Presentation (CTI)

OBJECTIVES: Meta-analyses (MAs) are central to evidence synthesis, making their accuracy critical. However, it remains unclear how frequently retractions are due to computer-aided or computer-generated content, and to what extent these cases overlap with issues such as data fabrication, plagiarism, or paper mill involvement. We analysed the Retraction Watch database to investigate any emergent trends.
METHODS: Using the Retraction Watch database, retracted MAs from 2010 onwards citing “computer-aided content or computer-generated content” were identified via keyword filtering. Retraction reasons and temporal patterns related to publication and retraction were assessed using graphical visualisation methods. Additionally, published retraction notices were reviewed for explicit mention or suspicion of Artificial Intelligence (AI) involvement.
RESULTS: 9121 articles were retracted for “computer-aided or computer-generated content” including 114 MAs (1.25%). In 84% percent of these MAs paper mill involvement was noted as a co-retraction reason; other co-reasons concerned references, data, results, or conclusions. Published retraction notices for 99% MAs also mentioned “incoherent, meaningless and/or irrelevant content” and “inappropriate citations.” Whilst explicit AI usage was not reported, the reasons cited seem to align with generative AI misuse. Retractions increased sharply from May 2023, plateauing by April 2024, coinciding with the public release of common large language models. Most retractions occurred via mass retraction events, with publishers investigating specific Special Issues of various journals. The mean time from publication to retraction was 519 days.
CONCLUSIONS: Retractions of MAs for computer- aided or computer-generated content rose sharply with generative AI’s emergence, with the journal’s published retraction notices highlighting features common with AI-generated text, such as incoherent content or inappropriate citations. Although incidence of these retractions has decreased more recently, it remains unclear whether this reflects improved pre-publication detection, or more sophisticated AI usage. These findings underscore ongoing concerns about computer-generated content in retracted MAs in the post-generative AI era.