Cyberbullying (CB), or online bullying, remains a pressing concern, particularly among youth. Yet current automated detection or moderation systems fall short in addressing the social, linguistic, and ethical complexities of this issue. This doctoral dissertation advances the methodological and ethical foundations for CB detection by foregrounding the lived experiences of young people in the development of Artificial Intelligence (AI) systems designed to safeguard them. It addresses six key research gaps in the field: (i) the lack of quality youth-centered datasets in English; (ii) the absence of ethical frameworks for involving minors in CB dataset creation; (iii) insufficient human evaluation standards that account for ecological and social dimensions of CB; (iv) limited use of multi-task learning for modelling different types of aggression and user-roles; (v) the vulnerability of natural language processing (NLP) systems to obfuscated and adversarial language used by youth; and (vi) the limited investigation into how vision–language models handle age-differentiated content moderation criteria. This interdisciplinary doctoral dissertation explores these gaps by leveraging computational and social science frameworks and involving minors, young adults and experts in bullying behaviour across different phases of this research. This results in (a) co-creating ecologically valid cyberbullying conversational datasets; (b) obfuscated or perturbed language resources for CB detection; and (c) an unseen image-text based “meme” dataset with age-differentiated labels. This dissertation also benchmarks different classification paradigms for detecting both hurtfulness as a precursor to CB detection and different user-roles (e.g., enabler, defender, bully, victim) in CB contexts. Thereby attempting to address who (user-role) said what (is it hurtful?), within youth cyberbullying conversations. Furthermore, this dissertation also evaluates open and closed source large language model (LLM) reasoning capabilities to extract explicit and implicit features within potentially harmful image-based content (memes). By integrating perspectives from computational linguistics, human-computer interaction, and the social sciences, this interdisciplinary doctoral research lays the groundwork for the ethical inclusion of youth perspectives to inform the development of context-aware cyberbullying detection systems. Content Warning: This doctoral dissertation contains examples and illustrations of antisocial online behaviour like cyberbullying that may be distressing and could potentially re-trigger or re-traumatize individuals. Reader discretion is advised.