AI Adoption Raises Employment in Some Countries Not Others

The AI Employment Paradox: Why Some Workers Win and Others Lose

In 2022, economists Alexandre Georgieff and Raphaela Hyee did something rare. They looked at the data on AI and employment across 23 OECD countries and found that the answer to "does AI kill jobs?" is not yes or no. It is "it depends on who you are and what tools you have."

Their paper, published in Frontiers in Artificial Intelligence, is a cold splash of reality in a debate that has been mostly hot air. The authors took an existing measure of how exposed different occupations are to AI—the Felten, Raj and Seamans indicator, which tracks how much an occupation relies on abilities where AI has made the most progress—and applied it across countries. Then they looked at what happened to employment in those occupations.

The headline finding is unsettling in its nuance: there is no clear relationship between AI exposure and employment growth overall. None. But when you split the data by how much computer use an occupation already requires, two completely different stories emerge.

The Great Divide: High Computer Use vs. Low

For occupations where people already work heavily with computers, greater exposure to AI is linked to higher employment growth. The authors found that in these jobs, AI seems to act as a complement, not a replacement. Workers with strong digital skills can use AI to do more, better, faster. They shift toward higher value tasks. Productivity rises. Demand for their labor increases.

But for occupations where computer use is low, the opposite happens. There is suggestive evidence of a negative relationship between AI exposure and growth in average hours worked. These are jobs like manual labor, retail, hospitality, and other roles where digital skills are minimal. For these workers, AI does not amplify their abilities. It automates parts of their work without giving them new tools to adapt.

Georgieff and Hyee (2022) describe this as a kind of digital skills filter. Workers who can interact efficiently with AI capture the productivity gains. Workers who cannot may find themselves displaced or working fewer hours.

This is not a story about AI being good or bad. It is a story about who gets to ride the wave and who gets drowned by it.

How They Actually Measured This

The paper is methodologically careful, which matters because the AI employment debate has been plagued by sloppy thinking. Georgieff and Hyee did not invent a new measure. They adapted an existing one from Felten, Raj and Seamans, which scores occupations based on how much they rely on abilities that AI has recently improved. Think pattern recognition, language processing, visual classification. Then they mapped these scores onto employment data from 23 OECD countries.

The key innovation was splitting occupations by computer use intensity. They used data on how frequently workers in each occupation use computers on the job. This simple split revealed the hidden pattern. Without it, the overall relationship looked flat. With it, the two opposing forces became visible.

The sample covered a wide range of countries: from the United States and Germany to South Korea and Mexico. This cross country approach is important because it shows the pattern is not unique to one labor market or policy environment. It appears to be structural.

What This Means for Different Types of Workers

Imagine two workers. One is a data analyst in a financial firm. She spends her day in spreadsheets, databases, and visualization tools. When AI tools become available, she can use them to automate routine data cleaning and focus on strategic analysis. Her value increases. Her employer wants more of her.

The other worker is a hotel cleaner. His job involves physical tasks: making beds, scrubbing bathrooms, vacuuming carpets. AI does not give him a better mop. It might automate the scheduling system that tells him which rooms to clean. It might power a robot that vacuums hallways. But it does not make him more productive at the core tasks. If anything, it reduces the need for his labor.

The same AI technology produces opposite outcomes for these two workers. Not because one is smarter or more deserving. Because one has the digital infrastructure to absorb and benefit from the technology, and the other does not.

Georgieff and Hyee (2022) offer a mechanism for this divergence: "Partial automation by AI increases productivity directly as well as by shifting the task composition of occupations toward higher value added tasks." In plain English: AI does not replace entire jobs. It replaces some tasks within jobs. Workers who can pick up the remaining higher value tasks win. Workers who cannot, lose.

Why This Challenges Both Optimists and Pessimists

The AI employment debate has two camps. The optimists say AI will create new jobs and boost productivity, just like previous technologies. The pessimists say this time is different and mass displacement is coming.

This paper suggests both are right, but only for specific groups. The optimists are describing the experience of high computer use workers. The pessimists are describing the experience of low computer use workers. Neither camp has a complete picture because they are looking at different populations.

The paper also challenges the idea that AI will naturally lead to a "hollowing out" of middle skill jobs. That pattern happened with previous automation technologies like industrial robots. But AI is different. It affects cognitive tasks, not just manual ones. The divide is not about skill level per se. It is about digital literacy and the ability to work alongside software.

A highly skilled surgeon with poor digital skills might be more vulnerable to AI displacement than a moderately skilled programmer who lives inside a terminal. The key variable is not education level. It is computer use intensity.

The Countries That Benefit vs. Those That Don't

Because the paper covers 23 OECD countries, it allows for some cross country comparison. The pattern holds broadly, but with variation. Countries with higher overall computer use in their workforce see more of the positive employment effects from AI. Countries where many workers have low digital skills see more of the negative effects.

This creates a feedback loop. Countries that invest in digital infrastructure and training see AI as a net positive for employment. Countries that neglect digital skills see AI as a threat to jobs. The technology itself is neutral. The context determines the outcome.

The United States, with its high computer penetration and flexible labor market, appears to be in the first group. Some European countries with more rigid labor markets and lower digital adoption in certain sectors may be in the second. The authors do not make strong causal claims about country level differences. The sample size is too small for that. But the pattern is suggestive.

What This Research Does NOT Prove

This is a correlational study. Georgieff and Hyee (2022) are careful to note they cannot prove causation. It could be that occupations with high computer use were already growing for other reasons, and AI exposure is correlated with that growth without causing it. The authors call their findings "suggestive evidence," not definitive proof.

The paper also does not measure the quality of the jobs created. It looks at employment growth and hours worked. But a job created in the gig economy with no benefits is different from a stable full time position. The authors do not distinguish between these.

There is also the question of time horizon. The data covers a period when AI was advancing rapidly but had not yet reached the capabilities of large language models like GPT 4. It is possible that newer AI systems, which can handle more complex cognitive tasks, will shift the pattern. The paper is a snapshot, not a prophecy.

Finally, the measure of AI exposure is based on occupational abilities, not actual adoption. An occupation might be highly exposed to AI in theory but not actually using AI tools in practice. The paper assumes that exposure leads to adoption, but that link may be weaker in some contexts.

These limitations do not invalidate the findings. They just mean we should treat them as a strong signal rather than a final answer.

The Policy Implications Nobody Is Talking About

Most policy discussions about AI and jobs focus on universal basic income or retraining programs. This paper suggests a more targeted approach is needed.

The workers most at risk are not the lowest skilled in absolute terms. They are workers in occupations with low computer use. A construction worker might have more physical skill than a data entry clerk, but the data entry clerk is in a high computer use occupation and therefore more likely to benefit from AI. The construction worker is in a low computer use occupation and more likely to be hurt.

This means retraining programs should focus on digital literacy, not just general education. Teaching a hotel cleaner to use scheduling software might be more valuable than teaching them a new trade. The goal is to move workers from low computer use occupations to high computer use ones, even if the new occupation is not much higher in traditional skill level.

It also means the digital divide is not just about access to the internet. It is about the ability to work with software in a way that makes you complementary to AI. Workers who can code, use data tools, or operate digital platforms will see their value rise. Workers who cannot will see their value fall.

What This Actually Means

• ▸Digital literacy is the new dividing line in the labor market. AI does not replace all workers equally. It rewards those who can work with software and penalizes those who cannot. The gap between high and low computer use occupations will likely widen.

• ▸Retraining programs should target computer use, not just skills. Teaching a displaced worker a new trade is less useful than teaching them to work with digital tools. The goal is to move them into occupations where AI is a complement, not a competitor.

• ▸Countries that invest in digital infrastructure will see AI as a job creator. The same technology that displaces workers in a low digital adoption country can boost employment in a high adoption country. Policy choices matter more than the technology itself.

• ▸The AI employment debate is asking the wrong question. It is not "does AI destroy jobs?" It is "which jobs and for whom?" The answer depends on the worker's digital environment, not just their occupation or education level.

• ▸This pattern may intensify with newer AI systems. If future AI can handle more complex cognitive tasks, the advantage of workers with digital skills may grow even larger. The window to act on these findings is now, not later.

The Georgieff and Hyee (2022) paper does not settle the AI employment debate. It reframes it. The question is no longer whether AI will help or hurt workers. It is which workers we choose to help and how we equip them to survive the transition. The answer is not written in the code. It is written in the policy decisions we make today.