Why Has the Number of Scientific Retractions Increased?

The number of scientific retractions has roughly doubled every decade since the year 2000. That single fact lands hard, but it tells you almost nothing useful on its own. Is science getting more corrupt, or is it getting better at cleaning house? Both of those stories fit the same number, and pulling them apart is exactly what Steen, Casadevall, and Fang set out to do. Their dataset includes two thousand and forty-seven retracted articles indexed in PubMed, spanning publications from nineteen seventy-two through twenty twelve. They hand-curated each one, separating retractions for fraud, which includes fabrication and falsification, from everything else grouped as error, including scientific mistakes, plagiarism, and duplicate publication. The tool they used to distinguish "more bad science" from "better detection" is elegantly simple: the clock. Time-to-retraction is the interval between when a paper is published and when the retraction notice appears. If that clock is running faster over time, the most natural interpretation is that journals and institutions are getting quicker at catching and pulling bad work, not that misconduct itself has surged. And the clock has changed dramatically. Across all two thousand and forty-seven articles, the mean time-to-retraction was about thirty-three months. But when you split the sample by publication date, something striking emerges. For the seven hundred and fourteen articles published between nineteen seventy-three and two thousand two, the average time-to-retraction was just under fifty months, which is over four years. For the one thousand three hundred and thirty-three articles published after two thousand two, it was under twenty-four months, roughly half. That difference carries a p-value below 0.0001. Think about what that halving means. Potentially flawed work that once sat in the literature for four years before being withdrawn is now being caught and pulled in about two. That's a real shift in how quickly the system responds. And it contributes directly to the spike in annual retraction counts — faster detection means more retractions accumulating in any given window of time, even if the underlying rate of bad papers hasn't changed. The paper also tested whether journal prestige drives faster retraction. High-impact journals have more readers, more scrutiny, and more people inclined to flag problems. There is a correlation — for misconduct specifically, the relationship between journal impact factor and time-to-retraction reaches statistical significance, with higher-impact journals retracting sooner. However, the effect is tiny. Impact factor explains roughly one percent of the variance in time-to-retraction. Prestige matters a little, but it doesn't explain the bulk of the change. Now here's where the story gets more complicated. Faster detection can explain why retractions are accumulating faster, but it can't explain why the raw number of retractable papers seems to have grown. For that, Steen and colleagues identified two additional mechanisms, and both of them change the shape of the problem in ways you wouldn't immediately predict. The first is definitional. The universe of behavior that qualifies as retractable has expanded. The earliest retraction in the PubMed sample was a paper published in nineteen seventy-three and retracted in nineteen seventy-seven. But the first article retracted specifically for plagiarism was published in nineteen seventy-nine. The first retraction for duplicate publication didn't appear until nineteen ninety. These aren't corner cases — they're entire categories of offense that simply weren't treated as grounds for retraction before. Plagiarism and duplicate publication now account for a meaningful share of modern retractions, but they were invisible in the earlier record not because no one was plagiarizing, but because the community hadn't yet decided that retraction was the remedy. When you widen the definition of what counts, you get more things counted. Some portion of the retraction surge is a change in the denominator, which is the set of papers that can be retracted, not just the numerator. The second mechanism is about who is getting retracted, and this is the most counterintuitive finding in the paper. You might assume that the rise in retractions is driven by serial bad actors — a handful of fraudsters with dozens of papers each inflating the counts. In the early era, from nineteen seventy-two to nineteen ninety-two, that was partially true. Authors with more than five retractions accounted for thirty-four point five percent of all retracted papers in that period. From nineteen ninety-three to twenty twelve, that same group accounted for only seventeen point eight percent, even though the raw number of their retractions rose from eighty-one to three hundred and twenty-three. The serial offenders got more prolific in absolute terms, but the field around them grew faster. What actually drove the recent surge is single-retraction authors. From nineteen seventy-two to nineteen ninety-two, forty-six percent of retracted papers were written by someone with just one retraction to their name. From nineteen ninety-three to twenty twelve, that fraction rose to sixty-three point one percent. The problem didn't concentrate — it spread. The two groups also differ in how long their papers survive before being caught. Papers by authors with more than five retractions took an average of just over fifty-two months to be retracted. Papers by authors with five or fewer retractions were pulled in about twenty-eight months on average. Serial offenders' work lingers longer in the literature. Among fraudulent papers retracted sixty months or more after publication, only ten point four percent were written by single-retraction authors. The long-delayed fraud is almost entirely a serial-offender problem. Single-retraction authors, by contrast, are disproportionately retracted for plagiarism, duplicate publication, and error, which are the newer, expanded categories. To estimate how many retractable papers might still be lurking undetected, Steen and colleagues built a simple correction model. The corrected eventual number of retractions equals the number already retracted divided by the cumulative probability that a retractable paper would be detected by that point in time. How many papers should we expect, given what fraction of retractable papers we think we've caught so far? The model suggests the current count substantially understates the eventual total, particularly for recent articles, which haven't had enough time to accumulate scrutiny. The paper is careful here. The apparent shortening of time-to-retraction for recent articles may partly be a measurement artifact. If many recently published papers that will eventually be retracted haven't been retracted yet, then the recent era looks faster than it actually is. That caveat doesn't erase the finding — the trend is real — but it means the full picture is still developing. Pull the threads together and you get a three-part explanation for why retractions have risen. Journals and institutions are retracting papers faster; the time-to-retraction clock has roughly halved since two thousand two. The scope of retractable offenses has expanded; plagiarism and duplicate publication are now recognized grounds for withdrawal in ways they weren't before. And the pool of retracting authors has broadened; the surge is driven by many authors with single retractions, not by a concentration of serial offenders. That third finding carries a note of genuine concern. The growth in single-retraction authors is consistent with lower barriers to publishing flawed work in the first place. More papers are entering the literature with problems that warrant eventual withdrawal, and more of those papers are coming from authors who presumably didn't intend to commit fraud. Error and plagiarism dominate in this group. The paper doesn't speculate on causes. What it does say is that the behavior of authors and institutions has changed in ways that can't be explained by serial misconduct alone. The optimistic reading is real and supported. Rising retraction counts can plausibly mean a more vigilant system is catching more errors. The correction mechanism is working faster. Offenses that once stayed in the literature for years are being removed in months. That's the self-correction engine running. But the engine is having to work harder because there's more to correct, and more of it is coming from a broader, more dispersed population of researchers than anyone would have predicted from looking at the early record alone. This lecture was created by ennepō. Go to https://ennepo.ai to Discover, Create and Follow the latest research in your field. Read when you can. Listen when you want to.