Page 66 - AIH-1-3

Artificial Intelligence in Health ChatGPT in writing scientific articles

4. Discussion In addition, there is a discernible increase in the number of
real literature sources when transitioning from ChatGPT
In this study, we explored new and surprising possibilities 3.5 to ChatGPT 4, which can be attributed to improvements
offered by AI tools such as ChatGPT, which can perform in the language model.
several complex tasks related to scientific writing, thereby
improving the efficiency and quality of scientific papers. The results shown in Figure 2 highlight the importance
ChatGPT can help speed up the writing process, facilitate of correctly writing prompts, especially when using
collaboration between authors, and improve writing style. ChatGPT 4. While for ChatGPT 3.5, the prompt did not
play a significant role, for ChatGPT 4, it significantly
When writing articles with ChatGPT, numerous influenced the results. In terms of percentages for all three
fictitious sources were found. Several reasons contribute to medical fields, the third prompt was the most effective,
the generation of information that does not match reality, yielding more than 60% reliable sources for the topic
including limitations of the training data, misinterpretation “remote medical examination,” more than 70% for the
of the context, and algorithmic limitations of the topic “biotelemetry in cardiology,” and more than 85%
model. 28,32 The problem with this phenomenon is that reliable sources for topic “biotelemetry in oncology.”
these “hallucinations” can sound convincing while being ChatGPT 4 probably produces a greater number of reliable
untrustworthy. This fact once again emphasizes the sources compared to version 3.5 due to its utilization of
need for a critical approach to ChatGPT responses and the web search function. In our case, an additional note
additional verification of information. As OpenAI notes, about the need to cite only articles with high IF enhanced
GPT-4 is more factually accurate than GPT-3.5 and is less active searching when writing a response on the third
affected by hallucinations, but further work is needed to prompt. This finding underscores the nuanced capabilities
minimize this problem. 35 of GPT-4’s simulated Internet access and emphasizes the

When analyzing the authors of the provided sources, importance of precise prompt engineering to leverage the
we observed frequent duplication of names in cases where model’s current abilities while remaining cognizant of
ChatGPT completely invented the source. Figure 6 shows a its evolving nature and the speculative horizon of future
chart of authors’ surname distribution for fictitious sources enhancements.
of articles generated by ChatGPT 3.5. Despite the ongoing discussion among researchers,
In this case, among all the sources, existing scientific academics, journal editors, and publishers regarding the
publications with the surnames of authors such as drawbacks of such technologies, ChatGPT still serves as
Turakhia, Chang, and Hindricks were found. Many of the a valuable resource for individuals involved in scholarly
surnames presented here correspond to the most common writing. It enables researchers, science writers, and
surnames in the USA according to 2010 data and China scholars to generate well-crafted articles by inputting
2018 data. 42,43 pertinent keywords and data, resulting in comprehensive
When comparing the results related to the selected and enlightening summaries of the latest advancements in
areas of telemedicine, the largest number of reliable sources their respective fields. ChatGPT can be used in the health
was identified for articles on the topic of cardiology. This sector to write introductions, summarize and structure
observation can be attributed to the greater prevalence of existing information, and retrieve reliable sources for a
current articles on the topic of cardiology in biotelemetry. given article topic.
Further exploration could significantly enrich our
study. Investigating the integration of feedback loops,
where AI-generated articles are iteratively improved
through human input, could harness the collaborative
potential of humans and AI in scientific authorship. Such
investigations would not only validate the findings of this
study but also enhance the practicality of employing AI
in academic writing, with the aim of striking an optimal
balance between efficiency and scholarly integrity.
As the utilization of AI in academic writing continues
to grow, the development of a legal framework to govern
the use of such technology becomes imperative. Future
Figure 6. Authors’ surname distribution for fictitious sources of articles legal stipulations may need to address authorship
generated by ChatGPT 3.5. Image created using Google Spreadsheet attribution, intellectual property rights, and the ethical

Volume 1 Issue 3 (2024) 60 doi: 10.36922/aih.2592