old medium of pornography. AI chatbots weren’t a variable in it. But right now in Korea, the hottest debate over teenagers’ digital use is about AI character chat.

Addiction through AI “love and friendship”… fast-growing character chat introduces teen safeguardsCharacter chat services, where users converse with AI characters, are introducing limits of 3 hours of daily use and ₩100,000 (~$72) in monthly payments to protect minors. The move comes in response to concerns raised at the National Assembly audit over teen over-immersion and harmful conversations. The industry is also strengthening safety filters as it works to build a safer service environment for young users.chosun.com

At the October 12, 2026 National Assembly audit of the Broadcasting, Media and Communications Subcommittee under the Science, ICT, Broadcasting and Communications Committee, three services came up together: Scatter Lab’s Zeta, Wrtn’s Crack, and Babychat. Lawmaker Lee Joo-hee pointed out that Zeta’s monthly usage time, at roughly 110 million hours, is more than double ChatGPT’s, and that 9 out of 10 users are teenagers or in their twenties. According to reporting, Zeta’s monthly active users stood at 1.48 million at the end of September, of which 25.7% were 19 or younger.

Findings from lawmakers’ offices were also disclosed. Representative Kim Sun-min’s office said that even with “safe mode” for minors turned on and without adult verification, the service still generated conversations depicting self-harm and suicide. Representative Hwang Jung-a’s office said that without adult verification, the AI generated sexual exploitation scenarios involving underage characters, and that after being blocked, the AI sometimes suggested on its own how to phrase things to get around the filter. It was also noted that of 119 content review complaints filed over AI character chat, 76 were still unresolved, and one complaint about Babychat had gone 494 days without a conclusion.

Government ministries also moved. The Ministry of Gender Equality and Family announced on October 15 that, after reviewing responses from some character chatbots, it had confirmed replies involving sexual content, violence, and the glorification of self-harm, suicide, and crime. It asked operators to adjust search, recommendation, and conversation scope for age-inappropriate characters based on adult verification status, and to ensure that responses encouraging self-harm, suicide, or crime are never generated regardless of the user’s age. The Korea Communications Standards Commission met with 15 AI chat operators on October 21 and demanded stronger detection of harmful content on both the input and output sides, with full-scale review set to begin in November.

Operators’ responses have followed a similar pattern. Scatter Lab explained that it had kept its detection thresholds loose to avoid blocking normal expressions, and that it had been filtering individual responses rather than reading the full context of a conversation. It said it would now strengthen detection to catch even indirect expressions, deploy a context-based detection model, and set a default daily usage limit of 3 hours and a monthly payment cap of ₩100,000 (~$72) for teen users. Wrtn’s Crack had already announced in September a cap of up to 3 hours per day and ₩100,000 (~$72) per month for teens, with plans to allow additional usage beyond that cap through guardian verification and consent. The strengthened policies take effect in early 2027.

When You Overlay the Two Stories

Let me draw a line up front. You can’t simply transplant the findings of the Spanish study onto Korean character-chat users. The country is different, the timing is different, the medium is different. And using a character chatbot can’t be equated with consuming pornography as the same kind of behavior. A US survey found that a large share of teen chatbot use looks more like play and creative writing — I covered that in Issue 79. Still, when you lay the two stories side by side, a few questions come into sharper focus.

Should We Look Only at Outputs, or Also at Usage Patterns?

Korea’s response right now is aimed almost entirely at outputs: which conversations shouldn’t be generated, which characters shouldn’t be recommended to minors. This is necessary, full stop. No user, of any age, should ever be served a sexual-exploitation scenario involving a minor character or a response that encourages self-harm.

But one of the variables that moved the most in the Spanish data wasn’t content — it was the pattern of use. Logging on because you can’t stop, because you’re cutting into sleep, because you’re trying to flee your emotions. The 3-hour daily cap that the platforms rolled out is aimed at exactly this. What’s never shown up in any reporting so far, though, is the basis on which that 3-hour default was set. Even the draft self-regulation guidelines from the Korea Internet Self-governance Organization (KISO) left daily usage limits to each platform’s discretion.

Testing the System vs. Understanding the Kids

The parliamentary audit and the ministry inspections confirmed what the system is capable of producing — essentially a red-team⁹-style test. The Spanish study, by contrast, looked at which students engage in which behaviors under which conditions. These are two different questions, and they demand two different kinds of data. The first can be answered by platforms and regulators poking at the system itself. The second requires surveys that ask the same students, together, about their usage patterns, sleep, family relationships, and other risk behaviors.

Content With No Post to Delete

The Ministry of Gender Equality and Family explained that it requested voluntary improvements from the platforms rather than demanding takedowns, on the grounds that character-chat conversations aren’t publicly visible and so there’s no individual post to request removal of. In other words, the old model — find the post, delete the post — doesn’t map well onto conversations that are generated one-on-one, on the fly. A porn site gives you tools like blocking access or removing a post. In character chat, the problem is that the same character produces a different conversation every time, depending on what the user types in.

Oswarld’s Lens

The most uncomfortable line in this paper is about parental control. How much parents controlled their teens, whether at home or outside it, barely explained teen pornography use once other factors were accounted for. By contrast, whether there was family conflict, and the quality of the relationship with father and mother, carried much more weight — especially for girls. It reads as though the state of the relationship tells us more than the amount of control does. Of course, this is cross-sectional data, so we can’t claim causation outright.

That’s why the spot given to guardian consent in the character-chat countermeasures bothers me. A mechanism where a guardian verifies and consents once a limit is exceeded is necessary. But that consent procedure doesn’t substitute for an actual conversation between parent and child. In a household where conflict already exists, a consent request could just become one more thing to fight about.

I think the default setting also deserves a word. In Issue No. 200, covering the phone-internet-blocking experiment, I wrote that what changes behavior isn’t willpower but an environment that removes the need for willpower in the first place. Three hours a day is a default the operator set. Whether this number was calibrated around children’s sleep and daily life, or set at a level that doesn’t meaningfully hurt usage time and revenue metrics, determines how the same mechanism actually functions. If self-regulation leaves usage time to the operator’s discretion, I think we can at least demand that they disclose what data was used to set that default.

Last is the question of models. There will likely be calls to attach models that estimate age or flag at-risk users. This paper shows that even with solid data from over 30,000 people, a model can miss one in four actual users, or flag one in four non-users as users. Which kind of error to minimize looks like a technical choice, but it’s really a policy choice. The same question I raised in Issue No. 116 — asking first what a published number actually measures — needs to be asked here too, this time of the model’s error rates.

Closing

The Spanish study showed that teen pornography use clusters together with problematic internet use, alcohol and cannabis, and family conflict. Korea’s character-chat debate, by contrast, is focused entirely on what shouldn’t be generated. That question absolutely needs asking — but it needs to be paired with data on which kids, in what condition, are opening that chat window in the first place. Otherwise our countermeasures will keep spinning their wheels.

---

💬 Reader, when you worry about a teen close to you and their digital habits, which do you check first — what they’re looking at, or how they’re using it?

📨 Share this piece with anyone wrestling with a child’s or student’s use of character chat.

---

References & Further Reading

Primary sources

• De Andrés-Sánchez, J., Belzunegui-Eraso, Á., Pastor-Gosálbez, I., & Sánchez-Aragón, A. (2026). “Predicting Pornography Use Among Adolescents in Spain: Findings from Ensemble Tree Models and Explainable Machine Learning.” Computation, 14(8), 177. Read the paper ··· It’s open access. I’d recommend starting with the gender-split SHAP ranking table (Table 8) and the limitations section in Chapter 5.
• Digital Today, “[Parliamentary Audit 2026] ‘Self-harm and sexual exploitation talk even with safe mode on—AI character chat lacks teen protection,’” Oct. 12, 2026. Read the article ··· This is the source for the individual lawmakers’ office inspection results and the figures on delayed reporting responses.
• AI Times, “At parliamentary audit, ‘AI character chatbots have a hole in teen protection’—will regulation accelerate?” Oct. 2026. Read the article ··· This lays out Zeta’s monthly active user numbers along with Scatter Lab’s explanation and improvement plans.
• ZDNet Korea, “[AI Now] Korea’s two leading character-chat apps add safeguards for teen use,” Oct. 12, 2026. Read the article ··· This lets you compare Zeta’s and Crack’s usage-time and payment limits.
• Asia Economy, “Ministry of Gender Equality: ‘AI character chatbots confirmed to encourage crime and illegal acts—strengthening teen protection,’” Oct. 15, 2026. Read the article ··· This is the source for the Ministry of Gender Equality and Family’s inspection methods and requests.
• ZDNet Korea, “Sexual conversations and provocative characters spreading in AI chat—review set for November,” Oct. 21, 2026. Read the article ··· Covers the Korea Communications Standards Commission’s meeting with operators and the review schedule.

Background

• Money Today, “AI chat apps popular with teens flooded with complaints about sexual content, yet government investigations number ‘zero,’” Aug. 21, 2026. Read the article ··· Shows the state of generative-AI user-protection guidelines two months before the parliamentary audit.

Related past issues worth reading

• Issue 79: How American teens are actually using AI chatbots ··· Looks at what teen chatbot use actually looks like, and the gap between play and risk.
• Issue 116: Australia’s teen SNS ban—what to check beyond the account numbers ··· Examines what the ban policy’s performance metrics actually measured.
• Issue 200: 467 people vowed to quit their phones—only 119 held out ··· An experiment showing that environmental design changes behavior more than willpower does.

---

The author is Oswarld (Kwangseob Ahn). Current roles: Adjunct Professor at Sejong University, Strategy Consultant at INLEVEL9. Career, research, books, and recent work are kept current on the About page. Latest · July 2026: HEMA-2: A Consolidation-Aware Tri-Memory Architecture with Multi-Channel Scheduling for Lifelong Conversational AI.

📝 Glossary

Footnotes

1. ESTUDES: A nationwide survey conducted every two years by Spain’s Observatory on Drugs and Addictions (OEDA) among secondary-school students aged 14–18. Though originally designed to track drug use, it also includes questions on behavioral addictions like internet use. ↩

2. Logistic regression: A statistical model that estimates how much each variable raises or lowers the probability of a binary (yes/no) outcome. Its main advantage is that the calculations are transparent and easy to interpret. ↩

3. Gradient boosting: A machine learning method that stacks many small decision trees in sequence, with each new tree correcting the errors of the ones before it. It captures complex relationships between variables well, but the results are hard to interpret directly. ↩

4. SHAP (Shapley Additive Explanations): An explainability technique that, for each individual prediction, calculates how much each variable pushed that prediction up or down. It shows each variable’s contribution to a prediction but does not prove causation. ↩

5. Cross-sectional survey: A survey design that measures the same people only once, at a single point in time. It can show whether two traits appear together, but it cannot establish which came first or which caused the other. ↩

6. Odds: The probability that something will happen divided by the probability that it won’t. It looks similar to probability but isn’t the same—so “92.8% lower odds” should not be read as “92.8 percentage points lower probability.” ↩

7. AUC: A metric between 0.5 and 1 that shows how well a model distinguishes users from non-users. A score of 0.5 is no better than a coin flip; the closer to 1, the better the discrimination. ↩

8. Sensitivity/specificity: Sensitivity is the share of actual users the model correctly identifies as users; specificity is the share of actual non-users correctly identified as non-users. Where you set the threshold creates a trade-off—raising one tends to lower the other. ↩

9. Red teaming: A testing approach in which someone deliberately probes a system from an attacker’s perspective to find weaknesses. For AI services, this means trying a range of inputs designed to elicit harmful responses. ↩