Recent Highlights in Dental AI Research (July 1, 2026 – October 1, 2026)
Every three months, the AAAI-D Council for Research and Evidence reviews recently published literature and selects studies that may be of particular interest to clinicians, educators, researchers, and others following developments in dental artificial intelligence. Selection is based on factors such as study design, methodological approach, clinical relevance, or the questions a study raises for future investigation. These summaries are intended to encourage further reading rather than replace the original publications.
1. Mobile AI Triage in Community Oral Cancer Screening
Hao-Yun Liu, et al. JMIR mHealth and uHealth. 2026. DOI: 10.2196/101721
What they did: A prospective field study placed a mobile AI triage system, with its settings fixed in advance, into community oral cancer screening run by trained non-specialists. It covered 4,283 images from 602 participants, and three specialists adjudicated the reference standard.
Why it matters: Real-world evaluations remain scarce in dental AI. High specificity combined with a modest positive predictive value shows how screening performance can change once a tool leaves the development setting.
2. AI-Assisted Caries Detection Among Early-Career Dentists: A Multicenter Reader Study
Yujia Wu, et al. International Dental Journal. 2026. DOI: 10.1016/j.identj.2026.109812
What they did: Early-career dentists at several centers read 402 panoramic radiographs with and without AI assistance. Their results were compared against AI alone and an expert reference standard.
Why it matters: Most caries studies report what the algorithm does alone. This one measured what happens when clinicians use the tool. Sensitivity rose without any loss of specificity, and reading time went down.
3. Randomized Trial of AI-Enhanced Clinical Reasoning Training in Orthodontic Residency
Qiannan Niu, et al. Frontiers in Medicine. 2026. DOI: 10.3389/fmed.2026.1852507
What they did: Seventy-four orthodontic residents were randomized to an AI-enhanced interactive virtual case system or to conventional case-based learning. Outcomes were objective case-analysis scores and learner experience.
Why it matters: Randomized evidence in dental education is limited. The study offers a controlled comparison for programs deciding whether to add AI-based training tools.
4. AI vs. Clinicians in Predicting 10-Year Tooth Loss
Pablo Antonio López-Galindo, et al. Journal of Clinical Periodontology. 2026. DOI: 10.1111/jcpe.70191
What they did: AI predictions of tooth loss over 10 years were compared with those of 12 periodontists and 11 general dentists on 300 teeth whose actual outcomes were known.
Why it matters: Long-term prognosis is one of the hardest calls in periodontics. Clinicians did better at a fixed decision threshold, and the AI only matched them once each group’s threshold was optimized. Predictive ability was modest for both, which shows how much the reported performance depends on threshold choice.
5. ML Prediction of Treatment Response in Masseter Myofascial Pain
Alshaimaa Ahmed Shabaan, et al. BMC Oral Health. 2026. DOI: 10.1186/s12903-026-09491-0
What they did: Machine learning models were built in a multicenter cohort of 1,181 patients with DC/TMD-diagnosed masseter myofascial pain to predict treatment success after trigger point injections. The models were then tested on an external dataset.
Why it matters: Most studies use AI to describe a condition. This one predicts how a patient will respond to treatment. Reports in orofacial pain are uncommon, and external validation makes this one more useful.
6. Multi-Institutional Bone Age Assessment from 3D Cervical Vertebrae
Wen Tang, et al. International Dental Journal. 2026. DOI: 10.1016/j.identj.2026.111122
What they did: Machine learning models estimated bone age from 3D cervical vertebral shape on 702 CBCT scans from five institutions. They were tested on a separate external set and compared with clinicians.
Why it matters: Accuracy fell from internal to external testing but stayed high. That gap is the kind of evidence on generalizability that skeletal maturity tools for orthodontic timing need.
7. Multimodal LLMs vs. Radiologists for Incidental Findings on Panoramic Radiographs
İsmail Çapar, et al. Healthcare. 2026. DOI: 10.3390/healthcare14182948
What they did: Three multimodal large language models and two oral and maxillofacial radiologists looked for nine types of incidental findings on 500 panoramic radiographs. The reference standard was built from CBCT.
Why it matters: Clinicians increasingly use general-purpose chatbots to read images. This study shows how far they still are from specialist performance.
8. LLM Hallucinations and Clinical Risk in Endodontic Irrigation Questions
Damla Erkal, et al. Journal of Endodontics. 2026. DOI: 10.1016/j.joen.2026.08.033
What they did: Four general-purpose large language models answered 100 literature-based questions on endodontic irrigants. Two blinded experts scored accuracy, readability, type of hallucination and potential clinical risk.
Why it matters: The study linked hallucinations directly to clinical risk instead of scoring accuracy alone. It found a strong correlation between the two, which suggests an approach other chatbot evaluations could adopt.
9. Systematic Review: Safety and Workflow Burden of LLM-Generated Medical Reports
Jie-Lin Huang, et al. Journal of Medical Internet Research. 2026. DOI: 10.2196/97007
What they did: A systematic review with risk-of-bias assessment pooled the evidence on using large language models to write imaging reports, including dental imaging. It looked at expert acceptance, productivity, errors and safety.
Why it matters: AI documentation tools are arriving in clinics, so evidence on safety and workload matters as much as evidence on efficiency. The review found the evidence mixed and at high risk of bias, and it does not yet support routine use.
10. Algorithmovigilance: Monitoring Drift in a Population Tooth-Loss Prediction Model
Quang Tuan Lam, et al. Diagnostics. 2026. DOI: 10.3390/diagnostics16162569
What they did: A model predicting severe tooth loss was trained on 2016 BRFSS survey data. It was tracked across survey cycles through 2024 and validated externally on NHANES to check for performance drift and shifts in which predictors mattered.
Why it matters: Very few dental AI studies follow a model after it is built. This one adds evidence on long-term monitoring, a topic likely to grow as AI tools are deployed.
Prepared by the AAAI-D Council for Research and Evidence
The Council periodically reviews recently published literature to identify studies that contribute to the evolving evidence base in dental artificial intelligence. Inclusion in this summary does not constitute endorsement by the American Academy of Artificial Intelligence in Dentistry (AAAI-D).