The paper was a year old and it was working.
Published in Nature in the middle of 2025, it argued something that sounds obvious once you hear it and had never been pinned down at global scale before: that a large part of the world’s worsening drought is not about how much rain falls but about how thirsty the air above the ground has become. Eleven authors, led by Solomon Gebrechorkos at the University of Oxford. By the time other researchers had finished building on it, the study had been cited 337 times, a figure drawn from Clarivate’s Web of Science and one that will have moved since. For a single paper inside twelve months, that is a serious result.
Last month it was retracted, at the authors’ request. All eleven agreed to it.
The reasons, set out in the retraction notice, are not dramatic. There are four of them and not one involves fabricated data, a hidden dataset, or anything a reader would recognise as a scandal. They are the kind of mistakes that live inside analysis code, the sort a working scientist reads and immediately feels a small cold sensation about, because they have very nearly made all four themselves.
Start with what the paper was actually claiming, because the whole thing rests on a phrase most people have never encountered.
When you picture a drought you probably picture missing rain. That is half of it. Soil loses water in two directions: rain puts it in, and evaporation takes it out, with plants pulling more of it up and releasing it through their leaves. How fast that outgoing side runs depends on the state of the air. Warm air can hold much more water vapour than cool air, and as air warms, the gap between how much moisture it could hold and how much it actually holds tends to widen. A wider gap means the air pulls harder on every wet surface it touches, including soil, leaves, reservoirs and skin. Scientists call that pull atmospheric evaporative demand. The informal version, the thirst of the air, is close enough to be useful.
The consequence is that a place can receive the same rainfall it always did and still dry out, because the air above it is now removing water faster than it used to. Drought indices that only count rainfall miss this entirely. The index the team used, the standardised precipitation evapotranspiration index, is built specifically to account for both sides of the ledger.
The paper ran that accounting across more than a century of data, with detailed attention to the four decades up to 2022, and the headline numbers were striking. Drought severity globally had increased, and by their estimate the thirst of the air accounted for roughly forty percent of that increase. They found the area of land experiencing drought had expanded by about seventy four percent in the five years to 2022 compared with the preceding decades, that moderate to extreme drought touched about thirty percent of global land area in 2022 alone, and that evaporative demand rather than falling rainfall was responsible for most of that expansion.
It mattered outside the literature too, which is part of why the citation count climbed so fast. If drought is being driven substantially by evaporative demand rather than rainfall, then the practical picture changes for anyone planning around water. Rainfall forecasts stop being a sufficient early warning. Irrigation demand rises in years that look normal on a rain gauge. Reservoirs lose more to the air than their operating assumptions expect. Regions that are not getting drier in terms of precipitation can still be getting drier in terms of what actually reaches a root. That is a useful and slightly unsettling idea, and a paper that quantifies it at global scale is exactly the kind of thing other researchers reach for.
Those were the numbers that travelled. They were also the numbers the errors touched.
The first error is about the shape of the planet. Climate data arrives on a grid, boxes of latitude and longitude laid over the globe. Because the Earth is round, those boxes are not the same size. A cell near the equator covers far more ground than a cell of identical degree width near a pole, where the lines of longitude converge. If you want a global average of anything, you must weight each cell by the actual area it represents. If you simply take the mean of all the cells, you have quietly given high latitude boxes the same vote as tropical ones many times their size. The retraction notice states that area weighted grid cells were not used when the study wide averages were calculated, which affected numbers in the main text and in three figures.
The second error is a version number. The team used a model of evaporation called GLEAM, which had moved from its third version to its fourth. In one of the supplementary figures, values from the earlier version were used instead of the later one, which changed what the figure showed about the drought affected area in 2022. Anyone who has ever kept two folders of output from two runs of the same model, one of them stale, knows exactly how that happens and how invisible it is afterwards.
The third is a timestepping mistake in the code that worked out how much regional land was in drought. When the data were split into periods, the dates were not parsed the way the authors intended, so some values ended up assigned to the wrong stretch of years. The notice ties it to the division between the long baseline period and the recent five year window, which is precisely the comparison the seventy four percent expansion figure rested on.
The fourth is the least glamorous of the four and possibly the most instructive. Very dry places break drought indices. If a region receives almost no rain in a normal year, a percentage based measure of rainfall deficit can produce absurd values from tiny absolute changes. So analysts mask those regions out, excluding land below a rainfall threshold, in this case an average of under one hundred and eighty millimetres a year. The team did that for one panel of their first figure and did not apply the same mask to the panel beside it, from which several of the reported percentages came. Two panels, printed next to each other, describing slightly different worlds.
It is worth saying why none of this was caught before publication, because the answer is structural rather than anyone’s fault. Peer review, as practised at almost every journal, means a small number of researchers reading a manuscript. It does not mean anyone downloads the code, obtains the input datasets, which can run to terabytes, and reproduces the figures. Almost nobody has the time or the compute to do that unpaid, and journals do not require it. Reviewers assess whether the method is sound, whether the interpretation follows, and whether the result is plausible. All three can be true of an analysis with an unweighted average buried in it. The errors here were not conceptual, and conceptual errors are what review is built to catch.
That is the entire list. An unweighted average, a stale model version, a date handling bug, and a mask applied inconsistently between two panels. None of it is exotic. All of it is the residue of a large analysis assembled over a long period by people passing files and code between institutions.
Here is the point where it is worth being careful, because it is the part of the story most easily got wrong.
It is not publicly established who found these errors. The journal attached a note to the paper in the spring of last year telling readers that the conclusions were subject to criticisms and corrections that editors were considering. That wording tells you criticism existed and reached the journal. It does not say who raised it, what they said, or which of the four items in the eventual notice came from outside rather than from the authors’ own re checking. Gebrechorkos has characterised the problem as a coding error rather than a difficulty with the data or the methodology, and that characterisation is his. Every published account of how the errors came to light traces back to the authors. Anyone telling you a specific outside critic caught it is going beyond the record, and so would this article if it did the same.
What is documented is what the authors chose to do about it, and that choice is the reason this is a story rather than a notice.
A retraction is not the only option available when a paper turns out to contain mistakes. There is a correction, sometimes called an erratum or an author correction, which leaves the paper standing and amends the specific values that were wrong. It is faster, cheaper in reputational terms, and much quieter. It does not appear in the databases that track retractions. It does not get written about. For a set of errors that the authors believe changed their percentages without overturning their conclusion, a correction is the obvious defensive move, and plenty of papers with problems of this scale have taken exactly that route.
They went the other way. The notice retracting the paper carries the agreement of all eleven authors, and according to reporting on the case it is the first retraction any of them has on their record.
The practical cost of that is easy to underestimate. A retraction attaches permanently to every author’s publication list. It follows them into grant applications, hiring panels and collaborations, where it will be seen by people who will not read the notice closely enough to register that no misconduct was alleged, because none was. The word does most of its damage on its own. A researcher assessing an unfamiliar name does not usually distinguish between a paper withdrawn for calculation errors the authors themselves reported and a paper withdrawn for something worse. That distinction matters enormously and travels poorly.
Then there are the 337 citations, and the honest answer about them is that nothing automatic happens. There is no mechanism that reaches into other researchers’ published work and flags a reference. Some of those citations will be passing mentions in introductions, where the retraction changes nothing material. Others will have used the paper’s specific figures as an input, a comparison, or a justification for a modelling choice, and those authors now have to decide for themselves what to do. Most will do nothing, not out of carelessness but because there is no clear process to follow. Retraction databases and reference managers are getting better at surfacing these flags, but the work of untangling a widely cited result stays distributed across hundreds of people who never made the mistake.
Gebrechorkos’s position, as reported, is that the revised numbers will come out slightly lower than the published ones while the central conclusion holds: that warming has made global drought worse, and that the thirst of the air is a substantial part of why. He has said a revised manuscript was already prepared, incorporating additional datasets, with the code checked thoroughly. Nature’s chief editor for physical sciences, Karl Ziemelis, indicated that any resubmission would be assessed entirely independently of the retracted version.
That last detail is the sharpest thing in the whole business. It means a paper that had already passed peer review at one of the most selective journals in science, been published, been read, and been cited hundreds of times, goes back into the queue with no standing whatsoever. No credit for having been there. No shortcut through the process it already survived once. Reviewers who may reject it. A year or more of waiting, at minimum, on work that is now genuinely more solid than the version that made his name.
There is a temptation to file that under scientific self correction working as designed, and the temptation should be resisted a little. The system did not catch these errors before publication, and it took more than a year to resolve them afterwards. What worked was narrower and more human than the system: eleven people looked at four unglamorous mistakes in their own code, worked out that the cheap fix was available and chose the expensive one instead, and put their names on the notice saying so.
The retracted paper is still on the journal’s site, marked, readable, with its numbers now formally off the record. The corrected analysis exists as a manuscript. Somewhere in a review queue it is waiting for strangers to decide whether it is good enough, which is exactly where its author put it on purpose.