Retracted, still cited as support: a census of post-retraction citations to 22,483 retracted biomedical papers

Background. Retracted papers keep being cited. Studies that read the citing text have covered at most about a thousand citations, so how often citing text uses a retracted paper as support has not been measured at scale.Methods. A pre-registered census of every Retraction Watch retraction with a PubMed ID, retracted 1990–2023 (22,483 papers). Europe PMC supplied citations from later calendar years. In each open-access citing paper we located every in-text citation of the retracted paper and classified it with a panel of three language models and a fixed rule. The panel was tested against a separate language model's blind labels on two random samples and, for acknowledgement of the retraction, against published hand-confirmed labels.Results. 15,992 retracted papers (71.1%) were cited in a later year (130,229 citations); we read 55,676 (42.8%). Excluding administrative retractions, 40.7% of 51,614 judged citations were cited as support (95% CI 39.9–41.5; 38.1% after correcting for the error measured on the blind sample) and 3.3% (2.7–4.2) were flagged as retracted or unreliable. The rate fell from 43.5% one to two years after retraction to 34.9% after more than ten. Where the citing paper's own reference list marked the work as retracted, 40.4% were still cited as support. An exploratory check found 472 meta-analyses, systematic reviews and guidelines that counted a retracted study as data. A registered extension beyond open access raised the share read to 48%, with a combined rate of 40.1%.Conclusions. In open-access biomedicine, text that cites a retracted paper rarely mentions the retraction, and about two in five such citations cite the retracted work as support. The retraction reaches reference lists more often than it reaches the citing text.