Despite only achieving specialty designation by the American Dental Association in 2020, oral medicine has long served as a pioneering discipline at the crossroads of dentistry and medicine and of clinical care and academia. Clinical oral medicine encompasses the diagnosis and management of medically related diseases, disorders, and conditions affecting the oral and maxillofacial region as well as the oral health care of medically complex patients. Beyond clinical care, oral medicine boasts a particularly robust relationship with academia and a history of seminal contributions to dental education and research. A 2024 study identified that 47% of board-certified oral medicine specialists practice in the dental school setting compared to 30% in hospitals and 19% in private practice, reflecting one of the highest academic-focused practice distributions among the dental specialties.

Many oral medicine specialists partake in both clinical care and academic pursuits, and both branches of oral medicine are amenable to growth and optimization by artificial intelligence (AI). An earlier article of this issue, “Artificial Intelligence and Its Applications in Oral Medicine–Part 1” discusses AI applications related to direct patient care in oral medicine. However, indirect patient care, student education, and research are also data-driven fields that have the potential to gain from AI tools that circumvent human fatigability, time constraints, and resource limitations.

Indirect patient care refers to health care services performed without face-to-face patient contact. Patient education is a service that is critical to building patient rapport and ensuring positive treatment outcomes. While patient education is frequently performed in-person, the increasing recourse of patients to online resources along with limited clinician time and manpower means that virtual patient education has become a common reality. A surplus of inaccurate publicly accessible medical information in the virtual world, though, often poses a risk to constructive patient education in the absence of clinician guidance. The ability of AI to replace or supplement the clinician in patient education as an indirect patient care service may, therefore, prove valuable to explore. Practice management, which includes the completion of administrative tasks such as encounter documentation and billing, remains another aspect of indirect patient care that may benefit from AI applications. The exigencies of practice management often detract from clinician time and resources, not only affecting career satisfaction but also impacting patient care downstream.

Outside clinical care and within the academic setting, oral medicine specialists are often involved in predoctoral and postdoctoral student education and research. Oral medicine education has seen major transformations over the years, with increased elicitation of active learning methodologies rather than a passive, top–down pedagogical approach. Continued challenges in oral medicine education include limited predoctoral clinical oral medicine exposure as well as a shortage of board-certified oral medicine faculty in certain institutions. Despite a strong focus on research by many oral medicine specialists, there still remain gaps in the literature on numerous topics, with time and funding constraints often representing the main barriers to traversing this gap.

AI applications in oral medicine targeting indirect patient care services such as virtual patient education and practice management and academic activities such as student education and research are truly necessary complements of those targeting direct patient care. Enhancements in patient education and practice management can elevate patient care outcomes by raising patient and clinician satisfaction, respectively. Given the representation of oral medicine in the academic setting and the pipeline of education and research to clinical practice, advancements in education and research can also translate into effectively improved results for patients in addition to benefits for students, academicians, and clinicians.

This study reviews the applications of AI in oral medicine patient education, practice management, student education, and research, navigating the benefits and limitations of AI applications in each domain. Fig. 1 demonstrates these domains of AI applications.

Domains of artificial intelligence applications in oral medicine related to indirect patient care and academia.

Patient education is an important aspect of patient care. Yet with long appointment waiting times, limited time to ask and answer questions in office, or restricted access to care, patients often seek out medical information on their own. The Internet and social media have made it relatively easy to access medical information for the public. However, medical information present in the cyberspace, even information contained in some medical and dental journals, often suffers from inaccuracies and deficiencies. And as health literacy differs from patient to patient, elite medical and dental journals may not be an easily understandable information source for many. It is particularly challenging to find trustworthy online information on rare or complex oral and maxillofacial diseases and conditions as well as on the role of oral medicine specialists in managing them. Investigating and establishing reliable resources for oral medicine patient education can serve to reduce delayed diagnoses, increase patient confidence and compliance, and improve overall patient outcomes, especially for socioeconomically underprivileged populations.

AI in the form of large language models (LLMs) has been considered as a possible resource for patient education. LLMs have the distinct advantage of being able to analyze vast volumes of textual and visual data. Prompted by open-ended queries, LLMs can generate textual responses that reflect natural human speech. LLMs can generate responses freely or perform various manipulations on provided text, such as summarization, reorganization, simplification, and translation. These responses are generated and displayed instantaneously, eliminating the time-intensive perusal of other online resources. Most importantly, LLMs can customize their responses to their audience, with varying diction, syntax, and tone. These characteristics poise LLMs as possibly revolutionary patient education tools.

Several studies have evaluated the potential of LLMs in oral medicine patient education specifically or in oral medicine literacy in general. These studies have focused on oral cancer, ^,,, oral mucositis, and medication-related osteonecrosis of the jaw (MRONJ).

Hassona and colleagues investigated ChatGPT’s ability to answer patient questions on early detection and diagnosis of oral cancer. The authors evaluated the usefulness, quality and reliability, understandability and readability, and actionability of ChatGPT’s responses to 108 common patient questions derived from an expert panel of oral medicine specialists and the Google-based online search tool People Also Ask (PAA). They found that 75% of the responses were “very useful,” with none being “not useful,” 96.3% of the responses were of “good quality and reliability,” and 70% of the responses were adequately understandable, yet most responses had poor readability and poor actionability. This demonstrates that while generally useful and understandable, ChatGPT responses may not necessarily empower patients to make medical decisions.

Alsayed and colleagues investigated the quality of information from ChatGPT 4 in the fields of oral surgery, preventive dentistry, and oral cancer. A total of 20 questions on oral cancer were developed by experienced clinicians and a custom standardized scoring rubric was designed to evaluate ChatGPT’s responses to these questions. The authors found that the accuracy, completeness, relevance, and clarity and comprehensibility of ChatGPT’s oral cancer information were all above 66%. Significantly, 53% of responses to the oral cancer questions were found to have potential risks and limitations. The risks and limitations identified raised caution regarding reliance on ChatGPT for oral cancer patient education.

Diniz-Freitas and colleagues evaluated the accuracy and readability of ChatGPT 4 and Gemini in answering oral cancer questions from laypersons. A total of 15 questions were selected by 2 researchers from Google searches on frequently asked oral cancer questions and subsequently directed to ChatGPT 4 and Gemini. The authors assessed the accuracy and readability of these responses. Using a 4 point scale for accuracy, they found that 65% of ChatGPT 4 responses were correct and complete, 20% were correct but insufficient, 15% included correct and incorrect/outdated information, and none were completely incorrect. Gemini produced 80% correct and complete responses, 20% correct but insufficient responses, and none in the other 2 categories, revealing greater accuracy than ChatGPT 4. Flesch Reading Ease Scores were not statistically significantly different between ChatGPT 4 and Gemini (both between the 10th and 12th-grade levels), but using the Flesch-Kinkaid Reading Grade Level, the authors found that ChatGPT 4 generated responses at a 10th-grade level on average while Gemini generated responses at a 12th-grade level on average. While ChatGPT 4 may have greater readability than Gemini, both tools were assessed as having limited readability for laypersons.

Maniyar and colleagues investigated the accuracy, completeness, and potential for misinformation dissemination by ChatGPT 3.5 on the topic of oral cancer. A total of 25 primarily patient-oriented and partially provider-oriented questions on various aspects of oral cancer were developed from a search of online resources. Using a 4 point grading system, the authors found that 80% of ChatGPT responses were “comprehensive/correct,” 16% were “incomplete/partially correct,” 4% were “completely inaccurate/irrelevant,” and none were “mixed with accurate and inaccurate data/misleading.”

Hunter and colleagues investigated the quality and readability of online resources for oral mucositis compared to ChatGPT-generated responses. A total of 46 commonly asked patient questions derived from PAA were extracted, and the authors evaluated Google’s own links provided to answer those 46 PAA questions and ChatGPT’s responses to the top 10 questions. ChatGPT was asked first to answer questions without an audience specified and then asked to generate responses at an eighth-grade reading level, which the authors defined as universal readability. The authors found that ChatGPT generated factual and accurate responses to all questions in both iterations. While only 6% of Google-provided links and none of the initial ChatGPT responses were scored as universally readable based on the Flesch Reading Ease Score and the Flesch-Kinkaid Reading Grade Level, 20% of ChatGPT responses generated specifically for an eighth-grade audience specified were scored as universally readable.

Çoban and Altay evaluated the quality of ChatGPT 3 responses to questions about MRONJ. An experienced oral and maxillofacial surgeon generated 120 common patient questions, including case scenarios related to stages 0 to 3 of MRONJ. The authors found that ChatGPT 3 responses had a mean 3.9 score on the Global Quality Scale, showing moderate quality.

Overall, the studies on LLMs in oral medicine patient education show promising results but are limited in scope to select oral and maxillofacial diseases. Beyond the limited scope of the research, there remain several deficiencies in the use of LLMs for patient information. LLMs have been found to provide misleading, incomplete, or inaccurate medical information in some cases. The data accessible to and analyzed by LLMs changes over time, as revealed by significant differences between different ChatGPT versions, introducing the potential for the dissemination of outdated information. Biases in the data available may also exist. Unlike medical providers, LLMs are also not able to provide personalized medical care as they do not possess the full, unique medical profile of each patient and therefore lack the requisite level of contextual knowledge to render patient-centered decisions. This limits the generalizability of LLM-generated medical information and can lead to harmful consequences.

Prospective trials evaluating the impact of LLM-assisted patient education compared to provider-directed patient education, actual patient perceptions of LLM-generated medication information, and clinical workflow implications of LLM use are warranted to gauge the clinical utility of LLMs in oral medicine patient education. If validated clinically, LLMs can enhance the efficiency of oral medicine practices by reducing the time and workload of in-person patient education in addition to overcoming certain cultural and socioeconomic gaps in oral medicine.