A large language model (LLM) is a very large ‘autocomplete’ trained to predict the next most likely word, with no medical ‘brain’ intrinsic to it. The article, based on the AI Lab session ‘AI fundamentals for busy cardiologists’ presented at EuroPCR 2026 in Paris, France, argues that when an LLM gives a poor answer, the problem is often the user’s prompt. These tools are already in daily use for tasks like discharge letters, guideline checks, or translation, yet adoption has proliferated faster than training. A large systematic review of 519 studies found that only 5% used real patient-care data, 44.5% tested examination-style knowledge, and 84.2% addressed question-answering. While some LLMs reach passing or near-passing scores on selected cardiology board-style examinations, this reflects ‘exam’ evidence, not ‘bedside’ evidence. The article offers one practical way to narrow the gap between widespread use and limited proof of appropriateness at the bedside.