It is 1 a.m. somewhere, and a second-year dental student is arguing with a chatbot about a swollen lower molar. The chatbot never sighs, never checks the clock, and never runs out of cases. The obvious question is whether that student is actually learning anything — and a new meta-analysis is the first serious attempt to answer it with randomised evidence rather than enthusiasm.
The 30-Second Version
- A 2025 meta-analysis in the International Dental Journal pooled six randomised controlled trials of AI-assisted problem- and case-based learning in dental and medical education.
- Against traditional lectures, AI-supported learning produced a clear gain in knowledge scores (standardised mean difference 0.70, 95% CI 0.24–1.15).
- Against human-facilitated small-group teaching without AI, the advantage shrank to 0.31 and was not statistically significant (95% CI −0.33 to 0.66).
- The honest caveat: only six trials, all rated “some concerns” for risk of bias, and four measured learning immediately after a single session.
It sounds like a settled win for the machines. But the interesting result in this paper is not the headline — it is the comparison the headline leaves out. Wei and colleagues at the University of Hong Kong, publishing in the International Dental Journal, ran a PROSPERO-registered systematic review and meta-analysis asking whether AI-powered problem-based learning (PBL) and case-based learning (CBL) actually improve what students know, how they reason, and how satisfied they feel. The answer depends entirely on what you compare AI against — and that single distinction should shape how dental schools spend the next decade.
The study, in one glance
This was a PRISMA 2020 systematic review searching PubMed, MEDLINE, the Cochrane CENTRAL register and Web of Science, with a supplementary Google Scholar sweep for grey literature. Forty-one full-text articles were assessed; 35 were excluded. The six that survived were all randomised controlled trials. Four of them, with 110 students in the AI-assisted arms and 102 in the controls, provided objective examination data suitable for pooling on knowledge acquisition. Heterogeneity was low (I² = 20%), so a fixed-effect model was used.
The promise: a tutor that scales, and never gets tired
Give the technology its due. Problem- and case-based learning are widely considered the best things dental education does — students work through realistic clinical scenarios in small groups, building reasoning rather than memorising. They are also expensive, demand trained facilitators, and buckle under large cohorts. That bottleneck is precisely what AI targets.
The tools tested were more varied than “students used ChatGPT”. They included COMET, an intelligent tutoring system built at Thammasat University Dental School in Thailand that generates Bayesian-network tutoring hints for second-year dental students; CC-Cruiser, an image-based diagnostic platform; and LearnGuide, a customised ChatGPT build. The efficiency gains can be striking: authoring one problem scenario took roughly a person-month in the original COMET system, but only four to five hours in its successor, METEOR. Students liked it, too — pooled satisfaction across three trials reached a standardised mean of 0.7 (95% CI 0.47–0.92).
But did it beat a good human tutor?
This is where the story turns. When AI-supported learning was compared with traditional lectures, the effect was solid: SMD 0.70 (95% CI 0.24–1.15, p = 0.003). When it was compared with problem- and case-based learning run by humans without AI, the effect fell to 0.31 with a confidence interval of −0.33 to 0.66 — an interval that comfortably includes zero, at p = 0.07. In plain terms: the evidence that AI outperforms a well-run tutorial is not there yet.
⚠ A “46% improvement” is not 46% more marks
The pooled result is a standardised mean difference of 0.46 — a moderate effect size expressed in standard deviations, which the paper summarises in words as a 46% improvement. It is easy to read that as “students scored 46% higher.” They did not. Effect sizes and percentage gains are different currencies, and conflating them is how a reasonable finding becomes a marketing claim.
An AI tutor is less like a better teacher and more like a gym that’s open at 3 a.m. The hours are unbeatable. Whether you get stronger still depends on what you do once you’re inside.
The evidence base is also thinner than the confident conclusion suggests. Every one of the six trials was rated “some concerns” on the Cochrane RoB 2 tool; only one was preregistered; five described random allocation without explaining how randomisation was done. There were too few studies to test for publication bias at all. And crucially, four of the six measured outcomes immediately after a single learning session — which tells you something about short-term recall and almost nothing about whether a student becomes a better clinician. The authors themselves raise the possibility that novelty inflated the satisfaction scores.
Who checks whether the AI tutor is right?
Clinical reasoning — arguably the whole point of case-based learning — could not be pooled at all. The three trials that measured it used three incompatible instruments and pointed in different directions: one found AI-tutored students gained about as much as human-tutored ones (p = 0.058, not significant), while LearnGuide showed a large advantage over traditional PBL (SMD 1.34) and ChatGPT-assisted PBL a more modest one (SMD 0.90). One of those trials relied on students’ self-assessment while they knew which group they were in — a design that invites exactly the answer the researchers hoped for.
The authors are refreshingly direct about what holds this together: the facilitator. Tutors did not merely teach students to operate the system; they verified the accuracy of what the AI produced and supplied personalised feedback. That unglamorous job is why “AI replaces the tutor” is the wrong reading of this paper. A generative model that invents a plausible-sounding periodontal classification is a hazard, not a shortcut — and someone qualified has to be in the room to catch it.
Who gets the AI tutor — and who gets the lecture?
The equity picture cuts both ways. If AI-supported case-based learning genuinely beats lectures, the schools that stand to gain most are the ones running large cohorts on thin budgets, where small-group teaching was never affordable. That is a real democratising possibility.
But the studies also flag the cost of entry: building case models takes time and expertise, hardware requirements are non-trivial, and knowledge bases must be kept current or students learn outdated dentistry with great efficiency. The plausible outcome is uncomfortable — well-resourced schools keep their human tutors and add AI, while under-resourced schools swap lectures for chatbots and call it parity. On this evidence, that second group would still be better off than with lectures alone. It just would not be the same education.
What this means for you
If you’re a patient
Your dentist’s training increasingly includes AI tutors, and on current evidence that is neither a red flag nor a guarantee. What still matters is supervised clinical hours and examined competence — not whether the case discussions were run by a person or a program.
If you’re a clinician or educator
The defensible move is replacing passive lectures, not facilitators. If you adopt an AI tutor, budget for the human who verifies its output, and measure retention months later rather than at the end of the session.
The bottom line
This meta-analysis does not show that AI teaches dentistry better than people do. It shows that AI-supported case discussion beats sitting passively in a lecture theatre — which, honestly, is a low bar that dental education has been trying to clear since long before machine learning existed. The technology is a genuine assistant, not an oracle: it scales the format we already know works and removes the excuse that good teaching is too expensive. The tutor’s job does not disappear. It shifts from delivering content to checking what the machine said — and that may be the more demanding role.
Frequently asked questions
Does AI actually help dental students learn better?
Compared with traditional lectures, yes — a 2025 meta-analysis of six randomised trials found AI-assisted problem- and case-based learning improved knowledge scores with a standardised mean difference of 0.70 (95% CI 0.24–1.15). Compared with human-facilitated small-group teaching without AI, the advantage was smaller (0.31) and not statistically significant.
Is an AI tutor better than a human tutor for dental students?
Not on current evidence. The subgroup comparing AI-assisted learning with human-led PBL/CBL produced a 95% confidence interval of −0.33 to 0.66 (p = 0.07), meaning the true difference could plausibly be zero. Human facilitators also remained essential for verifying the accuracy of AI-generated content.
What does a standardised mean difference of 0.46 actually mean?
It is an effect size measured in standard deviations, not a percentage. An SMD of 0.46 is conventionally read as a small-to-moderate improvement. It does not mean students scored 46% higher, and the paper’s plain-language phrasing of “46%” is easy to misread that way.
Is this evidence strong enough to redesign a dental curriculum?
Not on its own. Only six trials met inclusion criteria, all were rated “some concerns” for risk of bias, just one was preregistered, and four measured learning immediately after a single session. There is no evidence yet on long-term retention or clinical performance.
Which AI tools were tested in these trials?
Four of the six trials used intelligent tutoring systems: COMET (developed at Thammasat University Dental School), CC-Cruiser (an image-based diagnostic platform), LearnGuide (a customised ChatGPT build), and ChatGPT itself, which was used to generate open-ended questions and simulated patient interviews.
For a look at how the same promise-versus-proof gap plays out in diagnosis rather than education, see our analysis of AI-powered smartphone cavity detection.
Source & author credit
This article interprets, and does not reproduce, the following peer-reviewed study. All figures are the authors’ original findings.
Wei H, Dai Y, Yuan K, Li KY, Hung KF, Hu EM, Lee AHC, Chang JWW, Zhang C, Li X. AI-Powered Problem- and Case-based Learning in Medical and Dental Education: A Systematic Review and Meta-analysis. International Dental Journal. 2025;75(4):100858. DOI: 10.1016/j.identj.2025.100858
ORCID — Kar Yan Li: 0000-0002-7259-6611 · Kuo Feng Hung: 0000-0002-3971-3484 · Xin Li: 0000-0002-6979-1037
Source article published under a Creative Commons Attribution (CC BY 4.0) licence. Protocol registered with PROSPERO (CRD42024626780). Decadentry is an independent educational publication and is not affiliated with the study’s authors.
Leave a Reply