A teenager arrives at a border with no birth certificate, no passport — nothing but the word “seventeen.” A caseworker orders a panoramic X-ray, and one question hangs over the room: are those wisdom teeth young enough to make this person a child in the eyes of the law? For decades that judgment fell to a trained examiner squinting at root tips. Now researchers are asking whether AI dental age estimation can make the call instead.
The 30-Second Version
- A new study taught AI to find wisdom teeth on panoramic X-rays and judge how far they’ve developed — the backbone of estimating whether someone has crossed the legal line of 18.
- Spotting the teeth was easy: 94.5% accuracy. Reading their exact stage of development was much harder — about 73%.
- At the 18-year threshold the model looked respectable overall (81.5%), but it correctly flagged only about 1 in 4 of the people who were genuinely 18 or older.
- The honest caveat: one clinic, one population, no external validation. The authors call it a tool you can audit, not a verdict you can trust on its own.
It sounds like the tidy forensics of a crime drama. But look closer. In Diagnostics (2026), Melis Büşra Aşkın, Ayşe Bulut and Gökalp Çınarer built a two-stage deep-learning system and tested it on hundreds of panoramic radiographs from a single Turkish dental faculty. Their unusually self-critical results expose a gap between what AI can see and what the law actually needs to know.
The study, in one glance
The team used YOLO segmentation models in two steps. First, the AI located the four third molars — teeth 18, 28, 38 and 48 — across 737 X-rays. Second, on 695 X-rays, it sorted each tooth’s development into four grouped Demirjian stages (AB, CD, EF, GH), the standard ladder from early crown to closed root tip. Then came the part that matters in a courtroom or an asylum interview: a separate model trained to answer one binary question — is this person under 18, or 18 and older?
What AI genuinely gets right
The localization result is impressive and worth crediting. Across the test set the model found third molars with 94.5% accuracy, a mean average precision of 0.98, and tight spatial overlap with the human annotations (Dice 0.91) — just 4 false positives against 256 correct hits. Better still, the design is transparent by construction: because it segments the actual tooth and then labels the stage, an examiner can see exactly which structure drove the estimate. That “show your work” quality is precisely what forensic and medico-legal settings should demand — an auditable second opinion rather than a number from a black box. Done this way, AI can shave hours off large caseloads and blunt the observer-to-observer disagreement that has long dogged manual staging.
So why can’t AI dental age estimation just tell us your age?
Because a tooth doesn’t grow in tidy steps. When the same system had to judge developmental stage rather than merely locate the tooth, its best model landed at 72.7% accuracy — and stumbled hardest on the transitional stages (EF), where a root is half-formed and the boundary between one stage and the next is a smear, not a line. It’s the same split Decadentry saw when AI staged gum disease from panoramic X-rays: sharp on the obvious landmarks, shakier on the fine-grained call. Teeth mature on a continuum; supervised AI needs discrete labels. That mismatch is baked in.
⚠ The number the headline hides
At the 18-year threshold, the model’s 81.5% overall accuracy masks a lopsided failure. For people who were genuinely 18 or older, it caught only 23.5% — missing more than three in four real adults — while identifying the under-18s almost perfectly (96.5%). In plain terms, the system leans hard toward calling people children. Depending on whether it’s used at a border or in a courtroom, that bias sends the error in a direction someone will have to answer for.
It’s a smoke alarm that almost never cries wolf — and sleeps through three of every four real fires. The silence feels reassuring right up until it means nothing.
The rest of the fine print compounds it. The data came from one university archive in one country, and third-molar development is known to vary with population, genetics, sex and jaw. There was no external validation, no patient-level guarantee that the same mouth didn’t appear in both the training and test sets, and a marked class imbalance the authors chose to interpret around rather than engineer away. Every one of those is a reason the 81.5% could shrink on a different population.
Who’s accountable when a tooth decides your age?
This is where the stakes stop being academic. Dental age estimates already feed real decisions about asylum, criminal responsibility and safeguarding — decisions that reshape a person’s life. If an algorithm is in that loop, someone has to own its errors. To their credit, the authors are blunt: this is “a visually auditable decision-support workflow rather than a stand-alone legal age determination tool.” The model can hand an examiner a labelled X-ray to check; it cannot, and should not, sign the paperwork. The moment its output is treated as a verdict rather than a prompt, accountability evaporates into the software.
Whose teeth trained it — and who gets misjudged?
An age model is only as fair as the mouths it learned from. Trained on radiographs from a single Turkish faculty, its numbers may not travel to a teenager who grew up on a different diet, in a different climate, with different genetics — often exactly the migrant and refugee populations whose ages are most likely to be contested in the first place. An error rate measured in a dental school is not the error rate at a border. Deploy the tool where the training data doesn’t reach, and the people least able to challenge the result are the ones most likely to be misjudged by it.
What this means for you
If you’re a patient
If your age is ever assessed from a dental X-ray — for immigration, sport, or a legal case — treat the result as an estimate with wide margins, not a fact. Teeth of the same age vary enormously. You can ask how the number was reached, whether a human reviewed it, and what the margin of error is.
If you’re a clinician
Use AI staging as an auditable second read, never a stand-alone ruling. Report uncertainty, keep a human examiner accountable for the final call, and demand external validation and calibration on your own population before anything approaches a legal decision. Watch the direction of error, not just the headline accuracy.
The bottom line
AI can now find a wisdom tooth better than most humans and show its work while doing it. But finding the tooth was never the hard part — knowing what its half-formed root means for a person’s rights is. On the question that actually carries legal weight, this system was four times better at spotting a child than confirming an adult. That’s a capable assistant handing an examiner a clearer X-ray. It is nowhere near an oracle that should decide whether someone is a child or an adult.
Frequently asked questions
Can a dental X-ray prove someone’s exact age?
No. Dental development follows a broad range, so a radiograph gives an estimate with real uncertainty, not a birthdate. It is most useful for judging whether someone is likely above or below an age threshold — and even then it comes with error bars.
How accurate was the AI at the 18-year cut-off?
Overall it reached 81.5% accuracy, but that figure hides a lopsided result: it correctly flagged only 23.5% of people who were truly 18 or older, while catching 96.5% of under-18s. It rarely called a child an adult, but it missed most real adults.
Why is finding the tooth easier than estimating its age?
Locating a third molar is a shape-detection task with clear anatomical edges, which AI handles well — 94.5% accuracy here. Judging developmental stage means reading a gradual continuum of root growth into discrete categories, a much blurrier and more error-prone job.
Is AI dental age estimation used in courts or immigration today?
Traditional dental age estimation is already used in some legal and asylum contexts. This AI system, however, is research-stage: single-center, one population, and not externally validated. Its own authors say it should not be used for stand-alone legal age decisions.
What is the Demirjian method the AI used?
It is a long-established way of scoring tooth maturity from stage A (early crown) to H (closed root tip). The study grouped those eight stages into four pairs (AB, CD, EF, GH) to give the AI steadier, more balanced categories to learn.
Source & author credit
This article interprets, and does not reproduce, the following peer-reviewed study. All figures are the authors’ original findings.
Aşkın MB, Bulut A, Çınarer G. Comparative Analysis of Third Molar Segmentation Performance Across Dental Developmental Stages and the 18-Year Age Threshold Using Deep Learning Models. Diagnostics (Basel). 2026;16(16):2636. DOI: 10.3390/diagnostics16162636
ORCID — Gökalp Çınarer: 0000-0003-0818-6746 (the study’s other authors have no public ORCID on record).
Published by MDPI under a Creative Commons Attribution 4.0 (CC BY 4.0) license. Decadentry is an independent educational publication and is not affiliated with the study’s authors. Reviewed against the primary source per our editorial standards.
Leave a Reply