2.1 Dental professions

The provision of dental care largely depends on face-to-face communication, clinical and radiographic examinations, and operative procedures. Therefore, LLMs will not substantially transform the everyday working life of dental practitioners, dental assistants, and hygienists. By contrast, advancements in LLMs could change the work and workload of administrative personnel—such as insurance claims administrators or clerical staff in dental offices—and alter the delivery of dental telemedicine.

2.2 Dental telemedicine

Tailored and fine-tuned applications based on LLMs have the potential to enhance dental telemedicine services when used in combination with dental health care personnel. LLM applications could efficiently collect patient information, analyze symptoms, and suggest potential diagnoses, which can then be reviewed by a human dental professional.¹⁰ By leveraging LLMs, dental telemedicine services could become more scalable, particularly in underserved areas where access to dental care is limited.¹⁰

Furthermore, LLMs could facilitate dental telemedicine consultations by providing real-time language translation services. This feature can remove language barriers that often hinder access to dental care, allowing dental professionals to effectively communicate with patients who speak different languages.

2.3 Clinical decision support

The accuracy of answers provided by LLMs is dependent on the quantity, quality, and type of data used in their training. The dataset to which a LLM has access needs to be comprehensive, up-to-date, and vetted to ensure that the whole body of evidence, including the latest scientific findings, is considered and to avoid bias from sources that contain misinformation or disinformation.^11-13 The knowledge base of ChatGPT only extends until 2021, rendering some prompts and queries useless. Moreover, ChatGPT is not designed to give medical guidance, which is why it is unsuitable for clinical decision support today. However, LLMs have considerable potential to become an additional tool for clinical decision support. For instance, LLMs that also take account of patients' electronic health record data may soon be used to enhance evidence-based selection of medical imaging exams.⁹ Additionally, LLMs have the potential to aid health care professionals in quickly finding pertinent information by summarizing extensive medical records, reducing the need for time-intensive chart reviews.⁹

2.4 Administrative work

Recently, Dr Clifford Stermer, a rheumatologist, has demonstrated how the use of ChatGPT makes writing preauthorization requests to insurance companies more efficient.⁹ This showcases the potential of LLMs in supporting health care professionals in routine written communications and record keeping, leading to increased efficiency in administrative work. Such time savings could free up health care professionals' time for other important tasks, potentially enhancing the quality of patient care and reducing costs. LLMs can also facilitate multilingual communication by translating texts, which is helpful to overcome language barriers in patient interactions and elsewhere.

2.5 Patient education

The internet has had a profound impact on patient education, making a vast amount of health information readily available. Wikipedia, for example, is the first port of call for many people worldwide seeking medical information.¹⁴ ChatGPT and similar LLMs are predicted to challenge the domination of web-based search engines such as Google in online query traffic. Microsoft has revamped its web search engine, Bing, including the addition of LLM chatbot functionality marketed as “the new Bing.” On February 6, 2023, Google announced Bard, a chatbot powered by LaMDA, its proprietary family of conversational LLMs. The ascendancy of LLM chatbots in the online search market has wide ramifications. ChatGPT is already capable of producing coherent and human-like conversational responses to various prompts and questions. In addition, unlike conventional search engines, it is not necessary to click on a website to get an answer. The dialogue-based, human-like interaction of ChatGPT and similar LLMs holds great appeal for users: Most simple prompts and queries will return an articulate and succinct reply. However, a major disadvantage of ChatGPT compared with web-based search engines is the impossibility to evaluate the credibility of the sources of its responses. Furthermore, ChatGPT sometimes wrongfully answers a question with complete confidence, its current knowledge base does not extend beyond the year 2021, and it cannot access the internet. Owing to these serious limitations in reliability, transparency, and knowledge, ChatGPT is unsuitable for providing health advice and guidance today. The answers to health-related queries online can have immediate health consequences.¹⁴ With the expected increase in the number of people requesting health information from LLMs, it is therefore vital to safeguard against inaccurate and biased responses to such queries.

2.6 Patient data privacy

The use of LLMs in dental medicine may involve the collection and storage of personal and medical information. This raises concerns about privacy and data security. Before LLMs are used in dental medicine, it is imperative to implement state-of-the-art data protection measures, including encryption of sensitive information, secure data transmission and storage, and access controls. The terms of service for ChatGPT permit OpenAI to collect and use log data, device information, and most importantly, usage data. To protect the privacy and confidentiality of patients' personal and medical information, no patient data must therefore be entered into ChatGPT.

2.7 Cybersecurity

There has been a uptick in cyber-attacks targeted at health care providers.¹⁵ The ransomware attack on the United Kingdom's National Health Service in August 2022 affected many of its services, including ambulance dispatch, emergency prescriptions, mental health services, patient referrals, and out-of-hours appointment scheduling.¹⁶ Such malware attacks put a spotlight on the vulnerabilities of health organizations. LLMs can potentially be abused to assist in writing phishing messages and developing malware.¹⁷ The companies and agencies involved in building LLMs must therefore take concerted efforts to thwart the risks of malicious use. Additionally, considering that threat actors are adept at exploiting technologies such as LLMs, it is crucial to fortify defenses against malware attacks to protect patient data and health care providers' operational hardware and software.¹⁸

2.8 Dental education

LLMs bring both opportunities and challenges to higher education, particularly in fields that heavily rely on written assignments.^{6, 19, 20} As LLMs respond differently to each interaction, traditional plagiarism checkers are becoming ineffective.¹⁹ To meet these challenges and to harness the potential of LLMs, some higher education curricula need to be tweaked and adapted. Given that dental schools primarily use oral exams, multiple-choice exams, practical assessments, and supervised patient treatments rather than essay type tests to evaluate dental students' knowledge and skills, LLMs do not present the same challenges in dentistry as they do in other academic disciplines. Nonetheless, it is important to increase dental students' literacy in AI applications relevant to dental medicine. For this purpose, a special curriculum has recently been proposed.²¹ The implementation of this core curriculum, or a modification of it, at dental schools worldwide merits serious consideration.

2.9 Scientific writing

English is the lingua franca used in most dental and medical journals today. Drafting clear and concise reports and grant applications can prove to be a challenge, particularly for non-native English speakers.²² LLMs may be a useful tool for authors to make the writing in such reports more fluent, thus leveling the playing field for non-native English speakers.²³ There are, however, grave concerns that LLMs could promote flawed or even fabricated research.^{24, 25} In a recent investigation, available as preprint, it has been shown that scientific abstracts written by ChatGPT based on completely generated data evaded plagiarism detection and often fooled human reviewers.²⁶ Some publishers have already acted. For instance, Science, one of the most highly ranked academic journals, has banned the use of text written by AI, machine learning, or similar algorithmic tools.²⁷ Springer-Nature, a leading publishing house, has also amended its guidelines.²⁸ The amended guidelines require that researchers report their use of LLM tools in the methods or acknowledgments section. Additionally, they prohibit listing LLMs such as ChatGPT as authors because LLMs cannot accept accountability for the work produced, which is an important requirement of authorship.

To distinguish between text written by humans and text written by LLMs, OpenAI launched a classifier tool on January 31, 2023.²⁹ However, according to OpenAI, the classifier tool only correctly identifies 26% of English texts written by a LLM and incorrectly labels 9% of texts written by a human as probably written by a LLM.²⁹ This underscores the importance of rigorous peer review to weed out unsound science and misinformation.

2.10 Scientific evaluation

This narrative review has major limitations owing to the paucity of high-level evidence and the lack of a systematic literature search. The conclusions are based mostly on expert opinions rather than robust empirical data. The lack of high-level evidence emphasizes the critical need for further investigation into the use of LLMs in health care fields. To ensure the accuracy of health information provided by LLMs, it is essential to subject their output to rigorous scientific evaluation and verification. By evaluating the performance of a LLM on a range of tasks and benchmark datasets, it is possible to understand the capabilities and limitations of a LLM.³⁰ This evaluation process can help identify areas that require improvement and develop strategies to optimize the performance of a LLM.³⁰ Output control is also critical for LLMs, especially in health care applications, to ensure that the models produce appropriate and accurate outputs. Output control may involve filtering out inappropriate or low-quality outputs and incorporating human oversight to ensure that LLMs produce high-quality results.