Вештачка интелигенција

A New Detector to Identify Scientific Papers Generated by Language Models

Summary

A new tool developed by researchers at the University of Kansas (KU) is providing chemistry with a new ally in the fight against the misuse of artificial intelligence in scientific journals. Heather Desaire, an analytical chemist at KU, and her […]

A New Detector to Identify Scientific Papers Generated by Language Models

A new tool developed by researchers at the University of Kansas (KU) is providing chemistry with a new ally in the fight against the misuse of artificial intelligence in scientific journals. Heather Desaire, an analytical chemist at KU, and her team have created a detector that claims to be 98-100% effective in identifying scientific papers generated by language models like ChatGPT. Their tool can assist scientific publishers in detecting and preventing the improper use of artificial intelligence in academic journals.

The KU detector was first introduced in June of this year when it was tested on scientific papers from the journal Science, demonstrating an accuracy of over 99% in recognizing text generated by ChatGPT. The team has now significantly expanded the scope of the tool by testing it on papers in the field of chemistry.

The detector was trained on 100 abstracts from 10 journals published by the American Chemical Society. They then asked ChatGPT to write similar paragraphs, and the results were impressive – the machine learning model accurately identified human-written passages 100% of the time, as well as those generated based on the report titles. When ChatGPT was trained on the abstracts, the recognition accuracy was 98%.

One of the researchers explains that the main motivation was to broaden the scope of journal selection and challenge more complex and diverse thematic starting points.

The detector passed a more rigorous test. It was put to the test with samples that were not used in training but were of the same nature as the training data – in this case, 150 abstracts from three other chemical journals that were not part of the original training. An improved version of ChatGPT was also used to enhance the generated artificial intelligence text. The detector successfully classified the new text with an accuracy ranging from 92% to 98% for the three journals.

This tool could be helpful in detecting suspicious papers, which would then be investigated by reviewers or journal editors to determine if the work is fraudulent. Experts in scientific research integrity emphasize the significance of this tool.

Elisabeth Bik, a microbiologist and scientific integrity consultant in the US, fully supports this study. She believes it is a valuable addition that can help journal editors review computer-generated submissions, similar to the use of plagiarism detection software.

Saniat (John) Sohrawardi, a Ph.D. candidate at the Rochester Institute of Technology in New York working on “deep fakes” detection, has some reservations. “No journal or scientific gathering should rely solely on this tool as a basis for rejecting a paper,” he notes. “However, I believe their work has merit as a preliminary observation if it is sufficiently effective and resource-accessible, but it must be emphasized that it should not be used as definitive proof for rejecting the work.”

Concerns about Exaggeration

Nevertheless, several experts in the field are skeptical about claims of high reliability in artificial intelligence detectors. Debby Cotton, the Director of Academic Practice at the University of Plymouth Marjon in the UK, highlights that previous tests of such detectors have shown high accuracy but rarely confirmed this when tested on a larger scale.

Cotton, who authored a recent study examining the use of artificial intelligence in higher education, states that in the case of this latest tool, the model was specifically trained for a narrow range, making it better than most others. However, she suggests it is usually very easy to bypass such detectors with superficial human correction, even mentioning a new service called Undetectable AI that helps authors conceal the use of artificial intelligence sources in their work.

Desaire now wants to determine the extent to which ChatGPT has penetrated research enterprise. “We have a detector that is useful for finding ChatGPT contributions to scientific writing, so the next step would be to apply it to academic scientific writing and see,” she says.

“And I don’t think anyone really knows how much ChatGPT is contributing to the scientific literature – is it zero, 20%?” Desaire ponders. She highlights that the goal of her future research will be to determine how prevalent the improper use of ChatGPT is in scientific literature.

FAQ: