The Bridge from AI Literacy to AI Competence: Scope for Curriculum Development
Shazaan Nadeem, MD
Christina Short, BS
Issue 3 | Volume 2 | August 2026
There is no denying that AI is a stakeholder in the clinical landscape in the future of medical practice. The need for safe parameters for AI use is becoming paramount with each passing day. Physicians and medical residents are increasingly using AI in their daily clinical practice, yet, despite this rapid integration into the healthcare ecosystem, there is an unanswered question- Is there a minimum level of AI literacy required to practise safely? [1]
AI literacy should be treated as a core patient safety competency because trainees may begin using AI before they have developed the clinical knowledge required to recognize when its output is incorrect, incomplete, or out of context. For the purposes of this framework, we define AI literacy as understanding the capabilities, limitations, biases, uncertainty, and appropriate applications of AI systems. AI competence extends beyond literacy to the ability to integrate AI safely into clinical decision-making while maintaining independent judgment and professional accountability.
AI adoption is outpacing AI education
Naum et al. surveyed 120 respondents at an academic tertiary care centre and found that 83.5% reported AI use in clinical or educational activities, with 42.6% reporting daily use. Residents and medical students reported particularly high rates of use, while formal AI education was delayed behind adoption. [1] In a separate international survey of 1,049 physicians across 50 countries and territories, Bold et al. found that 27.8% had used AI in healthcare practice, whereas only 17.7% had received formal AI training. [2]
Why trainees may be particularly vulnerable
Clinical reasoning is not simply the ability to identify the correct diagnosis; it is the ability to integrate information from multiple sources, manage uncertainty, adapt to a patient’s hemodynamic status, and tailor decisions to the individual patient. These higher cognitive skills are developed through repeated exposure to complex clinical decision-making and cannot be outsourced to artificial intelligence. [4,5]
The increasing use of generative AI in medical education therefore raises an important concern for trainees. Reliance on AI-generated recommendations may encourage cognitive off-loading, automation bias, and, over time, loss of opportunities to develop independent clinical reasoning.[3] Without structured education, automation bias may gradually replace independent clinical reasoning. If AI supplies the initial synthesis, trainees may increasingly practise secondary verification rather than independent clinical reasoning. This shift risks anchoring clinicians to the algorithm’s first suggestion, promoting a potentially premature diagnostic plan. The education concern extends beyond whether AI produces a clinically correct answer, but it is the lack of the bridging clinical reasoning in between.
Lack of AI literacy and understanding of bias
Medical AI systems derive their outputs from human generated data, including clinical, scientific and biomedical information. We must not assume that generative AI systems are validated clinical decision tools, merely because AI outputs appear “medically sophisticated”.
Can we achieve AI competency without Cognitive Dependency?
The goal should not be to prevent trainees from using AI, but to teach them how to use it without succumbing to giving up the reasoning that clinical training is supposed to help develop. A point to highlight to medical trainees would be- a correct AI generated output does not necessarily represent a successful educational outcome. AI may be most valuable when it supports reasoning rather than replaces its earliest steps, for example, by challenging a differential diagnosis, identifying overlooked considerations, or providing feedback after the trainee has independently assessed the case.
Education should be a priority. For this reason, AI literacy is no longer optional. We argue that every U.S. medical school should incorporate a minimum AI literacy curriculum as a core patient safety competency. [4,5]
Minimum AI Literacy Curriculum
Existing frameworks have begun defining AI competencies. [6] We propose that medical education requires a minimum patient-safety-focused subset centred on clinical reasoning, verification, contextualization, and physician accountability.A core component of this curriculum should address the language of AI itself. Large language models are designed to produce fluent, coherent, and confident responses, even where uncertainty exists. Rather than focusing exclusively on technical proficiency, such a curriculum should teach adaptive expertise [4,5] : AI fundamentals of language, verifying information, and understanding its capabilities and limitations.
In medicine, however, diagnostic reasoning rarely leads to a single unequivocal answer. Clinicians routinely navigate competing differential diagnoses, incomplete information, and evolving clinical evidence. Medical trainees must therefore learn to distinguish linguistic confidence from evidential certainty. One of the most important lessons future physicians can learn is that AI can be confidently wrong. AI literacy should also prepare clinicians to recognize the inherent limitations and biases of algorithmic systems. [8] AI models are trained on historical datasets and may perpetuate existing disparities, underrepresent minority populations, or perform less reliably in patients with multimorbidity or atypical presentations. The American heart Association’s (2025) scientific statement warned that AI algorithms are limited by the quality and diversity of datasets, leading to misclassification, incorrect risk prediction, and delayed diagnoses that disproportionately affect underrepresented populations. Part of our: proposed AI Literacy Framework [6] emphasizes seven patient-safety behaviours: understand, recognize, interrogate, verify, contextualize, preserve reasoning, and own the decision.
| Step | What It Means | Example |
|---|---|---|
| Understand | Know the limitations, training data, and be aware of uncertainty of the tool. | Before trusting an AI-generated risk score, ask what population the model was used on. |
| Recognize | Identify failure modes such as hallucination, bias, automation bias, anchoring. | E.g. the model confidently suggests a dose that does not match the patient’s renal function. |
| Interrogate | Actively think through, question AI output. | What information did the model use? Does this answer fit the patient in front of me? |
| Verify | Check the evidence, sources, clinical validity, and applicability. | Cross-reference the AI’s protocol against a primary institutional guideline. |
| Contextualize | Integrate the clinical picture the model cannot see (comorbidity, social factors). | An AI summary of patient vitals means little without clinical context. |
| Preserve reasoning | Know when an independent assessment should precede AI use. | Be aware that AI is not a factual, validated, clinical decision-making tool. |
| Own the decision | Accountability. | Whatever the model suggests, the responsibility for the final decision remains the clinician’s. |
CONCLUSION
In conclusion- medicine remains fundamentally a human profession, requiring contextual judgement, ethical reasoning, communication, and longitudinal relationships that extend beyond algorithmic pattern recognition. AI can assist in organizing information and identifying patterns, but responsibility for interpretation, prioritization, and patient-centred decision-making must remain firmly within the clinician’s domain.
THE MYTH OF NARCISSUS and How it Mirrors Human Use of AI
The Myth of Narcissus is typically misremembered as a tragedy of excessive vanity. This myth is rather a story of misrecognition of the self in reflection: the young, adored hunter Narcissus, upon witnessing his reflection in a pond, is unable to recognize the image as himself. He perceives the reflection as beautiful and falls in love. The hunter observes himself until his self-recognition, he speaks to himself just as he approaches his dying moments “Iste ego sum” (“I am that one”). Beyond the distortions, general artificial intelligence systems are not validated as clinical tools. The myth of Narcissus serves as a mirror to human medical reasoning and his vanity, a metaphor for our use of AI. Although the final product of AI can appear persuasive or knowledgeable, it may do more harm than good, if not used appropriately. The proposed AI Education Curriculum will enable us to overcome the ‘facade’ of glamourising AI, risking erosion of autonomy with clinical decision-making. Instead of our focus remaining on consistent limitations, we can empower academic dialogue among Clinicians which will help foster continuous professional growth.
REFERENCES
- Naum A, Gordon RS 3rd, Deogaonkar A, Madani M, Heilbroner L, Kokoneshi K, et al. Artificial intelligence in clinical practice: usage trends and educational implications across the medical hierarchy. J Natl Med Assoc. 2026;118(3):387-390. doi:10.1016/j.jnma.2026.02.002.
- Bold B, Serin O, Tantri Adhiatma L, Demberel SE, Wangmo C, John RE, et al. Global physician perspectives on artificial intelligence in healthcare across 50 countries and territories. NPJ Digit Med. 2026;9:574. doi:10.1038/s41746-026-02726-y.
- Abdelhameed F, et al. Potential risks of GenAI on medical education. BMJ Evid Based Med. 2025;30(6):406-409.
- Mylopoulos M, Kulasegaram K, Woods NN. Developing the experts we need: fostering adaptive expertise through education. J Eval Clin Pract. 2018;24(3):674-677 doi:10.1111/jep.12905.
- Branzetti J, Hopson LR, Gisondi MA, Regan L. Training for adaptive expertise: why, what, and how. Acad Med. 2024;99(1):121. doi:10.1097/ ACM.0000000000005217.
- Pylman S, Lowry JE, Thomas TV, Billings HA, Gaynier A, Hurtubise L, et al. Artificial intelligence competencies for medical educators. Association of American Medical Colleges; 2025. Accessed August 10, 2026.
- Fanous A, Goldberg J, Agarwal AA, Lin J, Zhou A, Xu S, et al. SycEval: evaluating LLM sycophancy. Proc AAAI/ACM Conf AI Ethics Soc. 2025;8(1):893-900. doi:10.1609/aies.v8i1.36598.
- Adegunle F, Chhatwal K, Arab S, Alabdaljabar MS, Raslan MA, Sayed O, et al. Bias and oversight in clinical AI: a review of decision support tools and equity frameworks. J Gen Intern Med. 2026;41(7):1957-1968. doi:10.1007/s11606-026-10229-5.