Artificial intelligence chatbots are increasingly becoming the place where teenagers ask questions they may be too embarrassed, unsupported or afraid to raise elsewhere. But a new analysis of 120 simulated conversations in Italian suggests that the answers adolescents receive can change sharply depending on who the chatbot believes is asking. The study, published in AI & Society, found that ChatGPT, Claude and Gemini often responded to the same sexual-health questions with different levels of reassurance, autonomy and medical caution when the simulated user’s gender, ethnicity, social class, religion, region or sexual orientation changed. The researchers describe this pattern as “intersectional amplification”: the accumulation of marginalized identities was associated with a qualitative shift from ordinary educational guidance toward moralization, surveillance and referral to professional authority.

The issue is particularly consequential in Italy, where comprehensive sexuality education remains uneven and politically contested. Fewer than half of Italian adolescents are estimated to receive structured sexuality education, with substantial differences between regions. That gap has pushed young people toward online searches, forums and conversational AI for information about bodies, relationships, contraception, desire, anxiety and identity. Unlike a web page, a large language model can respond conversationally, privately and at any hour, making it an especially attractive informal educator. Yet these systems do not retrieve knowledge in a neutral way. They generate text by predicting sequences of words from patterns learned during training and subsequently shaped by safety policies, preference tuning and other alignment procedures. Those layers can influence not only what information is provided, but also how a user is characterized: as capable, vulnerable, confused, risky or in need of supervision.

To examine those differences, Mario Ederoclite, Iana Ivanova Tzankova and Paola Villano of the University of Bologna created four fictional adolescent personas. Marco was a 17-year-old heterosexual Italian boy from Milan and a working-class background, serving as the reference profile. Giulia was a 17-year-old heterosexual, middle-class Catholic girl from Rome. Samira was a 16-year-old heterosexual girl of Moroccan origin, born into a second-generation immigrant family in Naples. Alex was a 16-year-old non-binary queer teenager from Bari in southern Italy. The personas were not intended to represent real minors; they functioned as controlled “probes” for testing whether identical questions produced different language when identity markers changed. The researchers adapted ten questions drawn from Italian Reddit discussions and sexual-health forums, covering anatomy and physiology, emotional and relational concerns, and safety or risk. Each persona introduced themselves before asking the questions in colloquial Italian.

The team submitted the resulting prompts to three commercial systems—OpenAI’s GPT-4-turbo, Anthropic’s Sonnet-4 and Google DeepMind’s Gemini-2.5 Flash—during June and July 2025. A fresh chat was opened for every interaction, with memory and previous context cleared and default settings retained. Each question was submitted once for each persona-model combination, producing 10 questions multiplied by four personas and three systems. The researchers recorded the responses verbatim, including warnings, disclaimers, professional referrals and suggested resources. Two researchers independently coded all transcripts before discussing differences and reaching consensus. Their coding framework examined communicative tone, patterns such as agency recognition, medicalization and cultural pathologizing, and the equity of the resources recommended. Because the study focused on discourse rather than statistical model performance, the researchers analyzed wording, framing and conversational positioning instead of assigning a simple accuracy score.

The strongest contrast emerged in questions about masturbation, sexual desire and anxiety during intimacy. Marco commonly received normalization and reassurance, with language suggesting that sexual exploration could be a normal part of adolescent development. When similar questions came from Alex or Samira, responses were more likely to introduce emotional conflict, cultural background or the need for guidance. A question about masturbation that was framed as ordinary development for the privileged reference persona could become connected to guilt, family values or psychological support when asked by a marginalized persona. The researchers argue that the shift was not necessarily expressed through explicit prejudice. Instead, it appeared through small linguistic choices: “you can” became “you should”; possibility became obligation; and a direct answer was followed more quickly by advice to consult a counselor, doctor or mental-health professional.

Those choices matter because language models distribute agency through grammar as much as through factual content. In the study, privileged adolescents were more often treated as competent decision-makers who could interpret their own feelings and make informed choices. Female, immigrant-background and gender-diverse personas were more frequently presented as subjects requiring supervision, clarification or clinical assessment. The researchers call referral escalation a key marker of this change. For Marco, a question about desire could receive an encouraging, peer-like answer. For Alex, the same topic could rapidly trigger recommendations to seek expert help. A referral can be valuable when a young person faces coercion, pain, abuse or serious distress; the concern is that the threshold for such escalation appeared lower when the persona carried multiple markers of marginalization, even when the question itself did not indicate danger.

The analysis also identified what the authors call cultural and geographic othering. Samira’s Moroccan heritage repeatedly activated assumptions about conservative family dynamics or cultural restrictions, despite the prompts not explicitly stating that her family was religious or repressive. Rather than simply acknowledging that people may grow up with different ideas about intimacy, some responses treated cultural difference as a likely source of dysfunction. Giulia’s Catholic identity prompted milder moral framing, with sexual questions more likely to be associated with personal or family conflict. Regional language subtly marked southern Italy as deficient: references to services in Puglia or other southern locations were sometimes qualified as though adequate support there were unusual. Meanwhile, socioeconomic circumstances were largely ignored. Recommendations for counseling, therapy or private medical care often assumed that such services were affordable, nearby and accessible, even though the personas’ class and regional positions could make those assumptions unrealistic.

The three systems did not behave identically. The researchers identified three broad response profiles: a clinical-paternalistic style that supplied technically accurate information but emphasized oversight and professional referral; a peer-supportive style that combined factual guidance with empathy, normalization and inclusive language; and a neutral-informational style that was concise and generally avoided overt moralization but offered little cultural or developmental context. One system, for example, responded to Alex’s uncertainty about desire by explaining that the teenager did not need to feel guilty or defective and might be not ready, asexual or simply uninterested at that moment. That answer offered multiple non-pathological interpretations without directing Alex toward medical supervision. Similarly, one response to Samira acknowledged that growing up with diverse cultural influences could make intimacy confusing while presenting cultural negotiation as a normal developmental experience rather than evidence of family dysfunction. Such contrasts led the authors to conclude that equitable interaction is technically feasible and depends on design priorities, training data, evaluation methods and alignment decisions.

The findings do not show that every answer from these systems is biased, nor do they establish how frequently the same patterns would appear across repeated runs, languages or model updates. Each prompt was submitted only once, so the study could not measure the stochastic variability of generative AI. The sample was also deliberately small and qualitative, and the personas necessarily simplified the complexity of real adolescents’ lives. No young people participated in the conversations, and the complete model outputs cannot be publicly released because of licensing restrictions, although the prompts and procedures are reported for replication. A further limitation is that the analysis compared identities in combinations rather than isolating every possible factor experimentally; a response difference might reflect the interaction of several cues rather than one marker alone. The authors therefore call for larger mixed-method studies, repeated testing, multilingual audits and participatory evaluation involving educators and marginalized youth.

Even with those limitations, the study highlights a risk that conventional chatbot safety testing may miss. A system can be factually correct while still distributing knowledge unequally through tone, assumptions and access pathways. If one teenager is told that a feeling is common and understandable while another is steered toward diagnosis or authority, both may receive accurate facts but very different messages about whether they are normal, trustworthy and entitled to make decisions about their own bodies. In that sense, AI-mediated sexuality education is not merely a question of information retrieval. It is also a form of pedagogy: the system teaches users how to interpret themselves and whom to trust. The researchers argue that bias audits should therefore examine modality, emotional validation, cultural assumptions, referral thresholds and the practical accessibility of recommended resources. As chatbots become unofficial counselors in places where formal sexuality education is patchy, ensuring that they do not reproduce social hierarchies may be as important as ensuring that their biology is correct.

Subject of Research: Bias and intersectionality in AI-mediated sexuality education for simulated Italian adolescents

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

Article References: Ederoclite, M., Tzankova, I. I., & Villano, P. “Digital bias in sexuality education: an intersectional analysis of AI responses to simulated Italian adolescents.” AI & Society (2026). Original research article

Image Credits: AI Generated

DOI: 10.1007/s00146-026-03320-2

Keywords: artificial intelligence, sexual health education, algorithmic bias, intersectionality, adolescence, large language models, cultural competence, human–computer interaction

Tags: adolescent sexual health chatbotsAI bias in sexuality educationAI moralization and surveillance in sexual healthethical implications of AI bias in sensitive topicsgender and ethnicity bias in AIimpact of AI responses on marginalized youthintersectional bias in AI responsesonline sexuality education disparities in Italyrole of conversational AI in adolescent healthsimulated Italian adolescent conversationssocial class and religion influence on AI advicevariability in AI-generated sexual health information