I have just spent two days at the Royal College of Radiologists' Global AI in Healthcare Conference, around the theme of safe and practical implementation. It was two days of dense, genuinely useful conversation, and rather than try to capture all of it, I want to set down the things that struck me most, particularly through the lens that matters most to us, which is education.

The real gap in AI education is delivery, not curriculum

The single most clarifying point of the conference, for me, was this. A great deal of what we need to teach radiologists about AI already exists in the curriculum. The foundations of machine learning, the basic statistics, the principles of testing AI in clinical settings, much of it is already written down. The problem is not that the content is missing. The problem is that it is not being delivered.

There are two structural reasons for that. First, there are not enough AI-competent faculty at local level to teach it, and it is genuinely hard to teach a subject your trainers do not yet feel confident in themselves. By some estimates around ninety per cent of consultants are not comfortable with AI, which tells you the scale of the training-the-trainers challenge. Second, none of it is assessed, and in any training system, what is not assessed tends not to be learned.

That reframing matters, because it changes the solution. If the gap is delivery rather than content, we do not have to wait years for a perfect new curriculum. We can start teaching what already exists now, and build the faculty capacity to do it. There was a strong sense through the conference that the specialty cannot afford to wait, because AI is being deployed in radiology today, and if radiologists do not lead on it, others will.

The risks education has to take seriously

What I appreciated was the honesty about the risks, rather than blind enthusiasm.

The one that stayed with me most concerns how trainees learn to see. As AI increasingly triages and filters cases, there is a real danger that trainees stop being exposed to the normal, routine studies on which pattern recognition is built, and are left seeing only the complex and the abnormal. If AI quietly removes the straightforward work, we risk producing radiologists who never built their foundations properly.

Closely related is automation bias. A trainee shown an AI's answer before forming their own tends to anchor to it, and may never develop the independent confidence to disagree. That is a serious concern, because the entire value of a radiologist in an AI world rests on being able to oversee, question and overrule the machine when it is wrong. There was a real call to make sure trainees still learn to read independently first, before the AI's opinion is in front of them.

The radiologist as gardener

One of the most useful framings of the conference came from Dr David Little , who described the future radiologist as an orchestrator of AI, the person who directs it rather than being replaced by it. I think that is exactly right, and it captures something important: the radiologist as the one who synthesises imaging, clinical context and AI outputs, holds and communicates risk and uncertainty, and carries the judgement that AI cannot.

Reflecting on it afterwards, I found myself reaching for a slightly different image that, for me, builds on the same idea. I keep thinking of the radiologist as a gardener.

A gardener does not grow the plants by force, and does not control everything. What a gardener does is exercise constant, patient judgement: deciding what to nurture and what to cut back, spotting a weed among the seedlings, pruning what has grown the wrong way, and shaping the whole towards something healthy and deliberate. AI in radiology produces a great deal of growth: findings, drafts, suggestions, flags. Much of it is genuinely useful. Some of it is a weed dressed up as a flower. The radiologist's role is to know the difference, and to tend the result with a skilled and discerning hand.

That is a more honest picture of what working well with AI actually feels like. It is not glamorous command, but judgement, care, and the quiet expertise to know what belongs and what does not. An old line was quoted at the conference that has aged well: AI will not replace radiologists, but radiologists who use AI may replace those who do not. I would add only that using it well is less like issuing orders and more like tending a garden, and it is exactly that discernment we should be teaching.

Using AI to learn, not to avoid learning

There was also a very practical thread running through the conference about how trainees can use everyday AI tools to study better, and I found this genuinely energising. Turning lecture transcripts into structured summary tables, generating tailored flashcards and memorable visual mnemonics, using tools that answer only from your own uploaded material so they do not hallucinate, even rehearsing your viva reasoning against an AI acting as an examiner. Used like this, AI is the best study partner a trainee has ever had.

Mind maps are a good example, and a personal one. I relied heavily on mind maps when I was preparing for my own 2A, and to some extent for the 2B as well, because laying knowledge out visually was how it finally stuck for me. Building them by hand took hours, which was part of the value but also a real limit on how many I could make. It was good to hear Alice Giucca describe how AI can now generate a mind map in seconds, from a topic, a transcript or a set of notes, leaving you free to spend your time studying it and reshaping it rather than drawing boxes and arrows.

There is a wider point hiding in that small example. If a tool can now do in seconds what used to take a person hours, then a great many products and companies built around providing that effort will have to adapt or be left behind. The organisations that thrive, in education as everywhere else, will be the ones that embrace these tools rather than resist them. That is precisely the spirit in which we are trying to build.

The principle that governs all of it is simple. AI helps you when it makes you think, and harms you when it thinks instead of you. That is a distinction I intend to keep at the centre of how we build.

The shortcut that costs you later

Not all of the conference was optimistic, and one session in particular, on the positives and pitfalls of AI in medical education, named a problem that is growing quietly and fast. AI is now being used to cheat. Not only in the obvious sense of answers fed into online exams, but in reflective portfolio entries written by a machine rather than a doctor, in supervisor reports, in assignments and commentaries that were never really the candidate's own work. It is a genuine threat to the integrity of how we train and assess doctors, and the response will rightly involve more verification, more face-to-face assessment, and real consequences for those who are caught.

But I want to say something to any trainee tempted by it, because the institutional response is not really the point.

If you use AI to shortcut your learning, the person you cheat is yourself. You may pass the exam, but you will not have learned the thing the exam was standing in for, and that gap does not disappear. It follows you into the reporting room, into the clinics, into the moment a patient is in front of you and there is no model to prompt you. You will be poorer for it, not richer, however the result reads on paper.

And there is something deeper still. The habit of taking the shortcut, of being dishonest when it is easy and unobserved, does not stay confined to one exam. It becomes a way of working, and eventually a way of being. People who build their careers on shortcuts tend, in the end, to be found out, and even when they are not, they have quietly settled for being less than they could have been. The ethos you choose now, honesty, effort, doing the real work, will shape the rest of your professional life far more than any single result.

This is exactly why the distinction I keep returning to matters so much. Used to help you learn, AI makes you better. Used to avoid learning, it hollows you out. The tool is the same. The choice is yours, and it is a choice about character as much as about study.

A regulatory aside that I found fascinating

One thread I had not thought much about before was regulation, and it has stayed with me. Clinical AI, the kind that informs a diagnosis or guides treatment, is regulated as a medical device. That brings a heavy and entirely appropriate burden of validation, governance and post-market surveillance, because the stakes for patients are high. Building genuine clinical AI is therefore far more complex than the impressive demonstrations might suggest.

What interested me is that AI used purely for education may sit in a different place. A tool that helps a trainee revise, or that generates teaching material later checked by a human, is not making clinical decisions about a patient, and so, as I understand it, it does not carry the same medical-device classification. I want to be careful here, because this is my own reading rather than a regulatory ruling, and anyone building in this space should take proper advice. But the distinction is an important one, and it carries a responsibility of its own. The fact that educational AI may be less heavily regulated does not mean it can be less rigorous. If anything, it places the burden of accuracy squarely on us, because there is no external framework forcing the quality, only our own standards and our own radiologists checking the work.

What we take from it

For Revise Radiology, the conference confirmed the direction we were already moving in. We have used AI seriously for some time, to map and build teaching from a case bank far too large to handle by hand, always with experienced radiologists checking the output. What is becoming clear is that using AI ourselves is no longer enough. The profession needs help closing the delivery gap, and an education organisation is well placed to be part of that.

That is why we are establishing an AI committee at Revise Radiology , bringing together people who understand both radiology and the technology, to guide how we build AI into our teaching and how we help others develop real AI literacy. I have some excellent people in mind, and we will announce them shortly.

The foundations already exist. What has been missing is delivery, and rigour, and the confidence to lead. Those are exactly the things a good education organisation should provide, and it is the work we are committing to.

With thanks

This piece grew out of the talks and conversations at the conference, and I am grateful to the speakers whose thinking shaped it. My particular thanks to Dr David Little , whose talks on the AI curriculum and the future of radiology training gave me much of the framing here, including the idea of the radiologist as the one who directs AI. To Dr Alice Giucca, whose practical demonstrations of how trainees can use everyday AI tools to learn and revise were genuinely inspiring. To Dr Ayo Oluboyede , for a clear-eyed account of what the day-to-day clinical impact of AI means for training. To Dr Daniel Fascia , whose session on the foundations of coding and AI creation prompted my reflections on regulation and medical devices. And to Dr David Marshall, whose talk on the positives and pitfalls of AI in medical education prompted the reflections on academic honesty above. Any errors or oversimplifications in how I have captured their ideas are entirely my own.