AI tool achieves 94% accuracy in telling apart fake from real research papers

New Delhi: Researchers have developed a tool that could tell apart an original research article from one created by AI-chatbots, including ChatGPT.

In a set of 300 fake and real scientific papers, the AI-based tool, named ‘xFakeSci’, detected up to 94 per cent of the fake ones.

This was nearly twice the success rate seen among the more common data-mining techniques, the authors from the State University of New York, US, and Hefei University of Technology, China, said.

“… we introduce xFakeSci, a novel learning algorithm, that is capable of distinguishing ChatGPT-generated articles from publications produced by scientists,” they wrote in the study published in the journal Scientific Reports.

For developing the AI-based algorithm, the researchers developed two distinct datasets. One of them contained almost 4,000 scientific articles taken from PubMed, an open database housing biomedical and life sciences research papers and maintained by the US National Institutes of Health.

The other consisted of 300 fake articles, which the researchers created using ChatGPT.

“I tried to use exact same keywords that I used to extract the literature from the PubMed database, so we would have a common basis of comparison. My intuition was that there must be a pattern exhibited in the fake world versus the actual world, but I had no idea what this pattern was,” the study’s co-author Ahmed Abdeen Hamed, a visiting research fellow at the State University of New York, said.

Of the 300 fake articles, 100 each were related to the medical conditions Alzheimer’s disease, cancer, and depression. Each of the 100 included 50 chatbot-created articles and 50 authentic abstracts taken from PubMed.

The xFakeSci algorithm was trained on the first dataset containing scientific papers and then tested for its performance on the second one.

“The xFakeSci algorithm achieved (accuracy) scores ranging from 80 to 94 per cent, outperforming common data mining algorithms, which scored (accuracy) values between 38 and 52 per cent,” the authors wrote.

xFakeSci was programmed to analyse two major features in the fake papers, according to the authors.

One was the numbers of bigrams, which are two words commonly appearing together such as ‘climate change’, ‘clinical trials’ or ‘biomedical literature’. The second was how those bigrams are linked to other words and concepts in the text, they said.

“The first striking thing was that the number of bigrams were very few in the fake world, but in the real world, the bigrams were much more rich. Also, in the fake world, despite the fact that were very few bigrams, they were so connected to everything else,” Hamed said.

The authors proposed that the writing style adopted by an AI is different from that of a human researcher because the two do not have the same goals while producing a piece on a given topic.

“Because ChatGPT is still limited in its knowledge, it tries to convince you by using the most significant words,” Hamed said.

“It is not the job of a scientist to make a convincing argument to you. A real research paper reports honestly about what happened during an experiment and the method used. ChatGPT is about depth on a single point, while real science is about breadth,” Hamed said.