AI Study Pinpoints the Strongest Predictors of Pornography Use Among Spanish Adolescents

A large-scale study from Spain has mapped, with unusual statistical precision, the factors that most strongly predict whether an adolescent reports having used pornography in the past month. Drawing on responses from 33,543 students surveyed in the 2023 ESTUDES national survey, researchers at Universitat Rovira i Virgili combined classical statistics with modern machine learning to rank the variables that carry the greatest predictive weight. Their conclusion is striking in its clarity: the single most powerful predictor is sex, with boys substantially more likely to report use, followed closely by problematic internet use, cannabis and alcohol consumption, family conflict, and certain risky sexual behaviours. The findings, published in the journal Computation, offer one of the most comprehensive portraits to date of how digital habits, family dynamics and health behaviours intertwine around adolescent exposure to pornography.

The research team, led by Jorge de Andrés-Sánchez of the URV’s Department of Business Management together with Ángel Belzunegui-Eraso, Inma Pastor-Gosálbez and Anna Sánchez-Aragón of the Department of History and Art History, all members of the university’s Social and Business Research Laboratory, set out with a deliberately modest goal. Rather than attempting to explain why adolescents turn to pornography, they wanted to determine which personal, family, digital and behavioural variables are most useful for predicting whether an adolescent reports having viewed or used pornography in the previous 30 days. That framing matters, because prediction and explanation are not the same thing, and the authors are careful throughout to keep the distinction in view.

Methodologically, the study is a showcase of how contemporary data science can be layered onto traditional epidemiological survey work. The researchers ran a conventional logistic regression alongside three gradient-boosting machine learning models: XGBoost, LightGBM and CatBoost. These ensemble tree algorithms are widely used in fields ranging from finance to medicine because they excel at detecting complex, non-linear relationships between variables that simpler models can miss. On top of the predictions, the team applied explainable artificial intelligence techniques, a family of methods that reveal how much weight each input variable contributes to a model’s output. Without such tools, machine learning models often function as black boxes, producing accurate forecasts while hiding the reasoning behind them.

The results were notable for their consistency. All of the models, from the simplest logistic regression to the most sophisticated boosting algorithms, produced broadly similar rankings of predictive variables. The machine learning systems did slightly improve some performance indicators, but the differences compared with logistic regression were small. The authors draw a practical lesson from this: more complex models should not automatically replace traditional statistical techniques. In cases like this one, where the underlying relationships appear largely additive and interpretable, a well-specified regression can deliver nearly the same insight with far greater transparency. The added complexity of ensemble methods, they suggest, buys only marginal gains here.

What the models did agree on is the hierarchy of predictors. The sex of the individual stands out by a wide margin as the variable carrying the most weight, with boys considerably more likely to report pornography use. Age is also associated with use, though its predictive importance is lower. After sex, the next most significant factor is problematic internet use, a construct defined by behaviours such as having difficulty staying offline, sleeping less in order to be online, neglecting obligations, or using the internet as an escape from worries or negative emotions. The researchers are careful to note that problematic internet use, as measured in the survey, does not equate to a clinical diagnosis of internet addiction; it is a behavioural profile, not a medical label.

Further down the ranking sit cannabis and alcohol use and family conflict, both of which add meaningful predictive signal. Certain risky sexual behaviours, such as having sex without a condom or experiencing a sexual encounter that the teenager regrets the next day, also contribute slightly to the predictions. Perhaps counterintuitively, parental control shows a much more limited effect once all the other variables are analysed together. In isolation, parental oversight might appear influential, but in a multivariable model its apparent power largely dissolves, suggesting that its observed association overlaps with other family and behavioural factors captured more directly elsewhere in the data.

One of the study’s most interesting contributions is its analysis of differences between boys and girls. When the researchers excluded sex from the analysis and ran the models separately for each group, the internal hierarchy of predictors shifted. Among boys, alcohol consumption emerged as the most important predictor of pornography use, whereas among girls it was problematic internet use. The weight of family conflict also varied between the two groups, while age-related differences proved smaller. This kind of subgroup analysis is exactly where explainable machine learning earns its keep, revealing patterns that a single pooled model would average away and that could easily be missed in a conventional analysis.

The authors are emphatic about the limits of what their models can claim. The study is observational, based on cross-sectional survey data, which means it can identify associations and predictive patterns but cannot establish cause-and-effect relationships. Whether problematic internet use drives pornography use, whether pornography consumption contributes to problematic internet habits, or whether both reflect some shared underlying disposition simply cannot be determined from this design. For the same reason, the researchers insist the models should never be used to diagnose, label or sanction specific adolescents. Their purpose, as de Andrés-Sánchez puts it, was not to single out individuals but to better understand which factors tend to be associated with this behaviour at the population level.

That understanding points towards a particular vision of prevention. The conclusions argue for programmes that combine sex and relationship education, digital literacy, critical reflection on pornographic content, substance misuse prevention, the promotion of safe sexual practices and the strengthening of family communication. The logic is straightforward: if pornography use clusters alongside problematic internet habits, substance use and family conflict, then interventions that treat it in isolation are likely to miss the mark. Effective prevention, on this view, has to address the broader ecosystem in which the behaviour is embedded rather than targeting a single variable.

The study’s authors summarise the message in terms that capture both its scientific and its social significance. Pornography, they argue, should not be addressed as a phenomenon disconnected from the other dimensions of adolescent life; it is part of a broader ecosystem of digital practices, family relationships, substance use and sex education. In an era when artificial intelligence is increasingly applied to sensitive social questions, this research offers a model of restraint as much as discovery: powerful algorithms, transparent explanations, honest acknowledgement of correlational limits, and a refusal to convert population-level patterns into individual judgments. What remains is for educators, families and policymakers to translate those population-level patterns into approaches that treat adolescent digital life as the connected whole the data show it to be.

Subject of Research: Predictors of adolescent pornography use identified through machine learning analysis of a large Spanish youth survey

Article Title: Researchers identify the main factors in pornography use among adolescents

Article References: Researchers identify the main factors in pornography use among adolescents. (n.d.). Original publication

Image Credits: AI Generated

DOI: Not provided

Keywords: adolescence, pornography use, machine learning, XGBoost, problematic internet use, substance use, family conflict, risky sexual behavior, ESTUDES survey, logistic regression, explainable AI, prevention