aceModerate evidence
More
AI beat dental students on root-canal diagnosis tasks, review of five studies finds
Study title: Artificial Intelligence in Endodontic Education: A Systematic Review with Frequentist and Bayesian Meta-Analysis of Student-Based Evidence
- Evidence type
- Systematic review or meta-analysis
- Sources
- 1 sourcePeer-reviewed
- Published
- 3 min read
- Truvace record
- TRV-2026-2829cite or check it
Study title: Artificial Intelligence in Endodontic Education: A Systematic Review with Frequentist and Bayesian Meta-Analysis of Student-Based Evidence
A systematic review and meta-analysis searched four databases in July 2025 and found five studies that compared AI with dental students on endodontic tasks. In the two studies with full data, AI beat students by a pooled Hedges g of 1.48 (95% CI 0.60–2.36). In the two chatbot studies with analyzable means, that came to about 20 percentage points higher accuracy.
Read the study abstract
Background/Objectives: Artificial intelligence (AI) is entering dental curricula, yet its educational value in endodontics remains unclear. This review synthesized student-based evidence on AI in endodontics, primarily comparing AI vs. students on diagnostic tasks as an educational endpoint and secondarily considering assessment tasks relevant to training. Methods: PubMed/MEDLINE, Embase, Scopus, and Web of Science were searched in July 2025. Eligible studies involved dental students using AI in endodontic tasks…
The gain
A systematic review of five studies found AI models beat dental students on endodontic diagnostic tasks, with a pooled Hedges g of 1.48 and roughly 20 percentage points higher accuracy in the two chatbot studies with usable data.
In brief
Examples include ChatGPT-4 at 99.0% accuracy against 79.7% for senior and 77.0% for junior students, and ChatGPT-4o at 91.4% against 79.5% for fifth-year and 60.8% for third-year students. Image-based networks also beat students, such as 85.0% versus 75.4% for working-length estimation. The authors rated certainty as moderate for diagnostic accuracy.
The review could not show that AI improves learning. No study used validated measures of engagement, motivation or trust, none ran pre/post educational comparisons, and none deployed AI as live feedback during training. The authors say educational roles for AI remain hypothetical until prospective studies test them.
Main points
- Only five studies qualified out of 448 records found; two gave full mean and standard deviation data for the main pooled analysis, one joined a sensitivity analysis and two were summarized narratively.
- Pooled effect: Hedges g = 1.48 (95% CI 0.60–2.36; I2 about 84%); adding a third study gave g = 1.45 (95% CI 0.77–2.14; I2 about 77%).
- Bayesian analysis gave a posterior mean of 1.48 (95% credible interval 0.98–1.99) for the two main studies, with a near-certain probability of AI superiority.
- Across all five studies, AI accuracy ranged from 71.0% to 99.0% while student accuracy ranged from 60.8% to 79.7%.
- In a C-shaped canal detection study, a ResNet-101 model reached 86.7% accuracy versus 74.0% for graduate students, 77.2% for novice dentists and 81.2% for specialists.
- In a pulp-exposure prediction study, AI scored 78.0% accuracy against 65–68% for students, and students guided by the model showed limited immediate improvement.
- Risk of bias was low in three studies and moderate in two; GRADE certainty was moderate for the diagnostic comparison and low to very low for learner-centered outcomes.
What the review found when it combined the studies
These are the review’s own pooled results, copied from its abstract. Truvace did not recalculate them.
Standardised difference in endodontic diagnostic performance, AI versus students (Bayesian, three studies)
Posterior mean SMD 1.49 (95% CI 1.05 to 1.96)
3 studies · search to July 2025
The review’s own words
When the study with imputed SDs [ 21 ] was included, the posterior mean was μ = 1.49 (95% CrI 1.05–1.96)
Diagnostic performance of AI versus dental students on endodontic tasks (sensitivity analysis with imputed SDs)
Hedges' g 1.45 (95% CI 0.77 to 2.14)
3 studies · search to July 2025
The review’s own words
A sensitivity model adding the study with imputed SDs [ 21 ] gave a very similar estimate (g = 1.45; 95% CI 0.77–2.14; I 2 ≈ 77%).
Diagnostic performance of AI versus dental students on endodontic tasks (primary analysis)
Hedges' g 1.48 (95% CI 0.60 to 2.36)
search to July 2025
The review’s own words
For the primary meta-analysis [ 20 , 23 ], the pooled effect was Hedges g = 1.48 (95% CI 0.60–2.36) with substantial heterogeneity (I 2 ≈ 84%).
Standardised difference in endodontic diagnostic performance, AI versus students (Bayesian, two studies)
Posterior mean SMD 1.48 (95% CI 0.98 to 1.99)
2 studies · search to July 2025
The review’s own words
yielded a posterior mean for the overall standardized mean difference of μ = 1.48 (95% CrI 0.98–1.99)
Code checks every number above against the review’s text before it is shown. The short labels are an AI model’s reading of that text, so the review’s own sentence sits under each one. A pooled figure is only as strong as the studies behind it. How Truvace uses pooled results
The full story
The review was registered in PROSPERO and followed PRISMA 2020. Two reviewers screened records, 36 full texts were assessed and 31 were excluded, mostly for lacking student participation or an educational context. The five included studies covered two chatbot studies on case diagnosis (ChatGPT-4 and ChatGPT-4o) and three image-network studies on working-length determination, C-shaped canal detection and pulp-exposure prediction.
The primary meta-analysis used a random-effects model with Hartung–Knapp adjustments, backed by a Bayesian model. Unweighted differences across all five studies averaged +15.0 percentage points, with +20.6 for chatbots and +11.3 for image-based networks. Excluding the largest effect (Qutieshat et al.) gave +13.6 points.
Higher accuracy has not been shown to produce better learning. The authors note that the pulp-exposure study found limited improvement when students were guided by the model, and they stress that explainability and trust matter for learning impact. They also cite one report in which ChatGPT-4 scored 72% on endodontic assessment items while GPT-3.5 scored as low as 25%, so model choice matters.
The authors call for prospective, multicenter trials comparing usual teaching with teaching plus AI, measuring exam scores, retention and trust. They also recommend faculty oversight of AI outputs and prior validation of any system against expert-level performance before it is used for instruction.
Sources
1 on file- Dentistry Journaldoi:10.3390/dj13110489
Talk it over
Who gets this benefit, and who was left out of the study or story?
AI beat dental students on root-canal diagnosis tasks, review of five studies findshttps://truvace.com/item/artificial-intelligence-in-endodontic-education-a-systematic-review-with-frequen
Who gets this benefit, and who was left out of the study or story?
The gain A systematic review of five studies found AI models beat dental students on endodontic diagnostic tasks, with a pooled Hedges g of 1.48 and roughly 20 percentage points higher accuracy in the two chatbot studies with usable data.
Answer in the threadHow do we know? What can this kind of evidence show, and what can’t it?
The evidence This rests on a systematic review or meta-analysis, from Dentistry Journal.
Answer in the threadWhat does it leave open? What else would you want to know before acting on it?
The record’s own caveat Only two studies fed the primary meta-analysis, heterogeneity was high (I2 about 84%), and no study measured learning outcomes, engagement or trust with validated tools or pre/post designs.
Answer in the threadHave you seen AI used this way where you study, work or live? Did it go the same way?
Answer in the thread
The questions are the same on every record. The details under them are taken from this record as written.
What readers say