AI Is Redrawing the Jobs Map. Much of the World Is Still Missing

America fears replacement, China fears joblessness, and Japan fears having no one left to hire.

March 5, 2026. Anthropic published a report on AI and jobs.

April 2026. OpenAI released one of its own.

June 30, 2026. OpenAI published another.

Four months. Two labs. Three reports.

I read all of them. After I finished, my back went a little cold.

It wasn't the numbers that scared me. In fact, all three reports were careful with their language. There's no sign of systemic unemployment rising. Something else got to me. When you put the three reports next to each other, you see it: AI replacement isn't the kind of switch I thought it was. It's transmission. And when it transmits to different countries, the map it draws looks completely different.

I live in Japan. None of the three reports mention Japan. China only shows up in a few lines of a Stanford AI survey report.

None of these three maps is mine.

Four months, three reports: AI replacement research is evolving

Let me line up the three reports on a timeline.

March 5: Anthropic went first with Labor Market Impacts of AI. They used a new metric they call Observed Exposure. The idea was simple. Don't just look at whether AI can theoretically do a job. Look at how AI is actually being used in real work. They combined Claude's actual usage data on Claude.ai and the API with the U.S. Bureau of Labor Statistics (BLS) occupational data. The result was an exposure score that actually lines up with BLS employment projections.

April: OpenAI's Office of the Chief Economist released The AI Jobs Transition Framework: Mapping AI's Near-Term Impact on Jobs. They thought Anthropic's approach wasn't enough. High exposure doesn't automatically mean a job will be replaced. A teacher has high exposure in ONET task ratings. (ONET is the occupational dictionary maintained by the BLS. Every occupation has a task list attached to it.) But the teacher still has to stand in the classroom. So OpenAI added two more dimensions on top of exposure:

One is Human Necessity. Does this work need someone physically present? For regulation, for relationship, for physical action.

The other is Demand Elasticity. When the price of the service drops, does demand go up?

When you layer those two on top of exposure, 900+ occupations fall into four categories.

June 30: OpenAI took the same methodology to the European Union, releasing The AI Jobs Transition Framework for the EU. They used Europe's own occupational dictionary, ESCO (2,609 occupations), and Eurostat (the EU's employment statistics, in the same league as the BLS) for the data, and ran the whole thing again.

Four months, two labs, three reports on AI and jobs, and the methodology kept upgrading each time. That already tells you something. AI replacement isn't a static conclusion. It's a moving research target. In March, they were still using a single metric. By April, they'd added dimensions. By June, they'd crossed the Atlantic. The pace at which these research tools evolve is itself a reflection of how complicated AI replacement really is.

OpenAI's chief economist wrote something in the report. The gist was:

"Near-term labor market impacts may be gradual. Long-term impacts could be large."

They admit they haven't finished drawing the map.

Two labs, four months, three reports. Three versions of the map. Each one more detailed than the last. But all of them still being drawn.

US view (1): Anthropic's "observed exposure"

Start with Anthropic.

One number jumped out: the actual exposure for computer programmers is 75%. Highest of any occupation. Then customer service representatives (around 70%), then data entry workers (67%).

What do these three jobs have in common? White-collar. Text-based. Remote-friendly. This is Anthropic's own "hardest hit" list.

But they also found the other end of the spectrum: 30% of workers have zero exposure. Their work doesn't show up in Claude's usage data at all. Cooks, motorcycle mechanics, lifeguards, bartenders, dishwashers, locker room attendants. All at 0%.

0% doesn't mean these jobs don't matter. It means these jobs either require being physically present, or they require real-time judgment. AI can't reach them yet.

There's another data point worth looking at. Anthropic compared their own platform data to the task-level scoring from Eloundou's 2023 paper. 97% of Claude's actual usage falls on tasks Eloundou scored as "theoretically feasible."

In other words, that paper from three years ago, the one a lot of people questioned at the time, was right. AI really did do the things it said it could do.

Anthropic also did a fine-grained breakdown. They took a category like "programmers and mathematicians" and looked at it on its own: theoretically feasible 94%, Claude's actual coverage only 33%. Translation: most of what AI can do, most people haven't actually started using yet.

Anthropic opened the report with an honest line:

"All past attempts to forecast AI's impact give us reason for humility."

Then they followed with this:

"The impacts of AI, however, might be less like COVID and more like the internet or trade with China."

AI's impact might not hit like COVID. More like the internet, or like the structural shift in US-China trade. Gradual, but persistent.

The internet took 15 years to eat retail. AI might be faster. But the path is the same.

Programmers at 75% exposure are already in the middle of that process.

US view (2): OpenAI's "four quadrants"

Now OpenAI's report.

They thought Anthropic's approach wasn't enough. Just looking at how AI is actually being used in the workplace isn't enough. You also have to ask: does this work need someone physically present?

So on top of exposure, they added Human Necessity and Demand Elasticity.

Human Necessity comes in three flavors:

• Regulatory / accountability: judges, lawyers, court reporters. Someone has to sign and be responsible.
• Relationship: teachers, nurses. Face-to-face trust and care.
• Physical: plumbers, physical therapists, firefighters. They have to actually be there.

Demand Elasticity asks: when the service gets cheaper, does demand go up?

• Firefighters: -0.3. You don't burn your house down more often just because firefighters are cheap.
• Home care: -0.7. You don't buy much more even if it's cheaper.
• Editors: -1.0. Cheaper means you'd hire a few more.
• Graphic designers: -1.5. Cheaper means you'd use them more.
• Software developers: high elasticity. AI gets cheaper, demand goes up.

When you layer those two on top of exposure, OpenAI sorted the 900+ occupations in the US, covering 152 million jobs, into four categories:

• 18% High Automation Risk. High exposure, low human necessity, low demand elasticity.
• 24% Job Restructuring. High exposure, high human necessity, demand can't keep up.
• 12% AI-Augmented Growth. High exposure, high demand elasticity. Cheaper AI means more demand.
• 46% Limited Near-Term Impact. AI can't reach them now. Don't treat that as permanent insulation.

When you read that 18%, you might breathe out. Not 80%. Not 50%.

Don't relax too fast. OpenAI then tells you 24% is "restructuring."

Restructuring isn't replacement. Your job is still there. But it's been reshaped by AI. What three people used to do, one person with AI can do now. A job that took 8 hours takes 4. The other 4 hours go to something else. This isn't "unemployment." This is the content of the work changing. But after the change, you might not need as many people as before.

OpenAI wrote a sentence themselves:

"An AI-exposed occupation may still remain human-led if regulation, trust, liability, care, or physical-world constraints keep a worker at the center of the service."

Translated: even if AI can do a lot of things, as long as regulation, trust, liability, care, or physical constraints keep a person in the middle of the service, the job won't be replaced.

Sounds good. But here's the problem. A lot of these constraints are slowly breaking down.

OpenAI also has a more striking chart. They call it Capability Overhang. The data:

• High automation risk occupations: theoretical exposure 91.0%, actual usage 22.8%. A gap of 68.1 percentage points.
• AI-Augmented Growth occupations: theoretical exposure 92.8%, actual usage 24.6%. A gap of 68.2 percentage points.

What does 68 percentage points mean?

In my own words: the fire is already at the door. You just can't see the smoke yet. This isn't "hasn't happened." This is "happening at a speed you can feel, and it's coming for you."

My wife used Claude Code to build a semi-automated product listing system for our e-commerce business. It pulls data from other platforms, converts formats, generates listings, waits for human confirmation, then batch-uploads. She doesn't know programming, has never been a product manager, but she did it.

Six months ago I wrote listings, wrote code, organized my daughter's grades by myself. Now I do those things with AI. It's not the same job anymore. But it still goes by the same name.

EU view: same methodology, different 14%

Four months after the US report, OpenAI took the same methodology to the EU.

Using Europe's ESCO occupational classification (2,609 occupations) and Eurostat employment data, they ran it again.

The results:

• 14% High Automation Potential. Four percentage points less than the US's 18%.
• 27% Job Restructuring. Three percentage points more than the US's 24%.
• 12% AI-Augmented Growth. Same as the US.
• 47% Limited Near-Term Impact. Same as the US.

Same lab. Same methodology. Same team of researchers. A four percentage point difference.

What are these four points? Statistical noise? No. They're structural.

OpenAI explained it themselves:

"Europe employs relatively more people in manufacturing, skilled trades, transport, care, education, and public-service occupations, where work is often place-based, physical, or tied to relatively fixed service volumes. The U.S. has relatively more employment in managerial, sales, and digitally deliverable business-service occupations."

Translated: Europe has more jobs in manufacturing, skilled trades, transport, care, education, and public service. These jobs are tied to a place, require physical presence, or are bound to fixed service volumes. The US has more jobs in management, sales, and digitally deliverable business services. AI rewrites those a lot faster than it rewrites the assembly line.

The EU report also included a piece of data the US version didn't have. A breakdown of EU occupations by "Human Necessity":

• 49% Physical Necessity (nurses making rounds, plumbers showing up at the door, teachers standing in classrooms. AI can't reach these things.)
• 28% Regulatory / accountability necessity (lawyers, customs officials, linguists)
• 9% Relationship necessity (crisis social workers, middle school teachers)
• 14% Residual (customer service reps, employment counselors)

Nearly half of EU jobs are physically out of AI's reach.

That's why Europe comes in four points lower. Not because AI is stronger or weaker. Because of what the people standing at the door actually do. In the US, AI replaces the person writing emails. In Europe, AI replaces the person reading the meter. The second job barely exists in Europe.

And here's something even more interesting: within the same EU, 27 member countries split apart dramatically. Pick three extreme countries and look:

Luxembourg is the outlier. Finance and professional services are dense. 22% AI-augmented growth, 41% restructuring. When AI comes to Luxembourg, it doesn't replace people, it slots them in differently.

Germany's 17% high automation (one of the highest). Manufacturing-heavy.

Romania's 59% limited near-term impact (the highest). The industrial structure hasn't caught up with AI.

The other 24 countries don't gap that wide. But when the same methodology gives you a 10+ percentage point spread inside one union, that's already enough to make the point.

You thought the EU was one report? It's 27.
You thought US vs EU was "two maps"? It's 29.

Two reports point to the same thing: capability isn't replacement

The two reports came at it from different angles, but they landed on the same conclusion: AI replacement isn't a switch. It's transmission.

Anthropic found that in the "programmers and mathematicians" category, theoretical feasibility was 94%, but Claude's actual coverage was only 33%.

OpenAI found that in high-automation-risk occupations, theoretical exposure was 91%, but actual usage was only 22.8%. A capability overhang of 68.1 percentage points.

Both reports point to the same thing. AI replacement is a transmission process, not a switch process. Capability is there. That doesn't mean it's already transmitting into the labor market. Transmission takes time. It takes companies rebuilding themselves. It takes organizations reshaping. It takes people learning to use the tools.

Anthropic also surfaced an early signal: for workers aged 22 to 25, the hiring rate in high-exposure occupations has dropped 14% compared to 2022. Just barely statistically significant.

Unemployment rate hasn't gone up. But 22-year-olds can't get in. This isn't a switch story about "AI replacing you." This is a story about the door closing first. A closed door is an earlier signal than a layoff.

Stanford's AI Index 2026 pushes it further. US employment of 22-25 year-old software developers is down about 20% from 2024.

Put the two data sets side by side. The 22-25 age band is the real leading indicator of AI replacement. The unemployment rate is a lagging indicator. Hiring of young people is the early signal.

If I'm hiring right now, I almost certainly wouldn't hire someone who only knows the old way. I'd hire someone who knows how to work with AI. But those people are rare. Can't find the right hire, the position stays open. From the statistics, that looks like "we just haven't filled the role," not "replaced by AI." But at the root of it, AI has made hiring more selective.

I have a rough read on where I sit. I'm a heavy user of Hermes and Codex. I run cross-border e-commerce. I write this blog. I do content. Drop me into OpenAI's framework. Content creation and operations are high-exposure tasks (ChatGPT can already write, Claude can already edit). But business judgment, customer relationships, cross-cultural communication are human-necessary. I land in the 24% Job Restructuring bucket.

My wife's e-commerce listing system. A one-person business. High-exposure tasks, very high demand elasticity. She lands in the 12% AI-Augmented Growth bucket.

Friends I know doing pure translation, pure illustration, pure copywriting. If they're employees, they land in the 18% high-automation risk bucket. If they freelance, they may actually land in 12%.

Same job title. Employee or self-employed. Completely different fate.

East is uncharted: three fears across three economies

The two labs' research, taken together, covers 1 US state-system plus 27 EU member countries. About 800 million people.

Out of 8 billion people on the planet, fewer than 1 billion have been mapped. The other 7 billion are walking someone else's road.

US and EU research assumes inflation, tight labor, and well-functioning institutions. China and Japan don't sit in that environment. The map doesn't copy-paste.

China: will AI be a "pressure amplifier" for deflation and unemployment?

I won't quote the official numbers on China's employment problem. I'll use three side-door confirmations: the size of flexible employment, the actual employment rate of college graduates, and social financing credit data.

First: flexible employment.

The China New Employment Forms Research Center's 2025 China Blue-Collar Group Employment Research Report, released in June 2026, shows that as of 2025, China's flexible employment population has reached 280 million, with projections to break 320 million in 2026, accounting for 44% of urban employment. It has shifted from a "supplementary form" to a "major pillar."

What counts as "flexible employment"? Delivery riders, domestic workers, ride-hailing drivers, couriers, truck drivers, livestreamers, cleaners, security guards. Work an hour a week and you count as employed. 320 million people, 44% of urban employment. Sounds "flexible," but put it another way. 44% of China's 725 million employed population doesn't have a stable job.

Among the flexible employment population, the share of people with a college degree or above is climbing every year. A lot of white-collar workers and college graduates who originally expected to enter IT, real estate, finance, or the internet sector, ended up in delivery, ride-hailing, or self-media after layoffs or a tightening job market. The lower-barrier industries.

In June 2026, transport regulators in multiple cities issued ride-hailing saturation warnings. In Guangzhou, Shanghai, and Sanya, the average daily orders per ride-hailing car dropped below 10. Drivers work more than 11 hours a day. After fees, monthly income dropped sharply. The food delivery industry is also seeing "hard to get orders" and "price per order dropping." The reservoir is full. Where does the water go?

Second: actual employment rate of college graduates.

The 2026 college graduating class is projected to reach 12.7 million, up 480,000 from the previous year, a new record. That's more than the entire population of Sweden.

The harder part: media reports that the actual contract signing rate for the 2026 graduating class is under 50%. That means 6 million-plus fresh graduates are unemployed or waiting at graduation.

How does the rest distribute? 25.1% of graduates plan to take the civil service exam, public institution, or teaching exam (up 2.6 percentage points from the previous year). State-owned enterprises are the top choice, with a 45.7% share. One central enterprise's 1,730 openings received more than 1.19 million resumes. About 700 applicants per opening.

And 70% of companies have already turned on AI resume screening. A 985-university diploma may just be a string of characters in front of an algorithm. Degree inflation is happening in real time. The old logic of "a master's degree equals a high salary" is fading. "PhDs delivering food" and "elite graduates scrambling for entry-level positions" are no longer rare.

Third: social financing credit data.

The People's Bank of China's May 2026 data shows household loans dropped 631.4 billion yuan in May alone. Short-term loans fell 694.2 billion. Medium- and long-term loans added only 62.8 billion. Household loans have been negative for the first five months. People aren't borrowing. They're not spending. They're not willing to add debt.

In the same period, the M2-M1 scissors gap widened to 3.4 percentage points. Corporate investment appetite is also weak. It's not that there's no money (the central bank has been printing, M2 growth is 8.5%+). People and companies just don't want to spend, don't want to invest.

Put the three data sets together. 320 million in flexible employment, 6 million fresh graduates unemployed, household loans falling for months. The picture they draw is consistent. In a deflationary environment, demand isn't expanding. There are no new positions. People can only crowd into the flexible employment reservoir.

This is a judgment I made before. China's AI replacement story isn't the same US/EU multiple choice.

The US and EU assume an inflation-plus-tight-labor environment. Whether AI comes or not is a multiple-choice question. China is a deflation-plus-youth-unemployment-plus-flexible-employment-explosion environment. Whether AI's arrival will make unemployment worse, that is the real question.

12.7 million fresh graduates with an actual signing rate below 50%. 25.1% of graduates scrambling for civil service exams (up 2.6 percentage points from last year). 45.7% choosing state-owned enterprises. One central enterprise's 1,730 openings receiving 1.19 million resumes. These numbers put together, what the youth unemployment rate for 16-24 year-olds actually is doesn't matter anymore.

In deflation plus youth unemployment, the AI replacement map might be drawn wrong. In a deflation environment, AI doesn't push prices down further. It just makes it harder for people to find work.

China installed 54% of the world's industrial robots in 2024 (per Stanford's AI Index). Manufacturing automation is already ahead. Does AI keep accelerating automation, or does it replace some of the limited white-collar jobs? This isn't the same dimension as the US/EU story.

Japan: is AI the "population antidote" for super-aging and labor shortage?

Now Japan.

Japan's employment story is the exact opposite of China's.

Unemployment averaged 2.5% in 2025. The OECD average was 4.9%. Japan is among the lowest in the world. Job openings-to-applicants ratio: 1.22. In plain language, 122 openings for every 100 job seekers. The average job seeker has 1.22 openings waiting for them. The labor force: 70 million. A record high since records began in 1953.

A 70-million labor force, the highest ever recorded. Sounds great. But Japan still doesn't have enough. Why? Because people aged 65 and over make up 29.3% of the total population (2024). By 2050, the number reaches 38.7%. One in every 2.6 people will be 65 or older. Population collapse isn't a trend. It's a countdown.

The 2025 Shunto wage increase was +5.25%. Shunto is Japan's annual spring labor-management negotiation. "Shunto up 5.25%" translates to what Japanese companies are willing to pay extra to attract people. It's a 34-year high. But what's behind it isn't economic boom. It's can't find people. Companies have to raise wages to keep people from leaving.

The non-regular employment ratio is 53.4% for women and 22.2% for men. Combined, more than 10 million women are doing part-time, temporary, or contract jobs. These unstable positions. Underneath Japan's "full employment" numbers, there's a structural decline in quality.

What I see from living in Japan: hiring is genuinely hard. Service businesses can't find people. Manufacturing can't find people. Even my wife can't find part-time help for her e-commerce business. Big Japanese companies are still using fax machines. Some traditional software giants can't even get their employees on the external internet. AI adoption is slow. But once Japanese companies start using AI, the replacement speed might be faster than the US. Because they can't find people. There's no fallback.

In aging-plus-labor-shortage Japan, the AI replacement map might be drawn unnecessarily. This isn't AI replacing people. It's people not being enough, and AI filling the gap.

The US is asking: will AI replace me?
China is asking: will AI make it harder for me to find work?
Japan is asking: can AI fill the jobs no one will take?

America fears replacement, China fears joblessness, and Japan fears having no one left to hire.

Three countries. Three AI stories.

Why the US/EU map doesn't "draw" for China and Japan

Put the four side by side:

The assumptions baked into the US/EU AI replacement research all start with the same premise. Inflation, tight labor, well-functioning institutions.

• OpenAI US assumes: CPS data + ChatGPT usage data.
• OpenAI EU assumes: Eurostat data + ESCO system.
• Anthropic assumes: BLS data + Claude actual usage.

None of these three assumptions hold for China and Japan:

• China doesn't have real-time labor-force AI usage data comparable to CPS or Eurostat.
• Japan doesn't either.
• In China's deflation + youth unemployment environment, AI's impact is fundamentally different from the inflation + tight-labor environment.
• In Japan's labor shortage environment, the direction of AI's impact (filling gaps vs. replacing) is also fundamentally different.

So the US/EU map doesn't "draw" for China and Japan. Not because the data is missing. Because the underlying assumptions are wrong.

China's problem is AI making unemployment worse. Japan's problem is AI not arriving fast enough to fill the gap. The same OpenAI report, ported over, won't draw an answer.

8 billion people. Fewer than 1 billion mapped.

Every data point, every framework, every map in these three reports was drawn by Anthropic, OpenAI, and (for one data point) Stanford over a four-to-six month period.

But the assumptions baked into these maps (inflation, tight labor, well-functioning institutions) don't hold in China and Japan.

AI replacement maps are drawn by foot. But the ground underneath is different. The map looks different.

China and Japan need their own AI replacement maps. Not copies of the West. Maps drawn from the structural differences. Deflation vs inflation. Youth unemployment vs labor shortage. Flexible employment vs full employment.

Do the US/EU conclusions hold for China? For Japan? For India? For Southeast Asia? For Africa? For Latin America?

I don't know. But I know one thing. By the time someone draws them, it may already be too late.

I don't live in the US. I don't live in the EU. I live in Japan. I know China a little better than most. The two labs drew maps for me. None of them was mine. Maybe that's fine. Maybe by the time I see my own map, I'll already be off the road.

What I'm doing for the next three months: keep running cross-border e-commerce + AI workflows in Japan with Hermes + Codex + Claude, and push my wife's e-commerce system from "semi-automated" to "fully automated."

The AI replacement maps will keep being drawn. But the pen isn't in my hand. I can only borrow it.

I look forward to the day a Japan-version or China-version AI replacement map appears, with the same granularity as ESCO. That map should start from a completely different premise. Not "what do we do when AI arrives." Instead: "in deflation, unemployment waves, aging, and labor shortage, what is AI actually doing, and what can it do."

Until then, we people living in countries outside the AI labs' map can only make do with someone else's, for now.