Abdul Mannan: Are We Teaching Doctors to Think or Just Trust AI?
Abdul Mannan, Consultant Haematologist, Director Haemophilia Centre, Haematology Lab Lead at Betsi Cadwaladr University Health Board, shared a post on LinkedIn about a recent article by Junseong Lee et al., published in Digital Medicine, adding:
“The AI was wrong.
Trainees accepted 52% of its incorrect suggestions.
That should concern anyone responsible for training the next generation of doctors.
A new study in npj Digital Medicine examined what happens when clinicians work with AI, particularly when the AI makes mistakes.
14 clinicians. 125 neonatal radiographs. 1,750 interpretations.
Two findings caught my attention.
52%: Trainees accepted incorrect suggestions from an otherwise reliable AI system.
91.7%: Three experienced neonatologists correctly overruled erroneous recommendations from a deliberately flawed AI system.
These were different experimental conditions, not a direct head-to-head comparison. And the specialist sample was small.
But the question this study raises is important.
What happens when doctors become so accustomed to AI being correct that they stop questioning it?
This was a radiology study, not a study of language models or haematology.
Still, imagine an AI confidently misinterpreting an APS result, overlooking an important molecular finding, or recommending anticoagulation without considering a bleeding risk.
Would the trainee challenge it?
Or accept it because the answer sounds convincing?
I think we need to change how we teach with AI.
Before showing trainees an AI-generated answer, ask them to commit to their own interpretation.
Then ask one more question:
“What evidence would convince you that the AI is wrong?”
A clinician who can use AI is useful.
A clinician who knows when to disagree with AI is safer.
The study does not prove that AI causes long-term loss of clinical skills. But it gives us a good reason to test how well doctors detect AI errors.
Are we training doctors to think with AI, or simply to agree with it?”
Title: Expertise modulates automation bias and sentinel behavior in human–AI collaborative diagnosis of neonatal pneumoperitoneum
Authors: Junseong Lee, Yeonsu Kim, Victoria Kim, Changhyun Park, Jeong Min Song, Jimin Kwon, Yoojin Nam, Patrick Lenehan, Dong Yeong Kim, Young Ah Cho, Pyeong Hwa Kim, Jae-Yeon Hwang, Jaeseong Lee, Byong Sop Lee, Euiseok Jung, Ah Young Jung, Jinwha Choi, Namkug Kim, Hee Mang Yoon
Find more posts featuring Abdul Mannan on Hemostasis Today.
-
Oct 26, 2026, 14:18Shadi Yaghi: Evidence Supports DOACs as an Alternative to VKAs for Cerebral Venous Thrombosis
-
Oct 26, 2026, 14:15Kale Gubae: Comparing the Effectiveness and Safety of Warfarin and Rivaroxaban in Atrial Fibrillation
-
Oct 26, 2026, 14:06Mokhtar Tazik: Evidence-Based Regulation to Strengthen Plasma Supply and Patient Safety
-
Oct 26, 2026, 13:57Gladys Delancey: Spotting the Warning Signs of Vein Disease
-
Oct 26, 2026, 13:36Kefah Mokbel: Targeted Breast Cancer Therapies Linked to Increased Thrombotic Risk
-
Oct 26, 2026, 13:32Jorge Enrique: Beyond Fibrin Targeting New Strategies for Thrombolysis
-
Oct 26, 2026, 13:23Priya Sriskandarajah: Treatment Challenges in Elderly and Frail Patients with Philadelphia-Negative MPNs
-
Oct 26, 2026, 12:48Milica Kuzmanovska: Latest Updates in PV, ET, and Myelofibrosis Management
-
Oct 26, 2026, 12:38Hytham Abu Aagla: Daratumumab and Transfusion Compatibility