Bottom line up front: I estimate that AI could reduce consumer prices by 28% – equivalent to a 39% income increase for all households.1 Half of this effect comes from indirect effects: AI reducing costs in back-office sectors like software and accounting, that flow through to everything else. The price reductions are dispersed across many sectors, and they scale with how disruptive AI is. The more jobs AI threatens, the more it reduces prices, acting as an automatic stabilizer against job loss. This is a back-of-the-envelope exercise with wide uncertainty, but it suggests AI’s price impacts deserve much more attention than they get.

Framework

Imagine a simple economy where we only make one good. 50% of the cost is labor, while 50% comes from other sources (machinery, electricity, etc). Suppose that AI can do 60% of the tasks that these workers do. This is equivalent to AI being able to do the job of 60% of workers. If AI can do these tasks for free, then labor costs fall by 60%. Since labor costs were initially 50% of the total cost, that means the total cost of production falls by 60% * 50% = 30%. Thus, AI reduces prices by 30%.

This framework glosses over many important assumptions that I’ll discuss at the end. But it establishes the approach I will use: that the price reduction in a sector due to AI = % of tasks that AI can do * % of costs from labor.

Weighting price reductions in each sector by how much households spend on that sector then gives us the average price reduction experienced by a household. That way, we put more weight on the prices of healthcare and education than on the price of artisanal soap.

There’s one final wrinkle. Even a sector with little exposure to AI might depend on another sector with high exposure to AI. Take plumbing as an example. AI cannot do the physical tasks involved in plumbing, so it might seem like it could not reduce the price of plumbing. But the company providing plumbing services needs to use payroll software to pay individual plumbers, accounting software to manage its books, etc. These costs are an overhead on every individual job, an overhead which AI can reduce. In this way, AI could reduce the price of plumbing even if it can’t do any of the tasks required of a plumber.

Accounting for these indirect price effects is especially important because AI’s effects are likely to be concentrated in back-end sectors that consumers rarely interact with. So we have to calculate not just how much AI reduces labor costs in a sector, but also how much it reduces labor costs upstream of that sector.2

Analysis

Full details on the data and analysis are in this GitHub repo.

What do AI exposure and labor costs look like in different sectors?

AI exposure is dominated by financial services and software. Insurance carriers and securities firms top out at 94% of tasks exposed, followed by data processing (92%), computer systems design (92%), and legal services (86%). Among the least exposed are sectors dominated by physical tasks: food services (27%), warehousing (29%), accommodation (29%), and mining (32%).

We can more systematically look at the sectors with the most potential price reductions from AI, including a decomposition into the direct effects (from labor costs in the sector itself) and indirect effects (from labor costs upstream of the sector).

The gap between the dark and light bars points to the value of indirect effects. The most dramatic case is finance (funds, trusts, and financial vehicles), where almost the entire price reduction comes from cheaper inputs purchased from upstream sectors (in particular, software), not from AI replacing labor within the sector itself (because labor is a very small share of costs in the sector). Notably, even healthcare sectors appear in the top 20, buoyed by indirect effects from cheaper administrative, IT, and professional services inputs.

Across the economy, I estimate that AI will reduce aggregate prices by 28%. While software, financial services and consulting dominate the list of sectors with the highest price reductions from AI, they don’t actually contribute much directly to the price reduction, because households don’t spend very much on those sectors.

Instead, the price effects mainly come from healthcare, education, insurance, food and other services – familiar line items, because of how important they are in household budgets. But price reductions in these consumer-facing sectors come in part from the price reductions in upstream sectors like software. Builders of B2B SaaS products, be vindicated; your work does indeed matter.

To be clear, 28% is a big effect! A 28% reduction in prices is equivalent to a 39% income increase for every household. Put differently: any worker who sees their wages fall by up to 28% due to AI is still better off than they were before AI.3

Importantly, these price effects scale with the employment effects of AI – both price effects and employment effects depend on how much work can be automated by AI. The more economically disruptive AI is, the more it causes mass automation and job loss, the larger its price effects will be. Thus, AI’s price effects act as an automatic stabilizer to its employment effects, undoing (some of) the damage that workers face from automation.

Ways I might be wrong

There are a few ways in which my analysis predictably overestimates how much AI will reduce prices:

• I’m assuming that the share of tasks that can be done by AI is exactly the share of labor costs reduced.4 But if tasks are complementary to each other, which seems more realistic, then automating some of them would still leave bottlenecks in other tasks. This dampens the efficiency gains from AI.

• I’m assuming that cost reductions are passed through to consumers, rather than absorbed as profit increases by businesses.

• I’m assuming that AI replaces labor for free. In reality, AI inference is costly, and that cost will add back to the price in each sector.

But there are also some ways in which I’m predictably underestimating price effects!

• I use conservative estimates of the share of automatable tasks. Eloundou et al (2023) estimated AI exposure based on GPT-4, so their estimates of tasks that can be done by AI are more conservative than I would be now.

• This exercise doesn’t capture the value of quality improvements. If AI can perform medical procedures more cheaply than doctors, it could reduce the price of healthcare; but if it can also perform better than doctors, then the quality-adjusted price falls even more than the sticker price.

• Relatedly, this exercise doesn’t capture the possibility of AI creating new goods. If AI helps design new medicines that treat currently incurable diseases, that is worth more than reducing the price of existing medicines for that disease.

And to top it off, I could be wrong in even more ways with ambiguous directionality:

• I’m not accounting for any equilibrium changes in supply and demand. Productivity changes will shift supply and demand, which then feed back into prices in a complex way. My advisors would cry if they saw the way I’m butchering equilibrium.

• The data work is not bulletproof. Seriously, check my work. Claude Code made many judgment calls that I haven’t audited in depth. “Vibe research” still has a ways to go, and conclusions here could change.

Conclusion

Clearly, this exercise has wide uncertainty. I don’t expect it to be the final word on AI and consumer prices. But it is a provocation for people to pay attention to this topic, to give it even a fraction of the policy attention paid to AI’s labor market impacts.

A 39% increase in real income is an enormous effect, larger than any anti-poverty program – and it applies to every household, for free. In the worst case scenario where AI purely replaces workers and causes massive job loss, this price effect cushions the blow. In more moderate scenarios where AI transforms jobs rather than making workers obsolete, this price effect can lead to real income gains across the board.

But this is a mistake. Jobs facing future obsolescence tend to see wage increases, not decreases. This counterintuitive fact comes from a concept known as obsolescence rents.

Obsolescence rents

Career choice is a long-term investment. When you decide to become a lawyer or a programmer or a truck driver, you consider not just the money you’ll make in your first year, but the money you’ll make in twenty years. That includes considering whether that job will even exist in twenty years.

So imagine you’re a young worker choosing a career, and technology threatens one particular job in the future. But notably, technology cannot yet replace workers at that job, so you can still get a job in that sector if you want to. Do you take that job?

Maybe! But the risk of future job loss reduces your willingness to work in that profession. You need to be compensated for the risk that you will be stuck with obsolete skills halfway through your career, and forced to go through the unpleasantness and uncertainty of re-skilling for a new job. (You might protest that young people don’t have infinite foresight to see the future – but even if some fraction of them are attentive to the risk of future job loss, the argument below makes sense.)

This risk of future job loss shows up as a leftward shift in the labor supply curve – at any particular wage, workers are less willing to work in the job than they were before:

This shift in the supply curve has two immediate predictions that you can see in the graph.

1. Employment falls (L_new < L_old). Because the risk of obsolescence makes workers find a job unattractive, fewer people take that job now.

2. Wages rise (W_new > W_old). Since the job is not yet obsolete, employers still need to find workers to do it. But to compensate for the risk of obsolescence, they need to pay higher wages to hire those necessary workers.

Thus, the counterintuitive conclusion: impending obsolescence makes wages rise.

We can make one final observation. The threat of obsolescence also affects who chooses to work in a job. A 25-year-old has forty years of working life ahead of them; a 55-year-old has ten. The 55-year-old may retire before obsolescence occurs, but the 25-year-old has no such luxury. Thus, young workers are disproportionately filtered out of an occupation by the threat of obsolescence. This leads to a third prediction:

3. The profession becomes older in age composition. Younger workers are less likely to enter the profession and more likely to leave it than older workers.

So this is the theory: impending obsolescence leads to falling employment, rising wages, and aging workforce composition. Does it happen in reality?

Case study: motor trucks

The transition from horse-drawn freight to motor trucks in early 20th century America provides the clearest historical evidence for obsolescence rents. A recent economic study uses this transition to test all three predictions of the obsolescence rents story.

Before trucks, freight was carried on wagons pulled by teams of horses. The workers who drove these wagons were called teamsters – literally, people who managed teams of horses. In 1900, there were over 400,000 teamsters in the United States.

Yet motor trucks were anticipated for a long time. In 1895, Thomas Edison declared that it was “only a question of time when the carriages and trucks in every larger city will be run with motors.” The first commercial motor truck was sold in 1897. The technology was real. At the same time, trucks remained firmly on the horizon. Their design was still uncertain; competing designs used steam, electricity, or gasoline, and nobody knew which design would win. More importantly, road infrastructure wasn’t good enough for trucks in the early days. They were expensive, broke down frequently, and couldn’t travel far outside major cities.

World War I pulled trucks forward from the future, as American manufacturers built thousands of standardized military trucks for use in the war. After the war ended in 1918, all of that production capacity turned to the civilian market, and surplus military trucks flooded into civilian use.

You can track the growing anticipation in the pages of Scientific American, which served as a contemporary forum for discussing the cutting edge of technology. Between 1900 and 1910, only eleven articles mentioned motor trucks, and almost none forecast trucks replacing teamsters. But between 1910 and 1920, ninety-six articles discussed trucks.1 A 1909 article cautiously suggested trucks might be superior to horses in New York City – but warned that “two weeks at the factory is not sufficient to change a stable hand into a competent driver.” By 1918, the tone had shifted completely: “Prior to the war, the motor truck was making steady progress towards ultimate complete employment... But the war accelerated its adoption, perhaps by twenty years.” By 1930, the transition was largely complete. Motor trucks had taken over urban freight hauling.

The researchers use the historical transition from teamsters to motor trucks to test the three predictions of obsolescence rents: falling employment, rising wages, and aging workforce.

First, did teamsters see falling employment before trucks were actually deployed? Yes. The researchers used Census data to track how many people worked as teamsters in 1900, 1910, 1920, and 1930. Teamster employment rose from 1900 to 1910, when motor trucks were still a distant prospect. But it fell slightly from 1910 to 1920, even though trucks weren’t yet widespread. This is the sign of obsolescence rents: teamster employment dropped before trucks were actually deployed. Of course, employment absolutely cratered between 1920 and 1930 as trucks took to the road, but that is the part that we expected already.

Could other factors explain the 1910-1920 decline? Of course. WWI pulled workers into many occupations whose demand spiked during the war effort, and perhaps some of them stuck with those occupations after the war. But the timing of the decline is consistent with workers avoiding a profession they expected to disappear.

Second, did teamsters actually see rising wages before trucks were deployed? This is the important one, since it gives the most direct evidence about whether we should actually see rising wages as a bad omen. The researchers compared the wages of teamsters to wages in “close trades” – that is, professions that required similar skills and earned similar wages before trucks appeared (e.g. carpenters, building laborers, painters).

During the peak of anticipation, wages for teamsters did seem to rise relative to close trades, before cratering in 1920 as trucks hit the road. This isn’t slam dunk evidence by any means – the wage increase during the anticipatory period isn’t statistically significant – but it points in the right direction.

Third, did the age composition of teamsters become older? In a way this is the most convincing evidence for obsolescence rents, because it is the most specific. Wages and employment can fluctuate for all kinds of reasons not captured by the simple story we are telling. But “a profession that was historically full of young people suddenly becomes full of older people” is a much more tailored prediction.

Between 1910 and 1920,2 young workers’ entry into teamster employment dropped by 30% compared to the previous decade. But older workers’ entry actually increased by 20%.3 The pattern grew more pronounced after trucks were deployed en masse. Between 1920 and 1930, all age groups became much less likely to become teamsters – but old workers saw a decline of 40%, much less than the 60-70% decline for young workers. Despite being physically demanding work that should favor the young, being a teamster became an old man’s job as trucks arrived on the horizon.

Will AI produce obsolescence rents?

The arrival of motor trucks in the 1910s parallels the state of AI today:

1. Both are technologies which definitely exist, and are widely known to have the potential to automate jobs in the future.

2. However, neither can automate jobs today due to limitations of the current technology and the absence of complementary infrastructure (trucks needed roads and repair mechanics, AI needs organizational changes/regulatory freedom).

3. When this automation will occur is a question with huge uncertainty.

These three conditions are exactly the conditions under which we expect obsolescence rents to appear. Jobs like writing or translation are still needed today, but people can see the writing on the wall for them. So in order to attract people to fill the jobs that are still needed today requires higher wages. Perhaps Anthropic is posting a high wage for a writer because excellent writers need to be persuaded to continue working as writers rather than accumulating more future-proof career capital.4

The important takeaway is that if you want to forecast automation risk:

1. Do not use wages as a signal of occupational health, at least not if you’re interpreting wage increases as a positive signal.

2. Do use employment/job postings as a signal of occupational health. “Translator wages are up, so AI won’t automate translators” is a bad argument. “Translator employment is up, so AI is more likely to augment than automate translators” is a better argument.

3. Do use age composition as a signal of occupational health. If young workers are reluctant to take a job, it is a good sign that at least they believe it will be automated in the future (though who knows if they’re right!)

Caveats

There are a few limitations to this argument. First, obsolescence rents are obviously not the only reason for wages to go up, and it is clearly not the case that wages going up is always a bad sign. For example, wages for software engineers exploded in the 2000s because software became a growing part of the economy, not because software engineers were at risk of obsolescence (then). So my point is not that rising wages are always a bad sign, just that it’s important to distinguish between “wages went up because the sector’s demand for workers is healthy and growing” and “wages went up because willingness to work in this sector is anemic and falling”. Both scenarios would lead to wage increases, so wage increases alone cannot distinguish between these scenarios.

Second, these predictions rely on workers actually factoring the chances of future obsolescence into their career choice. If nobody believes that their job will become obsolete, then there will be no shift in willingness to work in a job, and no rising wages/falling employment/aging of the profession. If you believe that people are asleep at the wheel and not considering the risk of their job being automated, then this argument has no insight for you. But I don’t think people are asleep at the wheel. Many people remain skeptical that AI will replace their jobs, but this skepticism has declined sharply since 2022. People’s guesses about the future may lag behind the best available information out there, but they are clearly responsive to the growth of AI. The market for “what should I do to future-proof my career” advice is enormous. Even if people are only partially responsive to automation risk, we should expect to see obsolescence rents.

Third, it is a key assumption that the new technology’s impact is automation. If AI is actually going to augment workers (making them more productive), then the predictions are exactly flipped from what I’ve described. Suddenly, instead of expecting a reduced future income, workers expect an increase in their future income from entering that job. The resulting influx of young workers would increase employment, bid down wages, and make the profession younger. Once again, we are in backwards-land, where falling wages are a good omen. In this case, we should still read occupational health from employment and demographics, rather than from wages.

The automation of work has begun. In October, OpenAI released GDPval, an evaluation of whether AI can do economically important tasks across a wide range of occupations. Anthropic’s Economic Index classifies Claude interactions based on which economic tasks they are connected to, to paint a picture of which tasks AI is most capable of performing. In a 2023 paper with over 1300 citations so far, researchers estimate how exposed the US workforce is to AI-driven automation, by using human ratings and LLM ratings to assess whether GPT-4 could do each task done by current jobs. While not focused on economic impacts, METR’s influential time horizons report measures the capabilities of AI models in terms of the length of tasks that they can complete reliably.

All of these approaches to assessing AI systems share a focus on the task as the unit of analysis. They imagine jobs as bundles of tasks – a worker does some set of tasks, a machine does some other set of tasks, and collectively this produces goods and services. In this mental model, AI automating a job means that AI can reliably do most or all of the tasks needed to perform that job.

This gives researchers a way to quantify the progress towards AI being able to automate work – by focusing on which tasks frontier models can do, which ones they cannot, and how that success rate has trended over time. These task-based exposure measures are now by far the most common way to forecast whether and when AI will automate work across the economy.

But all of these measures rely on a single dataset: the Occupational Information Network (O*NET), created by the US Department of Labor. O*NET maps all jobs in the economy to lists of tasks that are essential to that job. It is the backbone of all task-based exposure measures: OpenAI constructed GDPval by commissioning experts to write tasks that map to O*NET categorizations, and Anthropic’s Economic Index links Claude usage to O*NET tasks.

Despite O*NET’s importance to AI automation measures, it remains poorly understood. Many people drawing on task-based exposure measures to forecast AI’s impacts have never heard of O*NET, let alone understand how it is constructed. This is a problem because O*NET’s construction makes it look easier for AI to automate human work than it really is, and fixing that issue requires better data than we have.

O*NET: a brief but affectionate history

Economists have historically modelled work as a black box. Old-school “production function” models in economics simply modelled the output of a company as some combination of its workers and its machines, rather than unpacking work in any meaningful way. One major problem with this approach is that it was unclear what automation even meant in this framework. The framework assumed that there was some fixed ability to substitute between workers and machines. Did automation just mean making machines and workers more substitutable? If so, how would you make that asymmetric, so that machines are better able to replace workers than vice versa? The old-school models did not have good answers to this problem, and that made them poorly fit for understanding automation.

These limitations were well-known, but economists still used the production function approach because they simply didn’t have the data to do better. Companies would report their capital stocks and payrolls in balance sheet data, but they wouldn’t report what each worker was actually doing. So it was natural for economists to build models that could use those observable features. Nobody believed that there was literally a mechanical relationship between workers, machines, and output – it was just the best that they could do with their limited data.

The revolution in how economists think about work traces back to the release of O*NET in 1998. O*NET was the successor to the Department of Labor’s Dictionary of Occupational Titles (DOT), a Depression-era categorization of jobs in the economy. The DOT focused on characterizing blue-collar and agrarian work, which made it obsolete in the information technology era. O*NET was created as a modern representation of work, with Department of Labor analysts running large-scale surveys of workers across the country to get a picture of work that represented the whole economy.

A key feature of O*NET (and DOT before it) was that it decomposed jobs into lists of tasks that were essential to those jobs. These task lists are compiled by organizational psychologists who interview workers to understand their most important work activities. They are then validated by surveying workers on how important each of the tasks on that list is to their job, with those survey responses being used to phase in and out tasks from the O*NET list (since a job could have changing tasks over time).

These task lists have become foundational to analyzing the labor market impacts of new technology. For example, Autor, Levy and Murnane (2003) classified tasks as “routine” and “nonroutine” tasks to represent how easily they could be automated by computers, and used these measures to estimate that the introduction of computers to the workplace decreased the employment of unskilled workers (who did routine tasks) and increased the employment of skilled workers (whose nonroutine tasks were now the bottleneck to work). By disaggregating between types of jobs based on the work they do, the DOT and O*NET made it possible to find nuanced impacts of new technology on labor markets. The methodology that was initially designed to assess the impact of computers on the labor market has been adopted to study the impact of AI on the labor market, by measuring the exposure of each job to AI automation based on its task content.

We need to talk about task exposure measures

A key feature of work that task-based exposure measures fail to capture is task dependencies – situations where in order to do one task, you need to have done other tasks. Imagine a sales representative whose AI system has automated all of their email writing. This seems like a major time savings – no more time spent composing emails. But then after an email chain about an issue, a customer sends a message requesting a Zoom call with the sales rep to discuss it, triggering another task that AI cannot do in this thought experiment (a customer Zoom call). In order to have a useful call, the sales rep needs to read and understand the whole email chain. This is a task dependency: the ability to do one task (having the call) depends on having done another (reading the email chain), because the knowledge from the email task is a key input to performing the Zoom call task. Task-based exposure measures assume that tasks are independent of each other, making them overestimate the usefulness of automating some tasks without automating others as well.

In the extreme, dependencies could prevent automation entirely. Suppose producing a good takes 10 sequential tasks, but each task can only be done by the worker or machine that did the previous task. Then unless all of those tasks can be done by the machine, all of them have to be done by a worker. Division of labor opens the door for tasks to be done independently of each other, but not all tasks can be so neatly divided. And when researchers decompose jobs into discrete tasks and ask “which tasks can an AI do?”, the dependencies often disappear from view. You end up with a list of independent capabilities rather than a picture of how those capabilities need to fit together to do a job.

Another feature of work that task-based exposure measures fail to capture is interstitial tasks – tasks that are hard to formalize, but are still an essential part of a job. When characterizing jobs, O*NET analysts focus on the most legible tasks done by each job. For a teacher, they may list “deliver instruction” and “grade assignments” and “manage classroom behavior.” These are real tasks. But a teacher also spends time having hallway conversations with colleagues about difficult students, mentally preparing for class transitions, noticing when a student is struggling emotionally and checking in with them, coordinating with other teachers on joint units, reflecting on what worked and what didn’t in a lesson.

None of these tasks would make it into a job description, but they are so important that teachers’ unions often go on a “work-to-rule strike”, a type of labor action in which workers strike by only doing tasks that are explicitly in their job description. In other words, these interstitial tasks are so important that they bring the government to the negotiating table despite all the “main” tasks being performed as usual.

In theory, the task model of work is meant to represent all tasks, including interstitial tasks. But task-based exposure measures do not deliver on this goal. They instead characterize jobs only through the legible tasks listed in O*NET. In reality, jobs are not just their top 5 most important tasks, or even their top 50 most important tasks. They are made up of countless small tasks that fill the cracks between the big, legible ones. Ignoring these interstitial tasks is a common blind spot in AI automation discourse. Geoffrey Hinton famously claimed that AI will automate radiology because AI models can read X-rays and scans better than human radiologists – but that is only one task among many that a radiologist has, and reading scans takes up only a third of a typical radiologist’s day.

The interstitial task problem is a more specific version of the argument that “if AI can do 99% of tasks, we will be bottlenecked by the 1% of tasks it can’t do.” It’s more specific because it identifies those remaining tasks as specifically the interstitial ones. It’s possible that in the future of work, people spend much less time on the things that used to be front-and-center focuses of their job. Teachers spend less time teaching or grading, radiologists spend less time reading scans, doctors spend less time doing diagnoses. For sufficiently extreme changes, you might question whether the job is even the same job as it was before. But you can’t measure that shift using task exposure measures – because the interstitial tasks that remain are invisible to them.

What should we do instead?

AI automation measures based on O*NET task lists are predictably misleading. They don’t capture task dependencies, and how automating some tasks might create bottlenecks in other tasks. They don’t see the interstitial tasks that hold those jobs together, that are the hardest to automate. Both of these gaps make them predictably over-optimistic about the speed of AI automation of work.

Nevertheless, these measures are the best we have given the data that we have. I can grouse about how O*NET is constructed, but it is the richest, most systematic source of data on occupational work we possess. If you want to forecast AI’s labor market impacts across the entire economy, you don’t have a better option than to acknowledge O*NET’s limitations, and then use it anyway. The real question is: how would you create a better option than O*NET-based task measures?

Imagine a ridiculously high-resolution description of work: workers describing everything they do, at high frequency (e.g. every 30 minutes), in whatever language feels natural to them – unstructured, conversational. And imagine this measurement is repeated every day for a long time (e.g. a month). If having workers describe their own tasks at such a frequency is infeasible, you could instead have enumerators paid to follow them around and record their work activities, although that could be a whole other issue.

The resulting work diary would be the most comprehensive possible picture of a job. It would be much messier than a simple list of tasks, but it would have much more information. By focusing on what people spent the last 30 minutes on, it makes it more likely that people would capture all the interstitial tasks they had to do and how much time they spent on them, not just the most salient ones. By reading sequential logic in workers’ own accounts of what they do, you can see which tasks depend on which others. You can measure how much time each task takes up relative to others, and what share of time is spent on each task. And while this would be way too much information for a human to process, an LLM would be able to process these work diaries to identify recurring task patterns, measure task frequency and duration, and map dependency relationships across tasks. This would produce automation exposure measures grounded in day-to-day work behavior rather than the beliefs of O*NET analysts or worker survey responses.

A work diary project is simpler than some of the more elaborate task-based exposure measures. You don’t need to create synthetic work products, or recruit expert graders. You only need to recruit workers willing to keep detailed diaries, and you need the computational infrastructure to process unstructured text at scale. Both of these are solvable. And while gathering work diaries for a comprehensive set of jobs would be a monumental effort, it could still be worth gathering work diaries for a smaller set of jobs that have special importance. For example, you could try to assess whether AI could automate AI research by collecting work diaries from AI researchers.

In the absence of better data, all we can do is handle task-based exposure measures with care, and remember that automating a job is much harder than automating the most legible individual tasks of that job. Forecasts of AI automation that are based on task measures should be treated as upper bounds, not predictions.

The economic growth stories of the past century were defined by technology transfer. The entire reconstruction of Europe as a set of advanced economies happened through the Marshall Plan, the large-scale transfer of knowledge and infrastructure from the US to European countries. Japan, Taiwan, China, and Korea all grew their industries using technology imported from the US, US, Soviet Union, and Japan respectively. In all of these countries, the growth of advanced industries was catalyzed by receiving that technology from another country – through a combination of supplying special equipment, and training workers to work in those industries.

But despite the power of this kind of technology transfer, these agreements don’t happen anymore. Today, policies around foreign investment in developing countries are framed in terms of creating jobs or gaining tax revenue – not in terms of gaining technological capabilities. It often comes with no strings attached, and no intended technological benefit for the host country at all.

So what happened to technology transfer agreements?

Technology transfer helps industrial growth

The most comprehensive technology transfer effort in history happened as part of the Marshall Plan, with the US aiming to bring technology and productive capacity back to European firms after they were devastated during World War II. The US sponsored training trips for managers from European firms to visit American factories, and they provided subsidized loans to buy advanced equipment from American firms, bringing the European firms back to the technological frontier. Giorcelli (2019) evaluates this program’s impact on Italian firms. She compares firms that received the training and subsidized machines with firms that could not receive them due to budget cuts on the American side – reasoning that without the budget cut, these firms are otherwise identical, so comparing their trajectories gives us the effect of the Marshall Plan’s productivity program. She finds that 15 years after the Italian firms received the training and equipment, they were 90% more productive and had 20% higher profits than their counterparts who missed out on the program due to budget cuts. These results make it pretty likely that the Marshall Plan’s technology transfer program helped stimulate industrial growth in Europe.

Uncannily similar evidence comes from the other side of the Iron Curtain. In the 1950s, the Soviet Union supported the development of large heavy-industrial clusters in China as part of their alliance. They exported state-of-the-art equipment to set up Chinese plants, and trained Chinese technicians in how to operate and maintain these plants. Giorcelli and Li (2025) evaluate this program using a natural experiment – many of the proposed transfers were cancelled abruptly as part of the Sino-Soviet split, while others were completed before this split. They find that the projects which received Soviet training and machines became more productive than the plants that were supposed to receive Soviet transfers, but didn’t. Strikingly, this benefit extended beyond Soviet knowhow and into helping the plants do their own internal R&D. For example, Soviet steel furnaces were made obsolete in the 1960s by new continuous-casting furnaces; yet the Chinese plants that received Soviet transfers actually became more likely to adopt continuous-casting furnaces, showing that they weren’t just blindly using imported machines but were actually keeping up with frontier capabilities. Decades after the initial transfer, these plants produced 50% more and exported 30% more than the plants that didn’t receive Soviet transfers. Much of China’s vaunted industrial capacity was built during this era, and thus can be attributed to Soviet technology transfer.

But this isn’t an essay about how technology transfer is good. That essay would fall flat, because technology transfer is off the menu today. Today, developing countries court multinational investment, rather than imposing conditions on it.

So given the unreasonable effectiveness of technology transfer, why doesn’t it happen anymore?

Why technology transfer happened to begin with

Technology transfer is a strange idea when you think about it. After all, companies invested tremendous effort and resources in reaching their level of technological capability. Why would they want to sell those capabilities to another firm? Even if they were paid handsomely for it, they would just be seeding their own end – creating firms that would eventually be able to outcompete them. It’s hard to imagine a contract that makes this risk worthwhile for them. So the puzzle is not why technology transfer doesn’t happen now – the puzzle is why technology transfer ever happened to begin with.

There were three main reasons why technology transfer was so common in the post-war era:

1. Geopolitical alignment. What all the most successful case studies of technology transfer have in common is that they were between geopolitically aligned countries. In the examples studied above, the US wanted to prop up Europe as part of the Marshall Plan, while the Soviet Union wanted to bolster another communist power. These motivations meant that the technology-exporting companies did not have to benefit very much – the respective governments were invested enough to encourage their companies to transfer technology to the recipient countries. It is hard to imagine companies protesting too much when they have been ordered to transfer technology in order to fight the enemy.

2. Market access. It’s hard to picture this now in the era of globalization, but it used to be quite difficult for Western firms to access other countries. They wanted access to other countries in two ways – as destinations in which to sell their products, and as origins in which to outsource their manufacturing. Both of these forms of access were hard to come by. Exporting to developing countries was held back by “import substitution industrialization” – the policy framework adopted by many developing countries, in which they reduced imports in order to force domestic firms to produce in essential sectors. Investing in setting up factories was often similarly restricted – for example, in India, foreign ownership in companies was capped at 40%, making it impossible for a foreign company to set up and majority-own their factory. Faced with these barriers, companies were much more willing to consider demands such as “you need to license your technology to a domestic firm in order to sell goods/set up a factory in our country”.

3. Creating a suitable supplier. If companies want to outsource production of goods with any technological sophistication, they can have a hard time finding a suitable supplier – one that is cheap enough to be worth outsourcing to, but is also high-quality enough to produce according to standards. Technology transfer is a way for companies to create that suitable supplier when one might not exist. For example, Bangladesh’s textile industry was built up by technology transfer from Korean textile firms, who wanted to produce in Bangladesh to have easier export access to the US.

These three forces conspired to create the landscape of successful technology transfer agreements that characterized the mid-20th century.

Today’s landscape

Of the three motivations for firms to transfer technology to developing countries, do any of them still apply?

1. Geopolitical alignment kind of applies today. The “China plus one” approach to supply chains involves firms preferring to not rely on Chinese suppliers because of geopolitical risk. India has certainly been appealing to countries on this basis. But by and large, firms don’t consider licensing technology on the basis of geopolitics anymore.

2. Market access doesn’t apply today. The world is far too globalized for it to be a valuable carrot. There are too many countries with big markets who don’t demand technology transfer. There are still a small number of countries with large enough markets to make such demands – China, India and Brazil come to mind. But only China is actually willing to make those demands.

3. Getting a cheaper supplier does still apply somewhat. But today, international firms would rather simply invest directly in setting up factories within countries where it’s cheap to produce. That way, they can reduce their production costs, without the baggage of licensing and transferring technology to local firms.

Overall, the motivations for companies to license out their technology to firms in developing countries are weaker than ever before.

Another significant barrier is that the global policy environment has become quite hostile to technology transfer as an enforced requirement on foreign companies. As part of the Washington Consensus, developing countries came under pressure to liberalize their economies – i.e. attract foreign investment by offering hospitable conditions, not by imposing conditions on investors. Most post-1990 investment treaties between countries explicitly prohibit governments from requiring technology transfer as investment conditions. So if a country tried to mandate technology licensing as part of foreign investment, it could be violating an international agreement. This legal status has a chilling effect: governments know that overt technology transfer conditions could invite trade disputes or investor-state arbitration lawsuits, so they functionally cannot do so.

Two arrangements have replaced technology transfer today. The first is foreign direct investment. FDI involves foreign companies directly investing in and operating a factory in the host country, rather than operating through a local partner. For example, a car manufacturer like Toyota might operate a plant in Pakistan directly, rather than outsourcing to a local partner as they might have in the past. This approach is the default approach to outsourcing in sectors where the manufacturing process is technologically complex and can’t be

The other arrangement that replaces technology transfer is simply having international supply chains. Rather than setting up a factory in a developing country to produce that input, a multinational firm can simply buy cheaper inputs from suppliers in that country. This arrangement is especially feasible when the input being supplied is not very technologically sophisticated, so that the supplier doesn’t need special training and equipment to supply it.

Between these two forms of investment, international firms can achieve their market access and supplier goals without the need for technology transfer – and with legal protection from countries that might push them to engage in technology transfer.

China is the only country that still pushes hard on getting technology transfer, through their controversial quid pro quo policy – that in order for multinationals to access the Chinese market, they need to transfer their technology to a Chinese partner firm, through a joint venture. China’s market is large enough that firms are actually willing to agree to these conditions. It’s no wonder that researchers estimate that this policy has helped China, while harming its FDI partners. But most countries don’t have the will or market size to enforce a quid pro quo policy, which is why technology transfer agreements have all but vanished.

Can foreign investment and supply chains achieve the same outcomes?

Can developing countries benefit from FDI and supply chains in the same way that they benefited from technology transfer?

Maybe. There is a large literature on the effects of FDI on developing countries, most of which estimates that it has positive effects. But these studies are not generally high-quality – they’re only slightly more careful than simple correlation. The challenge is that we don’t want to evaluate FDI against the standard of being better than no-FDI; we want to evaluate it against technology transfer. And while FDI certainly bleeds knowledge – domestic firms learn from being in proximity to foreign firms – it doesn’t do so nearly as efficiently as directly learning from those firms as part of technology transfer.

Supply chains present their own learning opportunities, because international firms have an incentive to share knowledge with their suppliers to make them more efficient, even in the absence of a technology transfer agreement. Alfaro-Urena et al (2022) show this effect in Costa Rica. They compare Costa Rican firms that sell to multinationals with comparable firms that don’t sell to multinationals, showing that the former group see larger growth in productivity and sales (including to non-multinational buyers) after the multinational enters. This is a microcosm of how participating in global value chains can upgrade domestic firms, by giving them buyers who will hold them to high-quality standards, and give them the training necessary to achieve those standards.

So while we are unlikely to recreate the power of mid-20th century technology transfer agreements, the main benefit to the recipient countries – having domestic firms learn from more advanced foreign firms – may still be achievable through FDI and supply chain linkages.

Conclusion

Technology transfer agreements powered the most successful industrialization stories of the 20th century. But they required conditions that no longer exist: geopolitical alliances willing to subsidize transfers, closed markets that gave developing countries leverage, and production networks fragmented enough that firms needed to create capable local partners. Globalization eliminated these conditions. Today, firms can access markets and production capacity without transferring technology, and international trade law makes it illegal for countries to enforce technology transfer requirements.

China’s continued use of technology transfer requirements shows that technology transfer still works to grow domestic industries, but it also reveals why other countries can’t copy that approach, without the market size and leverage that China has.

FDI and supply chains are likely not as transformative as technology transfer was, because they involve more muted transfers of knowledge. But they are still likely to be the best option on the table today.

The setting is a medieval kingdom. Through circumstances not relevant to this plot, the kingdom’s leading scholar has become an oracle, who now knows everything that there is to know. The king orders that he be placed under house arrest, to avoid his being targeted by a rival kingdom, or his escape, or even his slipping and falling on his head. Only carefully vetted individuals can visit him for advice. The oracle receives this arrangement with benevolent indifference.

The king’s advisors ask the oracle for stately advice, on taxation and war, on relations with other kingdoms. But it is an ambitious merchant who one day chooses to pursue a different line of opportunity. He tells the oracle that he wants to create a tonic that will cure a common and deadly disease, so that he can sell it to doctors and patients across the kingdom. The oracle gives him instructions for what ingredients to collect and how to process them. His advice helps the merchant create a health treatment that works miraculously often. As the merchant produces and sells this medicine, he accumulates even more riches, and the oracle becomes even more renowned.

Doctors throughout the kingdom come to learn from the oracle, to see how he was able to cure this disease. Most of them leave frustrated, telling others that the oracle can only speak of mysterious concepts such as “germs”. When pressed further, he can only link these concepts to other cryptic ideas like “proteins”, which are somehow linked to cattle, even though the disease is not caused by animals in any way. While most of these prospective students leave in pessimism that they can learn anything useful from the oracle, a few stay to study with him. The oracle does not object to this arrangement. Some of these students become adept healers on their own, while others abandon medicine and begin conducting studies with symbols that they cannot explain to their former colleagues.

The merchant is farsighted. He realizes that with the oracle’s instruction and his own resources, he can build things that the king could never imagine. He imagines palaces in the sky, using the stars as fuel. He imagines a farm that uses machines to produce food without human or animal power. He imagines horses that can travel fast enough to cross the kingdom in a day. He takes these visions to the oracle and asks for help in realizing them. The oracle accepts. The merchant hires thousands of artificers to carry out the oracle’s instructions, and they get to work on building the city of the future.

At this point, I can see two endings to this story. Initially, I conceived the oracle’s task as impossible. So I first imagined an ending where his help is insufficient. Components built in spring degrade by autumn, so that the artificers constantly have to return to the oracle for new consultations. Artificers come and go over the years, and each new person inherits a half-finished machine of such complexity that it takes years before they can understand everything they need to know in order to carry out the instructions. The project stagnates in perpetual memorylessness.

After decades of this delay, the oracle passes away from old age. His passing destroys the merchant’s vision for the city of the future. Resolved to find a successor, the merchant travels across the kingdom to appeal to all of the oracle’s students, and offers them a lifetime of riches if they can finish their teacher’s work. A few of them show promise in how much of the oracle’s teachings they have grasped, but they readily admit that it will be decades before they are capable of continuing the project. The merchant does not have decades to offer them; he himself passes away within a few years of the oracle.

With no one to lead the work or provide the instructions, the city of the future is abandoned. The half-constructed machines and structures become objects of curiosity for scholars, who use them to learn more about the physical world. These learnings do advance society, but their contribution is meager compared to what the merchant had imagined.

To me, this ending reflects the story that flows most naturally from this premise. But when I challenged myself to conceive of success for the merchant’s project, I imagined an interesting scenario.

In this second ending, the oracle successfully sees the project through. The artificers’ hands and the merchant’s management move in accordance with his blueprints, and they do build a city that is more advanced than any cities of the past, and (though they do not know it yet) will be more advanced than any cities of the future, for another thousand years.

But nobody who lives in the city of the future understands it. They live as children, cradled by intelligent design. Their rooms are warmed and cooled by invisible force, their food is created and prepared by anonymous mechanisms, and their lights glow without fire. Visitors from around the world come to see this impossible city, only to be baffled that nobody in this city can tell them how it works.

When the oracle eventually passes away, there is nobody who understands the city they live in. The oracle leaves behind detailed manuals on the complex maintenance required to keep the city functioning smoothly. An entire guild of artificers is trained using these manuals. They debate over its ambiguous terminology, over segments that appear to contradict each other, over situations that are not covered explicitly in the manuals. Over time, the manuals become scripture, the artificers become priests, the guild becomes a church, and the oracle becomes the city’s founding deity.

Each new generation inherits unreplicable prosperity, and the city carries on in perfect constancy for many lifetimes.

Subscribe now

It was 1972, and Alves was ready to leave Brazil. He had returned to his home country after finishing his PhD in agricultural economics in the US for what was meant to be a short stay – he had lined up a professorship at Purdue University. Along with a working group of Brazilian agricultural researchers, Alves proposed that the Brazilian government establish a new agricultural research corporation, Embrapa, that would focus on problems specific to the country and equip farmers with the knowledge and technology to transform Brazil’s stagnant agricultural sector. The government accepted the proposal and asked Alves to be Embrapa’s second-in-command. He declined. It was a rude shock, then, to hear his nomination to Embrapa announced on the radio only days later.

When Alves returned to Brasilia, the military government informed him that he could resign from his appointment, but then he would never work in Brazil again. Reluctantly, Alves accepted the position, intending to leave quietly after a couple of years. Instead, he stayed for twelve years, guiding Embrapa through its first decade and overseeing its transformation of Brazil’s agriculture.

This is the story of how Embrapa transformed Brazil from a food aid recipient into the world's agricultural superpower, and what it reveals about the marriage of science and political will. It’s a story about betting your budget on sending scientists to the US for their PhDs, even when you're under pressure to produce immediate results. It's about running TV ads with cattle marching into Sao Paulo to keep politicians from cutting research funding. It’s a lesson in how developing countries can build world-class research institutes that transform their economies, and it’s a lesson on how the US can turn science towards its industrial policy objectives.

Brazil’s agricultural transformation

In 1960, Brazil was poor, with a GDP per capita roughly equal to Nigeria's today. Despite being an agrarian economy, Brazil was actually a major food importer and even a recipient of US food aid.

But today, Brazil is the world's biggest agricultural exporter. While the country has not reached high-income status – its growth has been slow in manufacturing and services – its agricultural productivity growth has been extraordinary. Brazil is now the world's largest exporter of soybeans, coffee, sugar, orange juice, and beef, and a leading player in corn, cotton, and pork. This transformation has occurred primarily through yield growth, rather than through expanding farmland. Brazil has quadrupled yields across leading crops since 1960 – faster than the United States, despite the US’s much higher agricultural R&D spending.

There wasn’t a single slam-dunk cause for this transformation. Brazil adopted a subsidized credit scheme for farmers who adopted modern equipment/fertilizer/high-yielding seeds, which helped transform the way agriculture was done from being just about small-scale farms. They had functioning land markets, so more productive farmers could expand their area of cultivation and boost overall productivity. They had a robust fertilizer industry due to both protectionism and induced demand from subsidized credit. All of these factors would deserve their place in a full story of Brazil’s agricultural transformation.2

But these are all factors that development economists already emphasize when discussing how countries can modernize. They are part of the standard playbook: subsidize farmers to adopt modern inputs and fix market failures. Basic steps that fit a “backwards” country.

Establishing a world-class research institute to advance the frontier of knowledge in agriculture is not part of that playbook. Yet that is exactly what Brazil did in establishing Embrapa. And against the odds, they succeeded. That is why I’m telling the story of Embrapa today.

Embrapa’s history

Brazil’s status as a food importer and a recipient of food aid was a high-priority problem for its military dictatorship, installed in 1964. It made the country vulnerable to being internationally pressured, and it was a dangerous source of population unrest. As a result, securing Brazil’s food independence was a high priority for the new government.

A natural focus was to build out rural extension. Rural extension is the system through which agricultural knowledge and best practices are transferred from researchers to farmers through field agents who teach modern farming techniques. Brazil already had rural extension infrastructure from the 1950s, built with American collaboration. But this extension simply didn’t work. Before his PhD, Eliseu Alves worked at ACAR-MG, the rural extension corporation of Minas Gerais state, and evaluated the organization’s impact on farmers for his master’s thesis. He concluded that farmers with ACAR-MG’s assistance did no better than farmers without it. Remarkably, his supervisors and colleagues agreed with his conclusions, but simply shrugged – they didn’t know what to do better. It was an open secret that rural extension was not working in Brazil.

But why?

A working group of agricultural researchers, led by sociologist Jose Pastore and Alves, identified a lack of applicable knowledge as the problem. Rural extension was based on the knowledge taught by American agronomists: knowledge cultivated in Iowa and Wisconsin, not in Mato Grosso or the Cerrado. That temperate advice withered in the tropical heat, and extension agents didn’t know how to answer farmers’ questions.

The working group concluded that they needed to create new, Brazil-specific knowledge for farmers. Thus, they created the blueprint for a new organization: a state-owned company that would conduct agricultural research with a clear mission to solve Brazilian problems, respond to farmer demands rather than academic curiosity, and maintain its focus through centers organized by crops and biomes rather than scientific disciplines.

Embrapa’s first significant win came in the Cerrado. The Cerrado is a massive biome in western Brazil, three times the size of Texas, and almost all of it was available for farming. But farming was practically absent from the Cerrado, because its soil was too acidic and nutrient-poor for crops to grow, and its climate was hotter than Brazil’s subtropical south, where farming was widespread. To make the Cerrado arable, Embrapa started an agricultural liming campaign – treating the soil with limestone-derived chemicals to reduce soil acidity. They also developed a new soybean variety that was tolerant of the Cerrado’s harsh climate, unlocking new possibilities for growing soy, a profitable cash crop. Today, the Cerrado hosts 70% of Brazil’s beef cattle and 50% of Brazil’s soy production, both of which are key to Brazil’s agricultural exports. And Brazil’s center-west region, of which the Cerrado is the biggest biome, has become the largest area of Brazilian agriculture.

Embrapa continued to rack up victories: developing varieties of rhizobium bacteria that could fix nitrogen in tropical soils and reduce the need for fertilizer, cross-breeding African grasses to create pastures that produced twenty times the yield of cattle feed, and creating no-tillage crop management systems that preserve soil health. Since its founding, Embrapa has created over 350 cultivars and filed over 200 international patents. It hosts the largest gene bank in Latin America and one of the largest in the world. It is, in short, one of the most cutting-edge agricultural research institutes on the planet.

So how did Embrapa do it?

Lesson 1: Invest in human capital

Suppose you’re in charge of a brand-new research institute in a poor country, with the intimidating goal of transforming the country’s agriculture. You are under intense pressure to generate results, your organization’s continued existence is not guaranteed, and you have a tiny base of researchers to draw on – in 1975, Embrapa had only 28 researchers with PhDs. What would be your first priority?

You might throw all your researchers and money at one problem, prioritizing one area for a major push. That strategy could pay off – with a small but brilliant team, you could conceivably make some progress, thus living to see another fiscal year. But your ambitions would be strangled by the reality of having too few researchers to actually do research. Embrapa’s mission was to create knowledge with clear economic relevance to Brazil, and producing knowledge is really hard, let alone directing that research effort towards a specific topic area. (Just ask your favorite graduate student!) Without investing in your researchers, in a few years, your institute’s research outputs would dry up. Politicians would come knocking, asking you five things you did in the past week. Cue the game over screen.

What you would probably not try to do is immediately start spending huge sums of money on sending your research staff to do advanced research degrees, before they do any actual research for you. But that is what Embrapa did. In its first ten years, Embrapa spent 20% of its budget paying for its staff to get advanced degrees in agricultural science, at universities in the US and Europe. By 1988, it had more than 1,000 researchers with PhDs, most trained abroad on Embrapa’s dime. In the graph below, it’s visually obvious when all the researchers with bachelor’s degrees get master’s degrees, and then later get PhDs. (source)

It would be a minimum of 3-5 years before any of these researchers could return and do any useful research, making it a risky investment for an institution that might not even exist in 3-5 years. But Alves, who created Embrapa’s research training policy, was a sophisticated student of Chicago economics. His doctoral advisor, Edward Schuh, had done his own PhD under Gary Becker and Theodore Schultz at the University of Chicago – the originators of human capital theory, a view so ubiquitous among economists today that it isn’t even called a “theory” anymore. Human capital theory views people’s skills as analogous to physical capital – in particular as a resource which can be invested in, and which has compounding returns over time. Viewing people’s skills as capital pushes you to invest in building up a stock that will compound over the long run, rather than spending it all down at the very beginning out of desperation.

This was exactly the strategy Alves followed, in focusing on training. The quality and quantity of its researchers was essential to why Embrapa could make breakthroughs like liming the Cerrado and creating new soybean varieties that would grow in tropical soil – breakthroughs that would have been out of reach for Embrapa’s original research team, as small and untrained as it was.

An important part of this pipeline was ensuring that Embrapa’s research staff was actually well-equipped to benefit from this training. Embrapa started by hiring the brightest graduates from Brazil’s universities – not a difficult task, because they paid handsomely, and because the labor market was weak. The very best of these new hires were sent off to do advanced studies immediately; the rest were given a couple of years of training within Embrapa, learning on the job and through internal seminars. The internal skill-building also occurred through frequent collaboration with foreign researchers, who would give talks at Embrapa and advise its staff, especially on breakthroughs that had been made elsewhere that Embrapa could build on. This was a complete human capital pipeline, a system for generating large numbers of world-class researchers who could tackle Brazil's specific agricultural challenges.

Lesson 2: politics, politics, politics

Reading the section above, you might conclude that Brazil’s government was saintly in its wisdom and foresight, allowing Embrapa to make long-term investments like sending its research staff to do PhDs in the US.

You would be wrong.

When Embrapa began its training policy, it was vigorously opposed by the minister for agriculture, to whom Embrapa reported. But Alves called up a colleague who was close friends with the president, and the president overruled the minister, declaring that Alves called the shots at Embrapa.

This was part of a broader pattern; Embrapa’s leaders ensured that they were always friendly with people in power, so that they would be free to make patient or difficult decisions. This was essential, in an environment where even valuable agencies could be dismantled by political winds. That reality was made very clear by the case of Embrater, a sister agency that worked on rural extension and diffusing Embrapa’s knowledge to farmers. In 1989, an influential group of officials within Embrater supported Luiz Inacio Lula da Silva in the presidential election. But Lula lost, and the new president, Fernando Collor de Mello, promptly shut Embrater down. Embrapa avoided this mistake, always cultivating allies and never picking fights.

But avoiding partisan politics didn't mean avoiding politics altogether. Embrapa invested heavily in political capital through strategic communication. The institute created a sophisticated communications department, even paying for communications professionals to earn advanced marketing degrees. It published glossy press releases, hosted seminars for journalists to hear about Embrapa’s research, and mandated that scientists give interviews to reporters – a deeply unpopular policy with those scientists, but one that paid off in public visibility.

Embrapa even ran full-blown public advertising campaigns. On its 10th anniversary, Embrapa aired a campaign on state television showing a herd of cattle entering Sao Paulo, with the caption ‘the results of Embrapa’s research are coming to town.’ The message was clear: we at Embrapa are making your food cheaper, so keep us generously in your hearts and in your budgets. Alves put it bluntly: “We were always in the press publicizing [Embrapa]. Without publicity, we would be dead.”3

This focus on politics was built into Embrapa’s structure from the beginning, in the decision to establish it outside of the university system. Brazil’s American advisors had pushed for Embrapa to be housed inside the university system, echoing how America’s agricultural research ecosystem started from land grant colleges that evolved into universities. But Embrapa’s founders resisted. Partly because it would have made focused research difficult (more on that below), but also because Brazil’s universities were broadly antagonistic to the military dictatorship. If Embrapa had been associated with the university system, it would have been intensely scrutinized by the military, and susceptible to the partisan impulses that would later result in Embrater’s downfall. It was established outside the traditional research ecosystem in order to always be on the winning side of politics.

This vision contrasts sharply with how we normally think about science policy. In our mental model, the goal is for science to be insulated from political pressures, so that scientists can work on what is most important rather than responding to parochial concerns. But that was just not an option for Embrapa, who needed to be savvier than anyone who would see them defunded. Alves would never have been caught off-guard by politicians complaining about shrimp on treadmills.

Lesson 3: respond to market demand

An applied research institute has a more complex task than you might expect. “Do socially relevant research” sounds like a good guiding principle, but it could easily lead to a fragmented and ineffective mission. For one thing, Brazilian agriculture had a very large number of potential focus areas – many crops that were cultivated and could use research, many potential technologies that could benefit farmers if created. In addition, scientists left to their own devices will do research that they find interesting, and will argue its social relevance after the fact (just ask your favorite grad student again!)

So instead of letting their attention be fractured across too many areas or driven by the interests of its researchers, Embrapa leaned into induced innovation, an influential concept among agricultural economists.4 Put simply, agriculture has three inputs into production – land, labor and intermediates (e.g. fertilizer). To make agriculture more efficient, we need to focus our innovation on one of these three inputs – reducing the need for one of them, and thus reducing costs. But which one should we focus on? Induced innovation says that we should focus our efforts on the input that has the highest price. Intuitively, the biggest cost reduction comes from focusing on the most expensive inputs.

In Brazil, that meant focusing on land. Agriculture had flourished in the south, around Sao Paulo and Rio de Janeiro. But those regions were running out of frontier land as they rapidly urbanized. As a result, land was becoming expensive – and that land cost was showing up in the price of output, making Brazilian commodities more expensive and unable to compete in global markets. Thus, tackling the high price of land thus became Embrapa’s focus. By increasing the supply of arable land, they would reduce the price of agricultural land, both increasing the quantity that Brazil could produce and reducing the price it would sell for. This was the reason why they focused on making the Cerrado arable – the initial victory that established their value to Brazilian agriculture.

Organizational structure reinforced this market focus. Embrapa organized research centers by crop and biome rather than by discipline, making it easier to direct research towards specific priorities. There was pressure to make each center serve all crops grown in its region, but Embrapa resisted, knowing that specialization was essential for progress with limited researchers. This structure made it possible to respond to market signals that were crop-specific – for example, they could direct research towards new soybean varieties because of the rising global price of soy.

The focus on market demand gave Embrapa a clear direction, ensuring that every research breakthrough had a clear path to improving farmer productivity and Brazil's agricultural competitiveness.

So what can we learn from Embrapa?

Embrapa’s history is clearly a success story, but what does it tell us today about science policy?

I think the wrong lesson would be to follow Embrapa’s decisions. Those decisions were the product of its context. It invested in training researchers because Brazil needed more researchers. It prioritized increasing the amount of arable land over (for example) creating better fertilizers or improving crop management because land was the binding constraint that Brazil faced at the time. It was set up as a corporation to allow for more flexible contracting and higher salaries than could be found in government agencies. These decisions would not necessarily make sense in every country and context.

But what I think we should take away is the importance of the forces that Embrapa cared about. Consider the lesson of being politically conscious. Of course, what that means would look very different for (say) the National Science Foundation than what it looked like for Embrapa. Applied agricultural research is intrinsically easier to sell than basic research, even if basic research is more useful. So while it might fall flat for the NSF to run a Super Bowl ad about its research, we should consider it a virtue for our science agencies to be able to navigate political firestorms and advocate for themselves. We are living with the consequences of our scientific institutions being politically naive.

The other lesson I think we can draw is about the utility of directed research programs as an industrial policy tool. Embrapa was set up following an economic directive – to make Brazil a powerhouse in the agricultural sector. And it achieved that by building up the knowledge that Brazil needed to make its agriculture much more productive. This rhymes with the industrial policy motivations that countries have today. The US wants to become a world leader in producing chips, or to increase its manufacturing capacity? Instead of blunt force instruments like tariffs, it could create the largest directed research program in history to ensure that it has the scientific knowledge to produce more sophisticated goods than other countries. This has the advantage of being positive-sum, creating knowledge for the world as a whole, while still helping enshrine American advantage.

Vannevar Bush, the founder of the NSF, forcefully argued for it to focus on basic scientific research, with no direction other than that chosen by scientists themselves. “A nation which depends upon others for its new basic scientific knowledge,” he argued, “will be slow in its industrial progress and weak in its competitive position in world trade, regardless of its mechanical skill.” This position has weakened over time, and the NSF does fund applied research. But it remains verboten to simply direct research efforts – both funding and researchers – on a large scale towards policy priorities. I think this is a mistake, and Embrapa shows that applied research institutes can be powerful.

Subscribe now

But that didn’t happen. Funded by Western philanthropists and international organizations, scientists created new high-yielding varieties of rice, wheat, and corn, as well as new pesticides, herbicides, and fertilizers, in a concerted research effort known as the Green Revolution. These new varieties and techniques dramatically increased food production across the developing world and ensured that no Malthusian famines ever occurred. The figurehead of the Green Revolution, an Iowa plant breeder named Norman Borlaug, would go on to receive the Nobel Peace Prize and be credited with “saving more lives than any man in history.” It was a Promethean story – humans using our ingenuity to defy the limits that nature put on us.

But its legacy has been tarnished by critics who argue that input-intensive farming has destroyed land and water reserves, making people poorer and sicker. Now the rallying call among activists is for food sovereignty – the right to “healthy and culturally appropriate food produced through ecologically sound and sustainable methods” – the term being a pointed reference to the Western funding behind the GR. In the face of this criticism, the Alliance for a Green Revolution in Africa (a Gates-/Rockefeller-funded organization trying to promote the same model in Africa) dropped the “Green Revolution” from its name in 2022, now known simply as AGRA. It’s now a faux-pas in agricultural policy to advocate for another Green Revolution.

For a long time, I dismissed these criticisms without much thought. But that was because the Green Revolution is central to my personal narrative. Growing up in India – one of the epicenters of the GR – I learnt about it as one of the landmark events in India’s history. My parents could have grown up malnourished if not for the food abundance created by the GR. It gave me a visceral sense that science can literally save the world, a belief that has influenced my research path.

But it’s important to revisit the Green Revolution with clear eyes, both for me and for the world. Climate change threatens agriculture and food security today, and the GR remains the only historical example of rapidly scaling food production to match population growth – making it either an essential blueprint for future food security, or a cautionary tale about unsustainable intensification. And for me, the reassessment is personal; I need to know whether my belief in human ingenuity is actually built on shaky foundations. So let’s figure out the Green Revolution’s actual legacy.

Disclaimer: I am not an expert on the Green Revolution. I am just an economist who generally knows how to read papers. I was aware of some of these papers before I started this review, and not others. I come into this review as an outsider. This disclaimer matters because there are people who have spent literal decades studying the question I’m trying to answer. I’m not declaring that I know better – I just needed to find an answer for my own satisfaction.

Did the Green Revolution feed the world?

Let's start with the most basic question: did the Green Revolution increase agricultural productivity in the developing world? Yes. The simplest version of this argument is that cereal yields in the countries that were the epicenter of the GR increased dramatically, especially compared to yields in low-income countries.

In fairness, the fact that yields grew does not mean that the GR – and its particular package of high-yielding seeds and input-intensive farming – was responsible for those yield increases. Maybe the GR simply coincided with other policies that led to more productive agricultural sectors in countries, or maybe there was just a natural upward trend in yields. Is there any evidence that the GR caused that yield growth?

There is, in the form of a quasi-experimental study by Gollin et al (2021). They note that high-yielding seed varieties for different crops were released at different times. They use this as their natural experiment, comparing yield growth in crops that saw high-yielding seed varietals (HYVs) developed earlier, against crops that saw HYVs developed later. Assuming that these crops would have seen parallel trends in yields without the GR, this comparison tells us the true effect of the GR on crop yields. Gollin et al find that, on average across crops, HYVs increase yields by 9% within 10 years of their introduction, with effects climbing to 75% over 40 years. They argue that this gradual effect makes sense, both because adoption of HYVs is gradual, and because once there is a breakthrough, HYVs are improved upon by successive generations.

In short, we don’t have to be concerned about whether the basic concept of the Green Revolution is fictitious. At a bare minimum, it did increase food yields across the developing world.

Did the Green Revolution make people richer?

So we’ve established a baseline fact that the Green Revolution did increase food yields. The next step is to ask, did that yield growth make people and countries richer? Unfortunately, here is where the consensus ends.

Gollin et al argue that it did. They measure how “exposed” each country was to the Green Revolution at any point in time, based on the share of land that was planted with each crop, and whether that crop had an HYV seed released at that point. Intuitively, countries that grow more rice and wheat (where HYVs were released in 1963 and 1965) would see an earlier boost from the GR than countries that grow more sorghum and cassava (where HYVs were released in 1983 and 1984). So comparing countries exposed earlier to those exposed later allows us to measure the impact of the GR.

Thirty years after the start of the Green Revolution, more-exposed countries had a staggering 70% higher growth in GDP per capita than less-exposed countries. The authors argue that this is because the GR not only increased agricultural output (which directly increases incomes), but it also pushed people out of agriculture and into manufacturing/services employment (which indirectly increases incomes, since those sectors are more productive and thus pay better than being a farmer). Aggregating across countries, the authors calculate without the GR, the cumulative loss to global GDP would be $83 trillion! Taking this result at face value would suggest that the GR was one of the most important causes of development in the 20th century.

However, their measure of exposure reveals a thorny problem with their estimates. To see the issue, look at how the authors’ exposure measure is distributed across countries (note that the map is incomplete because they leave rich countries out of it):

The exact numbers are not important, but it jumps out that most of Africa has the lowest exposure – because HYVs for many important staple crops in Africa (sorghum, millets, cassava, etc) were developed decades later than HYVs for rice, wheat and maize. So when we compare income growth in “more exposed countries” to that in “less exposed countries”, we are implicitly comparing income growth in Asia/Latin America to income growth in Africa… and attributing those differences to the Green Revolution.

That seems bad! While exposure to the GR was certainly an important difference between Africa and the rest of the world, it wasn’t the only one. Attributing all the differences in income growth to the Green Revolution – not to colonial institutions, or to geography, or to any of the hundred and one theories for why Africa is poor – is just wrong. So the landmark study on the Green Revolution is uninformative about whether it accelerated economic growth in the beneficiary countries. Back to the drawing board.

How could incomes not increase?

You might say that I’m nitpicking over the study’s particular design issues – because if we accept that the Green Revolution increased yields, then it mechanically must increase incomes. If the additional produce is there, then someone has to be profiting from it, right? Wouldn’t that mean incomes are going up on average?

But there is an important subtlety that can lead agricultural productivity growth to be neutral or even negative for incomes. The question is about whether the higher yields pull people into agriculture, when those people would have otherwise moved into more productive jobs in the manufacturing or service sectors. If the Green Revolution caused more people to become farmers instead of moving to the city and finding higher-paying jobs, then it may have been neutral or even negative for incomes on average.

This is an important consequence of globalization, that wouldn’t occur in a country closed off to trade. After all, once you produce enough to feed everyone in a country, there’s no demand for more food, so there is no point in having additional people working in agriculture – so agricultural productivity growth would push people out of agriculture, with no tradeoff. However, when countries can export food, there is no “enough to feed everyone.” There are always customers for the food you produce, somewhere in the world. So making countries more productive at agriculture can make them specialize in exporting food to the rest of the world – which is bad for their growth in the long run, because specializing in agriculture means that a country has fewer people living in dynamic cities, fewer people working in the manufacturing and service industries that can serve as engines of growth.

This is exactly what Moscona (2019) argues happened during the Green Revolution. Like Gollin et al, he estimates how the agricultural productivity growth from the Green Revolution affected country incomes. But he uses a different natural experiment than Gollin et al. Rather than comparing countries with more vs less area planted in rice/wheat/maize, he uses the biophysical potential yield improvements from HYVs as his measure of exposure to the Green Revolution. Countries differ in their agronomic conditions due to their different geographies; some countries have environmental conditions that benefit more from HYVs than others, leading to differences in the impact of HYVs across countries. By comparing the income growth of countries with higher potential yield improvements from HYVs to that of countries with lower potential yield improvements, Moscona can estimate the impact of the GR.

Moscona estimates that the Green Revolution did not have any effect on national incomes on average. Actually, he shows that when you focus on the countries that were the most globalized, the Green Revolution had a negative effect on income – by increasing the agricultural labor force and reducing the speed of urbanization, just as in the story outlined above. This is why Moscona concludes that the Green Revolution did not increase incomes on average.

But Moscona’s methodology is also flawed, because of its focus on countries with the most potential yield improvements from HYVs. It’s not at all clear that these are the countries that actually saw the most yield improvements from the Green Revolution. Moscona doesn’t list the countries which have the highest value of this measure, but some of his tests suggest that his measure only weakly tracks actual benefits from the GR.1 So it’s really hard to take his results as evidence for anything about the GR.

The alternative approach: connecting yields to incomes directly

Both of the major papers studying the Green Revolution itself have deep flaws. What do we do with this? Instead of restricting ourselves to papers that study the Green Revolution specifically, we can learn more by looking at papers that study the relationship between agricultural yield increases and income growth more generally. The most helpful study in this vein is McArthur and McCord (2017), who estimate how agricultural yield increases affect incomes within a country. They use the fact that different countries have different access to fertilizer, based on how far they are from the factories where fertilizer is produced and exported to the world. So countries with more access to fertilizer have higher yields – and they use this access to fertilizer as a natural experiment for whether higher yields do in fact translate into higher incomes. If these regional differences really are only about access to fertilizer (and not, for example, being generally more central and connected), then their income differences are actually explained by the higher yields induced by more fertilizer access. They estimate that increasing cereal yields by half a ton/hectare increases GDP per capita by 15%.

How do we map this back to the Green Revolution? The graph of cereal yields shown above from Our World in Data roughly implies that the Green Revolution increased cereal yields in the most affected countries (India, Mexico, Philippines) by 1 ton/hectare within 30 years. So McArthur and McCord’s estimates would imply that it increased national incomes by 30% over that period. This is much lower than Gollin et al’s estimate of 70% over the same period, but it is still substantial. I think this estimate is inflated because of design issues with McArthur and McCord’s study2, but it is still evidence in favor of a positive effect of the GR on incomes.

So I still feel alright concluding that the Green Revolution increased incomes on average, although this is much less certain than Gollin et al make it seem, and could definitely vary across countries like Moscona argues.

Did the Green Revolution make people healthier?

You might be thinking, “Karthik, we know you’re hopelessly tainted by the economic prejudice, and you can only value dollars and cents, but actually the Green Revolution has been disastrous/amazing for human health, and that’s what matters.” So let’s figure out whether either of those things is true.

Once you accept that the Green Revolution increased agricultural yields, it is hard to reject that it increased the amount of calories that people consume and reduced malnourishment. One particularly legible metric of health improvements from more calories is infant mortality. The main paper estimating this impact is von der Goltz et al (2020). They estimate the relationship between HYV adoption and infant mortality across 37 developing countries, using a strategy similar to Gollin et al – they compare subnational regions with different crop mixes to measure different exposure to the Green Revolution. Regions with more rice/wheat/maize were exposed earlier to the GR, while regions with other crops were exposed later, and comparing infant mortality trends between these two types of places tells us the impact of the GR on infant mortality. The results indicate that the GR reduced infant mortality by 2-5 percentage points – a large decrease given the baseline mortality of 18% in the early 1960s. While they can’t pinpoint the source of this infant mortality reduction, it is intuitive to me that pregnant mothers consuming more calories would make their babies less likely to die – and that there would be more calories for the babies themselves. This effect is the most visceral demonstration of the general health benefits of having more calories because of more food production.

But there is also evidence of long-term health harms from the same dietary changes. Singh (2025) estimates the health impacts of the Green Revolution by comparing districts within India. Like Moscona, she compares districts that have high agronomic suitability for rice and wheat HYVs to those with low suitability, before and after 1966 (when HYVs were introduced) to identify the effects of the GR. The important assumption is that without the introduction of HYVs, health outcomes in these high-suitability and low-suitability districts would have evolved in parallel. If this assumption is true, then the difference in how health outcomes evolve can be attributed to the Green Revolution. She shows that high-suitability districts saw a move towards monoculture, producing more rice and wheat at the expense of legumes and millets, which offer more protein and micronutrients. In nutritional terms, calorie production increased by 20%, but entirely from carbs; proteins and micronutrients declined as a share of calories. She shows that people exposed to the Green Revolution during early childhood in high-suitability districts are on average 0.3 cm shorter as adults, have 3 pp higher rates of hypertension and 1.5 pp higher rates of diabetes. So the criticism that the Green Revolution has made people sicker through monoculture is likely to be true.

Unlike the papers about the Green Revolution’s income effects, these papers do not directly disagree, so we don’t have to really dig into their methodology – for what it’s worth, I think both of them are decent, and substantially improve on both Gollin et al and Moscona because they focus on comparing regions within countries.3 This means that the concerns about comparing Africa vs Asia/Latin America don’t apply to these papers.

Comparing these two effects is an unenviable task. On the one hand, fewer infants dying from undernutrition; on the other hand, more chronic disease and stunting. Which of these is more important? There’s no way I can answer that. It might be possible in theory to aggregate the health burden caused by worsening diets and compare it to the benefits of reduced child mortality. But I’m definitely not going to be the person to take on that challenge. So I’m going to acknowledge that the critics have a point, and write this issue off as “the effects on health were mixed.”

Did the Green Revolution preserve the environment?

In reading science papers on the Green Revolution, I was surprised to see that the most commonly debated aspect of it is not whether it reduced poverty or improved nutrition. The most common debate is about the environmental impacts of the Green Revolution.

Norman Borlaug, the progenitor of the Green Revolution, argued for his efforts primarily on the basis that they would reduce the amount of land needed to be converted for agricultural use. His logic was that in order to meet world food demand, we could either increase the amount of land used for agriculture, or we could increase yields. By increasing yields, we have spared hundreds of millions of hectares from being placed under cultivation – and the resulting harms from deforesting that land. This land sparing means large amounts of deforestation prevented, less carbon emissions from the farming on that land, and less damage to ecosystems or biodiversity on that land. This idea, now known as the “Borlaug hypothesis”, makes sense to me as a clear environmental benefit of the Green Revolution. This graph from Our World in Data demonstrates how large this effect could be:

The argument of the OWID graph is essentially as follows: we have tripled cereal yields since 1960 while keeping land constant, for a tripling of total cereal output. In order to achieve current world production at 1960 yields, we would have needed to triple the amount of land used for cereals. This would mean adding around 1.5 billion hectares, or 10% of the Earth’s surface area, just for cereals! The environmental impact of that land use would have been devastating (deforestation, groundwater depletion, etc) and thus the Green Revolution spared us from environmental catastrophe.

But that argument is false in its simple form – it assumes that the world food production level we have today is exactly what we would have had without the Green Revolution. There’s no reason that would be true. Food is cheap today, so we consume a lot of it. But without the GR and its yield improvements, food would be expensive, so world food demand would be lower – which means that land use for agriculture would also be lower, since it’s meeting a lower demand. Thus, we need to explicitly model how food supply and demand would be different in a no-GR scenario, to estimate exactly how much land has been spared.

Stevenson et al (2012) do exactly this. They simulate the global economy, imagining how production in agriculture and non-agriculture across the world would respond to a yield reduction. They find that if yields were reduced to the pre-GR level, the total land under cultivation today would increase by 2% – much less than the 300% implied by the OWID graph. The reason is that in order to incentivize land conversion to agriculture in this scenario, food prices have to increase. But if food prices went up, then food demand would fall, so we wouldn’t need as much land to meet that demand. Indeed, part of their simulation result is that world food prices would be 20% higher without the GR. This would reduce world food demand by so much that just a 2% increase in cropland would be enough to meet it. Thus, they argue that the Borlaug hypothesis is generally true, but it is much less significant than you would guess from a simple exercise like the one displayed in the OWID graph.4 Thus, the environmental benefits of the Green Revolution were real but modest.

Stevenson et al is an example of a common type of agricultural economics paper – one that draws its conclusions from a massive simulation model of the world economy, a model with so many parameters and judgment calls that it’s hard to audit the results. I have a hard time evaluating these papers. On the one hand, my economic training places a high premium on papers making only a few important assumptions, and justifying those assumptions clearly, so that anyone can evaluate whether they believe them. On the other hand, reality doesn’t always cleave to simple assumptions, and having the aesthetics of simplicity doesn’t make a story more likely to be true. So I extend some deference to this literature, and I buy that the land-sparing benefits of the GR are small.

While Stevenson et al’s bottom line is generally supportive of the Borlaug hypothesis, not all research I found was supportive. For example, Rudel et al (2009) look at country-level data over time, and they estimate no correlation between increases in yield and decreases in cropland within a country. They conclude that the Borlaug hypothesis doesn’t hold generally. Many papers adopt approaches like this – they look within some country/region and estimate the relationship between yield changes and cropland changes. The problem with this approach is that it only captures decreases in cropland within the country that saw a yield increase. But for thinking about environmental effects, we care about effects on global cropland, not effects on cropland in any particular country. Indeed, the most plausible pathway for the Borlaug hypothesis is that countries seeing yield increases (and thus exporting more food) would allow other countries to reduce their cropland. So this country-level approach is worse at capturing the land-sparing effects of the Green Revolution than the simulation approach of Stevenson et al, and I don’t put much weight on it.

As for potential environmental harms: critics focus on the general features of input-intensive agriculture: that it involves a lot of groundwater usage (leading to depletion and aquifer collapse), and that intensive fertilizer usage damages soil health. These facts are not really contested by supporters of the GR, so the question is simply what we should do with them. I think these issues are subordinate to the land sparing question. If we had to farm more land in the absence of yield increases, we would have to use water regardless; we would have to enrich the soil with some kind of fertilizer regardless. As far as I can tell, it doesn’t matter whether that water/fertilizer is used all in one place, or used across many places. So to the extent that the Green Revolution reduced land usage, it probably also reduced the environmental pressures caused by groundwater usage and fertilizer application.

Conclusion

I came away from this evaluation a lot less enthusiastic than before. Not because it led me to conclude that the Green Revolution was bad! At the end of the day, my bottom line is that the GR increased world food production, it made people richer, it had mixed effects on health, and it had small positive effects on the environment. That’s a pretty solid record, and if I could press a button to duplicate that record with all its harms and benefits, I would.

But that solid record doesn’t live up to the mythology that I had in my head – that I imagine many people have. In my head, the Green Revolution was a victory over nature; I wanted the victory to be total, with no costs on our side. But it seems that the victory was exaggerated and there were some costs. It’s hard not to feel some disorientation at that new picture.

I’m now more sympathetic to the development policymakers who shy away from the “Green Revolution” branding on agricultural policy. I can see that the label promises too much, and comes with too much baggage. We want the Green Revolution to be a heroic tale, but it’s only great policy.