AI in endodontics: a patient guide

2026 evidence update: can chatbots diagnose a tooth from clinical records?

Editorial update — October 21, 2026. This Torabinejad study summary, related evidence-summary item and FAQ, and all seven translations await clinical review. The September 30 approval applies only to the earlier content.

Not reliably enough to replace your endodontist. In a retrospective study of 645 teeth, ChatGPT-5.2 and Gemini 3.0 agreed with expert reference classifications more often on individual tasks than on all required decisions together. This is not proof of safe autonomous diagnosis, superiority over clinicians, or improved patient outcomes.

Torabinejad and colleagues compared off-the-shelf consumer chatbots with calibrated endodontist reference classifications for teeth requiring initial root canal treatment across multiple centers [11]. The models evaluated recorded case information in a single pass through consumer interfaces; they did not examine patients. The supplied Journal of Endodontics paper is a journal pre-proof, accepted October 7, 2026, not the definitive version of record.

The endpoints must be kept separate. Reported individual-endpoint accuracies ranged from 74–88%. Agreement on all three clinical endpoints together—pulpal diagnosis, periapical diagnosis and treatment planning—was 61.1% for ChatGPT-5.2 and 59.4% for Gemini 3.0. Exact agreement on all four endpoints, adding case difficulty, was 47.3% and 46.2%, respectively. These are tooth-level results in the 645-tooth study, not interchangeable measures. The four-endpoint result is not a universal chatbot diagnostic error rate or an estimate of patient harm.

The reference was expert classification, not an infallible biological truth: each case’s analytic reference came from one calibrated endodontist, not panel consensus. Experts themselves varied, particularly on case difficulty. The retrospective, selected initial-treatment population, class imbalance and single-pass consumer-interface evaluation limit generalization; model versions and prompts also matter. A high overall accuracy can conceal weaker performance on uncommon categories. The study did not test whether using a chatbot improves healing, prevents harm or helps patients retain teeth.

The pre-proof also has reporting limitations: some paired model-comparison counts in Figure 3 conflict with its tables, and the screening flow differs from the Methods description. We do not use those disputed counts to rank the models. The supplied 48-page file includes supplement legends but not the underlying supplementary tables and figures, so those materials could not be independently checked. The percentages above are the author-reported aggregate endpoints, not reconstructed supplementary results.

This differs from Jalali and colleagues’ simulated oral-board questions [8], which tested GPT-4o and Gemini 2.5 Pro on three supplied case histories. Exam-answer scores and agreement on retrospective clinical records answer different questions, with different models and denominators; neither establishes safe independent care. Use a chatbot to prepare questions, not to decide whether your tooth needs treatment. Your endodontist must interpret symptoms, tests and imaging in context.

What the research can—and cannot—tell you

Imaging AI is not a chatbot

Imaging AI analyzes X-rays or CBCT scans for a defined task, such as marking a region that may need closer inspection. A generative chatbot produces text in response to a prompt. Fluent answers, or performance on a specialist board examination, do not establish competence to diagnose your tooth. Chatbots can invent facts and citations. Neither tool can perform your bite tests, assess symptoms in person, or take responsibility for treatment.

Potential benefits — and what they do not prove

An imaging tool might draw attention to a subtle finding or help organize a clinician's review. A chatbot might help you prepare questions or understand unfamiliar words. These are possible uses, not a guarantee of better care. More detections are not necessarily better: a useful tool must improve decisions without unnecessary scans, treatment, cost, or delay. Patient-important outcomes include pain, healing, quality of life, and keeping a tooth.

False positives and false negatives

A false positive flags disease that is not actually present; it can lead to worry, extra imaging, or unnecessary treatment. A false negative misses disease and may delay needed care. Accuracy depends on the task, image quality, the patients studied, and how the correct diagnosis was established. A high accuracy number from a selected dataset does not tell you how often the tool will be right for someone like you.

Why external validation matters

External validation tests a tool on patients and images that are separate from its development data, ideally from other clinics and scanners. Ask whether research included everyday clinical cases, different populations, and teeth with fillings or root canal materials that can create image artifacts. Validation of one task does not validate every use. Prospective studies in real care and comparisons with clinician care are needed to determine whether using a tool actually benefits patients.

What the JADA perspective supports

Swanson and colleagues discuss translating emerging technologies into clinical impact and evaluating patient-important outcomes [1]. Their September 2026 JADA paper is a translation perspective, not an AI diagnostic-accuracy study or a trial proving superior root canal outcomes. It supports asking how a promising technology becomes useful in practice; it does not justify an accuracy percentage or a promise of clinical benefit.

Does an AI alert mean my root canal failed?

No. Allihaibi and colleagues studied follow-up periapical X-rays of 376 treated teeth, using CBCT radiographic findings as the reference [3]. At tooth level, AI was more sensitive than the clinicians (67.3% versus 49.3%) but less specific (82.3% versus 92.5%): it detected more reference-positive findings while also raising more false alarms. These are study results, not this practice's performance. A persistent dark area can be shrinking during healing; an imaging finding alone is not proof of failure. The study did not show better healing or fewer unnecessary retreatments. Symptoms, examination, earlier images and follow-up all matter.

Can a consumer AI miss a serious complication?

Yes. Akyüz and colleagues tested three consumer image-and-text AI systems on 60 selected clinical cases against two endodontists' consensus [4]. Recognition of some lesions did not translate into reliable detection of complications: perforation sensitivity ranged from 0% to 22.7% across the tested systems. This was a small retrospective benchmark, not an independent-site clinical validation or a trial of patient outcomes. A confident chatbot response cannot rule out a perforation or a broken instrument, and its treatment advice should not replace assessment by an endodontist.

Can AI find hidden root canals?

Researchers are exploring this, but the study setting matters. Ji and colleagues tested image enhancement in 171 extracted teeth [5]. In a subgroup of 25 four-canal upper first molars, an additional canal was visible in 18 ordinary CBCT images, 23 enhanced images and all 25 micro-CT reference images. This laboratory comparison was not a clinical outcome study; even enhanced images missed canals. Turp and colleagues tested additional-canal detection in a held-out set of only 28 cases from one institution and scanner [6]. One case was missed. An internal test is not independent external validation. Neither study proves improved treatment success or justifies routine extra CBCT scans.

Why a better trauma prompt is not a safety guarantee

Ourang and colleagues compared chatbot answers to layperson and specialist-framed questions about four hypothetical dental-trauma scenarios [7]. Specialist questions supplied different clinical content as well as different wording, so the study cannot isolate the effect of phrasing alone. Some answers contained unsafe or fabricated advice. The 1,920 reported scores were rater-by-criterion assessments, not patient cases. Do not spend time refining a chatbot prompt instead of seeking prompt professional assessment for a dental injury.

Can convincing answers contain made-up citations?

Yes. In Jalali and colleagues' simulated endodontic oral-board study, each of the two tested models generated four fabricated citations [8]. The models answered questions about three supplied case histories; they did not examine patients or take the official board examination. A polished answer and an impressive score do not verify its references or establish safe independent clinical judgment. Open an important reference and check that it exists and actually supports the claim.

Can AI choose retreatment or microsurgery for my tooth?

The reviewed evidence does not establish that ability. Sabeti and Torabi's 2026 journal pre-proof is a scoping review and exploratory gap analysis, not a validated treatment-selection tool [9]. Across six prognostic studies, none reported external validation or formal calibration—checking whether predicted probabilities match observed outcomes. Accuracy at sorting cases is not the same as a reliable personal success percentage. No model provided validated predictions for both retreatment and microsurgery in the same scenario. Some inputs were only available after treatment or during follow-up, too late for an initial decision. Restorability, access to canals, surgical feasibility and your preferences still need clinician assessment. The pre-proof has a screening-flow discrepancy and unavailable supplementary tables; its proposed decision framework remains exploratory.

Privacy, consent, and accountability

Avoid uploading identifiable dental images, medical histories, or personal details to a public chatbot. Before an AI service receives your records, ask who receives the data, where they are stored, how long they are kept, whether they train future models, and what choices you have. WHO guidance emphasizes autonomy, safety, transparency, accountability, inclusion, and sustainability [2]. Your clinician remains responsible for interpreting findings and discussing uncertainty and alternatives.

How should I discuss an AI result with my clinician?

Narad, Sharda and Aasdhir's dental AI ethics narrative review discusses what AI can and cannot do, representative training data and fairness, privacy, and who is accountable when things go wrong [10, pp. 1000–1002]. Ask your clinician how an AI output fits your symptoms, examination and images rather than trusting an AI-generated explanation: even a plausible explanation made after the fact can mislead [10, p. 1002]. Ask how the tool performed on patients like you, and who will handle an error or concern. This review is not a diagnostic-accuracy or outcomes trial; it does not prove patient benefit or that any product is safe. This guide does not say our practice uses clinical AI.

Questions to ask your provider

What exact task is this tool designed for? Has it been tested independently on patients like me? What happens if it disagrees with your examination? Would its result change my treatment, and why? Could it lead to extra imaging or costs? Who checks its output? How are my records protected, and can I decline this use? A clear explanation of limitations is more useful than a best-case accuracy headline.

Common patient questions

Explore the evidence and your care

• Research: AI and vertical root fractures on CBCT
• Research: chatbots and the endodontic board examination
• Our technology and how we evaluate it
• How an endodontic diagnosis is made
• When CBCT imaging can help
• Research evidence and its limitations

Sources and evidence boundaries

1. Swanson et al. Driving emerging technology to clinical impact. JADA. 2026;157(9):957–967.

   Translation perspective; not an AI accuracy or clinical-benefit trial.

2. World Health Organization. Ethics and governance of artificial intelligence for health. 2021.

   Independent ethics guidance; not evidence of endodontic diagnostic accuracy.

3. Allihaibi, Koller & Mannocci. Diagnostic Accuracy of a Commercial AI-based Platform in Evaluating Endodontic Treatment Outcomes on Periapical Radiographs Using CBCT as the Reference Standard. J Endod. 2025;51:898–908.

   Retrospective radiographic comparison; sensitivity/specificity tradeoff, not proof of clinical failure or patient benefit.

4. Akyüz, Kıvırcık & Aslan. Artificial Intelligence-assisted Periapical Radiographic Assessment: Lesion Detection, Endodontic Complication Analysis, and Review of Clinical Treatment Recommendations. J Endod. 2026;52:1194–1205.

   Selected 60-case benchmark against expert consensus; missed complications and no patient-outcome trial.

5. Ji et al. Construction and Evaluation of an AI-based CBCT Resolution Optimization Technique for Extracted Teeth. J Endod. 2024;50:1298–1306.

   Extracted-tooth laboratory study with micro-CT reference; not external clinical validation or proven dose reduction.

6. Turp et al. Automated Detection of Middle Mesial Canals in Mandibular Molars on CBCT using nnU-Net: A Retrospective Diagnostic Accuracy Study. J Endod. 2026;52:1352–1358.

   Small 28-case internal held-out test from one institution/device; no independent external validation or demonstrated treatment benefit.

7. Ourang et al. Artificial Intelligence Chatbots’ Performance on Dental Trauma Case-based Queries: Examining the Effect of Prompt and Content Engineering. J Endod. 2026;52:1444–1452.

   Four hypothetical trauma scenarios; prompt and clinical content changed together. Not a study of injured patients.

8. Jalali et al. Artificial Intelligence Chatbots Taking American Board of Endodontics Simulated Oral Board Examination. J Endod. 2026;52:1148–1156.

   Three simulated cases; four fabricated citations per tested model. Not the official examination or validation of autonomous care.

9. Sabeti MA, Torabi G. AI Prognostic Models in Endodontics: Scoping Review and Clinical Gap Analysis for Retreatment vs Microsurgery Decision Support. J Endod. 2026. Journal pre-proof.

   Six prognostic studies, without external validation or formal calibration; exploratory treatment-choice framework. Supplied pre-proof has a screening-flow discrepancy and unavailable supplementary tables; final volume/pages not assigned here.

10. Narad C, Sharda K, Aasdhir SS. Ethical dimensions of artificial intelligence integration into dental practice. JADA. 2026;157(9):999–1004.

    Six-page narrative ethics review; not a diagnostic-accuracy study, patient-outcome trial, proof of benefit or product-safety evaluation.

11. Torabinejad M, Tordik PA, Alenezi M, Parirokh M, Dianat O, Foroughi M, Zegar Z, Oyoyo UE. Comparing Diagnostic Ability of Two Off-the-Shelf Large Language Models with That of Endodontists for Teeth Requiring Initial Root Canal Treatment: A Multicenter Retrospective Investigation. Journal of Endodontics. 2026. Journal pre-proof.

    Supplied journal pre-proof, accepted October 7, 2026. Retrospective expert-reference comparison of 645 teeth; individual and composite endpoints differ. Expert variability and single-pass consumer-interface testing limit interpretation. No autonomous-care safety or patient-outcome benefit established. Editorial addition awaiting clinical review.

The dental AI ethics article was read and verified as a narrative review. Its discussion informs patient questions; the article alone does not establish clinical benefit or product safety. The clinical approval of this guide and its limited scope are stated above, separately from source verification.