Bioethical Governance of Relational AI for Minors: Evidence–Regulation Alignment in Article 14 of the Proposed EU KIDS Act

Background: Relational artificial intelligence (AI), including AI companions and general-purpose conversational AI systems (chatbots), is increasingly used in forms of interaction that may be experienced as socially or emotionally meaningful. For minors, these systems can offer accessible opportunities for expression, identity exploration and perceived support, while also raising questions about anthropomorphic interpretation, emotional dependency, compulsive engagement, persistence and possible displacement of human relationships. The European Commission's proposed EU KIDS Act, COM(2026) 681 final, addresses this emerging field through a dedicated set of safeguards in Article 14. Objective: This study examines how far those safeguards correspond to relational-risk mechanisms identified in contemporary scientific and normative literature concerning minors, while keeping legal scope, pre-existing EU-law baselines and future questions of implementation or effectiveness analytically separate. Methodology: A prespecified protocol combined structured evidence mapping, doctrinal EU-law analysis, directed qualitative content analysis, bioethical interpretation and an evidence-to-regulation crosswalk. Scientific and normative/legal evidence were extracted in separate streams. The final scientific corpus comprised 36 active sources and 111 mechanism-level extraction rows, alongside 20 normative/legal sources. Article 14 safeguards were classified as Direct, Partial, Indirect or No identifiable correspondence. A separate baseline analysis examined Specification, Extension and Mixed relationships with the DSA, the 2025 DSA minors Guidelines, the AI Act and the GDPR. Results: All seven predefined relational mechanisms were represented in the final corpus, together with a separate potential-benefit domain. Nine scientific sources met the strict criterion for direct minor-specific empirical evidence, three provided separable minor findings within mixed-age samples, and three were child/adolescent-specific systematic or review sources. Four mechanisms showed Direct correspondence with Article 14, two Partial and one Indirect. None fell within No identifiable correspondence. The baseline analysis identified one Specification, four Extensions and two Mixed relationships. Scientific coverage was uneven, particularly for conversational memory/persistence, and no long-term causal harm study was identified within the strict direct-minor empirical subset. Conclusions: Article 14 corresponds to the seven predefined mechanisms in different ways and with different scope limitations. The findings should not be read as evidence that the proposed safeguards are complete or effective in practice. Important uncertainties remain around age-specific generalisability, long-term causality, the embedded-deployment limits of activation and disengagement safeguards, operational standards for testing and monitoring, and the longer-term relationship between relational AI and human relationships.