J Esthet Restor Dent - 2023 - Eggmann - Implications of Large Language Models Such As ChatGPT For Dental Medicine
J Esthet Restor Dent - 2023 - Eggmann - Implications of Large Language Models Such As ChatGPT For Dental Medicine
DOI: 10.1111/jerd.13046
REVIEW ARTICLE
1
Department of Preventive and Restorative
Sciences, Penn Dental Medicine, Robert Abstract
Schattner Center, University of Pennsylvania,
Objective: This article provides an overview of the implications of ChatGPT and
Philadelphia, Pennsylvania, USA
2
Department of Periodontology,
other large language models (LLMs) for dental medicine.
Endodontology, and Cariology, University Overview: ChatGPT, a LLM trained on massive amounts of textual data, is adept at
Center for Dental Medicine Basel UZB,
University of Basel, Basel, Switzerland
fulfilling various language-related tasks. Despite its impressive capabilities, ChatGPT
3
Department of Reconstructive Dentistry, has serious limitations, such as occasionally giving incorrect answers, producing non-
University Center for Dental Medicine Basel
sensical content, and presenting misinformation as fact. Dental practitioners, assis-
UZB, University of Basel, Basel, Switzerland
tants, and hygienists are not likely to be significantly impacted by LLMs. However,
Correspondence
LLMs could affect the work of administrative personnel and the provision of dental
Florin Eggmann, Department of
Periodontology, Endodontology, and telemedicine. LLMs offer potential for clinical decision support, text summarization,
Cariology, University Center for Dental
efficient writing, and multilingual communication. As more people seek health infor-
Medicine Basel UZB, University of Basel,
Mattenstrasse 40, CH-4058, Basel, mation from LLMs, it is crucial to safeguard against inaccurate, outdated, and biased
Switzerland.
responses to health-related queries. LLMs pose challenges for patient data confiden-
Email: [email protected]
tiality and cybersecurity that must be tackled. In dental education, LLMs present
Funding information fewer challenges than in other academic fields. LLMs can enhance academic writing
Freiwillige Akademische Gesellschaft;
Gottfried und Julia Bangerter-Rhyner-Stiftung fluency, but acceptable usage boundaries in science need to be established.
Conclusions: While LLMs such as ChatGPT may have various useful applications in
dental medicine, they come with risks of malicious use and serious limitations, includ-
ing the potential for misinformation.
Clinical Significance: Along with the potential benefits of using LLMs as an additional
tool in dental medicine, it is crucial to carefully consider the limitations and potential
risks inherent in such artificial intelligence technologies.
KEYWORDS
artificial intelligence, dental care, dental education, evidence-based dentistry, machine learning
This is an open access article under the terms of the Creative Commons Attribution-NonCommercial-NoDerivs License, which permits use and distribution in any
medium, provided the original work is properly cited, the use is non-commercial and no modifications or adaptations are made.
© 2023 The Authors. Journal of Esthetic and Restorative Dentistry published by Wiley Periodicals LLC.
terabytes of textual data. They are generative mathematical models of efficiently collect patient information, analyze symptoms, and suggest
the statistical distribution of tokens in the vast public corpus of human potential diagnoses, which can then be reviewed by a human dental
generated text, where the tokens include words, graphemes, individ- professional.10 By leveraging LLMs, dental telemedicine services could
4
ual characters, and punctuation marks. When given a prompt, such become more scalable, particularly in underserved areas where access
as “Who first described the use of phosphoric acid etching for dental to dental care is limited.10
bonding?,” the LLM provides the answer, “Dr Michael Buonocore,” Furthermore, LLMs could facilitate dental telemedicine consulta-
because the statistical assessment of the corpus of human generated tions by providing real-time language translation services. This feature
text indicate that this is most likely the correct reply.5 However, the can remove language barriers that often hinder access to dental care,
LLM does not possess actual knowledge of adhesive dentistry. The allowing dental professionals to effectively communicate with
capabilities of LLMs are nevertheless impressive. LLMs can generate patients who speak different languages.
fluent and coherent texts, answer questions, translate languages, and
perform other language-related tasks.
ChatGPT, developed by the US company OpenAI, part-owned by 2.3 | Clinical decision support
Microsoft, is a LLM with a conversational interface that is free to use
and easily accessible online. Since its launch in November 2022, the The accuracy of answers provided by LLMs is dependent on the quan-
ChatGPT chatbot has been used by millions, sparking both consider- tity, quality, and type of data used in their training. The dataset to
able excitement and serious concern.6–8 which a LLM has access needs to be comprehensive, up-to-date, and
LLMs have numerous potential applications in dental medicine, vetted to ensure that the whole body of evidence, including the latest
from streamlined dental record keeping to AI support in clinical deci- scientific findings, is considered and to avoid bias from sources that
sion making.9 While such applications may become available in the contain misinformation or disinformation.11–13 The knowledge base
near future, the current iteration of LLMs are already adept at text of ChatGPT only extends until 2021, rendering some prompts and
summarization and translation, making them useful tools for dental queries useless. Moreover, ChatGPT is not designed to give medical
practitioners and students, especially those studying in a non-native guidance, which is why it is unsuitable for clinical decision support
language. However, LLMs can also give entirely wrong answers, pro- today. However, LLMs have considerable potential to become an
duce nonsensical content, and present misinformation and disinforma- additional tool for clinical decision support. For instance, LLMs that
tion as fact, which causes serious concerns in critical fields such as also take account of patients' electronic health record data may soon
health care. be used to enhance evidence-based selection of medical imaging
Given the importance of emergent AI applications for patient exams.9 Additionally, LLMs have the potential to aid health care pro-
care, education, and dental research, the aim of this narrative review fessionals in quickly finding pertinent information by summarizing
is to provide an overview of possible applications of LLMs such as extensive medical records, reducing the need for time-intensive chart
ChatGPT in dentistry, their current limitations, and drawbacks. reviews.9
challenge the domination of web-based search engines such as that threat actors are adept at exploiting technologies such as LLMs,
Google in online query traffic. Microsoft has revamped its web it is crucial to fortify defenses against malware attacks to protect
search engine, Bing, including the addition of LLM chatbot function- patient data and health care providers' operational hardware and
ality marketed as “the new Bing.” On February 6, 2023, Google software.18
announced Bard, a chatbot powered by LaMDA, its proprietary fam-
ily of conversational LLMs. The ascendancy of LLM chatbots in the
online search market has wide ramifications. ChatGPT is already 2.8 | Dental education
capable of producing coherent and human-like conversational
responses to various prompts and questions. In addition, unlike con- LLMs bring both opportunities and challenges to higher education,
ventional search engines, it is not necessary to click on a website to particularly in fields that heavily rely on written assignments.6,19,20 As
get an answer. The dialogue-based, human-like interaction of LLMs respond differently to each interaction, traditional plagiarism
ChatGPT and similar LLMs holds great appeal for users: Most simple checkers are becoming ineffective.19 To meet these challenges and to
prompts and queries will return an articulate and succinct reply. harness the potential of LLMs, some higher education curricula need
However, a major disadvantage of ChatGPT compared with web- to be tweaked and adapted. Given that dental schools primarily use
based search engines is the impossibility to evaluate the credibility oral exams, multiple-choice exams, practical assessments, and super-
of the sources of its responses. Furthermore, ChatGPT sometimes vised patient treatments rather than essay type tests to evaluate den-
wrongfully answers a question with complete confidence, its current tal students' knowledge and skills, LLMs do not present the same
knowledge base does not extend beyond the year 2021, and it can- challenges in dentistry as they do in other academic disciplines.
not access the internet. Owing to these serious limitations in reliabil- Nonetheless, it is important to increase dental students' literacy in AI
ity, transparency, and knowledge, ChatGPT is unsuitable for applications relevant to dental medicine. For this purpose, a special
providing health advice and guidance today. The answers to health- curriculum has recently been proposed.21 The implementation of this
related queries online can have immediate health consequences.14 core curriculum, or a modification of it, at dental schools worldwide
With the expected increase in the number of people requesting merits serious consideration.
health information from LLMs, it is therefore vital to safeguard
against inaccurate and biased responses to such queries.
2.9 | Scientific writing
2.6 | Patient data privacy English is the lingua franca used in most dental and medical journals
today. Drafting clear and concise reports and grant applications can
The use of LLMs in dental medicine may involve the collection and prove to be a challenge, particularly for non-native English speakers.22
storage of personal and medical information. This raises concerns LLMs may be a useful tool for authors to make the writing in such
about privacy and data security. Before LLMs are used in dental medi- reports more fluent, thus leveling the playing field for non-native
cine, it is imperative to implement state-of-the-art data protection English speakers.23 There are, however, grave concerns that LLMs
measures, including encryption of sensitive information, secure data could promote flawed or even fabricated research.24,25 In a recent
transmission and storage, and access controls. The terms of service investigation, available as preprint, it has been shown that scientific
for ChatGPT permit OpenAI to collect and use log data, device infor- abstracts written by ChatGPT based on completely generated data
mation, and most importantly, usage data. To protect the privacy and evaded plagiarism detection and often fooled human reviewers.26
confidentiality of patients' personal and medical information, no Some publishers have already acted. For instance, Science, one of the
patient data must therefore be entered into ChatGPT. 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
2.7 | Cybersecurity guidelines.28 The amended guidelines require that researchers report
their use of LLM tools in the methods or acknowledgments section.
There has been a uptick in cyber-attacks targeted at health care pro- Additionally, they prohibit listing LLMs such as ChatGPT as authors
viders.15 The ransomware attack on the United Kingdom's National because LLMs cannot accept accountability for the work produced,
Health Service in August 2022 affected many of its services, includ- which is an important requirement of authorship.
ing ambulance dispatch, emergency prescriptions, mental health ser- To distinguish between text written by humans and text written
vices, patient referrals, and out-of-hours appointment scheduling.16 by LLMs, OpenAI launched a classifier tool on January 31, 2023.29
Such malware attacks put a spotlight on the vulnerabilities of health However, according to OpenAI, the classifier tool only correctly iden-
organizations. LLMs can potentially be abused to assist in writing tifies 26% of English texts written by a LLM and incorrectly labels 9%
17
phishing messages and developing malware. The companies and of texts written by a human as probably written by a LLM.29 This
agencies involved in building LLMs must therefore take concerted underscores the importance of rigorous peer review to weed out
efforts to thwart the risks of malicious use. Additionally, considering unsound science and misinformation.
17088240, 2023, 7, Downloaded from https://onlinelibrary.wiley.com/doi/10.1111/jerd.13046 by Turkey Cochrane Evidence Aid, Wiley Online Library on [07/03/2024]. See the Terms and Conditions (https://onlinelibrary.wiley.com/terms-and-conditions) on Wiley Online Library for rules of use; OA articles are governed by the applicable Creative Commons License
EGGMANN ET AL. 1101
21. Schwendicke F, Chaurasia A, Wiegand T, et al. Artificial intelligence 28. Editorials N. Tools such as ChatGPT threaten transparent science;
for oral and dental healthcare: core education curriculum. J Dent. here are our ground rules for their use. Nature. 2023;613:612.
2023;128:104363. 29. Kirchner JH, Ahmad L, Aaronson S, Leike J. New AI classifier for indi-
22. Baron TH. ABC's of writing medical papers in English. Korean J Radiol. cating AI-written text. Accessed February 1, 2023 https://openai.
2012;13(suppl 1):S1-S11. com/blog/new-ai-classifier-for-indicating-ai-written-text/
23. Kitamura FC. ChatGPT is shaping the future of medical writing but 30. Cascella M, Montomoli J, Bellini V, Bignami E. Evaluating the feasibil-
still requires human judgment. Radiology. 2023;230171. doi:10.1148/ ity of ChatGPT in healthcare: an analysis of multiple clinical and
radiol.230171 research scenarios. J Med Syst. 2023;47(1):33.
24. Else H. Abstracts written by ChatGPT fool scientists. Nature. 2023;
613(7944):423.
25. Gordijn B, Ten Have H. ChatGPT: evolution or revolution? Med
Health Care Philos. 2023;26(1):1-2.
26. Gao CA, Howard FM, Markov NS, et al. Comparing scientific How to cite this article: Eggmann F, Weiger R, Zitzmann NU,
abstracts generated by ChatGPT to original abstracts using an artifi- Blatz MB. Implications of large language models such as
cial intelligence output detector, plagiarism detector, and blinded ChatGPT for dental medicine. J Esthet Restor Dent. 2023;35(7):
human reviewers: preprint. 2022. Accessed January 30, 2023. doi:10.
1098‐1102. doi:10.1111/jerd.13046
1101/2022.12.23.521610v1.article-info
27. Thorp HH. ChatGPT is fun, but not an author. Science. 2023;
379(6630):313.