ChatGPT Utility in Healthcare Education, Research, and Practice: Systematic Review on the Promising Perspectives and Valid Concerns

1Artificial intelligence (AI) can be defined as the multidisciplinary approach of computer science and linguistics that aspires to create machines capable of performing tasks that normally require human intelligence. These tasks include the ability to learn, adapt, rationalize, understand, and to fathom abstract concepts as well as the reactivity to complex human attributes such as attention, emotion, creativity, etc..

2The history of AI as a scientific discipline can be traced back to the mid-XX century at the Dartmouth Summer Research Project on AI. This was followed by the development of machine learning (ML) algorithms that allow decision-making or predictions based on the patterns in large datasets. Subsequently, the development of neural networks (brain-mimicking algorithms), genetic algorithms (finding optimal solutions for complex problems by application of evolutionary principles), and other advanced techniques followed.

3Launched in November 2022, “ChatGPT” is an AI-based large language model (LLM) trained on massive text datasets in multiple languages with the ability to generate human-like responses to text input. Developed by OpenAI (OpenAI, L.L.C., San Francisco, CA, USA), ChatGPT etymology is related to being a chatbot (a program able to understand and generate responses using a text-based interface) and is based on the generative pre-trained transformer (GPT) architecture. The GPT architecture utilizes a neural network to process natural language, thus generating responses based on the context of input text. The superiority of ChatGPT compared to its GPT-based predecessors can be linked to its ability to respond to multiple languages generating refined and highly sophisticated responses based on advanced modeling.

4In the scientific community and academia, ChatGPT has received mixed responses reflecting the history of controversy regarding the benefits vs. risks of advanced AI technologies. On one hand, ChatGPT, among other LLMs, can be beneficial in conversational and writing tasks, assisting to increase the efficiency and accuracy of the required output. On the other hand, concerns have been raised in relation to possible bias based on the datasets used in ChatGPT training, which can limit its capabilities and could result in factual inaccuracies, but alarmingly appear to be scientifically plausible (a phenomenon termed hallucination). Additionally, security concerns and the potential of cyber-attacks with the spread of misinformation utilizing LLMs should also be considered.

5The innate resistance of the human mind to any change is a well-described phenomenon and can be understandable from evolutionary and social psychology perspectives. Therefore, the concerns and debate that arose instantaneously following the widespread release of ChatGPT appear to be understandable. The attention that ChatGPT received involved several disciplines. In education, for example, ChatGPT release could mark the end of essays as assignments. In health care practice and academic writing, factual inaccuracies, ethical issues, and the fear of misuse including the spread of misinformation should be considered.

6The versatility of human intelligence (HI) compared to AI is related to its biological evolutionary history, adaptability, creativity, the ability of emotional intelligence, and the ability to understand complex abstract concepts. However, HI-AI cooperation can be beneficial if an accurate and reliable output of AI is ensured. The promising utility of AI in health care has been outlined previously with possible benefits in personalized medicine, drug discovery, and the analysis of large datasets aside from the potential applications to improve diagnosis and clinical decisions. Additionally, the utility of AI chatbots in health care education is an interesting area to probe. This is related to the massive information and various concepts that health care students are required to grasp. However, all of these applications should be considered cautiously considering the valid concerns, risks, and categorical failures experienced and cited in the context of LLM applications. Specifically, Borji comprehensively highlighted the caveats of ChatGPT use that included, but were not limited to, the generation of inaccurate content, the risk of bias and discrimination, lack of transparency and reliability, cybersecurity concerns, ethical consequences, and societal implications.

7Therefore, the aim of the current review was to explore the future perspectives of ChatGPT as a prime example of LLMs in health care education, academic/scientific writing, health care research, and health care practice based on the existing evidence. Importantly, the current review objectives extended to involve the identification of potential limitations and concerns that could be associated with the application of ChatGPT in the aforementioned areas in health care settings.

8The current systematic review was conducted according to the Preferred Reporting Items for Systematic Reviews and Meta-Analyses (PRIMSA) guidelines. The information sources included PubMed/MEDLINE and Google Scholar.

9The eligibility criteria involved any type of published scientific research or preprints (article, review, communication, editorial, opinion, etc.) addressing ChatGPT that fell under the following categories: (1) health care practice/research; (2) health care education; and (3) academic writing.

10The exclusion criteria included: (1) non-English records; (2) records addressing ChatGPT in subjects other than those mentioned in the eligibility criteria; and (3) articles from non-academic sources (e.g., newspapers, internet websites, magazines, etc.).

11The exact PubMed/MEDLINE search strategy, which concluded on 16 February 2023, was as follows: (ChatGPT) AND ((“2022/11/30” [Date–Publication]: “3000” [Date–Publication])), which yielded 42 records.

12The search on Google Scholar was conducted using Publish or Perish (Version 8). The search term was “ChatGPT” for the years 2022–2023, and the Google Scholar search yielded 238 records and concluded on 16 February 2023.

13The records retrieved following the PubMed/MEDLINE and Google Scholar searches were imported to EndNote v.20 for Windows (Thomson ResearchSoft, Stanford, CA, USA), which yielded a total of 280 records.

14Next, screening of the title/abstract was conducted for each record with the exclusion of duplicate records (n = 40), followed by the exclusion of records published in languages other than English (n = 32). Additionally, the records that fell outside the scope of the review (records that examined ChatGPT in a context outside health care education, health care practice, or scientific research/academic writing) were excluded (n = 80). Moreover, the records published in non-academic sources (e.g., newspapers, magazines, Internet websites, blogs, etc.) were excluded (n = 18).

15Afterward, full screening of the remaining records (n = 110) was carried out with the exclusion of an additional 41 records that fell outside the scope of the current review. An additional nine records were excluded due to the inability to access the full text of these records, being subscription-based. This yielded a total of 60 records eligible for inclusion in the current review.

16Each of the included records was searched specifically for the following: (1) type of record (preprint, published research article, opinion, commentary, editorial, review, etc.); (2) the listed benefits/applications of ChatGPT in health care education, health care practice, or scientific research/academic writing; (3) the listed risks/concerns of ChatGPT in health care education, health care practice, or scientific research/academic writing; and (4) the main conclusions and recommendations regarding ChatGPT in health care education, health care practice, or scientific research/academic writing.

17Categorization of the benefits/applications of ChatGPT was as follows: (1) educational benefits in health care education (e.g., generation of realistic and variable clinical vignettes, customized clinical cases with immediate feedback based on the student’s needs, enhanced communications skills); (2) benefits in academic/scientific writing (e.g., text generation, summarization, translation, and literature review in scientific research); (3) benefits in scientific research (e.g., efficient analysis of large datasets, drug discovery, identification of potential drug targets, generation of codes in scientific research); (4) benefits in health care practice (e.g., improvements in personalized medicine, diagnosis, treatment, lifestyle recommendations based on personalized traits, documentation/generation of reports); and (5) being a freely available package.

18Categorization of the risks/concerns of ChatGPT was as follows: (1) ethical issues (e.g., risk of bias, discrimination based on the quality of training data, plagiarism); (2) hallucination (the generation of scientifically incorrect content that sounds plausible); (3) transparency issues (black box application); (4) risk of declining need for human expertise with subsequent psychologic, economic and social issues; (5) over-detailed, redundant, excessive content; (6) concerns about data privacy for medical information; (7) risk of declining clinical skills, critical thinking and problem-solving abilities; (8) legal issues (e.g., copyright issues, authorship status); (9) interpretability issues; (10) referencing issues; (11) risk of academic fraud in research; (12) incorrect content; and (13) infodemic risk.

19A total of 280 records were identified, and following the full screening process, a total of 60 records were eligible to be included in the review. The PRISMA flowchart of the record selection process is shown in Figure 1.

20Summaries of the main conclusions of the included studies regarding ChatGPT utility in academic writing, health care education, and health care practice/research are provided in Table 1 for the records comprising editorials/letters to the editors, in Table 2 for the records comprising research articles, commentaries, news articles, perspectives, case studies, brief reports, communications, opinions, or recommendations, and in Table 3 for the records representing preprints.

21A summary of the record types included in the current review is shown in Figure 2.

22One-third of the included records were preprints (n = 20), with the most common preprint server being medRxiv (n = 6, 30.0%), followed by SSRN and arXiv (n = 4, 20.0%) for each. Editorials/letters to editors were the second most common type of included records (n = 19, 31.7%).

23The benefits of ChatGPT were most frequently cited in the context of academic/scientific writing, which was mentioned in 31 records (51.7%). Examples included efficiency and versatility in writing with text of high quality, improved language, readability, and translation promoting research equity, and accelerated literature review. Benefits in scientific research followed, which was mentioned in 20 records (33.3%). Examples included the ability to analyze massive data including electronic health records or genomic data, the availability of more free time for the focus on experimental design, and drug design and discovery. Benefits in health care practice was mentioned by 14 records (23.3%), with examples including personalized medicine, prediction of disease risk and outcome, streamlining the clinical workflow, improved diagnostics, documentation, cost saving, and improved health literacy. Educational benefits in health care disciplines were mentioned in seven records (11.7%) with examples including the generation of accurate and versatile clinical vignettes, improved personalized learning experience, and being an adjunct in group learning. Being a free package was mentioned as a benefit in two records (3.3%, Figure 3).

24Ethical concerns were commonly mentioned by 33 records (55.0%), especially in the context of risk of bias (mentioned by 18 records, 30.0%) and plagiarism (mentioned by 14 records, 23.3%) among data privacy and security issues.

25Other concerns involved the risk of incorrect/inaccurate information, which was mentioned by 20 records (33.3%); citation/reference inaccuracy or inadequate referencing, which was mentioned by 10 records (16.7%); transparency issues, which was mentioned by 10 records (16.7%); legal issues were mentioned in seven records (11.7%); restricted knowledge before 2021 was mentioned by six records (10.0%); risk of misinformation spread was mentioned by five records (8.3%); over-detailed content was mentioned in five records (8.3%); copyright issues were mentioned in four records (6.7%); and the lack of originality was mentioned by four records (6.7%, Figure 4).

26The far-reaching consequences of ChatGPT among other LLMs can be described as a paradigm shift in academia and health care practice. The discussion of its potential benefits, future perspectives, and importantly, its limitations, appear timely and relevant.

27Therefore, the current review aimed to highlight these issues based on the current evidence. The following common themes emerged from the available literature.

28ChatGPT, as an example of other LLMs, can be described as a promising or even a revolutionary tool for scientific research in both academic writing and in the research process itself. Specifically, ChatGPT was listed in several sources as an efficient and promising tool for conducting comprehensive literature reviews and generating computer codes, thereby saving time for the research steps that require more efforts from human intelligence (e.g., the focus on experimental design). Additionally, ChatGPT can be helpful in generating queries for comprehensive systematic review with high precision, as shown by Wang et al., despite the authors highlighting the transparency issues and unsuitability for high-recall retrieval. Moreover, the utility of ChatGPT extends to involve an improvement in language and a better ability to express and communicate research ideas and results, ultimately speeding up the publication process with the faster availability of research results. This is particularly relevant for researchers who are non-native English speakers. Such a practice can be acceptable considering the already existent English editing services provided by several academic publishers. Subsequently, this can help to promote equity and diversity in research.

29On the other hand, the use of ChatGPT in academic writing and scientific research should be conducted in light of several limitations that could compromise the quality of research as follows. First, superficial, inaccurate, or incorrect content was frequently cited as a shortcoming of ChatGPT use in scientific writing. The ethical issues including the risk of bias based on training datasets and plagiarism were also frequently mentioned, aside from the lack of transparency regarding content generation, which justifies the description of ChatGPT, on occasions, as a black box technology. Importantly, the concept of ChatGPT hallucination could be risky if the generated content is not thoroughly evaluated by researchers and health providers with proper expertise. This comes in light of the ability of ChatGPT to generate incorrect content that appears plausible from a scientific point of view.

30Second, several records mentioned the current problems regarding citation inaccuracies, insufficient references, and ChatGPT referencing to non-existent sources. This was clearly shown in two recently published case studies with ChatGPT use in a journal contest. These case studies discouraged the use of ChatGPT, citing the lack of scientific accuracy, the limited updated knowledge, and the lack of ability to critically discuss the results. Therefore, the ChatGPT generated content, albeit efficient, should be meticulously examined prior to its inclusion in any research manuscripts or proposals for research grants.

31Third, the generation of non-original, over-detailed, or excessive content can be an additional burden for researchers who should carefully supervise the ChatGPT-generated content. This can be addressed by supplying ChatGPT with proper prompts (text input), since varying responses might be generated based on the exact approach of prompt construction.

32Fourth, as it currently stands, the knowledge of ChatGPT is limited to the period prior to 2021 based on the datasets used in ChatGPT training. Thus, ChatGPT cannot be used currently as a reliable updated source of literature review. Nevertheless, ChatGPT can be used as a motivation to organize the literature in a decent format, if supplemented by reliable and up-to-date references.

33Fifth, the risk of research fraud (e.g., ghostwriting, falsified or fake research) involving ChatGPT should be considered seriously as well as the risk of generating mis- or disinformation with the subsequent possibility of infodemics.

34Sixth, legal issues in relation to ChatGPT use were also raised by several records including copyright issues. Finally, the practice of listing ChatGPT as an author does not appear to be acceptable based on the current ICMJE and COPE guidelines for determining authorship, as illustrated by Zielinski et al. and Liebrenz et al.. This comes in light of the fact that authorship entails legal obligations that are not met by ChatGPT. However, other researchers have suggested the possibility of ChatGPT inclusion as an author in some specified instances.

35A few instances were encountered in this review, where ChatGPT was listed as an author that can point to the initial perplexity of a few publishers regarding the role of LLM including ChatGPT in research. The disapproval of including ChatGPT or any other LLM in the list of authors was clearly explained in Science, Nature, and the Lancet editorials, which referred to such practice as scientific misconduct, and this view was echoed by many scientists. In the case of ChatGPT use in the research process, several records advocated the need for the proper and concise disclosure and documentation of ChatGPT or LLM use in the methodology or acknowledgement sections. A noteworthy and comprehensive record by Borji can be used as a categorical guide for the issues and concerns of ChatGPT use, especially in the context of scientific writing.

36From the health care practice perspective, the current review showed a careful excitement vibe regarding ChatGPT applications. The ability of ChatGPT to help in streamlining the clinical workflow appears promising, with possible cost savings and increased efficiency in health care delivery. This was illustrated recently by Patel and Lam, highlighting the ability of ChatGPT to produce efficient discharge summaries, which can be valuable to reduce the burden of documentation in health care. Additionally, ChatGPT, among other LLMs, can have a transforming potential in health care practice via enhancing diagnostics, prediction of disease risk and outcome, and drug discovery among other areas in translational research. Moreover, ChatGPT showed moderate accuracy in determining the imaging steps needed in breast cancer screening and in the evaluation of breast pain, which can be a promising application in decision making in radiology. ChatGPT in health care settings also has the prospects of refining personalized medicine and the ability to improve health literacy by providing easily accessible and understandable health information to the general public. This utility was demonstrated by ChatGPT responses, highlighting the need to consult health care providers among other reliable sources on specific situations.

37On the other hand, several concerns regarding ChatGPT use in health care settings were raised. Ethical issues including the risk of bias and transparency issues appeared as recurring major concerns. Additionally, the generation of inaccurate content can have severe negative consequences in health care; therefore, this valid concern should be cautiously considered in health care practice. This concern also extends to involve the ability of ChatGPT to provide justification for incorrect decisions.

38Other ChatGPT limitations including the issues of interpretability, reproducibility, and the handling of uncertainty were also raised, which can have harmful consequences in health care settings including health care research. In the area of personalized medicine, the lack of transparency and unclear information regarding the sources of data used for ChatGPT training are important issues in health care settings considering the variability observed among different populations in several health-related traits. The issue of reproducibility between the ChatGPT prompt runs is of particular importance, which can be a major limitation in health care practice.

39Medico-legal and accountability issues in the case of medical errors caused by ChatGPT application should be carefully considered. Importantly, the current LLMs including ChatGPT are unable to comprehend the complexity of biologic systems, which is an important concept needed in health care decisions and research. The concerns regarding data governance, health care cybersecurity, and data privacy should draw specific attention in the discussion regarding the utility of LLMs in health care.

40Other issues accompanying ChatGPT applications in health care include the lack of personal and emotional perspectives needed in health care delivery and research. However, ChatGPT emulation of empathetic responses was reported in a preprint in the context of hepatic disease. Additionally, the issue of devaluing the function of the human brain should not be overlooked; therefore, stressing the indispensable human role in health care practice and research is important to address any psychologic, economic, and social consequences that could accompany the application of LLM tools in health care settings.

41In the area of health care education, ChatGPT appears to have a massive transformative potential. The need to rethink and revise the current assessment tools in health care education comes in light of ChatGPT’s ability to pass reputable exams (e.g., USMLE) and possibility of ChatGPT misuse, which would result in academic dishonesty.

42Specifically, in ophthalmology examination, Antaki et al. showed that ChatGPT currently performed at the level of an average first-year resident. Such a result highlights the need to focus on questions involving the assessment of critical and problem-based thinking. Additionally, the utility of ChatGPT in health care education can involve tailoring education based on the needs of the student with immediate feedback. Interestingly, a recent preprint by Benoit showed the promising potential of ChatGPT in rapidly crafting consistent realistic clinical vignettes of variable complexities that can be a valuable educational source with lower costs. Thus, ChatGPT can be useful in health care education including enhanced communication skills given proper academic mentoring. However, the copyright issues should be taken into account regarding the ChatGPT-generated clinical vignettes, aside from the issue of inaccurate references. Additionally, ChatGPT availability can be considered as a motivation in health care education based on the personalized interaction it provides, enabling powerful self-learning as well as its utility as an adjunct in group learning.

43Other limitations of ChatGPT use in health care education include the concern regarding the quality of training datasets that could result in biased content and inaccurate information limited to the period prior to the year 2021. Additionally, other concerns include the current inability of ChatGPT to handle images as well as its low performance in some topics (e.g., failure to pass a parasitology exam for Korean medical students), and the issue of possible plagiarism. Despite ChatGPT versatility in the context of academic education, the content of ChatGPT in research assignments was discouraged, being currently insufficient, biased, or misleading.

44As stated comprehensively in a commentary by van Dis et al., there is an urgent need to develop guidelines for ChatGPT use in scientific research, taking into account the issues of accountability, integrity, transparency, and honesty. Thus, the application of ChatGPT to advance academia and health care should be carried out ethically and responsibly, taking into account the potential risks and concerns it entails.

45More studies are needed to evaluate the content of LLMs including its potential impact to advance academia and science with a particular focus on health care settings. In academic writing, a question arises as to whether authors would prefer an AI-editor and an AI-reviewer considering the previous flaws in the editorial and peer review processes. A similar question would also arise in health care settings involving the personal preference of emotional support from health care providers, rather than the potential efficiency of AI-based systems.

46In health care education, more studies are needed to evaluate the potential impact of ChatGPT on the quality and efficiency of both educational content and assessment tools. ChatGPT utility to help in refining communication skills among health care students is another aspect that should be further explored as well as the applications of LLMs in the better achievement of the intended learning outcomes through personalized and instantaneous feedback for the students.

47The current review represents the first rapid and concise overview of ChatGPT utility in health care education, research, and practice. However, the results of the current review should be viewed carefully in light of several shortcomings that include: (1) the quality of the included records can be variable, compromising the generalizability of the results; (2) the exclusion of non-English records might have resulted in selection bias; (3) the exclusion of several records that could not be accessed could have resulted in missing relevant data despite being small in number; (4) the inclusion of preprints that have not been peer reviewed but might also compromise the generalizability of the results; (5) the swift growth of literature addressing ChatGPT applications/risks mandate the need for further studies and reviews considering that the search in this review was concluded on 16 February 2023; and (6) this systematic review was based on the screening and interpretation of a single author, which may limit the interpretability of the results; therefore, future systematic reviews should consider collaborative work to improve the quality and credibility of the review results.

48The imminent dominant use of LLM technology including the widespread use of ChatGPT in health care education, research, and practice is inevitable. Considering the valid concerns raised regarding its potential misuse, appropriate guidelines and regulations are urgently needed with the engagement of all stakeholders involved to ensure the safe and responsible use of ChatGPT powers. The proactive embrace of LLM technologies with careful consideration of the possible ethical and legal issues can limit the potential future complications. If properly implemented, ChatGPT, among other LLMs, have the potential to expedite innovation in health care and can aid in promoting equity and diversity in research by overcoming language barriers. Therefore, a science-driven debate regarding the pros and cons of ChatGPT is strongly recommended and its possible benefits should be weighed with the possible risks of misleading results and fraudulent research.

49Based on the available evidence, health care professionals could be described as carefully enthusiastic regarding the huge potential of ChatGPT among other LLMs in clinical decision-making and optimizing the clinical workflow. “ChatGPT in the Loop: Humans in Charge” can be the proper motto to follow based on the intrinsic value of human knowledge and expertise in health care research and practice. An inspiring example of this motto could be drawn based on the relationship between the human character Cooper and the robotic character TARS from Christopher Nolan’s movie Interstellar.

50However, before its widespread adoption, the impact of ChatGPT from the health care perspective in a real-world setting should be conducted (e.g., using a risk-based approach). Based on the title of an important perspective article “AI in the hands of imperfect users” by Kostick-Quenet and Gerke, the real-world impact of ChatGPT among other LLMs should be properly evaluated to prevent any negative impact of its potential misuse. The same innovative and revolutionary tool can be severely deleterious if used improperly. An example to illustrate such severe negative consequences of ChatGPT misuse can be based on Formula 1 racing, as follows. In the 2004 Formula 1 season, the Ferrari F2004 (the highly successful Formula 1 racing car) broke several Formula 1 records in the hands of Michael Schumacher, one of the most successful Formula 1 drivers of all time. However, in my own hands —as a humble researcher without expertise in Formula 1 driving— the same highly successful car would only break walls and be damaged beyond repair.