AI in health: more output, better science, better care?
On April 23, HINT.GENT hosted a debate evening on what AI is starting to change in science and healthcare. After an introduction by Prof. Tom Braekeleirs, Dr. Alexander Decruyenaere moderated a panel with Prof. Katleen De Preter, Prof. Erik Mannens, Prof. emeritus Mirko Petrovic and Prof. Marijn Speeckaert.
The discussion quickly moved beyond tools and technical performance. It became a conversation about whether more output actually leads to better science, and ultimately better care.
Debate panel: (left to right) Prof. Marijn Speeckaert, Prof. em. Mirko Petrovic, Prof. Katleen De Preter, Prof. Erik Mannens, and moderator Dr. Alexander Decruyenaere.
The discussion deliberately avoided the usual extremes. AI is not a miracle solution, but it is also not something science can simply ignore. The more useful conversation is in the middle: where does AI genuinely help, where does it create new risks, and what do we need to put in place to use it well?
That is also where HINT.GENT has a role to play. AI in health needs biomedical researchers, data scientists, clinicians, engineers, educators and people working on policy, ethics and implementation around the same table.
The first tension was prediction versus explanation. A model that predicts well can be useful, and in some situations, performance may matter more than complete explainability. If a system clearly helps more people, that cannot be dismissed. But healthcare and biomedical research are rarely only about prediction. Patients differ, context matters, and researchers often need to understand why a model sees a pattern, not only that it sees one. A black box can therefore be useful, but the harder question is when we are willing to trust it, and what kind of understanding we still need before using it in science or care.
The second tension was democratisation versus dilution. AI may also change who gets to participate in science, but not in a straightforward way. Several speakers pointed out that AI can be genuinely useful for researchers who are not native English speakers: it can help them write, translate, summarise and communicate more fluently. In that sense, it may remove a real barrier. Several speakers also warned that access to the best tools may become expensive, creating new inequalities between well-funded groups and researchers with fewer resources. Easier writing also means easier overproduction. The result may be more papers, applications and reports, without the same increase in scientific value.
The third tension was AI-generated content versus quality in peer review. Peer review was one of the more uncomfortable parts of the discussion. AI can help to summarise a manuscript or check specific points, but nobody argued that it can replace expert judgement. The harder issue is that the current rules do not always match actual practice. Authors already use AI to polish language or structure, while reviewers are often discouraged, or not allowed, to use similar tools. This creates difficult questions about transparency and fairness. AI detectors do not solve that problem: they are unreliable, and may especially affect non-native English speakers who use AI to make their work clearer. In practice, the question is less whether AI was involved, and more whether the author or reviewer still takes full responsibility for the final text and judgement.
The fourth theme, individual gain versus collective loss, brought the discussion back to daily research practice. AI is useful because it helps the individual researcher. It can speed up writing, coding, brainstorming, literature work and administration. For some senior researchers, it may even make hands-on scientific work easier again, because it lowers the threshold to restart a piece of code, revisit an idea or explore a question quickly. But the collective effect is less obvious. If many people use the same tools in the same way, scientific writing and even scientific thinking may become more uniform. The risk is not only more output, but more of the same output.
That does not mean AI should only be used for routine tasks. Several people argued that it can also be useful as a sparring partner: to challenge an idea, expose weak points, suggest alternatives or test whether something is actually new. Used that way, AI is less interesting as a machine that gives answers, and more interesting as something that forces better questions.
Education came back throughout the evening. Several speakers worried that students might start avoiding exactly the parts of learning that matter most: writing, reasoning, struggling with a problem, and learning how to judge an argument. AI can make these steps feel optional, especially when the output looks convincing. Still, banning or ignoring AI would not prepare students for the world they are entering. The real challenge is to teach them early enough how to use it critically, and to make sure they still understand the work they submit or rely on.
The evening ended with a simple question: with funding for a four-year PhD position, would the panel hire a human researcher or spend the budget on AI tools and tokens?
The answer was unanimous: hire the human.
It was a useful reminder. AI will become part of research and care, and in many places it already is. But it does not remove the need to train people, build expertise, make difficult judgements and take responsibility for scientific or clinical decisions.
For HINT.GENT, this is probably where the debate matters most. AI in health is not only a question for developers or AI experts. It also affects clinicians, researchers, teachers, students, reviewers, institutions and patients. Bringing those perspectives into the same conversation is precisely the kind of role HINT.GENT can play.