How can medical trainees use AI without losing critical thinking skills?

Can doctors of tomorrow learn with AI without losing critical thinking? The NEJM Review offers clever AI and new collaboration models to enable educators to use AI while protecting clinical skills.

Review: Educational strategies for clinical supervision of the use of artificial intelligence. Image credit: Antonio Marca / Shutterstock

In a recent review published on New England Journal of Medicineresearchers unravel the challenge of overseeing early career health learners using powerful leading language models (LLM) as educational aids. The findings of the review highlight the dangers of “deskills” that are overreliant on AI, erode basic clinical reasoning skills, and “masskills” are unable to develop important abilities in the first place, where trainees employ errors generated by AI and together with “never skills.”

In this review, we propose a structured education framework called “diagnosis, evidence, feedback, education, and recommendation” to counter these potential AI disadvantages by scaffolding critical thinking. The review also presents the “cyborg” and “centaur” models of human collaboration, encouraging clinicians to learn to learn to be critically involved with the output generated in AI, rather than trusting them undoubtedly.

background

Recent advances in artificial intelligence (AI), particularly computational and large-scale language models (LLM), are progressing at an astonishing rate. Large-scale language models (LLMs) such as Openai's ChatGpt and Google's Gemini are increasingly used in medical learning, creating both opportunities and risks for clinical reasoning. The ever-growing literature suggests that AI tools are essentially reshaping medical learning and practice.

However, integrating AI into clinical practice presents unprecedented opportunities and significant risks for medical education. While quick access to information and the ability to expand vast amounts of data into easily accessible summaryes, it could be essential for future medical education and practice, LLMS is known to simulate human-like reasoning to create a “agency look.” This can be extremely dangerous for inexperienced medical trainees.

Therefore, medical educators face novel and urgent challenges. It teaches and oversees trainees who may be more skilled at leveraging AI than educators themselves, creating a “specialty reversal” in which teachers become learners. The current review highlights three specific hurdles that AI must overcome before solidifying its role in ensuring a safer and healthier future: “Desk Tor”, “No Skills”, and “Skilling Miss”

About the review

This review aims to address urgent and important needs by conducting a comprehensive study of the scientific literature that explores the challenges and opportunities presented by AI in medical education. We collate and integrate the results of over 70 previous publications across existing educational theories, cognitive science, and new research into human interactions, and use these insights to develop a new conceptual framework for clinical supervision of AI.

Diagnosis, evidence, feedback, education (deft), and AI engagement (deft-ai) recommendations – an adaptive framework to promote critical thinking during educational conversations about AI use.

Cyborg vs. Centaur model: A new typology that explains two different modalities of human collaboration. These models are designed to help educators and learners adapt their AI use to specific clinical tasks and associated risks.

Check the results of the survey

This review identifies and addresses several cognitive traps imposed on medical education by today's AI era. “Cognitive offload” is a process that overly relies on AI for complex tasks such as clinical reasoning, highlighting the link to “automation bias,” the overly reliance on the subsequent AI output, and the failure to catch mistakes.

Surprisingly, cognitive offloading and automation bias are not just theoretical concerns. The survey found that more than a third of senior medical students were unable to identify false LLM responses to clinical scenarios. Another study reported a significant negative correlation between the frequent use of AI tools and critical thinking ability, mediated by increased offloading, and this effect was particularly pronounced among younger participants.

This review recommends addressing these concerns by developing and adopting the deft-ai framework, a structured approach for educators, in response to trainees' dependence on AI. This proposes leveraging important conversations that move beyond AI answers to investigate learner reasoning. Key questions include “What prompts did you use?”, “How did you validate the output that AI generated?” and “How did AI suggestions affect or modify the diagnostic approach?” Educators are also encouraged to use Sackett's framework (Quest, Acquire, Evaluate, Apply, Evaluate) to teach evidence-based assessments of AI output and effective, rapid engineering techniques such as thinking inference.

The review further emphasizes that the director must distinguish between what he evaluates. AI Tool itself Its evaluation Specific output. For example, you can use an institutional scorecard or model leaderboard to determine the tool, but you must apply evidence-based medical assessment procedures to individual outputs.

Finally, this review presents “cyborg” and “centaur” modes with clinicians. In centaur mode, tasks are strategically divided, so clinicians delegate AI to clear, low-risk tasks (such as data summary and drafting communications) while maintaining complete control over high-reach clinical decisions and decision-making. This mode is recommended when dealing with complex or uncertain cases.

In contrast, cyborg mode assumes that clinicians and AI will co-structure solutions for tasks at hand. This mode is efficient for low-risk, routine tasks, but is at a higher risk of automation bias when not used for continuous reflexive monitoring and justification.

This review also warns that performance heterogeneity and bias in LLMS can exacerbate health inequality. AI systems can reduce performance in certain groups, and uncritical adoption can widen the gap rather than close.

Conclusion

The current review concludes that AI integration into medicine and medical education is inevitable (and mostly beneficial), but its successful and safe adoption is not. Medical education emphasizes that we must actively address the risks of deskills, skiing, skills and skills by fundamentally changing how clinical reasoning is taught, especially against the context of AI. Critical thinking is the basis of “adaptive practice.” This is the ability to change efficient routines and innovative problem solving when faced with the unpredictability of AI.

In summary, this review shows that the ultimate goal is not to create physicians that rely on AI, but to develop clinicians who can skillfully and safely utilize as a powerful tool to reinforce their own expertise through the “validation and trust” paradigm.