Opinion Piece by Dr Colin John Greengrass, Senior Lecturer in Pharmacology & Academic Director of Technology Enhanced Learning, RCSI Medical University of Bahrain

A clinician's main protection against diagnostic error is the capacity to recognise that a diagnosis may be incorrect. This capacity may not always be preserved when Artificial Intelligence (AI) is introduced into clinical decision-making, and it is what the design of such AI systems should protect. In a paper published this month in Frontiers in Digital Health, the BRACE (Bounded Reciprocal Adaptation for Clinician Engagement) framework describes how.[1]

The foundation for the work was established in January, with a paper published in Frontiers in Digital Health examining how AI augments human cognition.[2] Clinicians hold only so much in mind at once; fatigue and interruption erode that capacity, and they default to pattern recognition, fast and usually accurate, but fragile when a complex case is encountered. That article described the ways clinicians reason through a case, and where AI could support each of them. This was recently applied to emergency care [3] in the journal Academic Emergency Medicine, with Croskerry, Campbell and Clancy, leaders in the field of diagnostic reasoning, as coauthors. Interestingly, the January article has since been cited 19 times and accessed more than 9,000 times, which suggests an appetite to examine AI use in diagnosis in terms of cognition, rather than software capabilities or accuracy.

Evidence shows that AI models now outperform physicians on diagnostic reasoning tasks [4], yet giving clinicians an AI tool does not reliably improve their decisions.[5] In one study of difficult cases, clinicians using an AI assistant felt more confident and found the work easier, but confidence rose without corresponding improvement in accuracy in one third of cases.[6]

Thus, although shifting cognitive effort of recall, synthesis and ranking to AI is often appropriate, metacognition, the awareness and active control of one's own learning and cognitive processes, should always remain part of human decision-making. Where metacognition is transferred to AI, the clinician stops judging independently whether the reasoning holds and whether the response fits this patient. The safeguard is then absent, while the appearance of careful reasoning remains.

This forms the augmentation paradox, with AI potentially making consultations faster and better, precisely by assuming the work through which a clinician sustains independent reasoning, detects errors and remains able to practise without AI. Three risks follow:

● Deskilling is the weakening through disuse of a diagnostic reasoning capability already held.
● Never-skilling is different: the capability never forms. AI supplies the response; the clinician or student never works the case through their own reasoning and never learns the diagnostic structure.
● Mis-skilling is the propagation of errors from inaccurate AI outputs into later reasoning, as though they were correct [1,7,8].

This framework addresses these risks. At the hospital or health-system level, BRACE introduces clear boundaries to decouple safe learning from formal performance reviews. As a result, where clinically safe, the clinician proposes a working diagnosis before the AI system provides its own, and what is uncertain stays visible. Importantly, the reasoning the system records, mistakes included, stays within a protected space meant for the clinician's own learning, and is not open to employers.

Disagreement is where the reciprocity is tested. When a clinician and an AI system reach different conclusions, BRACE treats the difference as informative in both directions. A substantive disagreement may indicate a possible error in the system, which the clinician can refer for expert scrutiny, or an opportunity to work that case through formatively. Neither route is an assessment of the clinician.

In training, simulation enables clinicians to meet these systems failing by design, working difficult cases without AI before doing so with it, and with AI, including confidently projecting incorrect answers. A forthcoming paper by the author examines how teams need to respond likewise within an AI-augmented environment, and how simulation can prepare them.[9]

This body of research has developed from asking 'what can AI do to support clinical reasoning?' to asking what it costs in independent and unaided performance over time. The ultimate aim is a reciprocal partnership in which judgement and accountability remain with the clinician, who can still reason independently when AI is absent, or its output is contested or incorrect.

References
1. Greengrass CJ. Rethinking AI in clinical decision support: a framework for reciprocal human-AI interaction. Frontiers in Digital Health. 2026;8:1887161. doi:10.3389/fdgth.2026.1887161
2. Greengrass CJ. Transforming clinical reasoning: the role of AI in supporting human cognitive limitations. Frontiers in Digital Health. 2026;7:1715440. doi:10.3389/fdgth.2025.1715440
3. Greengrass CJ, Clancy M, Campbell SG, Croskerry P. Transforming human-AI collaboration in emergency care: the role of AI in adaptive support across diagnostic reasoning modes. Academic Emergency Medicine. 2026;33(8):e70398. doi:10.1111/acem.70398
4. Brodeur PG, Buckley TA, Kanjee Z, et al. Performance of a large language model on the reasoning tasks of a physician. Science. 2026;392(6797):524-7. doi:10.1126/science.adz4433
5. Goh E, Gallo R, Strong E, et al. GPT-4 assistance for improvement of physician performance on patient care tasks: a randomised controlled trial. Nature Medicine. 2025;31(4):1233-8. doi:10.1038/s41591-024-03456-y
6. Ong KT-S, Seo J, Kim H, et al. Success and failure of human-AI collaboration in clinical reasoning: an experimental study on challenging real-world cases. International Journal of Medical Informatics. 2026;211:106342. doi:10.1016/j.ijmedinf.2026.106342
7. Berzin TM, Topol EJ. Preserving clinical skills in the age of AI assistance. The Lancet. 2025;406(10513):1719. doi:10.1016/S0140-6736(25)02075-6
8. Ke Y, Jin L, Ong JCL, et al. AI-induced never-skilling in medical education. Nature Medicine. 2026;32(6):1997-2006. doi:10.1038/s41591-026-04438-y
9. Greengrass CJ. Artificial intelligence in clinical simulation: a framework for team-level never-skilling, metacognitive independence and reciprocal design. Forthcoming, Frontiers in Medicine, 2026.