Researchers release Global Automation Atlas dataset spanning 18,797 tasks across 124 countries and 2.33 million comparisons to document automation exposure outside high-income economies

Two highlights:

#1: automation exposure rises with income.

But countries at low- and middle-income levels still differ substantially.

Explore the Atlas for more results: http://automationatlas.org

This includes our paper, open data+code and interactive visuals.

We couldn't do it all, so we released data and code, and linked our measures to common units of analysis to enable future research: occupation, industry, and skills.

#2: when a task is exposed to automation, is technology substituting core worker input, or mainly augmenting workers to do the task?

In most countries, substitution-oriented exposure is larger, especially in lower-income settings.

@Prashant_Garg_ very cool.

i'm curious which exact "automations" are we talking about here?

and lets hope you do not get cancelled for using wrong India map 😭😭

This work wouldn't be possible without my amazing co-authors @TommasoCrosta and @jasmin_baier.

We have a lot more planned, and would love your feedback!

@Prashant_Garg_ @ATabarrok @tylercowen

@Prashant_Garg_ Nice! Been waiting for people to do this - looks awesome!

@Prashant_Garg_ @TommasoCrosta @jasmin_baier It looks like it is currently correlated with manufacturing share? Whereas service workers are OK?

@Prashant_Garg_ @TommasoCrosta @jasmin_baier Thank you for this public good. If I want to know what tasks are exposed in a country, say Bangladesh, how would I go about it?

@Prashant_Garg_ @TommasoCrosta @jasmin_baier Thanks. It is really cool. However, I was interested in how exposed garment workers are. The garment and its ecosystem are the source of employment for 40 million. The percentage of female workers is already decreasing due to automation. I didn't find anything.

raw correlation between manufacturing and exposed task share is 0.28. After residualising both on per capita income it remains 0.27.

so manufacturing isn't all. We tried to understand what predicts exposure a bit with random forest (but not fully...).

Digital connectivity, human capital, governance, and income account for much of the cross-country variation in exposure

@Prashant_Garg_ why the heck is this not viral af this is so cool, incredible work!!

we did lots of validation checks (see bottom bit in https://automationatlas.org/methods/ page, details in paper). One method is to compare various measures we create to closest available external dataset. Others are to see consistency across models and to check internal reasoning provided by the model systematically

@adidshaft it's a very good question, and we consider all the major categories of automation. This is often overlooked, so we decided to explicitly measure the channel of automation for each task, in each country

@Prashant_Garg_ @KhoaVuUmn Another great project @Prashant_Garg_

@Prashant_Garg_ how do you define exposure? And how do you measure it changing over time?

@krisgulati Thanks!! Hopefully it moves us forward on this literature!

@Prashant_Garg_ Really interesting!

Question: why’d you decide to use Gemini 3.1 Flash Lite for the classification, specifically? Not saying it isn’t a valid choice, just curious about how you and your coauthors settled on it.

@joefrancis505 @TommasoCrosta @jasmin_baier You have one? I slowly fading into a supervisory role for my persona—letting AI take on some elements. It’s quite liberating to let go of one’s ego

@Prashant_Garg_ @TommasoCrosta @jasmin_baier It means that rich countries always have 40% of the population who can't be automated.

Baristas?