A US university team finds a new way to check whether a paper was authored by AI with 99% accuracy

News on June 8, according to Japan's "Asahi Shimbun" report, a research team at the University of Kansas in the United States recently discovered a new testing method that can accurately judge an article Whether the paper is created by ChatGPT, its accuracy can reach 99%.

Currently, users only need to enter a set of data and add the comment "Please use this data to write an academic paper." ChatGPT can create a paper that can be fake in tens of seconds, which makes many The author of the paper has used ChatGPT partially or completely to write the paper.

However, some academic journals and research institutions have warned that if these authors do not clearly state that they use ChatGPT when submitting, it will constitute plagiarism or other fraud.

For this reason, there has been progress in the development of tools to detect whether an article was generated by AI, but so far there has been no detection tool specifically for academic papers.

The research team at the University of Kansas focused on "paragraph complexity" and "common phrases" to train a new AI to check whether a paper was generated by ChatGPT.

According to reports, the research team selected 64 papers from the American academic journal "Science" within the six months to March 2023, and asked ChatGPT to generate abstracts of about 300-400 words for the papers. , 128 articles created by ChatGPT were finally generated and used to train this new detection AI model.

After training, the model can 100% identify the artificially created parts from the articles generated by AI, and the recognition accuracy of specific paragraphs in the text is 92%. Comprehensive The accuracy rate is 99%. The relevant findings have been published in the journal "Cell Reports Physical Science."

The research team found that ChatGPT rarely uses expressions such as "although... but... (but, however, although)". Instead, it likes to use words such as “others” and “researchers”.

On the other hand, humans are more accustomed to interweaving long sentences and short sentences, and the sentence structure is also different from ChatGPT's usual writing method.

IT House previously reported that a professor at Texas A&M University in the United States once entered student papers into ChatGPT to reversely verify whether these papers were also generated by ChatGPT. He also achieved good results and directly As a result, half the class failed the class.

IT Home attaches the paper link: Click here to go

The above is the detailed content of A US university team finds a new way to check whether a paper was authored by AI with 99% accuracy. For more information, please follow other related articles on the PHP Chinese website!