Medical Education in the Age of AI: A Mini Guide for Future Doctors and Faculty
In this article
⚕️ Why Should Medical Education Be Redesigned in the Age of AI?
Medical education has seen many technological innovations over the past 50 years, including the introduction of computers into clinical practice, the internet, electronic health records and simulation laboratories. However, artificial intelligence (AI) and particularly large language models (LLMs)are creating a revolution unlike any previous innovation. These systems not only make information easier to access, but also give the appearance of human-like reasoning .
A medical student can now obtain a differential diagnosis list, draft a treatment plan or prepare a patient information note within seconds. Although this speed and convenience seem appealing, they carry serious risks: deskilling (the erosion of existing skills), never-skilling (failure to develop basic competencies) and mis-skilling (learning incorrect recommendations). For example, studies show that a substantial proportion of students cannot recognize incorrect AI answers, potentially setting the stage for clinical errors.
The central question for medical education in the coming period is therefore becoming clearer:
👉 “Rather than banning AI, how can we teach its proper use and help future physicians become stronger critical thinkers and independent decision-makers?”
The article on which this discussion is based appeared in the New England Journal of Medicineon August 20, 2025: “Educational Strategies for Clinical Supervision of Artificial Intelligence Use”.
🧠 The Two Sides of AI in Education
✅ Opportunities
- Making clinical simulations more realistic
- Quick, easy access to information
- Reducing students’ routine workload to create time for higher-order thinking activities
- Structuring complex information and providing immediate feedback
⚠️ Risks
- Deskilling: The gradual erosion of newly learned clinical reasoning skills. Just as using Google reduces information retention, excessive reliance on AI can weaken students’ own reasoning muscles.
- Never-skilling: Failure to develop critical thinking skills in younger generations. Research shows that students who frequently use AI have lower critical thinking scores .
- Mis-skilling: Learning biased or incorrect AI outputs. For example, AI may overestimate pneumonia and overestimate heart failure in certain populations. Moreover, low-performing clinicians have been shown to achieve worse results than AI itself when they become dependent on it.
For example, when young physicians rely too heavily on device output when interpreting ECGs, it can prevent their own interpretation skills from developing. The same mechanism applies to AI: accepting an output without verification can undermine learning.
📖 Clinical Scenario
After seeing a patient in the outpatient clinic, an internal medicine resident takes out their phone and asks ChatGPT: “What are the most likely diagnoses with these symptoms?” They save the output and present it to their supervisor. The educator then faces a dilemma: “Should I turn this into an opportunity, or warn the student about relying on ready-made information?”
This is precisely where the DEFT-AI framework comes in.
📊 The Scale of the Risk in Numbers
- 35% of advanced medical students failed to recognize incorrect AI answers.
- Students who used AI more frequentlywere found to have lower critical thinking scores.
- In a randomized controlled trial, participants who used AI recommendations without questioning them performed worse than those who did not use AI.
🔎 DEFT-AI: Turn Student–AI Interaction Into a Teaching Moment
- Diagnosis/Discussion: Find out which prompt the student used and how they asked follow-up questions.
- Evidence: Assess whether the output aligns with guidelines and the literature, and ask whether alternative hypotheses were generated.
- Feedback: Reflect on the student’s strengths and weaknesses and clarify what needs improvement.
- Teaching: Reinforce critical thinking and evidence-based medicine principles, and discuss biases.
- AI Recommendation: Specify whether AI can be used with or without supervision according to the task’s risk.
🤖 Two User Profiles: “Cyborg” and “Centaur”
| Feature | Cyborg | Centaur |
|---|---|---|
| Definition | Works together with AI at every step | Divides tasks between AI and the human |
| Suitable Area | Low-risk, creative work (e.g., a patient education brochure) | High-risk diagnostic processes (e.g., ECG interpretation) |
| Risk | Deskilling and automation bias | Slower but safer |
Note: The article emphasizes that a physician’s ability to switch between modes according to task risk is critical for adaptive practice .
⌨️ AI Literacy and Prompt Engineering
AI outputs often involve a leap of faith with no visible reasoning behind it . Teaching “AI literacy” is therefore at least as important as teaching pharmacological or statistical literacy.
Example of a Good Prompt
“I am a pulmonology resident. I have a patient whose asthma is not responding to treatment. According to current guidelines, what are the most likely differential diagnoses? Please explain your reasoning.”
Example of a Poor Prompt
“Does this patient have asthma?”
The Difference: A good prompt provides context, an unbiased question and a request for reasoning; a poor prompt is leading and narrow.
📚 A Checklist for Evidence-Based AI Use
- Ask: Clarify the clinical problem.
- Seek evidence: Question the basis of the output.
- Critique: Is there bias, missing data or prejudice?
- Apply: Use the output for the individual patient, rather than blindly.
- Evaluate: What was the outcome, and was it verified?
🌟 Conclusion: The “Verify First, Then Trust” Paradigm
With proper training, AI is an ally that accelerates clinical reasoning rather than an enemy of medicine. To achieve this:
- Medical curricula should be redesigned and AI integrated into education.
- Educational institutions, healthcare systems and AI developers should collaborate.
- Student–AI interactions should be systematically monitored and evaluated.
Sources
Was this article helpful?
Your feedback helps us improve our content.
Reader comments
Share your comment
Your name and comment are published only after administrator approval. Your email address is never shown to visitors.
Be the first to comment.