AI Insights: Bot Time?
By Beth W. Orenstein
Radiology Today  
Vol. 24 No. 3 P. 6

Author Uses AI to Write Research Paper for Radiology  

The author of a peer-reviewed research paper about text-generated AI models published in Radiology online in early February is a human. The author, Som Biswas, MD, of the University of Tennessee Health Science Center College of Medicine in Memphis, states that upfront. However, in the introduction, Biswas also states that he used ChatGPT (Generative Pre-trained Transformer), an AI program developed by OpenAI, to compose several sections of the paper. (OpenAI is a private company backed by Microsoft Corp.)

Knowing that ChatGPT could compose short paragraphs, Biswas decided to test its usefulness in medical writing. He asked the AI tool four prompts and edited its answers as the body of the paper. The questions were: “What is ChatGPT? What is natural language processing? Use of ChatGPT in medical writing? Will ChatGPT replace [the] human medical writer?” He manually edited the final section, which discusses questions about ethics, copyright, and accuracy when using AI models.

Biswas’ conclusion: “Overall, while AI-powered language models like ChatGPT are powerful tools that could assist human writers in some tasks, as we move forward with adopting this technology, it will be important to consider the limitations and potential risks of using AI in the writing process.”

In an accompanying editorial, Felipe C. Kitamura, MD, PhD, of Universidad Federal de São Paulo, outlines potential roles in medicine for large language models (LLM), including creating drafts for clinical trial protocols, study reports, regulatory documents and patient materials, and translating medical information into a number of languages and laymen’s terms.

Biswas does not believe that ChatGPT or any other LLM could write a research paper entirely on its own. In an interview after his study’s publication, he says, ChatGPT could not even write a paper as “simple” as his. “My paper was written in parts by ChatGPT, [but] then I had to edit them, join them, and form a complete paper,” he says. If the research were more complex, such as a case-controlled or randomized-controlled study, Biswas doubts “it would be possible for ChatGPT to write up the whole paper correctly—at least currently.”  

Reduced Burden of Writing
The best role for a program such as ChatGPT would be to reduce the burden of writing/typing the labor-intensive clerical sections of a paper, including the introduction, abstract, summary, and highlights, Biswas says. The core of the paper, conceptualization, design, data collection, statistics, materials, methods, data analysis, results, and conclusions, “should be dealt with by humans,” he adds.

Biswas also sees ChatGPT interacting with patients. It could be used to create chatbots that patients use to book appointments and search nearby medical facilities for the services they need. Chatbots could even answer basic queries about diseases as long as the dataset fed to it is medically accurate, he says. Yet another possibility: assisting in research and training of residents—“of course, under the guidance of human mentors.”

Could ChatGPT and the like be helpful in simplifying radiology reports for patients? “This needs to be seen,” Biswas says. “Rigorous testing is needed before any medical application of ChatGPT.” However, he says, “there is potential, provided the dataset is accurate and someone (human) is responsible for it.”

Kitamura says anyone could ask a program like Chat- GPT to write a scientific article out of nothing. “It will do it,” he says, “and the result will sound like real science,” even if the content is fabricated. It would be better, he says, for researchers to ask a program like ChatGPT to write specific parts of their paper, including information about their real experiment in the prompt. The human author would have to “review the output to make sure the end result is what the researcher expected,” he says. That would be “a legitimate use, in my opinion.”

Kitamura, who is also the head of Applied Innovation and AI at Dasa, a large health care system in Brazil, says LLMs may be useful for creating drafts for clinical trial protocols, regulatory documents, and translating medical information into a number of different languages. All that could be done in real-time, he says.

Not Without Controversy
Biswas and Kitamura recognize that the use of AI in helping to draft research papers or any other materials is controversial. When used to help write research papers, “it can be difficult to determine the extent to which the large language model has influenced the results and conclusion of the paper,” Biswas says. “LLM generates text based on knowledge gained from massive datasets. If this dataset is biased, then it can affect the results and conclusions of the paper.”

The information that programs like Chat- GPT create “look a lot like real,” Kitamura says. “This is a significant concern if someone naively uses those tools.”

Both agree it is crucial for the author of a paper that uses LLM to state it upfront, as Biswas did in his paper for Radiology. “Human readers need to stay informed about the source of what they’re reading,” Biswas says. And, just stating that ChatGPT was used does not exonerate the author from writing responsibly and factually correct content, he notes. “Rather, it makes him even more responsible to carefully edit/curate the content once it’s generated by LLMs. As I said before, LLMs are only to reduce the labor of writing/typing.”

Several prominent publications, such as the JAMA journals and Science, have announced that they will not accept papers written with any help from AI models. Biswas believes that the blanket prohibition is unfair. “ChatGPT is just a tool to help writers (like Grammarly) produce work more efficiently and faster,” he says.

Supportive Colleagues
While Kitamura understands the journals’ prohibition on the use of AI to help write scientific articles, he does not believe it’s practical. “There is no way we can enforce that, since no tool can accurately tell if a text was written by AI or by a human being, at least for now,” he says. “I favor humans being in charge of reasoning and developing original ideas. These human authors could benefit from using AI text generators to expedite the writing process, provided they use AI as a tool. I disagree with using AI to reason or replace human judgment, at least with today’s tools.”

Biswas says the reaction to his Radiology paper was largely positive. “My colleagues were happy and proud of me, and that was what I expected,” he says. AI has come a long way in recent years, and its future role in radiology is going to continue to be “very exciting.”

Kitamura says most of his colleagues are impressed with ChatGPT, although some are afraid of it. After his editorial was published, he found his colleagues supportive of his view, as well. Only one friend thought the world was overreacting. “This friend thinks ChatGPT is not a big deal,” he says. 

— Beth W. Orenstein of Northampton, Pennsylvania, is a human freelance writer and regular contributor to Radiology Today. No AI was used in the writing of this article—unless spellcheck counts.