Conversational artificial intelligence compared with orthodontic reference standards for orthodontic information: A systematic review of accuracy, readability, safety, and certainty of evidence

Background Conversational artificial intelligence (AI) is increasingly used to obtain orthodontic information, but its reliability, completeness, readability, agreement with specialist judgment, and safety remain uncertain. Objective To systematically evaluate AI chatbots for delivering orthodontic information and clinical guidance, focusing on accuracy/reliability, completeness, readability, specialist agreement, and misleading information. Methods This systematic review followed PRISMA 2020 and Cochrane guidance. Embase, PubMed, Scopus, Web of Science, ClinicalTrials.gov, Open Science Framework, and Google Scholar were searched. Eligible studies assessed AI-generated orthodontic responses against specialists, reference answers, or structured assessment tools. Risk of bias was evaluated using the Joanna Briggs Institute checklist. Meta-analysis was not performed because of substantial heterogeneity in chatbot systems, prompts, comparators, scoring instruments, and outcome definitions. Certainty of evidence was assessed using GRADE. Results Sixteen analytical cross-sectional studies were included. Chatbots generated coherent and accessible answers to routine orthodontic questions, especially foundational concepts and patient frequently asked questions. However, performance varied by model version, prompt design, and clinical complexity. Important limitations were identified in treatment planning, individualized recommendations, contextual risk framing, and information completeness. Some studies reported incomplete or false claims regarding clear aligners, airway effects, stability, and surgery. Readability was generally favorable but did not ensure clinical adequacy. Risk of bias ranged from low to moderate, and certainty of evidence was very low. Conclusions AI chatbots may support routine orthodontic education under professional supervision, but current evidence does not support autonomous use. Standardized prompts, validated scoring tools, calibrated evaluators, and real-world outcome studies are needed.