Dispelling the magic of artificial intelligence in medical education

Abstract

We read ‘When I say… artificial intelligence’ by Bearman and Ajjawi and ‘A systematic review of large language models and their implications in medical education’ by Lucas, Upperman and Robinson with interest.1, 2 These articles describe what artificial intelligence (AI) is and how it can be implemented in medical education. Both articles approach AI as a black box whose inner workings cannot be understood, only implemented and the outputs evaluated. Unfortunately, this approach evokes Arthur C. Clarke's quote: ‘Any sufficiently advanced technology is indistinguishable from magic’.3 We write to advocate for dispelling the magical aura around AI in medical education. That clinicians must learn how AI works to use it in patient care has been argued elsewhere.4 Instead, we suggest that the knowledge needed for medical AI literacy is already at our fingertips.5 Consider this example. A woman presents to hospital with dyspnea and manifestations of emphysema, volume overload and anaemia. These physiologies all interact, making a unifying differential diagnosis difficult to resolve. However, if one considers each system separately, the problem becomes more tractable, and a final diagnosis found more readily. One can approach AI education similarly. AI is essentially a tool that receives an input, runs it through an algorithm and produces an output mimicking human performance. These outputs generally depend on the principles of statistics, computing and, for large language models, semantics. How much knowledge students need to judge an AI's appropriateness for a given clinical situation is not yet known.6 It stands to reason, though, that understanding AI's underlying ‘physiologies’ is a good starting place. Thankfully, we have seen these principles before. Biostatistics is already taught in most medical school curricula. While computer programming is reserved for subspecialty training like clinical informatics, logic and the critical appraisal of evidence forms a backbone of regular clinical practice. Similarly, semantics, the study of how words gain meaning from their context, is not taught explicitly to trainees. However, its concepts form the basis of how we understand modern diagnostic reasoning and underpin how we make decisions every day. What is missing then is not more coursework on AI. More coursework in such a rapidly changing field would require near-constant revision and further strain an already overloaded curriculum. Instead, AI concepts may best be tied into pre-existing teaching on biostatistics and clinical reasoning. By dispelling AI's magic and showing students it is just several familiar concepts working in concert, we can prepare them to understand both the AI we have now and whatever arrives in the years to come. Casey N. McQuade: Conceptualization; writing–original draft; writing–review and editing. Thilan P. Wijesekera: Conceptualization; writing–original draft; writing–review and editing. David Chartash: Conceptualization; writing–review and editing; writing–original draft. Data sharing not applicable to this article as no datasets were generated or analysed during the current study.

Ähnliche Arbeiten