> We wanted to see why Uzbekistan didn’t jump out, so we reproduced it in our comment (Extended Data Fig 1). It turns out that Uzbekistan wasn’t even the biggest outlier, but that the version they had published had the axes cropped so you couldn’t see the outliers (see red boxes in our version). This seemed indicative of a different issue, which is why we documented it in the comment.
Cropping the chart to hide the outliers is so bad that I can't tell if they're incompetent or malicious. I wouldn't be surprised if this is the kind of thing an LLM would produce in the hands of an operator not paying too much attention, but the paper was published in the time period before LLMs were everywhere in publishing.
Creating misleading data still took effort. Now there is none needed. Slop will simply flood the zone and good content will become increasingly rare. See the problem?
Artisanal human slop took time - hours, days, weeks - and some effort to produce. AI slop can be produced in seconds, minutes or hours, at the click of a button.
> Cropping the chart to hide the outliers is so bad that I can't tell if they're incompetent or malicious
It reminds me of the famous "hide the decline", when the climate scientists working on the famous "hockey stick" paper discussed how to hide in the graphs the recent decline of proxy temperatures while measured temperatures kept growing (which would put in question the general reliability of the proxies).
Figure 1a (the leftmost subfigure in TFA’s lead image) shows the data for Uzbekistan from the DOSEv1 dataset (green), DOSEv2 dataset (red) and World Bank (black). The authors of the retracted study used the DOSEv2 dataset in order to model climate effects on the economy at a sub-national level, as opposed to the country-level analyses used in prior work. However, it looks like the DOSEv2 data was just bad for all 14 provinces in Uzbekistan (a 90% drop in GDP for all provinces in 2020!).
The typical correlation between weather and the economy is going to be fairly noisy across the dataset, but if you have 14 extra datapoints all saying there’s a catastrophic GDP crash in one year together with some coincidental weather effect, that’s going to bias the model hard. Notably, they also extrapolate losses forward all the way to 2100, so the effects of such a bias will compound.
We need something akin to the international geophysical year, but for data integrity. Make it an interdisciplinary priority to clean house and root out papers that are hanging by a thread of included / excluded outliers, biased samples, and outright fraud. It would be humbling, but we'd be in much better shape afterwards.
All else aside I mean how can they even claim to predict what an economy will do in 100 years anyway, it's going to adapt to complex higher order effects. Maybe climate change will increase GDP of everyone has to hire a worker to fan them with palm leaves.
Comparison: 100 years ago there was a global gold standard, Germany didn't exist, only a few people had cars, there were no computer machines no matter how rich you were, no transistors, no TV but many people listened to broadcast radio instead, stock trading was also for rich people and nobody was using the market to see how well the economy was doing, science fiction was about going to Venus because it was thought to be more habitable than Mars, protons had only just been discovered but not neutrons yet, and east of the Mediterranean was the Ottoman Empire.
The article suggests it's unreasonable numbers in the original Uzbekistan data source and that other datapoints may have been worse, the authors just didn't correctly execute their basic checks.
"It turns out that Uzbekistan wasn’t even the biggest outlier, but that the version they had published had the axes cropped so you couldn’t see the outliers..."
They are dancing around the accusation to help authors save face. Extreme incompetence (as in don’t let these people near 100 feet of any Excel spreadsheet level) could be another explanation but given the cropping issue it’s probably intentional
They don’t specify, but based on the period they’re talking about I’d put money on it being related to the cotton scandal, to pripiski - that is, the Soviet tendency to make up production figures. When glasnost happened in ‘88 the fiction collapsed, although not immediately - most cotton producers continued to bullshit about their numbers until the mid 90s, while the industry dwindled due to lack of water for irrigation and desertification.
The cascading effect is what makes this particularly dangerous. One bad data point doesn't just produce one wrong conclusion, it gets cited, incorporated into meta-analyses, and eventually shapes policy. By the time someone traces it back to a cropped chart and a suspicious outlier, the conclusions drawn from it have their own citation momentum. The fix isn't just better peer review, it's making raw datasets reproducible enough that anomalies like a 90% GDP drop across 14 provinces get flagged automatically before publication.
You have just described 90% of climate science. It's solid at the foundation (the basic physics of the atmosphere, the predicted trends for the future) and mostly bullshit all the way down from there, each layer building on the uncertainties and biases of the previous.
> We wanted to see why Uzbekistan didn’t jump out, so we reproduced it in our comment (Extended Data Fig 1). It turns out that Uzbekistan wasn’t even the biggest outlier, but that the version they had published had the axes cropped so you couldn’t see the outliers (see red boxes in our version). This seemed indicative of a different issue, which is why we documented it in the comment.
Cropping the chart to hide the outliers is so bad that I can't tell if they're incompetent or malicious. I wouldn't be surprised if this is the kind of thing an LLM would produce in the hands of an operator not paying too much attention, but the paper was published in the time period before LLMs were everywhere in publishing.
This is partly why I don’t worry about “LLM slop” — we had plenty of artisanal human slop before.
Creating misleading data still took effort. Now there is none needed. Slop will simply flood the zone and good content will become increasingly rare. See the problem?
Well, you can also use LLM to find issues, can't you? If it's quick to introduce an error, it's also quick to find and fix the error.
Bonus - when you ask it to find an error but there are none, it will make one up!
Artisanal human slop took time - hours, days, weeks - and some effort to produce. AI slop can be produced in seconds, minutes or hours, at the click of a button.
> Cropping the chart to hide the outliers is so bad that I can't tell if they're incompetent or malicious
It reminds me of the famous "hide the decline", when the climate scientists working on the famous "hockey stick" paper discussed how to hide in the graphs the recent decline of proxy temperatures while measured temperatures kept growing (which would put in question the general reliability of the proxies).
Figure 1a (the leftmost subfigure in TFA’s lead image) shows the data for Uzbekistan from the DOSEv1 dataset (green), DOSEv2 dataset (red) and World Bank (black). The authors of the retracted study used the DOSEv2 dataset in order to model climate effects on the economy at a sub-national level, as opposed to the country-level analyses used in prior work. However, it looks like the DOSEv2 data was just bad for all 14 provinces in Uzbekistan (a 90% drop in GDP for all provinces in 2020!).
The typical correlation between weather and the economy is going to be fairly noisy across the dataset, but if you have 14 extra datapoints all saying there’s a catastrophic GDP crash in one year together with some coincidental weather effect, that’s going to bias the model hard. Notably, they also extrapolate losses forward all the way to 2100, so the effects of such a bias will compound.
We need something akin to the international geophysical year, but for data integrity. Make it an interdisciplinary priority to clean house and root out papers that are hanging by a thread of included / excluded outliers, biased samples, and outright fraud. It would be humbling, but we'd be in much better shape afterwards.
It can only be done by outsiders. Everyone involved is incentivized to hide mistakes and fraud.
Given what happened this past 2 weeks, I can't imagine that would be so popular right now
Good on them for the retraction. It's good to see science at work.
All else aside I mean how can they even claim to predict what an economy will do in 100 years anyway, it's going to adapt to complex higher order effects. Maybe climate change will increase GDP of everyone has to hire a worker to fan them with palm leaves.
Comparison: 100 years ago there was a global gold standard, Germany didn't exist, only a few people had cars, there were no computer machines no matter how rich you were, no transistors, no TV but many people listened to broadcast radio instead, stock trading was also for rich people and nobody was using the market to see how well the economy was doing, science fiction was about going to Venus because it was thought to be more habitable than Mars, protons had only just been discovered but not neutrons yet, and east of the Mediterranean was the Ottoman Empire.
Makes me wonder how many more papers out there have hard-to-pin-down errors like that
And how useful potentially AI could be to spot those (even if retrospectively)
I'm confused as to what the actual issue was. What was the data for which Uzbekistan was the outlier and why?
The article suggests it's unreasonable numbers in the original Uzbekistan data source and that other datapoints may have been worse, the authors just didn't correctly execute their basic checks.
"It turns out that Uzbekistan wasn’t even the biggest outlier, but that the version they had published had the axes cropped so you couldn’t see the outliers..."
> had the axes cropped so you couldn’t see
Almost sounds intentional...
> This seemed indicative of a different issue, which is why we documented it in the comment.
Yea, that different issue is fraud.
They are dancing around the accusation to help authors save face. Extreme incompetence (as in don’t let these people near 100 feet of any Excel spreadsheet level) could be another explanation but given the cropping issue it’s probably intentional
They don’t specify, but based on the period they’re talking about I’d put money on it being related to the cotton scandal, to pripiski - that is, the Soviet tendency to make up production figures. When glasnost happened in ‘88 the fiction collapsed, although not immediately - most cotton producers continued to bullshit about their numbers until the mid 90s, while the industry dwindled due to lack of water for irrigation and desertification.
This feels like it could be right, but then maybe its just interesting that there is exactly one outlier here like this, and not more?
My read is that the model had too much variance; more regularization was needed.
The cascading effect is what makes this particularly dangerous. One bad data point doesn't just produce one wrong conclusion, it gets cited, incorporated into meta-analyses, and eventually shapes policy. By the time someone traces it back to a cropped chart and a suspicious outlier, the conclusions drawn from it have their own citation momentum. The fix isn't just better peer review, it's making raw datasets reproducible enough that anomalies like a 90% GDP drop across 14 provinces get flagged automatically before publication.
You have just described 90% of climate science. It's solid at the foundation (the basic physics of the atmosphere, the predicted trends for the future) and mostly bullshit all the way down from there, each layer building on the uncertainties and biases of the previous.
Uzbecky nitwits