Four errors, one retracted Nature paper

The paper was "Warming accelerates global drought severity," by Solomon H. Gebrechorkos and ten co-authors, published in Nature on 4 June 2025, volume 642, pages 628 to 635. Its central claim, in the authors' own abstract, was this:

Our findings suggest that AED has increased drought severity by an average of 40% globally.

AED is atmospheric evaporative demand, the atmosphere's capacity to pull water out of soil, rivers and vegetation. Popular coverage translated it as "atmospheric thirst," which is a good phrase. The abstract carried two more headline numbers alongside the 40 percent: that over 2018 to 2022 "the areas in drought have expanded by 74% on average compared with 1981–2017, with AED contributing to 58% of this increase," and that 2022 saw 30 percent of global land area in moderate or extreme drought, 42 percent of which was attributed to increased AED.

The article record now carries a retraction banner dated 2 September 2026, and PubMed classes it as a retracted publication. Before that happened it was cited 337 times, according to Clarivate's Web of Science as reported by Retraction Watch.

I want to walk through the four errors one at a time, because the interesting thing about this retraction is not that it happened. It is that the failure mode is completely mundane, and completely portable to any lab that computes a number from gridded data.

How the errors were found

The note opens by describing its own provenance, which is worth noticing:

The authors have retracted this article after being notified of errors in the paper and independently identifying further issues.

Two channels, in other words. Somebody outside told them about something, and while looking into it they found more. That is the usual shape. Post-publication scrutiny rarely finds the whole set; it finds the loose thread, and the authors, who are the only people with the code and the intermediate files, pull it.

Error one: the missing cosine

This is the one I would teach from, because it is the most common quantitative mistake in the entire earth-observation and remote-sensing literature, and it is invisible on inspection. The note:

First, area-weighted grid cells were not used when calculating study-wide averages, an error that affected numerical values in the main text, Fig. 1b, and Extended Data Figs. 4 and 8.

Here is the geometry, and it takes one paragraph. Global gridded datasets are almost always stored on a regular latitude-longitude grid: every cell spans the same number of degrees north-south and east-west. But degrees of longitude converge as you go toward the poles. The physical area of a cell is proportional to the cosine of its latitude. A cell at 60 degrees north covers half the ground of a cell of the same size in degrees at the equator. At 80 degrees it covers about seventeen percent as much.

So if you load a global grid into an array and take the mean, you have not computed a global average. You have computed the average of the cells, which silently promotes the Arctic and Antarctic to several times their true weight and demotes the tropics. The fix is a single multiplication by cos(latitude) before averaging, and the reason it gets skipped is that nothing in the output looks wrong. The map still renders. The number still comes out plausible. The array does not know it represents a sphere.

I do a different kind of measurement for a living, microwave spectroscopy rather than hydrology, but the class of error is one I recognize immediately, and it generalizes past geography. Any time you average over a coordinate whose sample spacing is not uniform in the physical quantity you care about, the unweighted mean is the wrong estimator. Frequency-domain data binned linearly and interpreted per-decade has exactly this problem. So does anything sampled at uneven time intervals and then averaged as though the intervals were equal. The grid you stored the data on is not the space the data lives in.

Error two: the version mismatch

Second, results based on GLEAM v3 instead of GLEAM4 were inadvertently used in Extended Data Fig. 4: the two datasets are largely consistent for 1981–2021, but the area weighting based on GLEAM4 resulted in a reduced drought-affected area in 2022.

GLEAM is a global evaporation dataset, and this is a version-control failure of the ordinary kind: an intermediate product from an earlier run survived into a figure after the underlying dataset had been upgraded. Note the honest detail the authors include, that the two versions are largely consistent across the long record and diverge on the recent end. That is exactly the shape that makes this error hard to spot. If v3 and v4 had disagreed everywhere, somebody would have noticed the figure looked different. They agreed everywhere except the year that carried the paper's most quotable claim.

Also note that errors one and two are entangled: the notice describes the effect as coming from "the area weighting based on GLEAM4," so it is the combination of the corrected weighting with the newer dataset that moves the 2022 number. Bugs compound. Fixing one can change what the other was hiding.

Error three: the dates

Third, the underlying code included a timestepping error in the computation of regional drought-affected areas, meaning that dates were not properly parsed when subsetting the data for the 1981–2017 and 2018–2022 periods. This misclassification of time steps affected the regional averages in Extended Data Fig. 8h.

The paper's headline comparison is between a baseline period and a recent period. This bug is in the code that decides which time steps belong to which period. If you have ever worked with climate data files you know the ways this goes wrong: a time axis stored as "days since" some epoch, a calendar that is not the calendar you assumed, an off-by-one at the boundary of a slice, a string comparison that sorts "1981-1" after "1981-10".

This is the error the lead author pointed to when asked. He told Retraction Watch, in a statement sent through the University of Oxford, that the team had found "a calculation error" that was "not related to our data or methodology, but a coding error."

Error four: the mask nobody wrote down

The fourth item is different in kind, and it is the one I find most instructive.

Finally, it was not clearly explained how masking was applied in calculations. To clarify this point: the arid regions mask (defined as areas with mean annual precipitation < 180 mm) was applied to the SPEI time series in Fig. 1a but not Fig. 1b (the same distinction applies to Extended Data Figs. 4 and 8).

SPEI is the Standardized Precipitation Evapotranspiration Index, the drought metric the paper is built on. The authors give their reasoning: in very dry regions extremely negative SPEI values are common, associated with periods of zero precipitation, and including them "would have generated unrealistically low average values." So panel a masks them out. Panel b does not, and the note explains that too: "Because the % values in Fig. 1b are insensitive to negative outliers, the mask was not applied there."

Read on its own merits, that is a defensible analytical choice. Arid-region SPEI genuinely is pathological, and a metric insensitive to outliers genuinely does not need the same protection. The problem is not the decision. The problem is that a reader of the published paper could not have known the decision was made, which means a reader comparing panel a to panel b was comparing two different populations without being told.

I have opinions about this one. An undocumented mask is not a coding error, it is a reporting error, and it is the single most under-policed thing in quantitative papers. Every real analysis excludes something. Outliers, bad channels, saturated pixels, samples that failed a QC gate, regions where the model is known not to apply. Excluding them is usually correct. Not saying so in the caption is what turns a defensible choice into an unreproducible one, because the next person cannot recreate your population, and if they cannot recreate your population they cannot check your number. I made this argument at greater length in the report on reproducible versus replicable, and this is the cleanest example of it I have seen in a while.

"None of the paper's conclusions are affected"

So why retract, if the errors are arithmetic rather than argument?

The lead author's position, as given to Retraction Watch, is that the errors "changed the percentages reported, but none of the paper's conclusions are affected." He went further: "The new number will be slightly lower but the conclusion remains the same: that it is reasonable to conclude that global warming likely contributed to exacerbate global drought severity."

The note gives the journal-facing reason, and it is procedural rather than scientific:

Although the authors are willing to correct and clarify the above matters, given the scope of the necessary corrections and the need for additional revision and peer review, the authors have decided to retract the paper and to resubmit a fully revised and corrected version for peer review. All authors agree with the retraction.

Both statements can be true at once, and I think they are. The direction of an effect can survive while its magnitude does not, and a paper whose title is a direction but whose abstract sells a magnitude cannot be repaired with an erratum when the magnitude moves. Four corrections touching the main text and three figures is past the point where a correction notice is honest labelling; at that scale the corrected paper is a different paper, and a different paper needs review. The authors chose the slower, more expensive, more embarrassing route, and it was the right one.

The number that does not have an answer yet is how much lower "slightly lower" is. Until the revised manuscript is submitted and reviewed, 40 percent is withdrawn and nothing has replaced it. Nature's Chief Applied and Physical Sciences Editor, Karl Ziemelis, told Retraction Watch: "We have invited the authors to submit a new manuscript with appropriate/corrected calculations, but this is yet to be submitted." Anyone citing the 40 percent figure today is citing a number that no longer has a paper under it.

Peer review did not catch this, and could not have

The referee reports were published alongside the paper, and Retraction Watch read them. Their summary is that the two reviewers raised several questions and critiques, none of which related to the issues in the retraction notice. Referee 2 initially challenged the paper's fit for the journal, writing "I am asking myself if this paper is worth a publication in Nature, instead of a climate-based journal." After revision the same reviewer wrote "I am positively surprised by the new version of the manuscript" and concluded that "the paper is now ready, in my opinion for publication."

That is not a failure of those two reviewers. It is a structural fact about what peer review is. A reviewer receives a manuscript and figures. They do not receive the machine, the environment, the dataset versions, or the intermediate arrays. Nothing in a PDF distinguishes a global mean computed with area weighting from one computed without it. The numbers look identical in kind. There is no font for "this average is wrong by the cosine of latitude."

Which means that if your quality-control model is "peer review will catch it," your quality-control model has a hole in it precisely the size of your code. I have written before about what does and does not survive contact with someone else's machine, and the same boundary applies here: review checks reasoning, and only execution checks execution.

Fifteen months, and 337 citations

The timeline matters for anyone deciding how much to trust the recent literature. Published 4 June 2025. Expression of concern from Nature on 14 May 2026, per Retraction Watch. Retracted 2 September 2026. That is nearly fifteen months of live citation, plus a further three and a half months between the journal formally signalling doubt and the retraction landing.

Citation counts do not unwind. Some fraction of those 337 citing papers used the 40 percent figure as a premise, and none of them will be corrected. This is the same pattern I described in the report on the $38 trillion climate number, which Nature retracted in December 2025 after criticism of its methodology. Retraction Watch notes that the drought paper is the journal's seventh retraction of 2026.

I want to be careful about what that last statistic means, because it is the sort of number that gets weaponized in both directions. Seven retractions in a year at a journal publishing thousands of papers is not evidence of collapse, and a journal that retracts is doing the thing a journal is supposed to do. It is also not nothing. The honest read is that post-publication error detection is getting better and faster, and that the literature you are reading today contains errors that have not been found yet, in roughly the same proportion.

The four checks

Every one of these four has a cheap, mechanical guard, and if you compute numbers from gridded or time-indexed data you can add all four to your pipeline in an afternoon.

For the weighting. Assert that your weight array sums to the total area you think you are covering, and check it against a closed-form value for the sphere or region. If you never built a weight array, that is the finding. A second cheap tell is to compute your global mean twice, once weighted and once not, and log the difference. If it is large, you have learned something about your field distribution; if it is zero, your weighting is not being applied.

For the version. Record the dataset identifier and version in the output file, not in a lab notebook. Then have the figure script read that field and print it, so a version mismatch shows up as a caption that disagrees with the methods section rather than as nothing at all.

For the dates. After every subset, assert the count and the endpoints. A 1981 to 2017 monthly slice has 444 elements, it starts in January 1981, and it ends in December 2017. Three assertions, one line each, and a date-parsing bug becomes a crash instead of a result.

For the mask. Have the code that applies the mask also emit the sentence that describes it, including the threshold and the number of cells removed, and paste that sentence into the caption. If a panel has no mask, the emitted sentence says so. This costs nothing and it makes the panel-a-versus-panel-b inconsistency structurally impossible to ship in silence.

None of that is sophisticated. It is the computational equivalent of running a blank, and the reason to write it down is that the people who made these four errors are not careless. They are eleven authors at Oxford, Southampton, CSIC, UC Santa Barbara, Ghent, the UFZ, Leipzig, UKCEH, KAUST and Cardiff, publishing in the most scrutinized journal in science, on a topic under permanent adversarial attention. The errors still shipped, and they shipped in the plumbing, where nobody was looking.

What I did not verify

My primary sources are two Nature pages that I fetched and read directly: the retraction note, and the retracted article's own record page. Every passage I have quoted from the note is quoted from the note's full text as retrieved, not from a summary of it, and I diffed each quotation against that retrieved text before publishing. The abstract quotations come from the article record page.

I did not read the body of the retracted paper. It is behind a subscription and the article page returned only metadata and the abstract to me. So everything I say about what the paper claimed rests on its abstract and on the retraction note's description of its figures, not on the methods section. I have not seen the code, which is not public as far as I found, and I have not attempted to reproduce anything.

The statements from the lead author, from Nature's editor, and from the published referee reports all come from Retraction Watch's reporting, which I read in full as raw page text. I did not obtain those statements independently and I did not read the referee reports myself. The 337-citation figure is Retraction Watch citing Clarivate's Web of Science; I did not check it against Web of Science, and any citation count is a moving number in any case. The expression-of-concern date of 14 May and the count of this as the journal's seventh retraction of the year are likewise theirs.

Retraction Watch also quotes the New York Times's coverage of the original paper, saying the study found atmospheric thirst "played an even bigger role than previously thought" in drought severity. I have not opened the Times article and I am reproducing that phrase only as Retraction Watch reports it.

The explanation of latitude-dependent grid-cell area, and the four suggested checks, are mine, not the retraction note's. The note states that area weighting was not used and which figures it affected. It does not explain the geometry, quantify the resulting bias, or endorse any remedy. Do not read my worked numbers at 60 and 80 degrees as a statement about the size of the error in this specific paper, which I cannot know without the data.

Finally, this report will need a dated update when the revised manuscript appears. The authors say they have already prepared one. If it is published with a corrected headline figure, the right thing to do is to come back here and put that number in, in place, with a note saying what changed.