What does Retraction Watch's "computer-generated content" reason capture? A metadata analysis of 9,161 retractions
Year: 2026
Type: preprint
Abstract: Retraction Watch records a retraction reason for computer-aided or computer-generated content, and that reason is increasingly read as evidence about generative artificial intelligence in research. We asked what the reason captures when it is compared against the rest of the retraction record. From a snapshot of 71,448 records we retained the 66,013 classified as retractions and isolated the 9,161 (13.9%) carrying this reason. The category was absent before 2010, grew to 41.8% of retractions recorded for 2023 (95% CI 40.9–42.6), and one publisher accounts for 70.0%. No meaningful difference in median time to retraction was detected (difference 3 days, cluster-robust 95% CI − 17 to 43). Crude timing appeared far less dispersed (mean absolute deviation 268 versus 769 days, ratio 0.35), but this proved compositional: across 5,119 pairs matched exactly on retraction year, publication year and publisher the ratio moved to 0.97, leaving a residual of 5 days, and removing the dominant publisher from both cohorts reversed it. A paper-mill flag co-occurred in 71.1% of the cohort against 9.1% elsewhere; stratification attenuated a crude risk ratio of 7.78 to 1.43 (cluster-robust 95% CI 1.36–1.52) with extreme heterogeneity (Cochran Q = 248.2 on 7 degrees of freedom), so no common association holds. A crude post-retraction citation deficit of − 9.5 percentage points attenuated to − 1.3 (95% CI − 3.08 to 0.51) once pairs were matched on citations accumulated through the retraction year and on publisher. Both indicators come from one Reason field, so these are co-coding patterns and cannot establish what the articles contain.
Source: Research Square
Author: Jonas Mandalunes
Institution: AMA Computer University
Language: English
Subfield: Education
Field: Social Sciences
Domain: Social Sciences
Open Access status: green
OpenAlex ID: W7204457624