As a first-year fellow, I faced the responsibility of relaying difficult news to a parent with a child in the neonatal intensive care unit (NICU): an infant of 28 weeks’ gestational age whose head ultrasonography showed a small, grade II intraventricular hemorrhage. This finding is not infrequent in our world, and a seasoned neonatologist may not have to think twice before going into the room to have a conversation with the family. But as a new fellow, I was struggling with how to best communicate with the anxious new parents who were overwhelmed with medical terminology and grieving the loss of the uncomplicated newborn period they had imagined. I knew the key points: the hemorrhage was limited and the prognosis was generally favorable, but uncertainty remained. Still, I wanted to speak in a way that was both medically accurate and deeply human. As is my habit, I opened an artificial intelligence (AI) medical language model and asked it to summarize the long-term prognosis, outline how to discuss uncertainty without instilling fear, and suggest ways to communicate empathy clearly and directly.The conversation with the family felt calm, honest, and grounded. We talked about what we knew, what we did not know, and the plan for monitoring over time. I emphasized that we would be there with them, not just now, but moving forward. Afterward, a bedside nurse, someone who has witnessed countless versions of these conversations, pulled me aside to say, “You handled that beautifully. They really felt heard.”I remember feeling grateful. And then, unexpectedly, a little unsettled. I had used AI as part of my preparation. Had I “cheated”? Was I supposed to rely on only what I carried in my head? On my own ability to find the right words?The truth is that using AI did not replace my medical knowledge, judgement, or empathy. What it did was support those things. It helped me organize a complex emotional and clinical task so I could show up more fully for my patient’s family. If a tool can aid clinical work and help providers care more compassionately, the question is no longer whether it should be used, but how to do so safely.Given the ubiquity of this technology, it is no surprise that trainees are already turning to it for support.1,2 But how can young physicians leverage AI to enhance patient care and accelerate learning without replacing the foundational development of clinical reasoning and family-centered communication? And, as we navigate this shift, how do we recognize the risks of misinformation? To answer this, AI must be treated not as an answer key, but as a complement to the profession, ensuring that while medical tools may change, the wisdom and humanity in medicine remain the compass.The comfort and effectiveness of using large language models (LLMs) in medicine increases as trainees strengthen their own clinical knowledge throughout residency. Unlike a standard search engine where one types a keyword and retrieves a list of links, interacting with an LLM relies on “prompting,” which is the process of giving the model specific instructions, context, and constraints to generate a tailored response. The quality of the output depends heavily on the specificity of this input, often requiring an iterative conversation to refine the results.3,4 However, technical proficiency in prompting is insufficient without a robust medical foundation. Deep knowledge of physiology and disease pathology acts as the ultimate guardrail, enabling the physician to recognize when a response, however confident, defies biological plausibility, applies information out of context, or lacks the necessary clinical nuance.4With this critical lens applied by an experienced user’s inputs, medical LLMs can be utilized to refine differential diagnoses. By inputting a de-identified, complex clinical presentation, trainees can ask the model to broaden the differential, engaging in an iterative process that challenges it to consider rare etiologies or conflicting symptoms, an application where recent studies have shown LLMs can perform comparably to physicians in specific contexts.5–7 The goal of such integration is not to outsource the cognitive work of diagnosis, but to provide a robust sounding board that compels the trainee to critique and sharpen their own analytical skills.Of course, medical LLMs can be leveraged to do more than simply supplement an expert’s clinical judgment. As in my experience, these tools also offer a unique venue to simulate potentially difficult conversations, allowing clinicians to rehearse by asking the model to play the role of a parent with specific concerns or varying levels of health literacy. This in turn provides a safe and easily accessible environment for communication training.8 Additionally, AI facilitates the rapid synthesis of nuanced, patient-specific prognostic data,9 which, together with the clinician’s generalized professional knowledge, assists in tailored decision-making.While the potential benefits of these tools are extensive, so are the pitfalls. A primary risk is the phenomenon of “hallucination,” where an LLM generates plausible sounding but entirely fictitious information. Because these errors are often presented with confident, academic formatting, they can be harder to detect than standard internet misinformation. Trainees must learn that an LLM is a language predictor, not a truth engine; it does not “know” medicine but merely predicts the next likely word. Therefore, verification must be treated not as an optional step, but as a core competency. Referenced studies should never be accepted at face value, as AI can fabricate titles and authors that appear legitimate. Finally, clinicians must recognize that AI inevitably reflects the biases of their training data; if left unchecked, these tools can reinforce existing inequities in health care, making it the clinician’s duty to critique outputs rather than blindly accept them.Trainees occupy a unique space: we are newly joining fields of medicine at the very moment medicine itself is transforming. We are early adopters, optimists, and critical observers. This position gives us both responsibility and opportunity. As trainees, we can advocate for educational curricula that incorporate AI literacy while also advocating for model transparency and an equity-centered design.While AI can streamline the gathering and organization of information, it is our presence that gives meaning to that data. The role of the pediatrician remains, at its core, relational: to care for children and their families in a way that honors their values and experiences. If we fail to preserve this, we risk gaining efficiency while losing the essence of our profession. The task ahead, therefore, is to pair the vigilance of a conscientious user with the deep commitment I felt in that NICU room, where the goal was never just to deliver a prognosis, but to ensure a parent felt truly heard. By doing so, we ensure that as medicine evolves, we deliver care that is both more informed and more deeply human.
Lima et al. (Wed,) studied this question.