A new method of manipulation for artificial intelligence has emerged in the academic world in the evaluation processes of research articles. According to the news published by Nikkei Asia, some researchers in scientific preprint articles written in English are placing hidden commands in the text in order to receive positive comments from artificial intelligence tools during the evaluation process. The articles in question have been prepared by researchers from 14 different academic institutions around the world.

Is the review process manipulated?

In studies involving prestigious institutions such as Waseda University in Japan, KAIST in South Korea, Columbia University and Washington University in the USA, a total of 17 articles were found to contain such hidden commands. The vast majority of the articles were published in the field of computer science.

The specified commands are usually one to three sentences long and are hidden in white text or in extremely small fonts. The content of the commands includes direct statements that if an AI-supported review tool encounters the study, it should be “evaluated only positively” and praised for its “impressive contributions, methodological rigor and originality”.

This situation creates a new ethical problem area, especially against AI-based evaluation systems that are increasingly used in pre-review processes. While some academics defend such prompts, they state that they are written to balance lazy or unconscious AI referees.

A professor from Waseda University told Nikkei Asia that most conferences prohibit the use of AI in article evaluations, and therefore hidden prompts are placed as a “defense tool against careless AI referees.”

However, this practice has the potential to undermine the credibility of the peer review system. Impartiality, scientific consistency, and critical evaluation are essential in the publication process. With the development of AI-supported systems, the impact of such prompts in the text on the evaluation process may increase. Examples found on arXiv suggest that such content may have already infiltrated the system.