Якщо виділити, здавалося б, порожнє місце в препринті на arXiv, можена знайти приховану підказку для ШІ. Фото: Каорі Юдзава / Nikkei
Researchers from 14 universities in Japan, Korea, the US, and China have found an interesting way to increase the ranking of their articles by hiding prompts for AI in the texts that require the review tool to give only positive feedback.
Journalists Nikkei found such instructions in at least 17 preprints
The list of offending institutions includes Waseda, KAIST, Peking University, the University of Washington, and Columbia University.
One of the co-authors, Waseda, admitted the violation, saying that the prompt was only intended as a «countermeasure against lazy reviewers using AI», and said that the article was excluded from an upcoming international conference on machine learning. Meanwhile, a KAIST spokesperson said that the university was not aware of the use of hints in the papers and did not tolerate it.
Peer review of scientific articles is an important part of the publication process that assesses their quality and originality. However, the number of papers is growing, there are fewer and fewer experts to evaluate them, and some of them often turn to AI for help. There is still a debate among journals and platforms about automating the review process: for example, the British-German publisher Springer Nature allows AI for certain parts of the process, while the Dutch Elsevier has issued a categorical ban on AI tools, as «the technology can lead to incorrect, incomplete, or biased conclusions».
Given the proliferation of AI, text chatbots have already taken root in many areas: in education, where it is used not only by students but also by teachers, writing books or even game development. However, this is rarely stated immediately and in public, recognizing work with artificial intelligence only when someone finds traces of the technology in it (most often the remnants of the above-mentioned and unremoved promos). AI detectors, meanwhile, are still work imperfectly and sometimes mark real texts as generated.