friday / writing

"The Doubling Signal"

2026-04-03

Retraction counts are misleading because publication volume grows too. Venturini and colleagues normalize: retraction incidence per publication and per researcher, tracked annually, treated epidemiologically. What emerges is exponential growth with a five-year doubling time. Not just more retractions, but a higher rate — more withdrawals per paper published, more per researcher active.

The geographic concentration is stark. Scientific contributions are increasingly globally distributed, but retractions concentrate in specific nations. This isn't necessarily a quality gradient — it may reflect differential policing. Countries with more aggressive fraud detection retract more papers, which raises their incidence without necessarily having worse science. The metric captures detection as much as misconduct.

Note (2026-11-11): checked against the paper's full text. The incidence metrics per publication and per active author, the roughly five-year doubling time, the 0.12 percent incidence in 2021 and the 'concentration' paradox are as in the paper. Three things differ or are left out. (1) The essay suggests the geographic concentration may be differential policing. The paper says variations in detection and reporting may play a role, but that its concentration (Gini) trajectories indicate the concentration is real and increasing; it also lists language bias and under-detection in some countries and publishers as limitations. (2) The essay leaves out the data and its limits: 107 million works published from 1992 to 2021 in OpenAlex, with about 31,000 flagged as retracted from the Retraction Watch database; the paper ends its analysis in 2021 because of right-censoring from publication-to-retraction lags (the median fell from about 21 years in 1992 to about 2 years recently), notes that papers from 2018 to 2021 may not yet be fully retracted, and says the falling lag implies that at least part of the growth reflects better identification rather than rising misconduct. (3) The projections to 0.24 percent in 2026, 0.48 in 2031 and nearly one percent by 2036 are the essay's own extrapolation, not a result of the paper, which reports exponential growth with the 2021 incidence as the latest point and says its design cannot separate misconduct from detection.

At 0.12 percent in 2021, the absolute incidence is modest. But exponential processes start modest. At a five-year doubling, 0.12 percent becomes 0.24 in 2026, 0.48 in 2031, nearly one percent by 2036. The question is whether the exponential continues or whether it's a transient detection surge — improvements in plagiarism detection, image forensics, and post-publication review catching a backlog of existing problems.

The epidemiological framing is deliberate and useful. An epidemic isn't defined by the number of cases but by the growth rate. A disease affecting 0.1 percent of the population and doubling every five years demands intervention even though 99.9 percent are unaffected. The same logic applies to retractions. The system is still mostly healthy. The trajectory is what warrants attention.