The clinical blind spot: artificial intelligence risk awareness and clinical preparedness among child and adolescent psychiatrists

Introduction

Artificial intelligence (AI) has entered the daily lives of children and adolescents, creating new demands on child and adolescent psychiatry. These technologies may give rise to clinical challenges such as AI-reinforced maladaptive beliefs, AI-related self-harm content, and deepfake-based victimisation. Despite this growing exposure, it remains unclear how prepared child and adolescent psychiatrists are to recognise and manage these issues. In this study, we explored clinicians’ awareness of AI-related risks, how often they screen for them, and how prepared they feel to respond.

Methods

In this cross-sectional survey, 50 child and adolescent psychiatry residents, specialists, and academic clinicians completed a structured questionnaire. The survey assessed professional experience, personal and professional AI use in the previous three months, self-reported AI-related clinical encounters in the previous 12 months, routine enquiry about patient AI or chatbot use, and assessment of AI-related risks. Conceptual knowledge was measured with five multiple-choice items covering hallucination, the ELIZA effect, jailbreak, deepfake technology, and algorithmic mirroring, yielding a score from 0 to 5. Perceived competence in risk assessment, psychoeducation, and intervention or referral was rated on 5-point Likert scales. Knowledge scores were compared between residents and non-residents using Mann-Whitney U tests. Self-rated competence across the three clinical domains was compared using the Friedman test.

Recruitment is ongoing, and these results are preliminary.

Results

Participants were mostly residents (n = 40, 80%) and women (n = 41, 82%); mean age was 29.7 ± 2.5 years. Frequent or routine professional AI use was reported by 39 participants (78%). Only one clinician (2%) routinely asked about patient AI or chatbot use, and four (8%) routinely assessed AI-related risks; 45 (90%) did neither. Twenty-one participants (42%) reported at least one specific AI-related clinical encounter in the previous 12 months. The most common were AI-related self-harm or suicide risk (n = 8, 16%), reinforcement of eating-disorder thoughts or behaviours (n = 8, 16%), deepfake or voice-cloning victimisation (n = 7, 14%), and AI-reinforced delusional thinking (n = 6, 12%). Ten participants (20%) had not enquired or were unsure.

The mean knowledge score was 3.88 ± 1.42, with item-level recognition ranging from 70% for the ELIZA effect to 84% for deepfake technology and algorithmic mirroring. Knowledge did not differ significantly between residents (n = 40, mean = 4.00) and non-residents (n = 10, mean = 3.40; U = 252.5, p = 0.179). Self-rated competence was highest for psychoeducation (3.24 ± 0.89), intermediate for risk assessment (3.16 ± 0.87), and lowest for intervention or referral (3.00 ± 0.78; Friedman χ² (2) = 7.35, p = 0.025). Forty-nine participants (98%) indicated at least some need for dedicated training or clinical guidance.

Conclusions

Many clinicians are already encountering AI-related issues in practice, yet very few are routinely assessing for them. Confidence is lowest when it comes to managing or referring these cases, and overall training in this area appears limited. Almost all participants indicated a need for dedicated training or clinical guidance. These early findings point to a gap between emerging clinical needs and current practice in child and adolescent psychiatry.

References

No references