Scientists have successfully generated an entire research paper in less than an hour with the help of ChatGPT, an artificial intelligence (AI) tool that can comprehend and produce human-like text. While the paper was well-written and insightful, the researchers highlight the need to address various challenges before the tool can truly be considered helpful.
The aim of the study was to explore ChatGPT’s abilities as a research co-pilot and initiate a discussion about its advantages and limitations, according to Roy Kishony, a biologist and data scientist at the Technion — Israel Institute of Technology in Haifa. Kishony emphasizes the necessity to find ways to harness the benefits of AI while minimizing the drawbacks.
Kishony and his student, Tal Ifargan, a data scientist also from Technion, obtained a publicly available data set from the US Centers for Disease Control and Prevention’s Behavioral Risk Factor Surveillance System. This database contains health-related telephone survey information from over 250,000 individuals, including their diabetes status, fruit and vegetable consumption, and physical activity.
The researchers tasked ChatGPT with generating code that could help them uncover patterns in the data for further analysis. Initially, the chatbot produced faulty and ineffective code. However, after the scientists conveyed the error messages and requested correction, ChatGPT eventually generated usable code to explore the data set.
With a more structured data set in hand, Kishony and Ifargan turned to ChatGPT for assistance in defining a study goal. The tool suggested an investigation into how physical activity and diet impact diabetes risk. ChatGPT then provided additional code and presented the results: increased consumption of fruits and vegetables, along with regular exercise, is associated with a lower risk of diabetes. The researchers further prompted ChatGPT to summarize the key findings in a table and write the complete results section. Step by step, they instructed ChatGPT to compose the abstract, introduction, methods, and discussion sections of a manuscript. Lastly, they had ChatGPT refine the text. Kishony explains, We composed [the paper] from the output of many prompts. Every step is building on the products of the previous steps.
Despite ChatGPT successfully generating a coherent manuscript with sound data analysis, Kishony acknowledges that the paper was far from flawless. A major concern was ChatGPT’s tendency to fill in gaps by fabricating information, a phenomenon known as hallucination. For example, it created false citations and inaccurate content. Tom Hope, a computer scientist at the Hebrew University of Jerusalem, points out that the paper claims to address a gap in the literature, a phrase commonly used in research papers, but inappropriate in this context. Hope explains that the finding is not something that’s going to surprise any medical experts and is far from being innovative.
Kishony also worries that such tools could enable researchers to engage in dishonest practices like P-hacking, wherein scientists test multiple hypotheses on a data set but only report statistically significant results.
Furthermore, Kishony raises concerns over the potential flood of low-quality papers inundating journals due to the ease of producing papers using generative AI tools. He argues that his data-to-paper approach, with human oversight involved at every stage, can ensure that researchers can easily understand, verify, and replicate the methods and findings.
Vitomir Kovanović, an AI technology developer for education at the University of South Australia, emphasizes the need for greater transparency regarding the use of AI tools in research papers. Without this transparency, assessing the accuracy of a study’s findings becomes challenging. He states, We will likely need to do more in the future if producing fake papers will be so easy.
Shantanu Singh, a computational biologist at the Broad Institute of MIT and Harvard, suggests that while generative AI tools have the potential to expedite certain time-consuming research tasks such as writing summaries and generating code, writing entire papers using this technology may not be a practical approach at present. Singh points out that detecting hallucinations and biases is challenging for researchers.
In conclusion, the successful generation of a research paper with ChatGPT highlights the potential for AI tools to accelerate the research process. Nevertheless, significant obstacles must be overcome, such as hallucinations and potential ethical concerns. Researchers argue that human oversight and transparency are crucial to ensure the integrity and quality of scientific publications.
