One note on measurement. The average unemployment rate vary significantly across groups. Some groups always have higher unemployment rates than other. So instead I rank each group’s current rate against that group’s own record since 2003. Fifty is a typical year for them, and near zero is about as good as it has ever been. That puts a machinist and a software developer on one scale.

Start with something I’ve written about before.

These are workers aged 22 to 34, split by whether they finished college. For twenty years the two lines moved together, through the housing bust and the pandemic. In 2024 they came apart. Young graduates now sit at the 70th percentile of their own history. Young non-graduates the same age sit at the 25th. Keep the scale in mind throughout: lower is better.

That matters for how we explain the graduate story. If this were about inexperience, or a soft economy, or a general pullback in hiring, the orange line would have moved too. It didn’t. Something is hitting one kind of worker far harder than the other.

Widen the lens to everyone over 22.

Every group without a bachelor’s degree is in the bottom third of its own history, and both groups with one are above their median. The levels still run the usual way, with dropouts at 6.4 percent unemployment and advanced degrees at 2.5 percent, so nobody has traded places. Workers with some college but no degree are the worst-positioned non-graduates, at the 33rd percentile.

Now sort everyone by occupation instead of education.

This covers every worker at every education level. The four lines moved as one from 2008 through 2023, then fanned out. Blue-collar and manual-service work, the largest group at 65 million people, is at the 18th percentile, and STEM is the only group above its median.

Occupation by occupation, the ordering is almost too neat. The bottom is physical and in-person work: production, protective service, cleaning, construction, transportation, healthcare support. The top is knowledge work. Nobody sorted these jobs by whether they happen at a desk, but that is roughly what came out. It also explains the some-college anomaly, since that’s the group most likely to hold an office job without a degree.

Splitting each occupation group by education complicates this in a useful way. Among non-graduates, those in sales and office jobs are the clear laggards at the 40th percentile, against the 9th for management and skilled professional roles and the 22nd for blue-collar work. The task divide is real inside the non-degree workforce.

The divide also has a geography, and it runs with density.

Non-metro America is at the 12th percentile and has barely moved since 2023, while the fifty largest metros have climbed thirty points. Office work is a metropolitan product, and the places with least of it have felt least of this.

State by state, the map is blue. Forty-seven of fifty-one jurisdictions sit below their own median, and the darkest run through the industrial Midwest and the Southeast. Ohio, Indiana, Georgia, Tennessee and North Carolina are near their best readings in two decades. This map would have looked unimaginable in 2010, when those were the states people wrote elegies about.

Grouped into regions, four of them covering more than 160 million workers all sit between the 19th and 38th percentile, moving together as they have for twenty-three years. This isn’t a few lucky states pulling an average around. Only the Washington area breaks away, for reasons I wrote about in the last Labor Matters.

Put the cuts together and the weak spot in this labor market has a precise profile: educated, metropolitan, desk-based. Outside that profile, most of the country is somewhere between fine and historically fortunate.

Part of the non-graduate strength is probably supply rather than demand. Immigration enforcement has thinned the pool of less-educated workers over the past year and a half, and the construction and factory-building boom has soaked up much of what remains. Data centers, after all, are built by electricians, so the same technology that may be cooling desk work is hiring hard hats. What supply can’t explain is the graduate side: fewer workers at the bottom says nothing about why unemployment is rising at the top.

I’ll keep writing about young graduates, because what’s happening to them is real and it’s early. But the same data shows tens of millions of people in the best position they’ve held since before the financial crisis. Young graduates are 20 million workers, about one in eight. They are worth writing about. They are not the labor market.

A note on method

Every series is a 12-month moving average, since monthly CPS estimates for subgroups are noisy. The 2003 baseline cuts both ways: it contains two downturns that hit manual work hardest, so an 18th-percentile blue-collar reading partly reflects how bad the baseline was, while rising attainment means today’s non-degree population is more negatively selected, which biases against the finding. And the CPS wasn’t collected in October 2025 during the shutdown, so each window here contains one interpolated month.

The lines crossed in 2024, and then the regional line went nearly vertical, peaking at 6.4% in early 2026 (way above the Great Recession peak!), before settling at 5.7% in July, against 3.9% in the rest of the country. Both series are 12-month averages for workers aged 22–34 with a bachelor’s degree or more. Then, in the first weeks of 2025, the new administration began cutting: layoffs, buyouts, hiring freezes, and reductions in contractor and research spending.

I define the region as DC, Maryland, and Virginia combined, because many of the people working in the federal government live there. The comparisons below measure the change from a base period of January 2023 through December 2024, which ends three weeks before the first layoff announcements, to the most recent 12 months of data, August 2025 through July 2026.

Look at who got hit. If this were an ordinary regional slump, every group would suffer together, and by roughly as much as their national counterparts. That’s not what happened. The region’s young graduates entered the period at exactly the national rate for their group, 3.2%, and now stand nearly two points above it: 5.6% versus 3.8%. Older graduates went from below the national rate to above it. The gaps among workers without degrees moved far less.

Because these groups start from very different baselines (young workers without degrees always have higher unemployment than mid-career professionals), percent changes are the cleaner comparison, and they show a clear education gradient. Unemployment rose 72% for the region’s young graduates and 54% for older ones, against 18% and 28% for the same groups nationally. Workers without degrees saw increases of 27–30%, against 4–11% nationally.

That pattern fits the mechanics of federal personnel policy. Hiring freezes and probationary terminations closed the entry point, which in the federal ecosystem is degree-gated, so the heaviest damage fell on the young graduates who would have been the next hires. Buyouts and reductions in force fell on incumbents, in professional and support roles alike, which is why older workers show a meaningful excess too. The groups most exposed to federal employment deteriorated the most; the group least exposed, young workers without degrees, shows the weakest and shakiest result of the four.

It is worth noting the recent improvement in the 12-month averages, from 6.4% at the start of 2026 to 5.7% by July. That may be an early recovery. It may instead mean that some unemployed graduates gave up on the region and took their job search elsewhere, which lowers local unemployment without creating a single job. The CPS can’t easily tell these apart, though migration data eventually will.

How does this end? Labor markets tend to find their way back to normal unemployment rates, and much of the adjustment here will come through migration: highly educated young people leaving the region for markets that are hiring. So regardless of whether the next administration reverses some of the cuts in federal employment, the 2025 shock will have long-term consequences. The young graduates who leave will build their careers, and their networks, somewhere else.

The first chart shows unemployment rates for a 5-year age group, 22 to 26 year olds, by education, smoothed as 12-month moving averages to remove seasonality. For thirty years these lines kept a stable order: more education, less unemployment. Part of that order has now inverted. Young workers with a bachelor’s degree or more are at 6.1 percent, above the 5.4 percent for those with some college or an associate degree, and the once-enormous gap with high school graduates (8.0 percent) keeps shrinking.

The second chart compares each group’s current rate to its own unemployment-rate history since 2003, expressed as a percentile. Young BA+ workers sit at the 75th percentile of that history and are still climbing. Every other group sits near the bottom third of its distribution. For young people without a degree, this is a decent labor market by historical standards.

The third chart looks only at graduates and splits them by age. The deterioration follows the career ladder exactly. Graduates 22 to 26 are at the 75th percentile of their history, graduates 27 to 34 at the 66th, and graduates 35 and older at the 54th, right around normal. The problem is not college graduates in general. It’s the bottom rungs of the white-collar career ladder.

Keep in mind that all of this is happening while the overall unemployment rate is 4.1 percent, low by any historical standard. This is not broad weakness. Layoffs remain rare; it’s hiring that is depressed, and a low-hiring market is hardest on people trying to get in the door. You can see it in the flows: Handshake reports entry-level postings down 15 percent over the past year, with applications per posting up by more than a quarter.

These charts pose a sharp question: why is the deterioration hitting 22 to 26 year olds with a BA and not 22 to 26 year olds without one? Any explanation that blames youth in general, such as inexperience or the usual cyclical sensitivity of young workers, fails this test, because young non-graduates are doing fine. The causes have to run through the specific jobs graduates take, and I see two that do.

First, AI is likely part of the story. Entry-level knowledge work, such as drafting, coding, research, and basic analysis, is precisely what generative AI does best, while the jobs young non-graduates hold, in health care support, hospitality, construction, and the trades, involve hands and presence that AI can’t replace yet. And whatever the mix of reasons behind it, the hiring freeze is concentrated in exactly the sectors that absorb new graduates: tech, consulting, and finance stopped hiring while in-person service sectors kept adding workers.

In its latest update, using ADP payroll records through June 2026, the Stanford “canaries in the coal mine” study finds employment of 22 to 25 year olds in AI-exposed occupations is now 19 percent below where it would be had it kept pace with less-exposed peers, a gap that operates through reduced hiring rather than layoffs and appears only where AI substitutes for human tasks.

Second, supply. The share of each cohort holding a bachelor’s degree has been rising for decades, so every year more graduates compete for the professional entry jobs a degree is supposed to unlock, while young workers without degrees have grown scarcer, tightening the market they compete in. That trend is too gradual to explain a break in the last three years by itself. But it’s the background condition that made the graduate market vulnerable: a long-building oversupply that the hiring freeze and AI turned from pressure into inversion.

The immigration crackdown has sharpened that scarcity of lower education workers: undocumented workers compete mostly in lower-education labor markets, so collapsing inflows and rising removals tighten exactly the jobs young non-graduates hold while leaving graduate labor markets untouched.

The weak situation of young college grads is happening in an economy where corporate profits are at record highs and employers are barely laying off anyone. History shows that major payroll cuts happen mostly during and after recessions. Imagine how bad this could get if the macro picture changes.

Rising real yields alone don't settle the question. Creditors demanding extra compensation for absorbing a flood of Treasury debt would push real yields up too. But if fiscal panic were the dominant story, you would expect at least some confirmation in other markets. Instead, high-yield spreads sit near historic tights and the S&P 500 made a record high this month.

What we do see is an economy investing through every channel at once. Corporate bond issuance is running at nearly $1.7 trillion through July, up 27%, with Amazon, Alphabet, Meta, and Oracle alone borrowing roughly $194 billion to fund data centers and the power to run them. Venture funding hit a record $510 billion globally in the first half, topping all of 2025. Big tech is rotating cash flow from buybacks into capex and R&D. Add electrification and reshoring, and business investment's share of GDP is climbing after 15 years of secular stagnation.

When desired investment surges relative to savings, real rates rise. That's textbook economics, not distress.

Yields rising because capital finally has abundant productive uses is a fundamentally different world from yields rising because creditors doubt America. The evidence says we're living in far more of the first than the commentary acknowledges.
Sometimes higher rates are the sound of an economy that believes in its own future.

Its key evidence: 7 million registered and practical nurses hold an active license, while there are only about 4 million nursing jobs, suggesting “well over two million licensed nurses are not currently working as nurses for patients.” The article got a lot of attention, but I think its key numbers are misleading.

The article gets a lot right about why bedside nursing has become a harder job, and its reporting on working conditions is worth your time. But the two numbers behind the headline don’t show what they seem to. There are three problems with them.

The 7 million counts licenses, not people. The figure comes from the National Council of State Boards of Nursing, whose table is titled “Number of Active RN and PN Licenses by State” (7,046,231 as of July 31, 2026). A nurse licensed in both New York and Florida shows up twice. NCSBN itself makes the distinction: its 2020 workforce report counted 4.95 million active RN licenses but, “after adjusting for nurses with multiple licenses,” only 4.2 million actual RNs, a 15 percent difference. Using that adjustment as a rough guide, today’s 7 million licenses belong to roughly six million people.

The 4 million counts the wrong jobs. The BLS figure covers two occupation codes: registered nurses and licensed practical nurses. But nurse practitioners, nurse anesthetists, nurse midwives, and nursing professors sit in separate codes, even though all of them must hold an active RN license, and many nurses working as administrators, researchers, or consultants are likely classified elsewhere too. Subtracting these jobs from the licensed population doesn’t measure nurses on the sidelines. It measures the distance between two definitions.

Many people with a nursing license have always worked in non-nursing jobs. Is the current number unusually large?

The source’s own survey answers the question more directly. NCSBN also runs the biennial National Nursing Workforce Survey, which asks license-holders whether they’re working in nursing. In 2024, 87.7 percent of RNs and 88.4 percent of practical nurses said yes, at or near the highest rates the survey has recorded. So no: the share of license-holders not practicing is near its record low.

The question of whether there is a nursing shortage depends on the gap between the number of people working in nursing jobs and the number of nursing positions demanded by employers. I don’t think the article answers or even tries to answer this question. I don’t know the answer to that, but the decline in retention the article documents works in the direction of creating shortages.

None of this weakens the article’s real argument, that pay, staffing, and working conditions are driving nurses away from the hardest jobs. Improved working conditions will help. The pipeline is growing too — NCLEX passers are up 34 percent since 2016 — but if there is a real shortage, it may need to grow faster still. Either way, these numbers are a reason to take the shortage seriously, not a reason to stop saying it exists.

Data notes: License counts from NCSBN’s National Nursing Database tables (July 31, 2026). License-to-person adjustment applies rates from NCSBN’s 2020 National Nursing Workforce Survey (Journal of Nursing Regulation, April 2021); treat the six million as approximate. Employment rates from the 2024 survey (Journal of Nursing Regulation, April 2025); NCSBN’s press release cites the practical-nurse full-time share (70.6 percent); the survey’s Table 71 puts their total employment in nursing at 88.4 percent. Job counts (BLS) and NCLEX pass counts as reported by Time.

There is some text here but focus on the charts!

Second Quarter GDP numbers were published yesterday. The economy grew 1.5% in Q2 and 2.1% over the past year. Every broad measure of demand landed between 2.1% and 2.6%, and the private economy grew 2.5%. That’s solid growth in a year of tariffs, the Iran war, federal cuts, and almost no job growth, with PCE inflation near 4% over the past year and eroding purchasing power.

· A year ago, many economists were very pessimistic about the US economy. They were watching Trump when the story was AI.

· AI capability keeps advancing, and companies, investors, and individuals are acting on it. AI’s footprint in the economy, investment, capacity, and usage, will very likely keep growing for years.

· Over the past year, AI-related investment contributed about 0.9 points to GDP growth, more than 40% of all growth and nearly the dot-com record. The buildout is visible well beyond the investment data: electricity delivered to commercial customers is growing two and a half times as fast as total generation, and advanced manufacturing is pulling away from the rest of the factory sector.

· Tech profits went through the roof, reaching a record 1.77% of GDP, and profits elsewhere grew solidly too, giving companies the cash to fund the next round of investment.

· Strong profits and optimism about AI’s future drove stock prices sharply higher. Household net worth now sits near 600% of GDP, far above any historical norm.

· That wealth, more than the labor market, is what’s carrying the consumer. Spending grew a completely normal 2.3% and contributed about 1.5 points to GDP, despite almost no growth in jobs or real income.

· The main story: a historic economic transformation is underway, and companies, investors, and households know it. The past year proved that tariffs, a war, and federal cuts are too small to derail it.

Introduction

Yesterday’s advance estimate showed the economy growing 1.5% in the second quarter, and 2.1% over the past four quarters. Now consider what the economy absorbed during that year: tariffs, the Iran war and its energy shock, federal workforce cuts, immigration restriction, almost no job growth, and historically weak consumer sentiment.

The chart below shows the past year on seven measures, each stripping out a different piece of the accounting noise. Final sales of domestic product removes inventory swings. Gross domestic purchases counts what Americans bought regardless of where it was made, removing trade. Final sales to domestic purchasers removes both. Final sales to private domestic purchasers also removes government, leaving the private-demand core. Every measure lands between 2.1% and 2.6%. Despite everything on that list, growth was solid, and the private economy, GDP minus government, grew 2.5%, faster than the total.

And that happened during a price shock. Tariffs are passing into goods prices and the Iran war pushed energy costs up sharply. PCE inflation is running near 4% over the past year and hit 5.1% annualized in Q2.

Two engines, nothing else

Decompose the past year’s growth and the picture is stark. Consumers contributed about 1.5 points, AI-related investment about 0.9 points, and everything else combined, housing, other investment, trade, government, subtracted. Outside of these two engines, the economy has been shrinking for a year and a half.

The AI buildout

Every category where AI capex lives is growing at multiples of GDP. Investment in computers is up 44% over the past year. Data center construction is up 18%.

Add it up and investment in the categories most directly exposed to the AI buildout contributed 0.93 points of the past year’s growth. Not every dollar in these categories is AI, so call 0.9 points an upper bound on direct attribution. But the acceleration is concentrated exactly where the buildout is occurring.

Only once in the modern data has this bucket carried the economy like this: 1999. The dot-com peak was 1.06 points. We’re just below it and rising.

One honest caveat. These are gross figures. A lot of the hardware is imported, and imports subtract from GDP, so the gross figure overstates the domestic contribution. But subtracting the full import bill overcorrects. AI-related goods and services are also exported, mostly chips and hardware, plus foreign purchases of US AI services. And imports are valued at the port, while investment is valued installed in the rack: the freight, the margins, the installation work, the building itself, the software and the R&D are all domestic. The net domestic contribution is smaller than 0.9 points but solidly positive.

This is just the beginning

The buildout is not a one-year spree. AI capability keeps improving, and adoption is following: about half of American adults now use AI chatbots, a fifth of businesses deploy AI in production, and the federal government just opened its doors to enterprise AI. The spending plans are already on the books. In their own earnings guidance, the four largest cloud providers are planning roughly $725 billion of capital spending this year, 77% more than last, and reported revenue at the leading model companies is roughly tripling year over year. None of that guarantees the contribution to growth rises forever; investment adds to growth only while it keeps accelerating. But AI’s footprint in the economy, investment, capacity, usage, is very likely to keep growing rapidly for years.

You can see it in the power lines

Electricity delivered to commercial customers, the category that includes data centers, is growing two and a half times as fast as total generation, and the gap opened exactly when the buildout started.

And in the factories

The buildout does not stop at the data center door. Much of what fills those buildings and powers them, the servers and chips, the cooling and electrical machinery, the battery and vehicle technology, comes out of exactly the factories that are pulling ahead.

US advanced manufacturing is booming, and almost no one predicted it. Since 2022, four subsectors, Chemicals, Machinery, Computer & Electronic Products, and Transportation Equipment, have driven essentially all of the net growth in US manufacturing, and they’re still accelerating. They keep pulling away month after month while the rest of the sector sits flat. The divergence is consistent with the same investment boom, though AI is one of several forces lifting these sectors, alongside reshoring, defense demand, and pharma. It’s advanced, knowledge-intensive production, it carries some of the largest multipliers of any sector, and it’s the part that’s surging.

Methodology: Monthly seasonally adjusted IP indexes from the Federal Reserve. The two aggregate lines are Törnqvist chain indexes built from the Fed’s published Relative Importance Weights, rebased to January 2016 = 100, shown as trailing 12-month moving averages.

The boom is profitable

Tech-industry profits reached 1.77% of GDP in the first quarter, a record, and nearly double their 2020 share. And it’s not just tech: total corporate profits rose 12.8% over the year through Q1, and 9.6% excluding tech. Profits matter for this story three times over. They drive the stock prices behind the wealth effect that is carrying the consumer. They may partly reflect the efficiency gains AI is already delivering in the industries deploying it. And they let firms fund the next round of data centers and chips out of internal cash flow, making the buildout far less dependent on credit markets.

The consumer refuses to break

The chart below shows the four-quarter growth rate in consumer spending. At first glance, it doesn’t look especially interesting. But then consider the headwinds consumers have faced over the past year, as economists have repeatedly highlighted: tariffs, uncertainty, historically low consumer sentiment, almost no job growth, almost no growth in real disposable income, and the Iran war.

And yet, look at the chart again. It looks completely normal. How come?

I don’t have the full answer, but the wealth effect is certainly part of it. Household net worth used to fluctuate mostly between 325% and 400% of GDP. Since the mid-1990s, that relationship has broken upward. As of the first quarter, it sits near 600% of GDP, and it kept climbing this past year even as the labor market stalled. The climb is mostly equity gains, concentrated in the AI-adjacent stocks that record profits and AI optimism keep pushing higher.

That matters because when households are sitting on that much wealth, and watching it grow, spending no longer depends only on current labor-market income. Asset gains, home equity, retirement balances, and balance-sheet strength can all support consumption even when income growth looks soft. And because there’s typically a lag between wealth accumulation and spending, the gains of recent years are still feeding through now. One caveat: the wealth effect is concentrated among older, higher-income, asset-owning households.

The bottom line

The AI buildout is demanding more than the economy can currently supply, and the data show it on every front at once: record investment, visible power demand, an advanced manufacturing boom, record tech profits financing the next round, and equity wealth carrying the consumer. That set of forces absorbed tariffs, an oil shock, and a stalled labor market, and still delivered 2+% growth, with private demand growing almost 4% annualized in Q2. A year ago, the median forecast in the Philadelphia Fed’s Survey of Professional Forecasters called for just 1.6% growth in 2026, and those shocks were supposed to produce a major slowdown if not a recession. The past year settled it: against a transformation this large, they were too small to derail it.

What is the message? The main story of the next 12 months is AI, not Trump.

Data: BEA advance estimate for Q2 2026 (July 30, 2026), BEA underlying detail tables, Federal Reserve G.17 and Z.1. Profits, net worth, and industry data through Q1 2026 where noted. Consensus forecast: Philadelphia Fed Survey of Professional Forecasters, August 2025. Hyperscaler capital spending: company earnings guidance.

Thirty years ago the productivity gap between the US and Europe was small. Since 1995, American output per hour has grown 69%. The euro area managed 29%. And the gap is widening faster now than at any point in between.

The divergence

The top panel tells the long story. The US pulled ahead during the IT boom of the late 1990s, while Europe never turned the same technologies into productivity gains. Both economies slumped in the 2010s and the gap stopped growing. Since 2019 it has opened up again: US output per hour is up 12%, the euro area’s barely 2%, and still below its 2021 peak.

The bottom panel is why I think this moment is different. Rolling six-year growth smooths out the pandemic distortions. The US is growing at its fastest sustained pace since 2010, the euro area at its slowest on record. The distance between the lines is the widest in thirty years of data, wider than at the height of the IT boom. Back then Europe was at least growing.

No European exception

Maybe the euro-area average hides a success story. It doesn’t. Every one of the five largest European economies has stalled. Britain has done best and is still thirty index points behind the US (its data runs through 2024). Germany’s export machine has added 2% in six years. French productivity is below its 2019 level, despite cheap nuclear electricity and record employment, so neither the energy shock nor job creation explains the stall. Europe now gets jobs or productivity, one or the other; the US got both. Italy has grown 5.5% in thirty years. Not per year, in total. Export-led, state-led, services-led, in the EU and out: same result everywhere. The problem runs deeper than any one country’s policy mix.

What explains the gap?

Start with the most basic fact: America produces the technology and Europe buys it. The world’s most valuable companies (Nvidia, Microsoft, Apple, Alphabet, Amazon, Meta) are all American technology producers. Europe’s largest, SAP and ASML, are a fraction of their size. When a German firm runs on Azure it captures some of the gains from using the technology; the far larger value of producing it stays in America. In the late 1990s roughly half of America’s MFP acceleration came from inside IT production. It’s happening again: since 2019, US productivity in information industries is up 32%.

But the American surge is broad, and that’s the part Europe should find alarming, because it’s the fixable part. Since 2019 US productivity is up 28% in professional services and 23% in wholesale and retail trade, sectors both economies share. Germany’s best major sector managed 9%. French professional services fell. Same technologies, available on both continents, converted into efficiency in one and not the other, just like the internet era. The usual suspects: less intangible investment, slower reallocation from weak firms to strong ones, and a fragmented market that shrinks the payoff to any big digital investment.

The growth accounting in the third chart sums it up. Since 2019 the US has added 1.9 points of productivity growth a year, four times the UK, its nearest pursuer. Little of it is capital deepening (workers getting more tools); most is multifactor productivity, the efficiency with which capital and labor combine. The American MFP contribution alone, 1.3 points a year, beats any European country’s total productivity growth. Germany’s MFP has nearly vanished. France’s is negative.

The counterarguments

Two objections deserve an answer. Some economists think the post-2019 surge is pandemic reallocation that will fade. The six-year window already smooths the pandemic, and the surge sits in technology-intensive sectors, not reopening ones. And Europe’s numbers are temporarily depressed by labor hoarding through the energy shock. True, mostly in Germany, and too small: it might explain the last two years, not the last thirty. These objections trim the gap at the edges. None touches the trend.

The future

· The past six years look less like an anomaly than a preview. AI is even more American than the internet was. The high-margin layers of the stack — frontier models, chips, hyperscale cloud — are overwhelmingly American; Europe’s strongest card is ASML. Market value per employee at the leading AI firms runs to tens of millions of dollars. Europe won’t capture much of it.

· Investors are betting the divergence spreads beyond tech. Eli Lilly (over $1 trillion) is worth five Novo Nordisks. JPMorgan (about $0.9 trillion) is worth more than HSBC, Santander, UBS, and BNP Paribas combined. America’s chip-and-data-center buildout, now the biggest driver of US investment growth, has no European equivalent.

· Meanwhile Asia is taking the industries Europe still leads. China’s exports are flooding Europe. Sinopec has replaced BASF as the world’s largest chemical company; BASF is shrinking Ludwigshafen while building a €10 billion complex in China. BYD became Britain’s best-selling EV brand this year despite tariffs. Volkswagen is suffering. Chinese and Korean firms already dominate batteries, solar, and shipbuilding.

· Run the last six years forward for thirty and American workers would produce roughly twice as much per hour as Europeans. Trends that long never run in straight lines. But the forces behind this one are strengthening, not fading.

· And productivity gaps become salary gaps become brain drains. US tech employers already pay senior engineers two to four times European rates. Every year the gap widens, the pull on Europe’s best people gets stronger.

Productivity is the main determinant of living standards; run these growth rates for another generation and Europe may no longer afford the social model that defines it. And the scary part: I’m not sure what could stop these trends from continuing. Can you?

Data note: Charts 1 and 2 use GDP per hour worked, constant prices, PPP-converted, through 2025 (UK: 2024). Chart 3 and the industry figures use value added per hour worked through 2024 (Italy, Spain: 2023; UK industries: 2023). The two bases can differ, most visibly for Italy, where value-added productivity rose through 2023 while GDP-based productivity fell. Market capitalizations as of late July 2026. Productivity data: OECD Productivity Database.

I am usually the last person to underestimate AI’s economic impact. But in this case, there may be better explanations.

A “business application” is not a business. It is a request for a business tax ID from the IRS, paperwork whose appeal can also be affected by tax rules, government programs, or the desire to keep a side hustle separate from personal finances. So the count can rise for reasons that have little to do with the creation of viable new businesses.

There was a genuine jump in entrepreneurship during the pandemic. But that does not mean every subsequent increase in applications reflects the same thing.

Sort the applications and the 2025 boom looks different. Applications of the kind that typically become employers were flat from 2024 to 2025. The entire increase came from the kind that, historically, rarely end up hiring anyone.

The timing is suggestive. Applications began rising around March 2025, when Treasury scrapped a new rule requiring many companies to report their beneficial owners to the federal government. They accelerated during the second half of the year, following a tax bill that made the small-business deduction permanent and added a deduction covering some self-employed workers who earn tips.

Maybe AI, online selling, and more one-person businesses are part of the story. It is certainly plausible. My hypothesis is that rule changes explain a meaningful share of the 2025-2026 boom.

What is your explanation?

Perhaps the The New York Times can investigate further and get back to us?

Explore the tool: https://bgi-law-school-rankings.netlify.app/

By Gad Levanon, Joe Winkelmann and Mels de Zeeuw

This brief introduces some of the data available in the BGI Law School Outcomes Tool and presents charts that show some of the interesting results. The data come from the Burning Glass Institute Career History Dataset, which includes where each individual attended undergraduate and law school, their employment history, and an estimated salary for each employment experience.

Some law school rankings answer one question: which schools produce graduates with the highest salaries. That question mixes several forces together: who the school admits, where graduates work, whether they enter public service or private practice, and what local labor markets pay. The BGI Law School Outcomes Tool lets users pull those forces apart. This report walks through eight patterns that emerge when we do.

The tool is built around four salary rankings of graduates ten years out of law school: a raw mean (R1); a residualized mean controlling for race and gender (R2); a residualized mean controlling for race, gender, and the school’s median undergraduate GPA and LSAT (R3); and a residualized mean adding state of school (R4). Each ranking answers a slightly different question. The tool lets users switch between them, filter schools, and see how the picture changes. This report focuses on R1 and R3 because the gap between them is where most of the interesting variation lives.

Different assumptions about what to measure produce different answers. This report and the tool that accompanies it are an attempt to show the size of that effect, not to publish a single new hierarchy that supersedes the existing ones.

These adjusted rankings should not be read as pure causal estimates of what a school adds. They are descriptive measures that show how salary outcomes change after accounting for observable student credentials and demographics.

The eight patterns are organized in five sections.

• A. Why raw salary rankings look the way they do. (Chart 1)

• B. What the picture looks like after adjusting for incoming-student credentials. (Charts 2, 3)

• C. Where the highest-credential-adjusted schools are concentrated. (Chart 4)

• D. How career choice shifts the picture. (Chart 5)

• E. What graduates actually do. (Charts 6, 7, 8)

A. Why raw salary rankings look the way they do

1. Student credentials are strongly correlated with salary

Across the 165 schools in our sample, median LSAT score is correlated with mean salary at r ≈ +0.79. Each LSAT point is associated with roughly $2,600 in additional salary ten years out. Most of the school-to-school variation in raw salary moves with student credentials, so a raw salary ranking is largely a ranking of who each school admits. The chart highlights the T17 — the T14 schools plus Vanderbilt, UCLA, and UT Austin — a grouping used throughout this report.

What differs across schools after adjusting for incoming-student credentials is the question the rest of this report and the tool address.

B. What the picture looks like after adjusting for incoming-student credentials

2. Top 30 schools shift sharply between raw and credential-adjusted rank

Bold blue = R3 rank substantially worse than R1 rank (selection-explained). Bold orange = R3 rank substantially better than R1 rank (over-performs credentials).

All five top schools by raw salary are T14. After adjusting for race, gender, and school-median undergraduate GPA and LSAT, the top ten includes USC (#2), Santa Clara (#5), UC San Francisco (#7), and Vanderbilt (#8). Yale moves from #16 to #122, Michigan from #19 to #127, Berkeley from #13 to #110. The residual captures everything the model omits — geography, career choice, unmeasured student characteristics — not only what a school adds.

3. Most top-30 public university law schools rank lower once credentials are removed

For most top-30 public university law schools, the credential-adjusted rank is lower than the raw rank, with Michigan, Berkeley, and UT Austin showing the largest drops. A minority — Alabama, Arizona State, Indiana, Arizona, and UC Davis — rank higher after the adjustment. The public-school tier is internally heterogeneous, but the dominant pattern is downward.

C. Where the highest-credential-adjusted schools are concentrated

4. Top 50 credential-adjusted schools concentrate in five metros

Half of the top 50 sit in five metros: New York City (10 schools), Los Angeles (5), Chicago (4), the San Francisco Bay Area (3), and Washington DC (3). Location is associated with higher credential-adjusted salary ranks. The cross-sectional data do not identify a single mechanism for that association; the pattern is consistent with employer concentration, with self-selection of graduates into geography, with cost-of-living differences in salary, or with all three operating together. Whether the California schools’ adjusted ranks reflect the schools or state wage levels is what R4 — which adds state of school as a control — is designed to probe.

D. How career choice shifts the picture

A lower salary rank is not necessarily a worse outcome. In some cases it reflects graduates choosing lower-paid public-interest or government work.

5. Within the T17, public-service share is strongly negatively correlated with adjusted rank

Within the T17, the share of graduates in public-sector or non-profit work is negatively correlated with credential-adjusted salary rank (r ≈ −0.67). Yale sends the most graduates to public-service work — roughly 26% — with Georgetown, Stanford, and Berkeley next. Vanderbilt and Northwestern send the fewest and post the highest credential-adjusted ranks. Michigan and UT Austin have low adjusted ranks paired with only middling public-service shares, so career choice does not account for the full T17 picture.

The pattern is specific to the T17 sample. It does not generalize across the broader set of law schools.

E. What graduates actually do

6. T17 graduates: employer mix plus partner and public-interest flags

Employer mix on the left sums to 100% for each school. Partner and public-interest-role flags on the right come from job titles and SOC codes; they use a different denominator and are shown separately so they are not read as slices of the bar. Yale sends roughly a quarter of its graduates to public or non-profit work, the highest share in the T17. Northwestern and Vanderbilt have the highest private-sector shares. Chicago and Northwestern have the highest partner rates, both near 20%.

7. Top 30 public schools: the partner-rate ordering is different

Two things differ from the T17 picture. First, the public-sector share at the top of the list is larger and is dominated by a few schools near the federal government: Maryland sends roughly a quarter of graduates to public-sector work, comparable to Yale’s public-plus-non-profit share. Second, the partner-rate ordering does not track salary rank; even in the T17 that link is loose (r ≈ +0.21; see Chart 8). Georgia has one of the highest partner rates (~19%) but a middling salary rank, consistent with partnership at smaller regional firms rather than at the national firms that drive the T17 partner numbers.

8. T17 salary rank vs. partner rate

Chicago and Northwestern lead on both salary rank and partner rate (~20%). Columbia has the highest salary rank in the T17 but a moderate partner rate. This is consistent with Columbia’s salary premium being driven by high-paying associate and counsel roles rather than by partnership tenure, although the ten-year-out cross-section cannot distinguish a trajectory effect from a composition effect. UT Austin sits at the bottom of the T17 on salary rank; Yale, despite its large fall in the credential-adjusted ranking, sits mid-pack here on both measures.

Explore the tool: https://bgi-law-school-rankings.netlify.app/

Methodology

Sample

ABA-approved law schools in the 50 states and DC. Graduates from the 2000–2013 JD cohorts, outcomes measured ten years post-graduation. 165 schools after restricting to schools with at least 100 graduates observed in the data.

Rankings

Four salary rankings are computed for each school. Each takes the school-level residuals from a regression of individual outcomes on the listed controls, then percentile-ranks schools by their residual mean.

• R1. Percentile rank of mean salary. No controls.

• R2. Percentile rank of residualized mean salary. Controls: race, gender.

• R3. Percentile rank of residualized mean salary. Controls: race, gender, school-median undergraduate GPA and LSAT.

• R4. Percentile rank of residualized mean salary. Controls: race, gender, school-median undergraduate GPA and LSAT, state of school.

Demographic controls are individual-level (race, gender). Credential controls are school-level summaries (median LSAT, median UGPA of the 2024 incoming class). These are measured for a later class than the 2000–2013 outcome cohorts; school-level credential rankings are highly persistent over time, but the adjustment should be treated as approximate for schools whose selectivity shifted sharply over that period. The tool exposes all four rankings; this report focuses on R1 and R3.

Variable definitions

Public sector: NAICS code starts with 92 (Public Administration) or employer name contains one of: city of, county of, state of, department of, U.S., united states, federal, court, public defender, attorney general, district attorney, prosecutor.

Non-profit sector: NAICS code starts with 813 (Membership Organizations) or employer name contains foundation, association, society, charity, nonprofit, non-profit, united way, red cross, habitat, ACLU, legal defense fund. Excludes employers already classified as public sector.

Private sector: all employers not classified above. Public, non-profit, and private are mutually exclusive and exhaustive; they sum to 100% per school.

Partner: job title contains “partner”. Public-interest role: job title contains public defender, legal aid, human rights, civil rights, civil liberties, pro bono, immigration law, legal services, public interest, legal defense, advocacy; or SOC code 23-1023 (judges), 21-1021 (child, family, school social workers), 21-1093 (social and human service assistants). Partner and public-interest-role are occupation-based flags. They are not mutually exclusive with the employer categories and they do not sum to 100%.

Rank stability

Ranks shown in this report are point estimates. Small differences between adjacent schools should not be over-read; a gap of a few positions is unlikely to be statistically meaningful.

Explore the tool: https://bgi-law-school-rankings.netlify.app/

1. The exposure measures predicted the wrong casualties, in the wrong order. Of the two rankings on which much of the “AI and jobs” field relies, Felten–Raj–Seamans rates managers, analysts, and lawyers among the most at risk, and both it and Eloundou et al. score much professional work as exposed as clerical work or more. The newest federal occupational data tell a different story. The BLS Occupational Employment and Wage Statistics (OEWS), whose 2025 figures were released in May, show clerical and sales occupations breaking sharply downward while much of the professional and managerial core accelerated over the same period. When the exposure indexes are tested against realized employment change, one is modestly positively correlated with growth (+0.28) and the other is essentially noise (+0.05). If these scores measured near-term displacement risk, their relationship with employment growth should be clearly negative.

2. This means the rulers are broken, not that nothing is happening. A study that regresses employment change on these scores and finds “no measurable AI effect” may be reporting the blindness of the measure rather than the absence of an effect. If a ranking says a records clerk and an operations manager face equal displacement risk, it has misunderstood what determines whether at least one of those jobs survives automation. Analyses built on these rankings, a large share of the empirical literature, inherit that flaw.

3. But fall they did. What fell was clerical work, and the break was recent; much of the professional and managerial core accelerated. Customer service representatives, an occupation employing 2.9 million people, went from growing 3.2% a year in 2015–19 to shrinking 4.2% a year in 2023–25. Bookkeepers, receptionists, and much of the rest of SOC 43 show the same pattern, with wholesale sales representatives, another 1.3 million workers, close behind. These are not typists continuing a decline that began forty years ago. These occupations were adding workers only a few years ago and are shedding them now. The break survives industry adjustment and a battery of robustness checks: the occupations are shrinking within industries, not merely because employment is shifting away from the industries that employ them. Over the same period, much of the professional core moved in the opposite direction. Operations managers accelerated from 4.7% annual growth to 5.5%, engineers from 1.6% to 2.9%, life scientists from 2.0% to 4.1%, and lawyers continued to grow.

4. The dividing line is codifiability: how completely a job can be reduced to a specified and verifiable process. What matters is not just how many of a job’s tasks AI can touch but what remains after the touchable tasks are removed. Take those tasks out of a records clerk’s job and little of the job remains. Take the same share out of a manager’s job and the core job (judgment, responsibility, persuasion, coordination, and accountability) is still there. The exposure measures score judgment as just another potentially automatable input, while accountability and responsibility are poorly captured by task lists.

This is not the old routine-task distinction wearing new clothes. A job is vulnerable when its output can be specified, produced, and verified without preserving much human judgment, accountability, or relationship management. That is why translation, a skilled and nonroutine occupation by most traditional measures, still falls on the exposed side of the line: its output can increasingly be specified and evaluated without retaining the full human role. That is the central error in the exposure measures, and it is why they got the ordering wrong.

5. The professional exceptions confirm the principle. Graphic designers and translators declined as generative models arrived. Some finance jobs declined but not entirely due to technology. Within computing, programmers, web developers, and quality-assurance testers contracted, while software developers and data scientists continued to grow. HR specialists, market-research analysts, underwriters, and public-relations specialists stalled without contracting; HR specialists slowed from 6.6% annual growth to 1.6%. Across these cases, the casualty is the more codifiable slice of the profession (codifiable, not simple), while the neighboring role with a larger residual of judgment, ownership, or accountability kept growing. A general hiring slowdown has no reason to single out the more codifiable occupation within the same occupational family and, often, the same industries.

6. The consensus built around task-exposure rankings is wrong in both directions at once. Automation of most professional work will probably proceed more slowly over the next several years than these measures imply. Management, engineering, science, and law show little sign of displacement so far and, in many cases, accelerated. Meanwhile, the impact on codifiable work is not prospective. It is already in the data. The prevailing framework therefore appears to overstate the near-term vulnerability of many judgment-intensive professions while understating the disruption already occurring in clerical, sales, and procedural work.

7. The likeliest story is AI as an accelerant, partly hidden by the jobs it is also creating. Several of the clearest occupational declines began before the arrival of large generative models, so the mechanism is best understood as AI accelerating automation already underway, and only in some cases starting the process from nothing. The timing and breadth of the pattern are also consistent with the recent acceleration in labor productivity.

At the same time, displacement is partly masked in aggregate counts because the technologies doing the automating are also creating jobs, probably hundreds of thousands of them, including jobs in many of the same broad technical categories in which other roles are being automated. Software developers can keep growing even as programmers and testers decline because demand for building, deploying, and integrating the technology rises at the same time that the technology reduces labor requirements elsewhere. That makes the losses harder, not easier, to see.

8. Why OEWS. The OEWS covers roughly 1.1 million establishments, records occupations from employer payroll rather than worker self-description, and is large enough to examine detailed occupations crossed with industry. That feature allows the analysis to separate an occupation shrinking within industries from one merely being pulled down by shrinking industries. BLS appropriately cautions against comparisons over time. Caution remains important, but using crosswalks across vintages reduces the risk. The 2025 release makes this window newly readable, and almost nobody has read it this way.

9. Prediction. Each successive OEWS release, beginning with next year’s, will show the clerical decline deepening and the weakness spreading into paraprofessional and procedural professional work. Watch the stalls (HR specialists and underwriters) tip into outright decline, and pressure reach paralegals, junior analysts, and other roles whose output is relatively easy to specify and verify. Three occupations still growing or roughly flat in the current data that I expect to turn within two to three years are claims adjusters, compensation and benefits specialists, and cost estimators. Each is organized substantially around codifiable procedure, and each sits next to an occupation that has already stalled or declined.

These estimates should also be treated as a floor. Because the occupation-by-industry OEWS data pool three years of observations, the “2023–25” figures contain relatively little information from 2025 and none from 2026, precisely when adoption accelerated. Whatever the next release shows, it will still be catching up to what has already happened.

Method, briefly

Raw employment change mixes the occupation with its industries: a job can shrink just because its industry is shrinking. So for every occupation I strip out its industry mix, comparing its growth to the average growth of the industries it works in. What’s left, “industry-controlled” growth, is the part specific to the occupation. This removes sector-wide shocks, not every confound. Offshoring, cheaper labor, regulation, or a quietly relabeled job title can also produce a within-industry decline, so the results are consistent with technology, not proof of it. And nothing in the design can date a break to the arrival of generative AI (a broad white-collar hiring slowdown landed in the same years), so “AI specifically” is a harder claim than “technology broadly,” and I treat it that way.

One check on the headline finding: clerical work’s raw and controlled figures nearly coincide (−5.0% and −4.8%). If the decline were an artifact of clerical jobs sitting in shrinking industries, the control would have erased it, and it didn’t. I compare 2015–2019 with 2023–2025, drop the COVID years between as noise, and keep only occupations spread across enough industries for the adjustment to be identified. Occupation definitions changed with the 2018 SOC revision, so the windows are chosen not to straddle it: 2015–19 sits entirely on the old vintage, 2023–25 on the new one, and occupations are matched across the revision with a crosswalk. Where the revision split occupations too finely to match cleanly (much of computing), I don’t compare across periods. All figures are industry-controlled annualized growth, in percent per year.

The landscape

Every major occupation group (2-digit SOC), ranked by industry-controlled 2023–25 change:

Industry-controlled employment change by major occupation group, 2023–25, fastest-growing at top. Green = growth, red = decline.

Most 2-digit SOC groups grew: this is not an economy-wide employment problem. The declines are concentrated at the bottom: Office/Admin deepest, then Arts/Media, with Production and Transportation just below zero.

The white-collar map

Moving to more granular occupation groups: 3-digit SOC groups, and restricting to white-collar work and coloring each group by its family:

Industry-controlled change, 2023–25, for 29 white-collar occupation groups, colored by major group.

(These 29 are the white-collar groups spread across enough industries for the adjustment to work; single-industry occupations, most of healthcare and education, are missing because we can’t measure them, not because they’re safe.)

Clerical and creative work is on the deficit side; managerial, scientific, and analytical work is on the growth side. Office/Admin flipped from growth to decline, while the groups the exposure indices rank as most at risk barely moved or sped up. The actual ordering isn’t just uncorrelated with the exposure ordering. It’s close to inverted.

The measurement problem

The two instruments the “AI and jobs” literature actually runs on are the Felten–Raj–Seamans AI Occupational Exposure index and Eloundou et al.’s “GPTs are GPTs” language-model exposure score. If exposure predicted displacement, plotting it against realized change would slope downward.

Felten exposure (x) versus realized industry-controlled change (y). The fitted line slopes upward.

Eloundou LLM exposure versus the same change: no relationship at all.

Displacement would require a clearly negative correlation. Instead, Felten’s is modestly positive (+0.28); it flags the jobs that grew. Eloundou’s is indistinguishable from noise (+0.05). On the Felten index, the occupations rated most exposed (managers, analysts, lawyers) cluster among the occupations that grew, while the actual decliners sit at middling exposure. The Eloundou score does put some clerical work near the top, and it still nets to zero, because much of the professional growth scores high too. Both correlations are computed unweighted across the 29 groups; weighting by employment pushes Felten’s further positive and leaves Eloundou’s near zero.

There are two ways to read this. First, exposure is sign-blind: the scores measure whether AI can reach a job’s work, not whether the job contracts, so “replaces this worker” and “augments this worker” get the same high score, a limitation the Eloundou authors state themselves. Second, the ranking itself may be wrong. The inventories these measures are built on do list judgment and decision-making, but the measures score judgment as one more thing AI can touch, which pushes managers up the exposure ranking; and accountability, the other thing that separates a manager and most professionals from a records clerk, is a property of the position rather than a task, so it appears on no task list at all. Either way, a study that regresses employment on exposure and finds nothing has mostly tested its own measure. That verdict is about the scores as predictors of near-term displacement; they may still capture which jobs AI will change, as opposed to shrink. I’ll come back to what the rankings miss once the evidence is on the table.

What fell: the sign-flip

The clearest result is the contraction of codifiable office work. The point isn’t that clerical jobs are declining; they’ve been declining for forty years. What’s new is the sign-flip: occupations that grew through 2015–19 turned negative in 2023–25. Information & Record Clerks went from +0.8 to −4.4; Art & Design from +2.8 to −2.7; Office/Admin supervisors from +1.2 to −2.4, all relative to their own industries. That’s a recent break, not an old trend continuing.

Table 1 — White-collar occupation groups, both periods (industry-controlled, %/yr)

(Abridged to the informative rows; the full 29-group table underlies Figure 2. Here and in the tables below, Δ is computed before rounding, so it can differ from the displayed columns by 0.1.)

The individual occupations inside Office/Admin are the biggest jobs in the economy:

Table 2 — The largest Office/Admin occupations

Customer service representatives, 2.9 million people, swung more than seven points, from +3.2 to −4.2, and the decline is statistically robust. Bookkeeping clerks and admin supervisors show the same reversal; the supervisors fall because there are fewer clerks left to supervise. None of this needs an exotic story: chatbots and self-service portals deflect customer contacts, and accounting software automates the reconciliation clerks used to do.

One caveat: the generalist, Office Clerks, General (2.5 million), barely moved, and its change isn’t statistically distinguishable from zero. The occupations that fell hardest are the specialized, codifiable ones; the harder a role is to reduce to procedure, the better it held. Even the exception fits, though I concede this reading is post-hoc: secretaries, the icon of the previous automation wave, are one of the few clerical groups improving (−2.5 to +0.9). My conjecture is that the earlier wave already stripped out the codifiable tasks, and the secretaries who remain do judgment and coordination work.

Beyond the office: sales reps and object obsolescence

The biggest casualty outside SOC 43 gets the least attention: wholesale and manufacturing sales reps, 1.6 million workers across the two lines at the top of Table 3, disintermediated as procurement moves to portals and marketplaces. The percentage declines are modest, but the headcount makes this one of the largest absolute losses in the data, and nothing else obvious happened to wholesale selling. (Telemarketers, contracting around 14% a year, probably belong here too, but the estimate is weakly identified and I don’t lean on it.)

Table 3 — Sales and non-white-collar occupations with plausible technology causes

Below the sales rows, outside white-collar work entirely, sit some of the steepest technology declines in the economy, and the story there is not just AI. Something different is happening in some occupations, object obsolescence: technology doesn’t automate the task, it deletes the thing the job was about. Cashless payments and remote monitoring are emptying out vending-machine repair. Disposable electronics and streaming ended AV repair; digital documents and cloud computing ended office-machine repair.

The material-moving occupations at the bottom are the counterexample: hand laborers, stockers, and forklift operators, the jobs everyone assumes robots are taking, grew through 2015–19 and are basically flat in 2023–25 once industry is controlled; the warehouse sector expanded fast enough to offset whatever automation did inside it. The one physical case that survives the control is hand packagers, replaced by automated packaging lines.

What didn’t fall

The other side of the gap is just as plain. Management, engineering, and science accelerated through the same window: Operations Specialties Managers went from +4.7%/yr to +5.5, Engineers from +1.6 to +2.9, Life Scientists from +2.0 to +4.1, Lawyers from +1.8 to +2.5. Business Operations Specialists eased from +4.3 to +3.1, a slowdown from a boom. These are the occupations the exposure indices rank as most at risk, and they hold the growth end of Table 1. Whatever is thinning the clerical layer has not reached the jobs built on judgment, accountability, and ownership of outcomes, or not yet; the exceptions are why I add the qualifier.

The professional exceptions, sorted by evidence

Some professional occupations did get hit, and they share a shape. You have to read this at the occupation level; group averages hide it: Drafters & Engineering Technicians nets out to −0.2%/yr while inside it drafters fall at −3 to −8 under CAD and engineering technicians grow on infrastructure spending. The cases differ sharply in quality of evidence, so: clean technology mechanisms first, then the stalls, then the cases where another shock dominates.

Table 4 — Professional occupations with strong technology evidence

Graphic designers and translators are the cleanest generative-AI cases: the tasks map directly onto the technology, the timing matches, and there’s no competing shock. (A caution: OEWS misses freelancers.) The drafters and library technicians are partly older technology declines, not entirely news about AI.

Table 5 — The stalls: occupations that almost stopped growing

As evidence, the stalls are weaker than the declines; a slowdown from a boom can have many causes. As a signal, I think they matter more. These are exactly the roles where workflow automation is arriving (screening software in recruiting, algorithmic underwriting), and what they show is strong growth evaporating while headcount holds. If the technology story is right, the stalls are where the next declines come from. That’s why the forecast at the end is built around them.

Table 6 — Dramatic declines that at least partly belong to other stories

These rows carry big numbers, and they don’t entirely survive scrutiny. The lending cluster is partly the 2022–23 rate shock collapsing loan volume; there’s an automation layer beneath it, but the two can’t be separated here. The film and TV decline coincides with the 2023 strikes. Merchandise displayers reflect store closures. I keep them in, labeled, so you can see I’m not cherry-picking.

The last exception is the most technical field there is: computer occupations, the obvious place to look for AI displacing its own creators. Software employment isn’t collapsing; it’s recomposing.

Table 7 — Declining computer & math occupations, 2023–25

(This table shows one period only: the 2018 SOC revision renumbered and split the computing occupations, so a clean 2015–19 baseline can’t be reconstructed for this field.)

The labels need translation: a “Computer Programmer” writes code to a specification someone else designed, a “Software Developer” designs the software, and coding to spec is the slice easiest to codify, offshore, and automate. The build-and-maintain roles decline while over the same period Data Scientists grew about 17% a year, Information Security Analysts about 5%, and Software Developers about 2%. Two caveats keep me from calling even this “AI”: the 2022–23 sector-wide layoffs are only partly removed by the industry control, and part of the programmer decline is the long reclassification of “programmer” into “developer.”

Across all of these, the casualty is the procedural slice of the profession, and the judgment core beside it kept growing.

Most of the errors run one way

Nearly every material limitation here points toward understating a technology effect, not manufacturing one. The industry control itself removes the biggest channel: a technology effect that thins a whole sector’s occupations evenly is subtracted by construction, and part of an occupation’s own decline can migrate into the industry term. The data lag the technology: OEWS pools three years of industry by occupation information, so the “2023–25” figures barely contain 2025 and none of 2026, precisely when adoption accelerated. OEWS counts employees, not freelancers, so a designer who goes independent registers as a decline; that one runs the other way, and it’s why I treat the translator and designer declines as upper bounds until the household surveys, which do capture freelancers, confirm them. And the numbers are net, not gross: new technologies create jobs as well as eliminate them, often inside the same occupations, so a net decline understates the gross displacement underneath it. Almost everything this instrument misses points the same direction. The estimates above are a floor.

Method: industry-controlled growth is the occupation fixed effect from a weighted regression of log employment change on 2-digit-NAICS industry and occupation, using BLS OEWS national data, with occupations harmonized across the 2010→2018 SOC revision; figures are employment-weighted to the occupation group and annualized. Coverage is ~46% of employment (occupations identified across enough industries); single-industry occupations are excluded, not declared safe. This is correlation, not causation.

Technical appendix: the estimating equation is a weighted least-squares regression of the log employment change of occupation-by-industry cells on fixed effects for 2-digit-NAICS industry and detailed occupation, weighted by base-year employment. The occupation fixed effects are re-centered to an employment-weighted zero, so each figure reads as growth relative to the average job. Cells below 50 employees are dropped and changes are winsorized at the 1st and 99th percentiles. An occupation is reported only when spread across enough industries to identify its effect, measured by the effective number of industries (the inverse Herfindahl of its industry employment shares, required to be at least 2).

For a generation before the pandemic, core PCE sat mostly between 1.5% and 2.5%—if anything, below target more often than above. Then it exploded to 5.6%. The disinflation that followed was real and came faster than the Fed had feared—the clearest sign its credibility survived the shock. But it stopped short. Core PCE has spent two years stuck near 3%, and lately it is drifting back up. It never closed the last mile to 2%.

With Powell’s chairmanship now ended, we can assess his record on the metric he chose for himself. At several critical junctures, the Powell Fed chose the more accommodative option: too slow to tighten in 2021–2022, too quick to ease in 2024–2025. That asymmetry is the story.

The First Error: Asleep at the Wheel

Inflation began accelerating in spring 2021. CPI hit 5% by May. The Fed’s response? Nothing. Rates stayed at zero. Quantitative easing continued at $120 billion per month. The Fed was adding monetary stimulus while inflation ran at multi-decade highs.

The justification was “transitory”—a word Powell used repeatedly until finally retiring it in December 2021. By then the damage was done. The Fed did not begin raising rates until March 2022, starting with a timid 25 basis points. The fed funds rate did not hit 1% until May 2022, 9-12 months too late.

The Taylor Rule had been signaling the need for hikes since mid-2021. Labor markets were extremely tight by late 2021—quits at record highs, wages accelerating—yet the Fed sat at the zero bound. Housing prices surged nearly 19% in 2021, fueled by ultra-low mortgage rates the Fed was actively maintaining, embedding shelter inflation into CPI for years to come.

A History of Wishful Thinking

One chart captures the Fed’s systematic bias better than any narrative. Track the Fed’s Summary of Economic Projections for core PCE over time—its forecasts for 2022, 2023, 2024, 2025, 2026, 2027—and every single line slopes upward. For every single year, the initial projection was too low and had to be revised higher.

Fed median core PCE projections by forecast year, per meeting. Updated through June 17, 2026. Source: FOMC Summary of Economic Projections.

How to read this chart: each line represents the Fed’s inflation forecast for a single calendar year—not the actual data. The horizontal axis shows when the forecast was made, meeting by meeting. A flat line would mean the Fed held its original call. An upward slope means it kept revising higher. Every single line slopes upward. The December value is very close to the actual realized inflation for that year, so the gap between where a line starts and where it ends in December is essentially the size of the miss.

This is not a one-time forecasting error. It is the same mistake repeated across five years and radically different economic conditions. The Fed has spent half a decade telling us inflation will return to 2% next year. It is always next year.

Forecasting is genuinely hard, and we should be honest about that. But the systematic direction of the errors is not random bad luck—it is bias.

The Second Error: Cutting Before the Job Was Done

If the first error is now nearly consensus, the second is more contested but may prove equally consequential. Between September 2024 and December 2025, the Fed cut rates by 175 basis points. It began with an aggressive 50-basis-point cut—a move that signaled urgency the economy did not warrant. GDP was growing above trend. Asset prices were at or near records. Inflation was above target.

The Fed’s justification was a softening labor market: slowing payrolls, rising unemployment, and the Sahm Rule triggering. But there is strong reason to believe these signals were misread. For two reasons.

First, throughout 2024 and 2025, GDP growth remained robust while hiring slowed. The conventional interpretation—the economy is weakening—assumes that jobs and output move together. But what if they were decoupling?

Productivity gains allowed firms to produce more with fewer workers. If this interpretation is correct—and the macro data is increasingly consistent with it—then weak payroll growth alongside strong GDP is not a recession signal. It is a structural transformation. The Fed treated it as cyclical weakness demanding rate cuts.

The broader picture told the same story. Capital expenditure on AI infrastructure—data centers, chips, cloud computing, power generation—was surging at a pace comparable to the internet and shale buildouts of prior decades. Household net worth rose sharply on the back of technology stocks posting dramatic gains in revenue and profitability. You do not get a recession in that environment.

Second, the rise in the unemployment rate in 2023-2025 was only partly a result of the weakening in job growth. The massive immigration from the southern border was also an important factor. Many of the new immigrants stayed in the US and joined the labor force within several months.

The Fed did not distinguish a labor supply expansion from a weakening economy. The unemployment signal was distorted by a supply-side shock the Fed was late to detect.

The Fed, reading payroll and unemployment numbers through a cyclical lens, missed the structural transformation happening beneath the surface. The result: 175 basis points of cuts that may have been built on a fundamentally wrong model of the economy, adding monetary fuel that pushed the return to 2% further out of reach.

The Verdict

A fair assessment requires noting what Powell got right. His March 2020 crisis response was fast, decisive, and probably prevented a financial catastrophe. Once hiking began in 2022, the pace was aggressive and ultimately brought inflation down from its peak. He also navigated intense political pressure from two administrations: Trump repeatedly demanded rate cuts both in his first term and after returning to office in 2025, at times publicly calling for Powell’s removal. Powell held his ground. That institutional independence mattered, and it deserves credit.

But on the Fed’s own stated metric—the one the institution chose, reaffirmed, and built an entire framework around—his tenure was a failure. Five-plus years above target is not bad luck. It is the cumulative result of a systematic dovish bias that consistently chose accommodation over restraint and hope over data.

Part of the explanation is institutional. Powell was not an economist by training, and in moments of uncertainty he leaned heavily on the staff and the profession’s consensus. From 2021 through 2025, that consensus was repeatedly too optimistic on inflation.

The consequences are visible. For decades the Fed worked to anchor inflation expectations. If people believe inflation will stay low, actual inflation is more likely to stay low. Unfortunately, households’ long-run inflation expectations have climbed to their highest in decades, recently near 3.5%—far above the range that held for a generation.

Kevin Warsh, now Chair, inherits an institution whose credibility on price stability has eroded. The 2% target increasingly looks like an aspiration rather than a commitment. Restoring it will require the willingness to be unpopular in the service of doing the job. That has always been the hardest part of being Fed Chair. For five years, it went undone.

The public was repeatedly told 2% was just around the corner.

It wasn’t. It still isn’t. And that is the Powell legacy.