[Artificial intelligence in child and adolescent psychiatry: A critical review of uses, evidence and relational stakes].

OBJECTIVES: To critically appraise the main current and emerging uses of artificial intelligence (AI) in child and adolescent psychiatry across three high-burden domains - autism spectrum disorder (ASD), attention-deficit/hyperactivity disorder (ADHD), and depression and suicide risk - and to examine the level of evidence, methodological limitations, and the ethical, safety and relational challenges specific to this population. METHODS: Critical review of the literature. PubMed/MEDLINE and reference documents from international bodies (WHO, UNICEF, UNESCO) and the European regulatory framework were searched for 2018-2025, combining terms for artificial intelligence, machine learning, digital phenotyping and conversational agents with autism, ADHD, depression, suicide, child and adolescent. Studies were grouped into three domains: diagnostic and screening support; digital and AI-supported interventions; detection of depression and suicide risk. Systematic reviews, meta-analyses, randomised controlled trials, landmark cohort studies and regulatory-cleared devices were prioritised. Synthesis is narrative and critical, without quantitative meta-analysis. RESULTS: In ASD, video-, sensor- and machine-learning-based systems support early detection; the only cleared device (Canvas Dx, FDA De Novo) reports high accuracy on determinate outputs in both pivotal and real-world studies, but leaves a substantial proportion of indeterminate results and lacks robust data on care-pathway impact. In ADHD, instrumented attention tests (QbTest) and a cleared digital therapeutic (EndeavorRx) show moderate effect sizes and unequal access. For depression and suicide, predictive models derived from electronic health records and digital phenotyping achieve fair discrimination (area under the curve often 0.80-0.86) but remain internally validated, with major concerns regarding bias, false positives and governance. Across domains, adolescents frequently describe conversational agents as available, non-judgmental and sometimes perceived as more "empathic" than human interlocutors, raising the question of a reconfiguration of the care relationship. CONCLUSIONS: AI offers genuine opportunities for detection, monitoring and some interventions in child and adolescent psychiatry, on the basis of evidence that remains limited and rarely tested in routine care. In a developing population, the integration of these tools must remain subordinate to clinical judgement, reinforced data protection and the centrality of the embodied encounter. AI should be conceived as an adjunct to care, not a substitute for the therapeutic relationship.