Paper mills may influence what AI systems tell users

Research on retracted papers highlights how problems in scientific publishing may extend to artificial intelligence responses, ERR reports. Anton Sokolov, a doctoral student at Tallinn University of Technology and researcher at the Tyche Institute, studies trust infrastructure, including research papers removed from databases.

Sokolov said the number of retracted studies shows how many problematic papers have been detected, not how many remain in scientific literature. Retraction Watch’s public database lists about 61,000 retractions from 2010 to 2025. About two papers per 1,000 are now retracted each year, he said, describing that figure as a measure of “cleanup” rather than the amount of problematic research still in circulation.

For his research, Sokolov compiled nearly 15,000 retraction records focused on computer science. Direct authorship-related misconduct accounted for less than 2% of the cases. But that does not mean authorship is rarely sold: publishers can usually document only what they are able to prove, while sales typically take place in private conversations and leave no clear trace in the published paper.

Retraction notices more often cite problems with peer review, which Sokolov said are mentioned in about seven out of 10 notices. A single notice may list several reasons. Computer-generated text is the most common reason for retractions related to computer science and artificial intelligence, accounting for about one-third of cases, he said.

The article follows a Novaator experiment published in early September, which found that co-authorship on a scientific paper could be bought through the paper-mill black market. The experiment took a few WhatsApp messages, 20 days and 800 US dollars to secure authorship of a paper ready for publication.