EAPE · ARTIFICIAL INTELLIGENCE & LEARNING

AI as a Catalyst for Paediatric Education

Generative AI can create new opportunities for practice and feedback. The educational gain depends on how we use it—and what we ask learners to do themselves.

4 MIN READ · EDUCATIONAL PERSPECTIVE

Imagine a trainee practising a difficult conversation with a simulated parent, then repeating it with a different concern or level of understanding. Imagine an educator adapting a fictional case for students at different stages of training. Generative AI makes these activities easier to prototype.

For paediatric education, the attraction is clear: more opportunities to practise, explore alternatives and discuss reasoning. These are promising uses, not guarantees of better learning. The central question is whether the activity develops a learner’s capability beyond the interaction with the tool.

Start with a learning objective

An AI activity should begin with a specific educational need. Does the learner need to justify a differential diagnosis, explain uncertainty to a family, identify missing information or give a structured handover? Once the objective is clear, the educator can decide what the AI should do and what must remain the learner’s task.

Useful starting points include fictional patient interviews, alternative explanations for different audiences, and draft questions that a teacher checks before use. The system can play a role in an exercise without becoming the authority that decides whether the clinical content is correct.

UNESCO’s guidance on generative AI in education emphasises human agency, appropriate pedagogical design and institutional validation. Applied to paediatric teaching, this supports a simple principle: choose the learning task first, then assess whether the tool serves it. ¹

Model performance is not the same as learner improvement

A strong answer from an AI system does not tell us whether a student has understood it, can identify its errors, or can perform without assistance. It also does not establish that a tool improves care for children.

Evidence from physician studies illustrates why the distinction matters. A 2024 randomised trial involving 50 physicians found no significant improvement in diagnostic reasoning when participants had access to an LLM alongside conventional resources. The participants were trained in family, internal or emergency medicine, and the tasks used clinical vignettes. ²

A separate trial of 92 physicians, published in Nature Medicine in 2025, found improved scores on simulated management-reasoning tasks with GPT-4 assistance, alongside longer time spent per case. Neither trial establishes improved paediatric learner competence or patient outcomes. Their different findings argue for evaluating the task, the users and the outcome—not assuming a universal benefit. ³

Keep paediatric context explicit

A case about a child should specify developmental stage and the roles of the child and caregiver. An exercise designed for adult medicine may overlook how communication, consent, family involvement and clinical interpretation change across childhood.

Educators should review generated cases for internal consistency, stereotypes, missing context and inappropriate certainty. Translated explanations need checking for meaning and suitability; fluent language alone is not a quality check. Any clinical assertions, references or suggested doses require independent verification against appropriate current sources before teaching use.

WHO’s guidance on large multimodal models identifies risks including inaccurate output, bias, privacy failures and automation bias. These concerns are especially relevant when apparently confident text is used to teach learners who are still building the knowledge needed to challenge it. ⁴

A practical exercise: commit, compare, critique

Here is one proposed classroom routine. First, give learners a wholly fictional case and ask them to write an initial interpretation without AI. Next, let the AI offer a competing explanation or ask questions from a simulated caregiver. Then ask learners to identify one useful contribution, one limitation and one claim that needs verification.

Finish with a faculty-led discussion and a short independent task. For example, the learner might explain which new information would change their reasoning. The purpose is to make the learner’s thinking visible, rather than reward polished AI-assisted prose.

For assessment, distinguish activities where AI is permitted from those intended to demonstrate unaided competence. Make the expected disclosure of AI use clear. Otherwise, it becomes difficult to know what the submitted work actually shows about the learner.

Build evaluation and safeguards into the activity

Use fictional cases wherever possible. Do not place identifiable patient information in public AI tools; follow institutional rules for any clinical-data use. Rare combinations of details can also create identification risks even when names have been removed. ⁴

Start with a small, supervised teaching activity. Assess factual accuracy, learner experience, accessibility and whether the intended skill transfers to an independent task. Record the tool and version used, because an activity’s behaviour may change after an update.

AI can help educators create practice opportunities. Its most valuable contribution will be demonstrated when learners become better at asking questions, checking evidence and explaining decisions—and can do those things when the tool is no longer beside them.

A thought to take into practice

Use AI to generate practice, questions and alternatives. Keep educators responsible for validation and require learners to demonstrate their own reasoning.