Can artificial intelligence help scientific research?
Can artificial intelligence help scientific research?
That’s a question Sentara Health doctors tackled in a recent series of scholarly publications.
With AI making inroads in a growing number of fields, the team of researchers see its use in academic research as “likely unavoidable.”
But according to them, AI still has big limitations when it comes to key research tasks. “It’s not ready for prime time by any means,” said Xian Qiao, MD, pulmonary and critical care medicine physician at Sentara.
Limitations of AI for scientific writing
Dr. Qiao and his colleagues tried using AI to write a series of scientific review articles. One issue they encountered? A tendency to invent information, which has been termed “hallucination.”
When the researchers used ChatGPT 4.0 to create references, for example, 16 percent did not actually exist. Another 54 percent were inaccurate or misattributed.
“The people who are champions of AI call it hallucinations,” said Thomas McCune, MD, nephrologist at Eastern Virginia Medical School (EVMS) and partner of Nephrology Associates of Tidewater. “Those of us who are looking at this and realizing the horror of it, we call it lies.”
Dr. Qiao (left) and Dr. McCune discuss their research.
Another big issue they found was poor analysis.
“ChatGPT 4.0 had a tendency to write broad, generalized statements without supporting facts,” wrote the team of researchers, who came from Sentara, EVMS, the Indiana University School of Medicine, and Mount Holyoke College.
They concluded that “it is not possible to write an accurate, well-informed, critical scientific review solely with ChatGPT.”
Utilizing AI as an assistant may be possible
Despite this, the team found that AI “could make parts of the writing process faster.”
For example, the AI-generated outline was “deemed acceptable with minor revisions.”
AI could also “be used as a source to overcome writer’s block” and be helpful for non-native English speakers, the team wrote.
Comparing three writing approaches
Dr. Qiao, Dr. McCune, and their colleagues, including Angela Toepp, PhD, Sentara’s Director of System Research, began wondering about AI’s usefulness for research while studying the long-term effects of Covid-19. Since the rollout of ChatGPT in late 2022, the tool has been used to ace the bar exam, compose articles, make music, and more.
“We started talking about AI, and how we can utilize it to look at different aspects of medical research,” Dr. Qiao said.
The team set out to compare scientific papers written with three approaches: human-only, AI-only, and a combination approach with both human and AI input. They used ChatGPT 4.0, the paid, premium version of ChatGPT.
But they encountered significant issues, from the extent of ChatGPT’s knowledge to the reliability of the tool.
The September 2021 knowledge cutoff for ChatGPT at the time meant the tool couldn’t write accurately about Covid-19. Much of the relevant scientific literature about the disease was published after that date. As a result, they abandoned the AI-only paper.
When they did use ChatGPT in the writing process, they encountered inaccuracies, plagiarism, and fabrications. This meant hours of fact-checking and editing. Because of this, the paper written with AI assistance required almost twice as much time to complete as the human-only paper.
AI will continue to develop and improve, Dr. Qiao said. But he foresees there will always be a need for a human gatekeeper. “At the end of the day, you still have to have a human component to all of this to review the product,” said Dr. Qiao.
By: Clancy McGilligan