Table 3.

Proposed clinical supervision framework for conversational AI use in orthodontic information.

Clinical use category	Examples of chatbot use	Evidence pattern across included studies	Main clinical concern	Recommended supervision level
Routine orthodontic education and general FAQs	Definitions of orthodontics; basic treatment process; appliance care; oral hygiene; diet; treatment duration; retention basics; frequently asked patient questions.	Several FAQ-based studies reported coherent, generally acceptable, or high-scoring responses for routine questions. Examples include Daraqel et al. for 100 general orthodontic questions, Arslan et al. for 24 common questions, and Zhou et al. for Chinese-language patient FAQs.	General responses may be fluent but still omit patient-specific qualifiers, alternative options, or limitations.	Adjunctive use is reasonable after orthodontist review or when embedded in clinician-curated patient education. Responses should be labelled as general information, not individualized advice.
Patient information requiring contextual qualification	Clear aligner FAQs; side effects of orthodontic treatment; retention and relapse; pain management; orthognathic-surgery recovery; appliance choice.	Evidence was mixed. Dursun and Bilici Gecer found generally accurate clear-aligner responses but variable reliability, quality, and readability. Vassis et al. found that patients often preferred AI-generated education, whereas experts judged the content insufficient for standalone education. Yurdakurban et al. reported good quality/reliability for orthognathic-surgery information but difficult readability.	Risk of incomplete counselling, difficult readability, underemphasis of warnings, and overgeneralized recommendations.	Use only after professional verification. Add disclaimers, explain uncertainty, and align outputs with institutional patient information and readability standards.
Evidence-sensitive or claim-sensitive topics	Claims about clear aligner effectiveness; airway effects; sleep apnea; root control; treatment stability; surgery need; TMD-related statements; comparisons between appliances.	Abu Arqub et al. found that only 58% of ChatGPT-generated clear-aligner statements were objectively true and 15% were false, including claims related to surgery need, airway function, root parallelism, sleep apnea, and stability. Makrygiannakis et al. also reported that LLM answers occasionally lacked comprehensiveness, scientific accuracy, clarity, or relevance.	Misleading or unsupported statements may affect expectations, consent, and treatment decisions even when the language appears confident.	Do not release directly to patients without fact-checking. Orthodontist review against current evidence is required before use.
Educational or knowledge support for trainees and clinicians	True/false knowledge testing; textbook-based orthodontic concepts; MARPE, CBCT, clear aligner knowledge; evidence retrieval for clinician-oriented questions.	Metin and Goymen found that chatbots generally outperformed students and dentists but remained below orthodontists, with important consistency differences among models. Naureen and Kiani found ChatGPT-4 more accurate than Google Bard for recent orthodontic modalities. These studies support educational adjunct use but not unverified authority.	Inconsistent outputs and model-dependent performance can create false confidence in educational or clinical learning contexts.	Use as a supplementary educational tool with answer keys, instructor verification, and explicit discussion of model limitations.
Clinical case interpretation and treatment planning	Diagnosis; individualized treatment recommendations; extraction or non-extraction decisions; impacted tooth management; mixed-dentition planning; surgical and airway-related planning.	Tanaka et al. showed significantly greater orthodontist agreement with an expert orthodontist treatment plan than with AI-generated plans. Hatia et al. reported high but incomplete accuracy/completeness for interceptive orthodontic clinical questions and scenarios. Alkhamees reported generally good ChatGPT-4 responses, but with variability and low inter-rater agreement.	Individualized decisions require examination, records, risk assessment, patient preferences, and specialist judgment; apparent agreement with one reference does not prove correctness.	Not appropriate for autonomous use. Chatbots may support discussion or drafting, but diagnosis and treatment planning must remain specialist-led.
Implementation safeguards across all categories	Model/version documentation; access date; standardized prompts; session resetting; evaluator calibration; readability checking; evidence verification; disclosure of limitations.	Across studies, heterogeneity in prompts, models, access dates, comparators, and outcome tools limited synthesis. GRADE certainty remained very low for all outcomes.	Model drift, prompt sensitivity, lack of reproducibility, and non-standardized outcome assessment limit confidence in direct clinical transfer.	Institutional governance is recommended: use predefined prompts, audit outputs periodically, document model versions, and ensure patient-facing outputs direct users to orthodontic consultation.

Abbreviations: AI, artificial intelligence; CBCT, cone-beam computed tomography; FAQ, frequently asked question; GRADE, Grading of Recommendations Assessment, Development and Evaluation; LLM, large language model; MARPE, mini-implant-assisted rapid palatal expansion; TMD, temporomandibular disorders. The framework synthesizes reported evidence patterns and does not replace clinical judgment or locally validated patient-information protocols.