We read with great interest the article by Kıyak et al,1 ‘“ChatGPT can make mistakes” warnings fail: A randomized controlled trial’. The study provides crucial empirical evidence that medical students significantly underweight AI diagnostic advice and that prominent safety disclaimers do not alter this behaviour. This challenges common assumptions about automation bias and refines our understanding of advice-taking thresholds in medical education, particularly within the framework of Judge–Advisor System (JAS) theory.1 We commend the authors for their rigorous experimental design. Their finding—that students' perceived credibility of ChatGPT may already be at a ‘behavioural floor’—is particularly significant. It suggests that for this population and task, simple cautionary cues have reached a point of diminishing returns.1 This insight is vital for educators and developers aiming to integrate AI tools into clinical training responsibly. We wish to extend the discussion by proposing a paradigm shift in how we conceptualize the educational role of AI-generated feedback. Rather than focusing primarily on mitigating overreliance through warnings, perhaps the greater pedagogical value lies in strategically leveraging AI disagreement to foster metacognitive skill development.2 The observed tendency for students in the warning arm to more frequently justify their decisions when retaining their original diagnosis hints at this latent potential. Could structured exposure to conflicting AI advice be deliberately designed as a ‘cognitive conflict’ exercise to stimulate deeper diagnostic reasoning? In this light, AI tools might be repositioned not as authoritative diagnostic aids, but as simulated ‘adversarial partners’ in clinical reasoning training. This aligns with established educational principles of desirable difficulties and constructive controversy, where engaging with opposing viewpoints strengthens argumentation, knowledge integration and cognitive flexibility.3 Future research could explore how scaffolded reflection prompts, specifically triggered by AI counterarguments, enhance diagnostic calibration and reduce cognitive biases more effectively than disclaimers alone. For instance, a learning module could require students to articulate their reasoning both before and after encountering AI disagreement, followed by guided comparison. This approach transforms the AI's role from a source of answers to a catalyst for critical thinking. It acknowledges that the real risk in education may not be overreliance on imperfect AI but rather the missed opportunity to use its capacity for generating alternative perspectives to train more resilient and reflective clinicians.4 Kıyak et al.'s work provides the foundational evidence that students are not blindly accepting AI advice, creating the perfect conditions to test such constructive adversarial uses.1 We thank Kıyak et al. for their valuable contribution and suggest that the path forward involves exploring how AI can be used not to provide answers but to provoke better questions and more robust reasoning processes among future clinicians. Yu Xiao: Conceptualization; writing—original draft. Yuan-Xin Guo: Writing—review and editing. Liang Liu: Investigation. Zhong-Rui Ma: Supervision; validation. We would like to thank Kıyak et al. for their work. The authors declare no conflicts of interest. Not applicable. Data sharing is not applicable to this article as no datasets were generated or analysed during the current study.
Xiao et al. (Mon,) studied this question.