LLMs in dental education: what a multi-country survey shows

Generative text tools arrived in dental schools the way fluoride varnish arrived in hygiene operatories: quietly, unevenly, and faster than the policy documents that were supposed to govern them. A cross-sectional, multi-country study published in Med Educ Online (2026) turns a survey instrument on that reality, asking students and educators across several countries how they are using large language models (LLMs) in their studies and teaching, and how they feel about it. The paper is worth reading not because it settles the question of whether these tools help people learn dentistry, but because it does something more modest and more useful: it takes an honest census of behavior at a moment when most institutional guidance is still in draft.

What a cross-sectional survey can and cannot establish

Cross-sectional means one point in time. Respondents report what they do, how often, and what they believe, and the researchers describe patterns across that snapshot. This design is well suited to estimating prevalence — roughly how widespread a behavior is within the sampled population — and to surfacing attitudes that institutions may not have asked about directly. It is not built to show that LLM use improves or harms learning. Any relationship between reported tool use and reported confidence, grades, or perceived competence is an association, and the direction of that association is genuinely ambiguous: students who already feel behind may reach for a chatbot more often, and students who feel fluent may reach for it differently.

Two further limits deserve naming. Survey participation in education research is usually voluntary, which means the people most engaged with a topic are the most likely to answer — a sampling pattern that tends to inflate apparent enthusiasm in either direction. And self-report on academic tool use is sensitive: respondents may understate behavior they suspect their program disapproves of, even in an anonymous instrument. Multi-country sampling adds breadth, but curricula, accreditation systems, language of instruction, and institutional rules differ enough that pooled figures should be read as a mosaic rather than a single number.

Why dentistry is a distinctive test case for these tools

Medical and nursing education have generated most of the early literature on LLMs in health professions training, but dentistry has a structure that makes the question sharper. Dental training braids three things together: a large didactic knowledge base, a psychomotor skill set built over thousands of repetitions, and a chairside judgment that develops only under supervision with real tissue in front of you. Language models touch the first of those directly. They can restate a concept in plainer terms, generate practice questions, help a second-language student parse a dense periodontology chapter, or triage a literature search before the student reads the actual papers.

What they cannot do is calibrate a hand. No model can tell a student what sound a high-speed handpiece makes when it is about to bind, how a periodontal probe feels dropping into a pocket, or when a patient’s hesitation means the anesthetic has not taken. Nor are general-purpose chatbots diagnostic instruments. The FDA maintains a regulatory framework for software as a medical device, and consumer text assistants are not cleared under it for diagnosis or treatment planning; that distinction matters when a student begins to treat model output as clinical authority rather than as a study aid. Fabricated citations and confidently wrong clinical detail remain a documented failure mode of the technology, and dentistry has plenty of near-miss facts — drug interactions, irrigation protocols, radiographic intervals — where “almost right” is the dangerous answer.

What programs can reasonably do next

The useful response to a study like this is not a ban, which the prevalence data in this literature consistently suggests would be unenforceable. It is disclosure and verification. Programs can ask students to state where a tool was used in a written assignment, then assess the thing the tool cannot fake: oral defense of a treatment plan, live case presentation, justification of a sequencing decision. Faculty can teach verification explicitly — pushing every model-sourced claim back to a primary source or to a real institutional document, such as an ADA policy or clinical guidance statement, an AAE position paper, or a Cochrane systematic review. Privacy deserves its own rule: patient identifiers and clinical photographs do not belong in consumer chat interfaces, a constraint that flows from the same recordkeeping obligations already familiar to any clinic.

There is also a curricular opportunity. A clinician experienced in full-arch planning will tell you that the hard part of full-mouth rehabilitation is not recalling material properties but sequencing decisions around a specific patient’s biology, budget, and tolerance. Teaching students to notice where a model is confidently indifferent to those variables is itself a lesson in clinical reasoning.

For patients, the practical takeaway is smaller than the headlines suggest and more reassuring. The clinician planning your care is still the one accountable for it, and you are entitled to ask how any software — imaging analysis, shade matching, scheduling triage, or a text assistant — informed a recommendation, and what a human verified afterward. A good answer will be specific. That specificity, not the presence or absence of artificial intelligence in the building, is the signal worth watching.