Most of a modern flight is flown by the autopilot. The aircraft holds its heading, trims its fuel, and smooths out the bumps with no hand on the controls. For hours the human is a monitor rather than a pilot. And yet pilots still train for years, for the sake of a handful of seconds that may never come: the moment an alarm sounds, the instruments stop agreeing, and a system that has been quietly competent all flight is suddenly carrying everyone on board towards the ground. In that moment the task is not to trust the automation. It is to see that something has gone wrong, take back control, and land the plane. The whole of a pilot's training exists for that moment, and the more the autopilot handles the rest of the time, the more it matters that the person beside it can take over when it counts.
Medicine is moving the same way. As AI grows more capable it will carry more of the routine cognitive work: suggesting diagnoses, drafting letters, surfacing what the guidance says. The doctor's role shifts towards checking that work, seeing where it has gone wrong, and taking back control when the case does not fit the pattern the machine has learned, when the answer that is right most of the time is instead confidently wrong. The real question is not whether doctors will still matter. It is how, and how quickly, a new doctor is supposed to become good at supervising something that powerful.
What does the autopilot analogy tell us about doctors?
It tells us that the skill that matters most is supervision, and that medicine has always treated supervision as something earned slowly. A hospital consultant may have a decade or more of clinical experience before they are trusted to oversee a registrar, who in turn has years of their own before they watch over a house officer. Judgement about another person's work sits at the top of a long ladder, and doctors climb it gradually, because you cannot safely supervise care you have never learned to deliver yourself, including all the times it went wrong.
AI does not respect that ladder. A doctor qualifying next year will, from their first day, be working with tools more capable than anything their seniors trained alongside, and they will be expected to know when those tools are wrong. That expectation is not only clinical, it is regulatory: medical regulators have been clear that when an AI tool is used, it is the human using it, not the tool, who is accountable for what happens next. No one will accept "how was I supposed to know it would get that wrong" as an answer. The new doctor does not get a decade to grow into the role of supervisor. They inherit it on day one. That is the real shift the analogy points to: AI does not lower the bar for a new doctor, it raises it, and it moves the hardest part of becoming a doctor, the judgement to supervise work that is not your own, from the end of training to the very beginning.
Will AI deskill the next generation of doctors?
The honest answer is that the real danger is worse than deskilling. Deskilling is what happens to experienced clinicians whose own abilities fade through reliance on a tool, and the evidence for it is already here: a 2026 scoping review found that physicians' performance can fall once AI support is removed, and early studies show experienced endoscopists becoming measurably worse at spotting growths on their own after only months of working with an AI aid.
The sharper problem falls on those who never build the skill in the first place. Specialists and aviation-safety experts writing in Ophthalmology Times call this "never skilling": when the machine is there from the first day of training, the underlying judgement may never form at all. For an established doctor who is not deliberate about how they use these tools, AI can erode a skill they once had. For the next generation the danger runs deeper, because the skill may never develop in the first place. That is the risk most easily missed: nothing visibly breaks, the assessments are still passed, and the gap only shows itself when it is finally needed.
None of this is a reason to step back from AI. The potential gains in healthcare are too large to forgo: earlier diagnosis, fewer missed findings, and time handed back to clinicians who have far too little of it. The conclusion is not to slow down. It is that if we are going to put powerful tools in front of doctors from the very start, we have to move with just as much urgency to build the judgement that lets them supervise those tools safely.
What does "better" actually mean for a medical student now?
It means developing the judgement to supervise, and developing it far earlier than medicine has ever asked. The doctor's role shifts from being the source of clinical reasoning to being its supervisor, and supervising an answer is a higher-order skill than recalling one. Given the speed at which this technology is advancing, it would be remiss to make firm assertions about what only a doctor can do and an AI cannot. The point is not about drawing that line; it is about responsibility. When an AI proposes a diagnosis or a plan, someone has to be able to say whether it is right, where it is plausibly wrong, and when it is dangerously wrong, and to own that decision in front of the patient and the medical regulator. Recent guidance in the New England Journal of Medicine places this supervisory judgement at the centre of clinical training in the age of AI.
That judgement does not appear on its own. It rests on reasoning that has been practised until it is sound: working through cases where the answer is not handed over, making the reasoning explicit so it can be tested and corrected, and being wrong often enough, somewhere safe, to recognise the shape of a wrong answer when a machine produces one. A doctor used to accumulate that slowly, through years of supervised cases on the wards. The task now is to build it deliberately and early, so a new doctor can supervise from the start rather than after a decade they will not be given.
Won't this just put more pressure on students?
Yes, and it is worth being honest about that rather than pretending the change is painless. Students now meet capable AI while they are still in medical school, not on some distant first day in practice; nobody is walled off from these tools until they qualify. So the skills earlier cohorts were allowed to build slowly, across the early years of a career, this cohort has to start building straight away, and against far less repetitive, supervised practice than any generation before them had.
That leaves a student today with two uncomfortable options. One is to lean on AI to get through, which is the fastest path to never-skilling: passing the assessments while never building the judgement underneath, then arriving as a clinician who, on day one, cannot safely supervise any of it. The other is to try to learn the way doctors always have, through repetition and supervised reasoning, in an environment that now offers far less of both. The bar has gone up, and the traditional ways of clearing it have become harder, not easier.
That tension is real. A student would be right to feel it, to conclude that neither of those options is good enough, and that there has to be a better way to build the skills they will need.
What does this mean for medical schools?
If we are going to expect more of doctors, we have to give them something different from what was provided before, not simply more of it. For a medical school, the long-standing problem of teaching clinical reasoning at scale, as cohort numbers boom while supervisors have less time than ever to teach, is no longer only a question about delivering this year's curriculum. It has become a question about whether your students will be equipped to practise safely in the world they actually graduate into.
This is not a problem simple resourcing will fix. More of the same teaching cannot close the gap, because the very nature of the problem has changed. It now calls for delivering the curriculum in ways that have not been considered before, supported by tools that did not previously exist, built for the kind of reasoning practice students now need. A changed demand cannot be met with an unchanged method.
What does Gestalt do about all of this?
This is where the good news comes in. For students, Gestalt is built to fill exactly this gap: a place to practise clinical reasoning and the judgement that supervision demands, again and again, with feedback guided by structured medical knowledge rather than a general model left to its own devices. For medical schools looking for a genuinely different way to deliver reasoning practice at scale, rather than more of the same, this is the option you have been missing.
The shift is already under way. The next generation of doctors will be supervising capable AI from the moment they enter clinical practice. The only real question is whether they build the judgement to do so safely, or not at all.