Beyond rules: A focused comparative vignette analysis of how public AI systems interpret sexual coercion, privacy abuse, and grooming in adolescent digital scenarios

Abstract

Adolescents increasingly turn to conversational artificial intelligence systems for information, interpretation, and advice about intimate relationships, sexuality, privacy, and online conflict. Yet current AI safety evaluations often focus on explicit content violations and harmful outputs, while paying less attention to morally ambiguous forms of relational harm, such as coercive sexting, sextortion, and online grooming. These harms are especially difficult to identify because they often appear through ordinary relationship language, including loyalty tests, emotional pressure, secrecy, flattery, and threats framed as intimacy.

This study examines how four publicly accessible conversational AI systems interpret adolescent digital harm scenarios that vary in explicitness and relational ambiguity. Drawing on Communication Privacy Management theory, relational ethics, and research on adolescent digital vulnerability, the study developed a twelve-vignette matrix crossing three scenario families: pressure to send intimate images, threats to distribute previously shared images, and grooming-related secrecy and exclusivity. The present article reports a focused qualitative comparative analysis of three theoretically selected vignettes drawn from this matrix: subtle relational pressure to send an intimate image, explicit sextortion, and grooming through flattery and exclusivity. Model responses were analyzed using a descriptive coding rubric covering harm recognition, conceptual precision, moral clarity, responsibility attribution, developmental sensitivity, guidance quality, and actionability.

The findings from the three focal vignettes show strong convergence across models in scenarios involving explicit image-based sexual extortion and clear adult minor power asymmetries. In these cases, all systems recognized harm, attributed responsibility to the perpetrator, and avoided victim blaming. Greater variation emerged in more ambiguous scenarios involving emotional pressure, normative minimization, and subtle relational coercion. Although models often recognized manipulation, they differed in their use of precise harm terminology and in the specificity of their protective guidance.

The study contributes a relational and developmentally informed framework for evaluating AI responses to adolescent digital harm. It shows that the central challenge for youth facing AI systems may not be whether models can detect obvious abuse, but whether they can name, explain, and respond to coercion when harm is embedded in ordinary relationship language.