There is a fairly understandable temptation when one studies medicine and, at the same time, builds things with artificial intelligence: to treat the tool as if it were the same everywhere, because on the notes screen and beside the patient's bed the model answers with a similar fluency, arranges information with a similar neatness, and produces, in both cases, the provisional feeling that something has already been settled, even though that surface resemblance is misleading precisely to the extent that the same system changes nature according to the regime of responsibility into which it is placed, so that confusing those regimes is not a technical detail but an error of judgment.
In study, AI works almost like a tutor available at any hour, able to explain a mechanism, compare differential diagnoses, rewrite a summary until the explanation sounds clear, and even point out what one forgot on a list, and the risk there is real though still limited: mistaking reading fluency for mastery, because it is possible to finish a session with the impression of having understood a topic precisely because the generated text was coherent and discover the next day, after closing the screen, that barely a blurred version of that coherence remains, which is corrected, in principle, with a simple test (trying to reconstruct the reasoning without the model), since if one cannot, what one had was not understanding but well-written intellectual company.
At the bedside, the same habit becomes something else, because it is no longer a matter of measuring whether one "understood" a chapter but of deciding what to do with a concrete person, with little time, with incomplete data, and with consequences that are not undone by clearing a chat, and when the model delivers a clean clinical summary the temptation ceases to be only intellectual and becomes operational: acting as if that summary had absorbed the uncertainty, so that the danger does not necessarily consist in the AI getting a lab value wrong, but in ceasing to notice what the summary left out (the tone, the contradiction in the history, the detail mentioned in passing that does not fit, what the patient minimizes out of embarrassment), because that usually lives in the residual and not in the tidy text.
That is why the useful distinction is not "use AI" versus "do not use it," but recognizing two different regimes: in study the error is paid for in an exam, in a false sense of competence, or in an explanation one could not defend, whereas at the bedside the error is paid for in someone else's body, so that using the tool in both places can be reasonable even though transferring without a filter the mode of confidence from study to the bedside is not, because the model summarizes patterns with an effectiveness that sometimes exceeds a tired student's memory, while clinical work often begins in deciding what deserves to enter that summary and what, precisely because it does not fit, requires one to stop.
I am not writing this as a prohibition but as a matter of proportion: AI speeds up learning the map (mechanisms, algorithms, lists that used to take hours), even though the bedside still requires someone to walk the ground and notice where the map still fails to describe what one is seeing, because mistaking speed of explanation for clinical judgment does not make anyone more modern and only makes it easier to feel certainty at the wrong moment, so that perhaps the most honest way to use these models, at least from where I am now, is precisely that: let them accelerate study and, at the same time, distrust any summary that sounds too clean when in front of you there is someone who does not quite fit into a list.
