Elsevier Medical Education presents an interview with Dr Philip Xiu, Honorary Senior Lecturer at the University of Leeds and Elsevier ClinicalKey Student UK: Assessment Editor-in-Chief and AI Board Chair, exploring how AI can enhance learning, improve assessment nuance, and support more equitable outcomes in medical education.
What clinical skills have been most critical to your success as an educator and a leader?
Deep listening
I think there are a couple of skills that are quite important. One of the foremost ones that comes to mind is listening. It sounds quite simple, but it’s the same skill that gets me through a 10-minute GP consultation in the morning, and that’s one that’s carried me furthest as an educator and healthcare leader. In the consulting room, if I’m not listening properly, I’ll often miss the diagnosis and the nuance of the patient’s agenda, if they have any. In a faculty meeting, if I’m not listening properly, I often miss the person behind the problem. And that’s what I’ve learned – the fact that the real listening isn’t about just hearing words, it’s about creating space for what gets to be said. That’s even more important when I’m in a meeting with the students and they tell me that they’re fine but actually, I’ve maybe noticed that their clinical reasoning has changed and they become more formulaic, or they have an element of memorization. Then that’s often, in my mind, someone who’s struggling with imposter syndrome or feeling out of place. Similarly, if I have a faculty colleague who may be from a different cultural background and goes quiet in a curriculum discussion meeting, that might signal we’re unconsciously designing for one type of learner from one specific type of background. I think that listening skill really helps me pick up those hidden concerns, especially from patients, about their symptoms. And similarly, those specific hidden concerns are what allow me to recognize when our own educational environment isn’t working for everyone.
Productive humility
I guess another skill that I can think of is about productive humility and being an educator, or a clinician, comfortable saying, "I don’t know, let’s find out together." To be fair, medicine has spent decades, hundreds of years, rewarding certainty. But I think the best educators would serve to model uncertainty well because those educators who are great, and certainly those who I role model, should say that it’s okay not to have all the answers. That’s where curiosity lives. This is especially important when we are working with a group of diverse student cohorts, because each of them can bring in different perspectives and can also challenge our existing assumptions.
You have extensive experience in personalized learning and have contributed to the development of numerous Elsevier publications, as well as tools such as StudyFinder AI and Assessment Creator within ClinicalKey Student. From that perspective, what traditional limitations of digital assessment do you think AI is addressing for the first time?
I think for a long time now, digital assessment has been quite good at certain things. One of the things that it’s good at is scale. The caveat of that is that it’s remarkably poor at other things, which is nuance. For example, faculty could deliver 1,000 MCQs to 1,000 students overnight however, this might just be measuring recognition instead of clinical reasoning. In medicine, that’s quite a dangerous gap to have.
AI shifts assessment focus from output to reasoning process
Now, what AI is doing for the first time, is closing that gap between what a student answered and why they answered it in that specific way. We can potentially generate follow-up items that can probe a specific reasoning pathway – not just any endpoint of the diagnostics. We can potentially adapt assessment difficulty in real time, matched to the aptitude of the student’s journey. We can also use AI to flag up learners who might consistently anchor on a fixed bit of information or misconception which, by the way, is exactly the sort of cognitive bias that can harm patients in my GP surgery on a Monday morning, as an example.
I would say that the old-style digital assessment is testing, “Did I get it right?” Whereas I think the AI-driven assessment is probably a more nuanced approach. It’s asking, “How did you think?”, “How did you get there?” That shift from an output to the process that leads to the output is what we need to focus on. And, for the first time, we can do this at scale without exhausting our faculty.
How is personalized learning evolving with the integration of AI, particularly in collaboration with faculty?
I think the biggest fear when AI first arrived is that it’s going to replace the educator, and I think the opposite is happening. What’s actually happening is it’s restoring the educator.
AI restores faculty to mentorship, not just administration
For a long time now, faculty has been buried under marking, assessments, content updates, responding to student queries. But AI is now actually shouldering a lot of the load which means that, actually, faculty can return to the part of teaching that really inspired them to do it in the first place. That includes mentorship, modelling, professional identity, and, for me personally, sitting in with a struggling learner and seeing their own development and their own pathway.
That’s what personalized learning means to me. It means much more than the computer system picking your next question. Now, hopefully, it means there is a continuous conversation between the learner, the content that they’re trying to tackle, and the educator themselves, and AI is really the connective tissue between all of them. I can see, as a faculty member, where my student is stuck, even before they tell me, and I could potentially intervene earlier and with much better data to back it up. And in all honesty, it does save the faculty time as well, because the learner has often self-resolved any content queries by their own AI searches anyway.
"Simply put, the AI can tell you who is failing and where they’re failing at, but I would argue only an educator can tell you why. And only that relationship between the educator and the student themselves can tell the student how the educator is able to help support them through that whole process."
Dr Philip Xiu, Hon. Senior Lecturer, University of Leeds | Elsevier Faculty Hub, June 2026
In this context, what is the role of the educator in assessment tools – such as tracking, continuous evaluation, remediation, and performance measurement? And what advantages does AI bring in strengthening that role?
I would say the role of the educator remains that of an interpreter. The data is not judgment because a dashboard telling me a student is in the bottom decile tells me almost nothing about why and how they’re struggling with the content. Maybe they’re struggling with their own internal confidence. Maybe they have other contextual issues, environmental issues, like caring responsibilities at home. Or if the curriculum was never designed with someone like them in mind.
I think that the AI’s strength is having pattern recognition across such longitudinal data sets, something that the human brain is not really built for. It could potentially show me, longitudinally, a learner’s reasoning errors clustering around specific clinical presentations or that their performance dips every time the assessment is delivered in a specific format. That’s real goldmine information for the educator. The remediation conversation with the learner then is, “Let’s sit down and work out what’s going on,” which can remain human, using the data as a linchpin.
Simply put, the AI can tell you who is failing and where they’re failing at, but I would argue only an educator can tell you why. And only that relationship between the educator and the student themselves can tell the student how the educator is able to help support them through that whole process.
"The AI is very much like a mirror. It’s not like a window. It will reflect whatever and whoever built it using whatever data set it has. So, if you want to reduce inequality, the people building AI tools will really have to look like the world that is meant to serve you."
Dr Philip Xiu, Hon. Senior Lecturer, University of Leeds | Elsevier Faculty Hub, June 2026
What role does technology play in reducing global inequalities in medical education? And more specifically, how can AI contribute to this?
Medical education has, for far too long, been almost like a postcode lottery, but on a global scale. This is because the students’ access to high-quality teaching, to simulation, laboratories, expert feedback, and up-to-date evidence has really depended on where the medical student happened to train. Technology, at its best, is the great equalizer here, because a medical student in Nairobi or Dhaka can now access the same evidence-base as one based in Cambridge or Boston.
AI scales expert feedback across geographies and languages
I think AI takes this further because it scales the scarcest resource in global medical education, which is the expert educated feedback. Now, you simply can’t ship 10,000 senior clinicians around the world but what you can do is build tools that can approximate very high-quality forms of feedback in any language, at any hour, and calibrate it to the local context and the local geography. And that is new.
The equity risk
But, and this really matters here, AI will only reduce inequality if it’s built equitably, because if it’s trained on data from just one patient population, it will perpetuate the inequities of that specific population data set. The AI is very much like a mirror. It’s not like a window. It will reflect whatever and whoever built it using whatever data set it has. So, if you want to reduce inequality, the people building AI tools will really have to look like the world that is meant to serve you.
As a highly experienced editor, what types of bias persist in educational content? Which has been the most difficult to identify or correct, and how can AI help address this challenge?
Visible bias in educational content
The biases that we frequently tackle are the visible biases. For example, as an editor, we see medical textbooks that only show dermatological signs on white skin. Or perhaps a question box, which has case vignettes, where every cardiologist is a man, and every patient with anxiety is a woman. We’re far from perfect, but we can see those biases now as educators.
Hidden biases in educational content – and how AI can help
The biases that persist are, I think, much more subtle and potentially more dangerous. So, we need to have an understanding of whose perspective, whose experiences, and whose ways of understanding are considered authoritative. For example, whose research are we using to teach medical students as evidence and as gospel? The second aspect of it is, which clinical narratives are framed as a classical case vignette of a typical medical condition, and which of those are presented as atypical, when the atypical label probably just means that that particular patient is not white, not male and not Western. Thirdly, we are very much aware there are conditions and whole swathes of populations and context which are simply absent from the medical curriculum themselves.
These biases are so hard to spot, not because of what’s written, but because of what’s not written. And that’s where AI can hopefully help, because this pattern analysis across the whole swathes of bodies of content can surface these absences of knowledge that no individual editor could ever see. So, the bias that you see is a problem – I agree with this. The bias that you can’t see is the hidden curriculum, and the AI’s gift to the editors is making the invisible biases more visible.
"…AI will not replace educators. However, I think AI is beneficial in rescuing the educators’ bandwidth. This is mainly because educators and teachers who use AI will have the time, the data and also the reach to do the human parts of the teaching better than anyone else."
Dr Philip Xiu, Hon. Senior Lecturer, University of Leeds | Elsevier Faculty Hub, June 2026
You manage multiple roles, including that of an educator. What advice would you give to teachers who are starting to incorporate AI into their practice?
Start with the teaching problem, not the AI tool
A couple of things. The first thing is that you start off with the problem – you don’t start off with the tool. Never ask, “How do I use AI?” Instead, you should be asking, “What part of my teaching life is failing my learners?”, “What part of my teaching life is draining myself?” Then, you look for the tool, whether that be AI or not, that helps with that specific problem. Otherwise, what you tend to find is that, especially newer educators, you’ll be spending months upon months playing with chatbots and you’re not actually resulting in any better teaching outputs.
Treat AI outputs with the skepticism
The second thing is you should stay skeptical but also curious. You should treat every AI output the way you treat a confident final year medical student on a ward round. They’re probably right, they’re occasionally wrong, but it’s definitely worthwhile double checking what they’re coming up with. And if anything, verification should now be a core teaching skill, both for you as the faculty member, but also to teach the learners themselves.
AI rescues educator bandwidth
The last one is probably the most important bit. It is to recognize that AI will not replace educators. However, I think AI is beneficial in rescuing the educators’ bandwidth. This is mainly because educators and teachers who use AI will have the time, the data and also the reach to do the human parts of the teaching better than anyone else.
Finally, a more personal question: how do you use AI in your day-to-day work?
I work as a general practitioner in family medicine, and I’ll be honest, AI has really changed my day-to-day working week. It’s much more than I initially expected and also not in ways that I imagined possible.
Clinical and educational use
With any clinical practice, I use it to summarize long discharge summaries before a specific consultation with the patients. That precious time that I get back within a 10-minute consultation really changes the quality and the amount of context I have with the patient in front of me. Within clinical education, I use it to draft case scenarios. I use it to stress test various assessment items. I use it to rephrase possible feedback so it just lands better with a learner who might be struggling as well.
The non-negotiable boundary: AI as input, human judgment as decision
But the key thing I want to reiterate is that I would never use any AI tool to make any decision process. I use it to ensure that it helps, so I can make the decision better, more robust, and more defensible. But still, in a clinic setting, when I see the patient in front of me, the diagnosis is still mine. Also, in any tutorial setting with a student, the judgments about a specific learner are still mine. The main thing is that the relationship I have with the patient and with the students is the one thing I will never outsource. What I would say is that the AI simply gives me back my time. Now, what I do with that time – whether I listen, I teach, I notice or I care – that’s still the job and that’s firmly within the realms of medicine.
Leave a Reply