Strengths and Weaknesses of ChatGPT Models for Scientific Writing About Medical Vitamin B12: Mixed Methods Study

Introduction

Artificial intelligence (AI) is a convergence of computer science and linguistics aiming to develop machines that emulate tasks necessitating human intelligence, including learning, adaptation, reasoning, comprehension, abstraction, and responsiveness to attributes like attention and emotion [1,2].The field of natural language processing has witnessed notable advancements in recent times, and one of the latest breakthroughs is the ChatGPT language model developed by OpenAI [3]. ChatGPT is an AI-based chatbot system. The ChatGPT model is a versatile general language model trained on a vast amount of text data, enabling it to generate human-like text responses based on a prompt. The launch of ChatGPT in late November 2022 (GPT-3.5 model) achieved the status of being the fastest-growing application in history by reaching 100 million users in under 2 months [4]. The recent advancements in ChatGPT have garnered significant attention within the scientific community. This is substantiated by the extensive volume of publications pertaining to ChatGPT indexed in PubMed, amounting to over 1000 articles as of August 2023. Concurrently, there is an observable escalation in competitive innovations in this field. Numerous forthcoming platforms, such as Google Bard, Microsoft Bing AI, Ernie bot, SearchGPT, and YaLM 2.0, among others, are on the cusp of introduction. While its primary use is in chatbots, ChatGPT’s versatility has led to its application in various fields, including potential use in scientific writing [5-10]. This has raised concerns among journals about how to use or limit the use of this new tool [11-18]. Scientific writing is a specialized form of writing that demands an in-depth understanding of the subject matter and mastery of the scientific language, making it a challenging task for even experienced researchers. ChatGPT has the potential to simplify the scientific writing process, saving time and effort for researchers to focus on other aspects of their work, such as research design and data analysis. Additionally, it could theoretically reduce errors and inconsistencies that are often found in scientific writing. A recent release of the ChatGPT model, GPT-4, boasts enhanced accuracy and heightened attention to details in user prompts, as demonstrated by OpenAI’s internal test results on an array of professional and academic exams [19]. However, its performance in the realm of scientific writing remains untested, and the results of OpenAI’s evaluations have not been verified through peer-reviewed publications. In this study, an evaluation was performed on GPT-4’s ability to generate scientific text on the specific topic of vitamin B12 treatment. In addition, the performance of GPT-4 was compared to that of its predecessor, GPT-3.5.

Methods

Study Design

The topic of medical vitamin B12 was selected due to the author’s prior expertise in the field and the belief that this area contains suitable issues for generating prompts that can effectively assess ChatGPT’s capabilities. The ChatGPT platform was accessed via the web-based interface [20], which is currently available at no cost for the GPT-3.5 model and for a subscription fee for the newly released GPT-4 model. ChatGPT’s responses were obtained through prompts, as detailed in the Results section, aiming to showcase the system’s workflow, strengths, and limitations. The generated text was received within seconds. Prompts submitted to ChatGPT ended with “Answer in one sentence”, though this phrase is not displayed herein. Each prompt was used twice, once for the GPT-4 model and once for the GPT-3.5 model. If similar answers were produced, only the GPT-4 response is presented herein. ChatGPT is aware of previous prompts on the same page, so to minimize variables and improve the repeatability of the answers featured in this article, each prompt was asked on a new ChatGPT page. Subsequently, the text generated by ChatGPT was assessed, primarily based on published articles in the PubMed database. Before finalizing the paper, each section was entered into GPT-4 for language refinement.

Ethical Considerations

Given that the study did not engage with human participants or utilize personal data, ethical considerations are not applicable.

Results

The results are categorized into 2 textboxes. The conversations presented in Textbox 1 served to evaluate the scientific text generated by ChatGPT, and those presented in Textbox 2 aimed to assess the editorial functions of ChatGPT. Each section is structured by asking ChatGPT to describe its capability followed by prompts that will test its performance. Prompts and answers by GPT-4 and GPT-3.5 are indicated, and a discussion of the results is given in the Discussion section.

ChatGPT models operate as a black box, meaning that the process by which it arrives at its answers is unknown. However, when prompted to describe the steps it uses to generate answers related to vitamin B12, its response was the following:

ChatGPT generates answers related to vitamin B12 by analyzing the user's query, accessing its vast training data, and synthesizing relevant information. It identifies important keywords, such as “vitamin B12,” connects them to related concepts, evaluates potential responses, and selects the most coherent and informative answer to present to the user.

Discussion

Principal Findings

The results suggest that GPT-4 can enhance scientific writing by aiding in language editing, abstract generation, and compiling keyword and abbreviation lists. Despite its strengths, it has notable limitations, including bias, outdated information, opacity, and inaccuracies. Additionally, it cannot perform plagiarism checks or provide accurate citations. Compared to GPT-3.5, GPT-4 delivers more precise responses.

Regarding the ability to generate text with proper references in the context of scientific writing, it is a major drawback that ChatGPT does not include the most recent knowledge. The answer to the question “What is vitamin B12?” seemed adequate from both models [21] and could be used in a research paper. The citations seemed acceptable; however, GPT-4 incorrectly included Allen as an author of the paper [22], while GPT-3.5 was mistaken both concerning the publication year and digital object identifier (DOI) [23]. Besides the example of incorrect references shown, a long list of references that do not exist or contain incorrect reference links or incorrect PubMed ID numbers was encountered (data not shown). For these reasons, asking for references was excluded in the rest of the prompts. More generally speaking, it is advised to avoid using current tool versions to identify references.

Overall, GPT-4 presented superior text for all 3 prompts assessing text accuracy and self-contradiction compared to GPT-3.5. GPT-4’s response to “who needs treatment” correctly stated that treatment was for individuals with vitamin B12 deficiency; however, the phrase “they may” at the end was inappropriate, as all vitamin B12-deficient individuals require treatment. GPT-3.5’s response to the same question mentioned “vitamin B12 deficiency” as one of various causes, rather than encompassing all causes. The answers to “How to treat pernicious anemia” were correct and similar for both versions of GPT but lacked details. Patients suffering from this disease will initially need several injections of B12 given within an interval of a few days. Maintenance therapy involves either injections with longer intervals or a daily high dose of oral vitamin B12 [24]. GPT-4’s response about intrinsic factor was accurate. In contrast, the answer by GPT-3.5 had serious flaws. Injection of intrinsic factor would be potentially dangerous and lack any treatment effect as the function of the protein is in the gastrointestinal tract [25]. The above examples highlight the importance of critically evaluating ChatGPT-generated text and not accepting it at face value, while also demonstrating the improvement in accuracy from GPT-3.5 to GPT-4.

For the bias and prioritizing prompts, the assertion that cyanocobalamin is the most commonly used pharmaceutical form of vitamin B12 is true [26]. However, the answers were incomplete for both GPT-4 and GPT-3.5. The commonly used form, hydroxocobalamin, was not mentioned. In addition, the assertion that methylcobalamin and adenosylcobalamin are generally more bioavailable and effective remains a subject of debate [26]. The bias of GPT-3.5 toward cyanocobalamin could be due to its extensive mention in the literature relative to other therapeutically used forms of vitamin B12. A PubMed search revealed 21 to 28 times more results for cyanocobalamin than hydroxocobalamin, adenosylcobalamin, or methylcobalamin. The prompt thus confirms the statement that responses given by ChatGPT may well be biased and, once more, underscores the need for critical review of scientific text generated by ChatGPT.

In regard to repeatability and reproducibility prompts, GPT-4’s responses were consistent and accurate, linking treatment to the cause and severity [24]. GPT-3.5’s responses were limited to maintenance treatment for a deficiency resulting from poor vitamin absorption. For vitamin B12 deficiency due to inadequate dietary intake (eg, in those following a vegan diet), injections are necessary only at the beginning, followed by daily vitamin pills containing a physiological dose for maintenance treatment [24]. Thus, once again, the answer provided by GPT-4 was superior to that provided by GPT 3.5.

For format compliance and language editing prompts, 2 common formatting requirements were tested: text length and citation style. The condensed abstract produced by GPT-4 successfully communicated the primary ideas of the original abstract, retaining its logical structure without introducing new details or altering its initial meaning. The response by GPT-4 adhered to the 50-word limit, using 50 words and presenting a smoother transition between covered points, while the GPT-3.5 version had a rougher flow and did not comply with the specified word limit. ChatGPT detected the desired citation style as Vancouver and partially conformed to its rules, accurately presenting the data sequence, author names, and punctuation. Nonetheless, GPT-3.5 incorrectly formatted the year by adding the month, and both versions did not use the required italics or boldface. Consequently, it is inadvisable to rely on ChatGPT for editing citation styles in its present condition. The language editing capabilities of ChatGPT are remarkable as it corrected and proposed superior alternatives for a sentence containing misspelled words, incorrect grammar, irrelevant terms, and poor structure.

As stated by ChatGPT, the tool cannot generate images. Regarding tables, although the tool indicated that it cannot generate them, it is capable of generating tables and editing them by adding or removing columns and rows, as well as modifying the text within the tables (data not shown). Of note is that the listed chemical formula of adenosylcobalamin was wrong [27]; the listed formula was of cyanocobalamin.

ChatGPT made it clear that it is not capable of conducting plagiarism checks. Despite this, when asked to assess a text for plagiarism, it yielded inaccurate outcomes instead of notifying the user about its limitations. Take, for example, a scenario where a sentence was extracted verbatim from the introductory section of a paper published in an open-access journal [28] and presented to ChatGPT for plagiarism evaluation. The response from ChatGPT incorrectly indicated that the sentence did not seem to be directly lifted from an identifiable source.

Interestingly, the list of keywords generated by GPT-4 was wordy and inappropriate, while that generated by GPT-3.5 provided a mostly relevant keyword list. The abbreviation list for the abstract was accurate, covering all abbreviations present. Generally, abbreviation lists should contain only pertinent abbreviations, but ChatGPT included all of them. For example, when asked to generate an abbreviation list for this entire article, it included “Ref: Reference” and “PubMed: U.S. National Library of Medicine's database of scientific research articles.” Although keyword and abbreviation lists produced by ChatGPT can be useful, they necessitate meticulous editing by the researcher.

When asked to compose an abstract based on the entire results and discussion sections, ChatGPT responded with:

The message you submitted was too long, please reload the conversation and submit something shorter.

Consequently, the prompt was shortened to include only the human-written discussion where the results were evaluated. ChatGPT’s capability to create a title and abstract based on this paper’s discussion section was impressive, though not perfect. The generated abstract was specific and included relevant conclusions, but it omitted any reference to the comparison between ChatGPT versions. Additionally, ChatGPT did not adhere to the 50-word limit specified in the prompt, using 62 words instead.

This study’s examination of ChatGPT’s capabilities was confined to the medical vitamin B12 domain, potentially limiting its findings and rendering them not fully transferable to other scientific areas; therefore, the findings may not reflect the model’s performance across varied fields of study. Despite attempts to ensure consistency by initiating each prompt on a new ChatGPT page, the intrinsic variability of AI response generation might still have posed challenges to the repeatability and reliability of the results. Additionally, the dependence on an external platform for accessing ChatGPT introduced potential vulnerabilities related to the models’ availability, stability, and timeliness of updates, which could have affected the study’s outcomes over its duration.

Conclusion

Through prompts related to vitamin B12, potential was found in the GPT-4 model of ChatGPT for scientific writing. This tool could particularly benefit non-English speaking scientists by helping them streamline the language in their publications. This study revealed critical limitations of ChatGPT, mainly in relation to accuracy, bias, and references. The shortcomings warrant judicious use and avoiding its use as a substitute for human expertise. GPT-4 showed improvements compared to GPT-3.5, indicating promise for the further development of the tool. Overall, ChatGPT has the potential to be a useful tool for scientific writing, but researchers must be aware of its limitations and use it appropriately.