Digital bias in sexuality education: an intersectional analysis of AI responses to simulated Italian adolescents

Abstract Large language models (LLMs) are increasingly used by adolescents to seek information about sexuality, relationships, and well-being. However, little is known about how these systems discursively construct and mediate such interactions. This study examines how three commercial LLMs (ChatGPT, Claude, and Gemini) respond to simulated Italian adolescent personas used as analytical probes for algorithmic bias. Drawing on 120 standardized Italian-language interactions across four personas differing by gender, class, ethnicity, and sexual orientation, the analysis investigates how identity markers shape the framing of sexual and affective health information. Qualitative analysis reveals recurrent patterns of medicalization, differential agency attribution, and cultural othering. As markers of marginalization accumulate, AI responses exhibit a systematic discursive shift from normalization toward surveillance framing and referral to professional authority. To account for this pattern, we introduce intersectional amplification as an analytical construct grounded in Italian-language interactions about adolescent sexuality with commercial LLMs, capturing a situated process of discursive differentiation within this sociotechnical configuration. Overall, the findings show that LLM-mediated sexuality education reproduces normative hierarchies embedded in training data, positioning adolescents differently along axes of privilege and marginality. Crucially, the study demonstrates that such inequalities are not imposed by technical constraints but reflect design priorities.