In June 2026, Elsevier hosted the webinar, AI in Health Education: Preparing Students for Real-World Clinical Readiness, bringing together three leading voices in health education to explore the growing role of AI in clinical training and practice. David Game examined the impact of AI on students academically and socially, Dr Philip Xiu explored how curricula and assessment must evolve in the age of AI, and Dr Ximena Alvira addressed AI’s role in clinical settings.
The session generated a wealth of interesting questions from attendees – more than we could address in the time available. Our panelists have since taken the time to respond to those we didn’t get to, and we’re pleased to share their answers here. Responses are attributed to each panelist individually, as each contributed where they felt best placed to do so.
Please note that some of the questions have been lightly edited for clarity and concision, while preserving the original intent.
Q1: Can you give an example of a teaching activity that would enhance students’ persistence and grit in the world of AI?
David Game responds:
We know that using AI as an answer delivery solution has the most corrosive effects on persistent learning. Therefore, AI activities should challenge students to bring knowledge to the activity. Any activities which require students to deconstruct, challenge, or reframe AI output would develop the grit and persistence that accompanies deep learning, which could be valuable.
Dr Philip Xiu responds:
Grit and persistence are exactly the same qualities that can get eroded when students become overly dependent on AI giving them instant answers. This is part of the mis-skilling that David was referring to in the webinar.
In the teaching activity sense, the AI could be used as a Socratic method of discourse with the student. In this activity, students are presented with a complex case, they work through it independently without the use of any AI tools and commit to a diagnosis and management plan in writing. Then they run the same case through an AI tool.
Their task is not to compare the notes passively but to identify where they disagree with AI and why and to defend their reasoning process. The aim is to not defer to the AI without justification, while also not dismissing it without evidence.
This discomfort of sitting with that disagreement and not immediately knowing who is right and who is wrong is precisely the thing that builds clinical persistence. There is always going to be uncertainty within medicine but students can learn that good clinical reasoning is about defending your position under uncertainty in shades of gray. That habit of mind will serve them long after any particular AI tool is superseded.
"AI activities should challenge students to bring knowledge to the activity. Any activities which require students to deconstruct, challenge or reframe AI output would develop the grit and persistence that accompanies deep learning."
David Game, SVP in Medical Education Partnerships, Elsevier | Elsevier Faculty Hub, August 2026
Q2: In my understanding, AI is a valuable tool that can support healthcare professionals in solving clinical problems and improving efficiency in day-to-day practice. However, many senior clinicians and academics appear reluctant to adopt or utilize AI in their practice and teaching. What would you recommend to address this resistance and promote the responsible integration of AI into healthcare and medical education?
Dr Philip Xiu responds:
We do see a lot of resistance from senior clinicians and academics, and this is a rational response in the face of such novel technology. The reason is the fact that they’ve spent decades honing their own global judgement with evidence-based medicine. They look at an AI tool that produces a recommendation without a clear explanation of why, and their understandable skepticism is simply good medicine.
I think we need to stop seeing this resistance as a problem to be overcome, and instead treat it as a prompt for better education and understanding.
With any new technology, one of the single most powerful leaders is a clinical champion. This is not an external technology enthusiast from the outside but a peer within the specialty who can say, "I’ve used this and here’s what it did for my patients and here is where I can still override it." This kind of social credibility is irreplaceable. As technology has moved on (for example, electronic prescribing, electronic health records, and clinical decision-making support), I’ve seen the teach-back effect in place. This is where senior clinicians are teaching trainees and engaging with the technology, which can deepen their own trust considerably. And when senior clinicians do that with AI (i.e., training trainees how to critically evaluate, spot bias, and maintain independent reasoning), it reframes AI from a threat to their own clinical authority into an extension of their own teaching role.
Within the medical education curriculum, there is clear evidence that AI education should not be siloed into a standalone isolated module. It should be woven within clinical-decision making and reasoning, medical ethics, straight from the outset. Medicine should involve AI-free reasoning so they are not in actuality producing medical graduates who cannot function without AI support. Just as hands-free cruise control is available to support drivers, the drivers should still need to pass their test without it.
The end goal of all of this is so that we have graduates who do not trust AI blindly, but are confident and critical users of its outputs.
Q3: I manage a journal club about pain in my university. Seeing how hard it is already to integrate the human aspect of the biopsychosocial model into conversations when students are taught to think in a biomedical way, how do you recommend introducing a critical conversation about AI without being ignored by students or labelled as the ‘fun police’?
Dr Philip Xiu responds:
This is such an interesting question. Some universities teach students to take history in the biopsychosocial medical model, and some focus more on the biomedical component of it. With a clinical history taken, having a holistic approach to the patient involved in a biopsychosocial approach can give much deeper rapport with the individual longitudinally as opposed to just a purely biomedical approach. And obviously, as you know, the psychosocial impact has as much, if not more, importance compared to the biological side of any medical condition, not least in management and planning discussions with their patients if they have their own health beliefs.
One way to introduce a conversation about AI is to put them in the position where they are consulting with an elderly individual with vague medical components, but whose overall problem relates to the psychosocial aspect of environmental health, low mood – given the context of recent bereavements – and financial uncertainties. Perhaps ask them to put the case into an AI tool of their choice and bring the relevant outputs to the session.
It’s possible the AI outputs will all converge into similar thought processes, which will be competent but perhaps clinically relatively thin. The aim is to put the focus back on the individual patient themselves, the impact on their life, their own concerns, and their own expectations. But you don’t need to necessarily point that out. You can just ask what is missing from what is being presented by the AI tools. And in that moment, hopefully the students will find it themselves.
What healthcare really requires of them is to recognize that the conversation belongs to the patient – it’s about the individual in front of them, not the AI output. Having that real-life discussion and sitting with shared uncertainty with patients, who are individuals with complex psychosocial overlay, is a more effective approach than deferring to AI.
"[Senior clinicians and academics] look at an AI tool that produces a recommendation without a clear explanation of why, and their understandable skepticism is simply good medicine. I think we need to stop seeing this resistance as a problem to be overcome, and instead treat it as a prompt for better education and understanding."
Dr Philip Xiu, Hon. Senior Lecturer, University of Leeds | Elsevier Faculty Hub, August 2026
Q4: How can we manage public reliance on AI and people increasingly consulting with AI on health concerns instead of visiting a doctor?
Dr Philip Xiu responds:
As shown from David’s data during the webinar, we are seeing that self-help and companionship are among the most common ways people are using AI.
I think it’s actually impossible to turn back the tide on the general public utilizing and consulting AI. What we can do is help the general public become more sophisticated consumers of AI health information and teach them to treat AI outputs the way that they might treat a knowledgeable friend’s advice. It is useful as a starting point but not a substitute for professional assessment. This really is a public health role in terms of AI health literacy and deserves a place in any health communication.
One of the reasons why people are turning to AI is because accessibility to healthcare advice, such as GP appointments, is genuinely difficult. If we want people to consult healthcare professionals rather than chatbots, then clinicians need to be equally accessible, which is currently impossible. The AI is partly filling a vacuum that healthcare systems have created. So far, this is a healthcare policy and resourcing problem as much as a technology problem.
In the UK, we already carry out healthcare screening programs – such as breast cancer, cervical and bowel cancer screening – using physical aids. One wonders whether, in the future, there will be AI symptom screeners that can prompt people who would have otherwise ignored their symptoms to seek care earlier. This is because the risk of under-consultation among under-represented groups is, in many of the diverse populations, just as serious as over-reliance on the healthcare system. I wonder whether the AI can then serve as a genuine public health function.
Q5: Given that large language model (LLM)-based chatbots such as Copilot, Gemini and ChatGPT are constantly ingesting and learning from data, how can we guarantee data security and confidentiality when using these tools?
Dr Ximena Alvira responds:
Honestly, with general-purpose standalone chatbots you cannot fully guarantee it. Copilot, Gemini and ChatGPT may retain and learn from what you submit, so entering identifiable patient information risks breaching confidentiality and data-protection law no matter how secure the interface feels.
My rules of thumb:
- never paste patient-identifiable data into a consumer LLM
- use only institutionally governed tools covered by a data-processing agreement
- de-identify inputs
- assume anything you type may be disclosed.
More fundamentally, a standalone LLM is not a clinical system, however secure or human it sounds, it was not built, validated or regulated for patient care. The “safer” architecture is a retrieval-augmented (RAG) system that draws only on an approved, access-controlled medical knowledge base, because it keeps data within a governed environment and lets you audit exactly what was used.
Q6: AI is often very certain of its answers but sometimes makes up studies and resources, which could influence the treatment/diagnoses of the patients. How can we combat this?
Dr Ximena Alvira responds:
This is the single biggest reason a standalone LLM should not be used for clinical purposes. These models generate plausible-sounding text, not verified facts. They will make up citations, authors and even identifiers that don’t exist. In a research or publication context that is an integrity failure, but at the bedside it’s a patient-safety risk.
There are four ways to combat this:
- Keep a human in the loop. A clinician must verify every output against primary sources.
- Prefer RAG systems that ground each answer in a verifiable, citable source rather than free-text generation, so you can click through and audit the reference.
- Treat any citation you cannot independently verify as fabricated until proven otherwise.
- Use your critical thinking abilities to discern between fake or real. And if you don’t have them, they need to be developed as a priority.
Q7: Are there any AI tools that are highly reliable for tracking new and up-to-date medicine, and are fully aware of recent research and clinical developments?
Dr Ximena Alvira responds:
Yes, there are AI tools that are highly reliable for keeping up with new and emerging medical evidence, but their reliability depends entirely on how they are built.
The ones worth trusting are developed by experienced, responsible organizations with the expertise, resources, and governance to stand behind them. That includes strong guardrails, clear policies for responsible AI use, and full transparency about how the system works and what it is based on. Equally important, they must have access, either proprietary or through formal agreements, to the highest-quality medical evidence, rather than relying on the open web.
"a standalone LLM should not be used for clinical purposes. These models generate plausible-sounding text, not verified facts. They will make up citations, authors and even identifiers that don’t exist. In a research or publication context that is an integrity failure, but at the bedside it’s a patient-safety risk."
Dr Ximena Alvira, Clinical & Research Manager, Elsevier | Elsevier Faculty Hub, August 2026
Q8: Why doesn’t the WHO or an equivalent medical authority develop a closed AI tool for medical sectors, fed with trusted, approved medical information, rather than relying on general-purpose tools like ChatGPT?
David Game responds:
There are definitely initiatives to build LLMs that better reflect regional or discipline-specific needs. South Africa is investigating the development of an LLM/AI tool that better reflects the realities of medicine in the South African context. As costs fall and more tools become available in open source or low-cost versions this will become more viable. However, the speed with which AI technology is progressing runs counter to the level of diligence and coordination that would be required if a model were to be built by WHO or any other transnational organization. But it is definitely a worthy goal.
Dr Ximena Alvira responds:
In fact, organizations such as the FDA and the European Medicines Agency (EMA) are already developing guidance and regulatory frameworks for the use of AI in healthcare. The FDA, for example, has issued specific guidance for AI-enabled medical devices and has already authorized more than 1,000 such products. However, regulating AI is particularly challenging because new tools emerge at an extraordinary pace, often faster than regulatory processes can adapt. Moreover, many AI applications used by healthcare professionals are not classified as medical devices and therefore are not required to undergo the same stringent approval pathways as diagnostic or therapeutic technologies.
As for creating a single WHO-managed AI trained exclusively on validated medical information, the concept is attractive, but maintaining such a system would require continuous updating, global consensus on evidence standards, multilingual content management, and substantial resources. In practice, it may be more realistic to establish clear governance frameworks for the safe use of AI tools rather than relying on a single global platform.
My view is that governance should not rely solely on international regulators. It also needs to be implemented locally through universities, hospitals, and healthcare organizations, aligned with national policies. Clinicians should be trained not only in how to use AI, but also in when its outputs can be trusted, how to verify them, and where its limitations lie.
Want to hear the full discussion? Watch the webinar recording to catch everything our panel explored during the session:
