By Sam Peak

Other states have opted to pass more incremental reforms. Tennessee, for example, initially tried to ban noncompetes for in the state, but ultimately settled for passing legislation banning them for employees earning less than . Similarly, Louisiana prohibiting noncompetes for interns and apprentices. Virginia also passed legislation for employees terminated without cause.��

While banning noncompetes for only vulnerable workers may seem like a sensible compromise, excluding higher-income professionals causes states to lose out on the lion’s share of economic benefits that come with a full ban. When free from noncompetes, top earners don’t just switch jobs, they also create jobs by launching their own startups.

Another common trend is for states to limit noncompete reform to specific industries or occupations. In addition to limiting noncompetes for laid-off employees, Virginia also for all licensed healthcare workers in the state, as did . Utah banned noncompetes both for and .

Other states have pursued more niche bans. Nebraska a law preventing healthcare staffing agencies from using noncompetes, while New Hampshire its law to prohibit them for physician assistants — an occupation previously excluded from the state’s healthcare worker ban. Iowa that only banned noncompetes for healthcare employees working at University of Iowa facilities. Maryland, meanwhile, for licensed architects — but only if the employer has at least 30 workers and has moved out of the state.��

While many of the noncompete reforms passed this year are limited victories, lawmakers can build on these wins next year with even bolder reforms. The Utah legislature, for example, has listed noncompetes as an , indicating that the issue is likely to be a priority for the next legislative session. The other states that have passed noncompete reforms should also pursue follow-up legislation that covers additional workers.

After all, even Washington’s complete ban on noncompetes was accomplished through a piecemeal approach across two pieces of legislation. The first bill was passed in 2020 and only banned noncompetes for employees earning less than $100,000.^[2] Despite initial concerns that workers subject to the ban could leak trade secrets, these fears proved . Shortly thereafter, Microsoft, one of the state’s largest employers, for most of its workers. Eventually, the momentum generated by the $100,000 ban emboldened state lawmakers to finish the job and enact a ban for all workers, regardless of income.��

Washington’s case is instructive. Implementing limited noncompete bans can serve as a vital first step toward a more comprehensive policy. As evidence mounts showing that noncompete clauses harm economic dynamism, all roads should eventually lead towards full bans.��

By Kenan Fikri, Catherine Lyons, and Phoenix Vu

Just in time to inform governors’ work nominating the next round of Opportunity Zone (OZ) census tracts this summer, the Treasury Department has released new data reporting that federal OZ tax incentives drove more than $112 billion worth of investment capital into more than 6,000 communities through the end of 2024.

The new statistics, released in an Office of Tax Analysis and derived from IRS tax filings, update previous estimates with two additional years of data to provide the fullest and most authoritative picture of OZ investment to date.^[1] The timing is propitious, landing shortly after the 90-day nomination window opened on July 1 for governors to designate the next round of OZ census tracts. The new designations will come into effect on January 1, 2027 and last for a decade.

Originally ��on Agglomerations, the Substack newsletter from the Economic Innovation Group.

By Adam Ozimek, Jason Harrison, and Nathan Goldschlag

To hear an extended conversation about the themes in this post, please subscribe and listen to our podcast,��The New Bazaar.

The China Shock refers to the widespread job loss that certain workers and communities suffered because of the rapid rise of trade with China in the 2000s.

The AI Shock refers to the same thing happening to white collar workers from the rise of Artificial Intelligence.

For now, the AI Shock is merely a��potential��shock. But a����of��,��, and other����have����worrying parallels between these two shocks.

We understand why. Across a������of��, researchers have found that the China Shock had big and enduring impacts on the local economies and workers who were most exposed to it.

Even more concerning, the extent of the China Shock’s damage was a surprise to economists and other experts. It had long been known that expanding trade will produce winners and losers, but economists tended to think that jobless workers would adjust by moving to the better job opportunities created in a growing economy. Outmigration and flexible wages would restore balance to local labor markets, while many factories would be repurposed to make different products or provide services.

For too many communities, however, especially certain manufacturing towns in the Midwest and South, those adjustments never came. Jobs were lost and not replaced. Jobless people who stayed in their towns simply stopped looking for work. Health, crime, and family formation deteriorated.^[1]

Will AI do to laptop workers what the China Shock did to so many factory workers? Hoping to avoid the same mistakes of the past, researchers and policymakers will be tempted to say yes.

But after a closer look at the details of the China Shock, combined with our understanding of workers exposed to AI, we arrive at a different conclusion.

The real lessons of the China Shock emerge not from its similarities with the AI Shock, but from its differences. And these lessons should mitigate concerns about a massive upcoming disruption, not exacerbate them.

But to understand how we discovered these lessons, we first need to explain how we compared the experiences of China Shocked workers to the potential outcomes for AI Shocked workers. We begin with our measures for each set of workers.

China Shock Exposure

Here is how we know.

A����published in 2021 by David Autor and his co-authors, building upon earlier work, produced a measure of trade exposure by industry — specifically, the increase in Chinese imports as a share of each industry’s output.

We largely follow this approach. But to arrive at a measure of China Shock exposure for��workers, we did two things. First, to measure industry exposure, we chose the time period of 1991 to 2007, a window that captures both the 1990s rise in Chinese import competition and the acceleration after China’s 2001 WTO accession. Second, we adapted the Autor measure — again, the rise of Chinese imports within each industry — by assigning exposure to individual workers based on the industries they worked in.

Our findings:

• For 82 percent of American workers — including everyone outside of manufacturing — there is no measured exposure. They worked in industries in which there was no rise in Chinese imports as a share of their industries’ output between 1991 and 2007.

• Another 14 percent of American workers faced positive but still modest exposure. The rise in Chinese imports was less than 15 percentage points as a share of their industries’ output.

• In contrast, roughly 5 percent of workers were in industries where the growth in Chinese imports was quite large — 15 percentage points or more — relative to output in their respective industries. For example, Chinese imports climbed by 15 points in electrical machinery and tires, 54 points in computers and related equipment, 70 points in toys and sporting goods, and 77 points in footwear. All are manufacturing industries.

It is this last category — the 5 percent most exposed workers — that we use as our comparison group in the rest of this analysis.

In other words, when we compare workers exposed to the China Shock to workers exposed to the AI Shock, we are specifically referring to the 5 percent of American workers who were most exposed to Chinese import competition between 1991 and 2007.

This group numbers about 6 million workers. They all worked in manufacturing, representing about a third of all workers in manufacturing industries. We explore their other characteristics in the sections to come.^[2]

AI Shock Exposure

There is no directly comparable dollar based exposure measure for AI.

Instead, we use the����in a 2024 paper by Tyna Eloundou and co-authors to identifying the set of workers, based on their occupations, that are most exposed to AI.

Eloundou and co-authors begin by measuring whether the use of AI, potentially with additional specialized software, could reduce the time to complete a task by at least half. To do this they prompt both expert human reviewers and an early version of OpenAI’s ChatGPT-4 to classify tasks based on this time reduction rubric, finding a high degree of agreement between them. For our analysis, we will refer to such tasks as those that can be “done by AI.”

Eloundou and co-authors then aggregate from the task level to the occupation level, computing the share of tasks that can be done by AI.

What is useful about this measure is that it focuses on what��could be automated, not what��has been automated��today. Because the AI shock is mostly yet to come, a forward looking exposure measure is more appropriate.

Another complicating factor is that it is not yet possible to know how much task exposure will lead to actual job loss. For the China Shock, empirical estimates show that the 5 percent most exposed workers were at serious risk of job loss. For the AI shock, however, there is no evidence about the share of a job’s tasks being done by AI that would constitute such risk.

Even if a very high share of an occupation’s tasks can be automated by AI, the occupation itself may survive as a new and different combination of tasks.^[3] As there is no obvious cutoff for which workers will face displacement (job loss) pressures, we consider multiple cutoffs for the pool of workers that are AI exposed. Table 1 below shows the percentage of the workforce employed in occupations at each respective amount of task exposure to AI.

Our narrowest definition of high exposure assumes that the AI shock only affects the workers at the most extreme end, where AI can replace 90 percent of their tasks. These workers represent less than 1 percent of the workforce.

Our broadest definition assumes the AI shock disrupts a much wider swathe of jobs, including everyone where at least half of their tasks could be done by AI, representing closer to one out of every four workers.

But as our analysis will illustrate, the results of our comparison between China Shocked workers and AI Shocked workers are consistent regardless of where we draw the line.

AI Shocked Workers: More educated, better paid

With highly exposed groups for our two shocks defined, we can explore how they are similar and how they differ. Two differences are especially notable.

1. Compared to China Shocked workers, AI shocked workers are far more educated.

Figure 2 shows that the majority of China Shocked workers have an education level of a high school degree or less, while only 6 percent have an advanced degree.

Using the broadest sample of AI Shocked workers — at least half of their tasks are exposed to AI — the majority of them have a bachelor’s degree or higher. Workers in this group are more likely to have an advanced degree (18 percent) than just a high school diploma or less (17 percent).

And the more concentrated the AI shock, the more educated the group becomes. Workers in the most exposed AI Shock group, representing less than 1 percent of the workforce, are three times more likely to have an advanced degree than to have a high school diploma or less.

In short, regardless of how broad or narrow we define the AI shock, it will affect a group of workers who are substantially more educated than workers who were most exposed to the China Shock. As AI exposure rises, so does educational attainment.

One of the reasons for the disparity is simply that average education rates increase over time, with the China Shock reflecting an early 1990s workforce. Between the early 1990s and early 2020s, the share of workers with a high school degree or less fell from about 48 percent to just over 30 percent. But even if we redo the analysis comparing shocked workers in the same period, the educational gaps persist.^[4]

2. AI Shocked workers also have higher earnings than China Shocked workers.

Regardless of which AI Shock exposure group we use, its share of the top third of all earners is much higher than the China Shock share, while the reverse is true for the respective shares of the bottom third of earners, as shown in Figure 3.

Only about a third of China Shocked workers were in the top tercile (top 33 percent) of overall earners during the time of that shock, whereas 43 percent of the broadest group of AI exposed workers made the top tercile. Meanwhile, when looking at the most exposed group of AI Shocked workers, nearly two thirds of them were in the top tercile.

In contrast to the heavy concentration of AI Shocked workers among the top tercile of earners, China Shocked workers were fairly equally distributed across the wage spectrum, with around a third of workers in each tercile.

Not all workers are equally vulnerable

One piece of evidence supporting this view is that the least educated are more likely to be unemployed. Census data going back over 80 years shows that unemployment for the least educated third of workers is always above, and usually at least double, that of the most educated third.

Another piece of evidence is that even among those workers who lost jobs from the China Shock, not all of them suffered equally.

Difficult adjustments and longlasting effects�� among workers with lower initial wages.^[5]��Low wage manufacturing workers in highly exposed industries lost about 1.2 years of initial earnings over the following 16 years relative to comparable workers in less exposed industries. But middle wage workers lost much less, and the estimated effect for the highest paid third of workers was essentially zero.

High wage workers were able to leave exposed firms and industries without large earnings losses. Lower wage workers, on the other hand, were more likely to remain at their initial employer until the firm underwent mass layoffs, pushing those workers into worse jobs or forcing them to exit the labor force entirely.

The proof that more educated and higher paid workers adapt better to shocks is not just limited to the China Shock research. A similar pattern occurred during the Great Recession. Workers with more education were����to lose their jobs initially, and when they got a new one it was less likely to be part time or to pay lower earnings. Looking at the longrun effects of local labor market shocks from the Great Recession, Danny Yagan����that the workers with the highest initial earnings were least likely to be jobless in 2015, long after the recession had officially ended.

The relationship between education and adaptability to local shocks has also been found in the��.

One reason that better educated, higher paid workers are more able to adapt is they are more��. Another reason is that their skills are more��.

Whatever the cause, the evidence on a wide range of economic shocks illustrates that when job loss is concentrated among lower wage and less educated workers, it is likely to generate different adjustment problems than a shock concentrated among workers with higher pay and more college degrees. We consider this an important reason to worry less about the AI shock mirroring the China Shock.

Where the shocks are

The disruptiveness of an economic shock is determined not just by which people get hit but also which places. Research has shown that one reason the China Shock was so disruptive was that it was geographically��.

Local labor markets adapt to small losses of employment all the time. Even outside of recessions, roughly 30 million jobs across the United States are����due to layoffs or business closures. But the national unemployment rate stays low as workers reallocate from shrinking sectors or firms to growing ones. This activity is simply part of the everyday churn of the economy.

When job losses are large and concentrated in a few specific places, in contrast, it becomes much more difficult for those local economies to adapt. Rather than Schumpterian creative destruction, the job loss is large relative to the local economy. The reallocation will take longer, and the struggle to adapt also can generate demand shortfalls as local spending pulls back and spills over into other sectors. Nationally set monetary policy will do little to stabilize local economies suffering from demand problems that arise from concentrated shocks.

If the few million jobs lost from the China Shock had been scattered evenly throughout the country, they would have represented just a small fraction of normal economic churn, and thus workers would have bounced back more easily. Instead, the shock was heavily concentrated in manufacturing towns in the Rust Belt and South.

The good news is, the evidence we have suggests that the AI shock will be far less geographically concentrated. Figures 5 and 6 below show the respective place level exposures to the China Shock and the AI shock across Commuting Zones (CZs). China Shock exposure is concentrated in a much narrower set of manufacturing heavy labor markets. AI exposure, in contrast, is more evenly dispersed throughout the country.

The difference is also obvious when looking at relative exposure in the hardest hit places, as shown in Figure 7. The 10 Commuting Zones most exposed to the China Shock average about 5.8 times the exposure of the average one; the next 40 about 3.2 times the exposure of the average one; and the next 50 about 2.1 times.

AI exposure is much flatter, with the top 10 CZs only about 1.2 times the average CZ.

Which types of places

How a place adapts to a shock also depends on the characteristics of the place.

��that local economies whose workers have higher average levels of education adapt much more easily to economic shocks than places with lesser educated workers. The reasons for their resilience go beyond the fact that higher educated workers themselves are more resilient. Places with more human capital tend to have faster rising populations, which helps to boost dynamism and the growth of underlying demand growth. A more educated population will also be more innovative and have more entrepreneurs who identify next best uses for displaced capital and labor.

When we compare China Shocked places to AI shocked places, it is clear that education levels are a major differentiator. Figure 8 below looks again at exposure ranked Commuting Zones, but it compares the average share of workers with a high school degree or less against the average share of those workers for all Commuting Zones in the relevant year.^[6]��(We use 1990 education data for the China Shock and 2024 education data for the AI Shock.)

The most China Shocked places had populations with lower than average education levels at the time, while the opposite is true for the most AI exposed.

Looking at the share of the population with bachelor’s degrees or higher, we see a complementary pattern. China Shocked places are slightly below average, while AI shocked places are way above average.

Missing Something?

We have presented the evidence for why the China Shock was far worse than the AI Shock is likely to be. What might the opposite case look like? We’re happy to play devil’s advocate, both because it’s a good intellectual habit and also because the case ends up looking so much weaker than ours.

One possibility is that the AI shock materializes much faster than the China Shock. If AI agents prove to be as capable as their creators suggest, displacing a worker whose primary tasks occur on a laptop could be as simple as installing and initializing the AI agent.

A second possibility is that the AI Shock ends up displacing a much bigger share of workers than we expect. In our most aggressive AI Shock scenario, 27 percent of workers would be exposed, 5.7 times larger than the 5 percent exposed to the China Shock. If most or all of these exposed workers lose their jobs — especially if, coinciding with the first possibility, they lose their jobs quickly — an utterly massive share of the labor force will find itself suddenly unemployed. The job losses and subsequent demand destruction could be enormous.

That’s the worst case scenario: AI hits faster and hits bigger than even pessimistic forecasts. We find this outcome unlikely, for two reasons.

First, rolling out AI at such scale requires significant computing hardware, which takes time to build. Data center construction is a real constraint and is already causing policy blowback. The prices of information processing equipment have��, rising at the fastest pace since 1959. The��current��AI buildout is already facing big obstacles. Imagine how much bigger they would be for the kind of buildout required to replace millions of workers.

Second, if the AI Shock were to produce vast job losses, the magnitude of the effect itself makes economic policy more powerful in offsetting the effects on demand. Recall that monetary policy, for instance, was of limited use for helping local economies destroyed by the concentrated China Shock. If the AI Shock proves to be bigger and more widespread, monetary policy regains in potency.

What is more, significant losses of jobs because of AI would be accompanied by rapid economic growth. Both economic growth and AI augmentation are helpful for smoothing the economic adjustment in a variety of ways, including boosting the demand for other types of workers.

Nevertheless, we admit that an AI Shock large enough to upend the labor market overnight would be unprecedented, which means that reasoning about it involves uncertainty in both positive and negative directions — about adjustment costs but also benefits, layoffs but also progress, unemployment but also innovation. The economy would be entering uncharted territory.

In such a world, comparisons to the China Shock, or to any other great historical shock, would therefore be pointless.

A tale of two shocks

In sum, the China Shock was disruptive because it 1) disproportionately hurt workers with less education, 2) was geographically concentrated, and 3) inflicted its worst damage on places with lower levels of human capital.

It is still early days for AI and so a large dose of humility is appropriate. But so far as we can tell, the AI Shock is aimed at the kinds of people and places that actually weathered the China Shock just fine.

The old cliché is that history does not repeat itself but does often rhyme. If the AI Shock turns out to be as bad as the China Shock despite such different underlying characteristics, it would turn the cliché on its head: an instance in which history did repeat itself but skipped the rhyming. It’s not impossible, but we find this outcome unlikely.

Appendix

In the preceding analysis, we compare the education levels of the workers most exposed to the China Shock to those who are most exposed to the AI shock. One complicating factor is that these two shocks are occurring at different points in time, and education levels among workers overall climbed.

Absolute increases in education over time are relevant for resilience, but we can also compare relative education rates by comparing each group of exposed workers to the average education rates of the overall workforce��from the same time period. In other words, we can compare the China Shocked workers to the average education of workers between 1991 and 1993. And we can compare the education of AI Shocked workers to average education levels of all workers between 2021 and 2022.

The results are directionally the same: Workers exposed to the China Shock are still less educated, while the Al exposed groups are more educated.

Specifically, compared to the overall 1991 to 1993 workforce, workers exposed to the China Shock are 10.4 percentage points more likely to be high school graduates or less and 5.9 percentage points less likely to have a bachelor’s degree or higher.

In contrast, the most AI Shock exposed groups are substantially more educated than the 2021 to 2022 workforce. Depending on the exposure group, they are 13.7 to 24.1 percentage points less likely to have a high school diploma or less, and 13.1 to 29.9 percentage points more likely to have bachelor’s degrees or higher.

Taking the z scores for each group shows consistent results, with China Shock exposed workers below their period average and every exposure level of AI shock above its period average.

Originally ��on Agglomerations, the Substack newsletter from the Economic Innovation Group.

By Nathan Goldschlag

If you are even just a casual user of the economic data produced by the government, or if you simply care about the integrity of the data, you should be distressed by a����announced by the Commerce Department last week.

The policy bans the use of something called noise infusion, a necessary tool used by statistical agencies, in particular the Census Bureau, to publish granular data about individuals and businesses without disclosing their identities, thereby protecting their privacy.

The likely reason for the policy change is that the Trump administration wants to improve the precision of certain data series. But the likely��result��of the policy will be that the government publishes less data, weakening the public’s ability to understand what is happening in their local communities and the American economy.

This consequence would be particularly damaging because individuals, businesses, and policymakers rely on these data to make decisions. Workers use these data to decide whether or where to move, which skills to acquire, and which jobs to seek. Businesses use these data to decide whether and what type of investments to make. And policymakers use these data to design policies and evaluate their effects on constituents. Without the data, they are all flying blind.

I agree with the goal of making data more precise. This just isn’t the way to do it.

Disclosure Avoidance Primer

An example is the County Business Patterns (CBP) data, which provides statistics on business activity by industry and geography. Consider the following three scenarios:

• There is only one brewery in a small county. If the CBP published the exact count of brewery employees in that county, it would be disclosing the information of one business (how many workers it employs), a clear violation of the law.^[2]
• There are two breweries in a small county, and the CBP again publishes the exact count of brewery employees. If I own one of those breweries, I could learn how many employees my competitor has, again violating the law.

• There are more than two breweries in a small county, but the CBP chooses not to publish the total number of brewery employees out of concern that it might compromise the privacy of the businesses. If I’m a prospective brewery owner, I may deem the project too risky to pursue without information about the market I’m entering.

These hypotheticals illustrate a core tradeoff of disclosure avoidance: Granular data is often more useful than aggregate data, but it also risks disclosing private information.

The stats agencies have several tools to manage this tradeoff, including coarse aggregations (releasing less data), cell suppression (removing certain statistics from the data that is released), and noise infusion (fuzzing the data).

It is the third option, noise infusion, that is being banned by the Commerce Department. When I say that noise infusion means “fuzzing the data,” I’m not using a technical term. What I mean is that noise infusion is a method of slightly altering granular data in such a way that the privacy and identity of individual people and businesses are protected in the statistics derived from them. The data are not so much altered, however, that they become unhelpful.

Another example helps to explain how it works. The CBP uses a type of noise infusion known as multiplicative noise, in which the employment of each business in a county is multiplied by a random number, increasing or decreasing it by a small amount. Go back to our hypothetical county, and this time there are 20 total breweries. Noise infusion allows the CBP to publish the total number of brewery workers in that county because it is not based on each brewery’s��exact��employment. But the data on��total��brewery employment in the county will be close enough to its actual value for it to be useful to anyone who needs this information to make decisions.

For any statistic that covers a large number of people or businesses, the noise cancels out and the statistic ends up very close to the actual value. For a specific industry in a small county with only five or ten businesses, the statistical and actual value may differ. But the differences tend to be small, and even in those cases it is nonetheless better to have��some��data rather than none. In the absence of noise infusion, for example, the stats agencies might be forced to only give the count of brewery employees at the state level. This information is far less useful to local policymakers and businesses — or to a prospective employer looking to build a new brewery in a specific county with unique business conditions that may differ from other parts of the state.

The practice of noise infusion became contentious around the 2020 decennial census because of the Census Bureau’s introduction of a new disclosure method called differential privacy, which is another type of noise infusion.

Differential privacy is similar to multiplicative noise infusion in that it adds noise to protect the privacy of people and businesses. But differential privacy is fundamentally different in that it gives the Census Bureau a better understanding (through complicated math) of how changes in the amount of noise added will affect the risk of disclosing too much private information.

The worry on the part of data users was that the Census Bureau would use — or misuse, in their view — differential privacy to err too far in the direction of protecting privacy, thus compromising the precision of the data.

Importantly, as I’ve already explained, differential privacy is not the only type of noise infusion used by the stats agencies. The County Business Patterns, Business Dynamics Statistics, and Quarterly Workforce Indicators have been using the simpler multiplicative noise infusion for decades. But the new policy from the Commerce Department would ban multiplicative noise infusion in addition to differential privacy.

Why I’m Worried

“Because of the fine detail offered by the published statistics and the confidential nature of the micro-data used to compile the statistics, confidentiality protection is a critical and integral part of the design of the QWI system. Only the application of state-of-the-art protection methods allows the Census Bureau to publish these statistics.”

Noise infusion makes possible the granular tabulations of QWI data, which are used to inform the public about employment, hires, separations, and earnings by industry, geography, and worker characteristics including age, sex, race, and education.

If the Commerce Department policy stands and noise infusion is banned, which disclosure avoidance tools could the agencies use?

The policy states: “Coarsening shall be the preferred category of Disclosure Avoidance methods for all statistical products.”

And: “Suppression shall be permitted as a last resort.”

As a reminder, coarsening means releasing less granular data, and suppression means no longer publishing certain statistics altogether.

And that is why you should be worried. Either the agencies decide it’s actually okay to publish data for the county with only a handful of breweries, or, more likely, they decide that there will simply be no county-level brewery data at all. The data might get rolled up to the state level, which is far less detailed and helpful.

Now what?

Remember again the fundamental tradeoff.

On the one hand, the statistical agencies must adhere to strict legal requirements that ensure the privacy of individuals and businesses. Those privacy guarantees, importantly, encourage survey responses by giving respondents confidence that their data won’t be made public.

On the other hand, the statistical agencies are charged with producing quality data about the nation’s people and economy. Excessive focus on disclosure risks can degrade the quality of data, reducing its value to the individuals, businesses, and policymakers that rely on them.

Outside of clearcut cases like our small county with just one or two breweries, the statistical agencies exercise significant discretion in terms of how strict the disclosure-avoidance protocols should be. In the brewery example, how many breweries does a county need for the data to be considered safe? If one brewery was much bigger than the others, could the smaller ones back out the data of the bigger one?

This ambiguity leaves a lot of room for choice. The agencies can choose to publish as much data as possible given legal constraints, or they can avoid the risk and design rules that reduce the amount of data released. The agencies often choose the latter.

The 2022 proposal to change the Current Population Survey Public Use microdata files is an instructive example. In early 2022, the Census Bureau announced a plan to more than double the population threshold for suppression and round wages. By December of 2022, the Bureau����those changes due to pushback from data users, finding alternative approaches less damaging to the usability of the data.

Given this tendency towards limiting data release, which I myself observed over years working within the statistical system,^[3]��it is entirely reasonable for policymakers to rebalance the scales and place greater emphasis on producing more and higher-quality data.

But if the administration’s goal is to pursue that rebalancing, banning a specific disclosure avoidance method is not the way to do it.

Without changing the laws governing privacy protections at the statistical agencies, or promoting a culture and leadership more focused on the quality and volume of data they produce, removing noise infusion as a disclosure avoidance technology will lead to��less��data overall.

Worse yet, the prospect of less data comes at exactly the wrong moment. There is growing����that we need significant investments in our statistical system to improve our measurement of AI’s impact on the economy.

I hope to be proven wrong, and that the rollout turns out to be more sensible than I’ve imagined it here. But this new policy has me worried.

Originally ��on Agglomerations, the Substack newsletter from the Economic Innovation Group.

By Kenan Fikri and Thomas Cronin

Another��, another reminder that healthcare has been the driving force behind U.S. job growth in recent years. Given that healthcare is used by everyone, everywhere, we initially assumed that all those new healthcare jobs would be evenly distributed across the map, varying only by rates of population growth or shifting demographics.

Not so. Healthcare job growth varies dramatically by state, as does the extent to which healthcare contributes to total job growth. And no state appears as reliant on the healthcare and social assistance sector to support its labor market as the nation's most populous and the world's fourth largest economy: California.

Another take on a "health-full" labor market

The total number of jobs in the United States increased by 4.7 percent from March 2022 to March 2026, but the differences across sectors were massive.^[1] Jobs in healthcare and social assistance increased by a whopping 16.3 percent, compared to just 2.9 percent in the non-health segments of the economy.^[2]��Artificial intelligence (AI) may dominate financial markets, but healthcare still reigns supreme in the labor market.

Scan the figures in the charts below, and it becomes apparent just how geographically disparate job growth has been over the past four years even in this quintessentially local sector.

California outpaces every other state with its 25.3 percent job growth in healthcare and social assistance from March 2022 to 2026 — and that growth wasn't being powered by a quickly rising population. The number of California residents increased by less than 1 percent between 2022 and 2025.^[3]

But what really distinguishes California is how it combines such massive growth in healthcare with such little growth in other parts of the economy.

California eked out total job growth of a mere 3.4 percent over the past four years, the bulk of which was racked up at the beginning of the period (see the top figure). Take out healthcare and social assistance, and employment across all other sectors in the state��declined��by 0.3 percent. Over the past four years, employment in the non-care portions of the economy shrunk in only three other states, led by DOGE-.

The growth gap between the health and non-health parts of the economy (the length of the bars in the above graph) is greater in California than in any other state.

Healthcare now accounts for 17.4 percent of all jobs in California, up three percentage points in four years. That is twice the rate at which healthcare is expanding its dominance over the labor market nationwide.

All subsectors go

Growth was solid across all four main subsectors of the California care economy. Employment in hospitals increased 5.2 percent, in ambulatory medical care (think doctors' and dentists' offices) by 15.7 percent, in nursing home facilities by 17.2 percent, and in social assistance by 27.2 percent over the past three years.^[4]

Even more granularly, 36 of the 39 narrow industries that make up the care sector added jobs over the most recent three-year window.^[5] In absolute terms, services for the elderly and disabled topped the list, adding a whopping 211,000 jobs (31.5 percent increase) and accounting for nearly half of all job growth in the state's care sector.^[6] Mental health practitioners led the way in percentage terms, with the ranks of California therapists nearly doubling (92.7 percent growth, or 27,212 new jobs).^[7]

These numbers are huge but not anomalous. Nearly every corner of the care economy is growing, and no single corner can explain the larger sector's might.

Nevertheless, it is striking how exceptionally low-paid these new jobs in services for the elderly and disabled are, coming in at $487 a week on average (or just over $25,000 per year) in the state.^[8] This is the lowest paid sub-sector of the entire care economy and one of the lowest paid in the entire economy, too.

Stepping back, it is also notable just how starkly the robust growth across basically every corner of the care sector compares to weak growth across the rest of California's economy.

Trading down?

Lay out all the pieces and it becomes clear that, as the California economy evolves, jobs in high-wage sectors are being replaced by jobs in low-wage sectors. Jobs are shifting from manufacturing, information (tech), and finance and professional services into much lower-paying care roles. States such as Idaho are adding jobs in both high- and low-wage sectors. California is undergoing a great labor market transformation towards lower-paid work.

Consider this rough estimate: Multiply the average weekly wage in each sub-sector in Q3 2022 by the number of jobs gained or lost over the subsequent three years (essentially an expanded version of the above chart with the most granular subsectors available), and California workers shed around $1.26 billion in inflation-adjusted earnings a week through this sectoral trading down. That equates to $65.4 billion annually, and this amount is actually an��understatement��of the composition change's effect because the total number of private sector jobs grew slightly (by 28,900, as shown in the rightmost column above) over this time period.

Diagnostic check

Could these findings be figments of the data — statistical or administrative blips that don't reflect economic realities on the ground?

Our analysis was conducted using two Bureau of Labor Statistics' (BLS) datasets, State and Area Employment, Hours, and Earnings (SAE); and QCEW. The latter, which is derived from state unemployment records, feeds into the former, which adds survey responses from BLS's Current Employment Statistics series to offer even more timely estimates. QCEW is derived from high-quality administrative data and forms the benchmark for numerous other labor market series. California's sheer size translates into high-quality data estimates across both sources.

Nor do there appear to be any California-specific biases in the largest subsectors driving the trends, such as Services for the Elderly and Persons with Disabilities (or NAICS 62412 in the classification system used by statistical agencies; see the footnotes). Concerns about how home health aides are scoped and categorized at the����— is each home they work in considered an establishment, or only the main employer or agency home office? — do not apply to this exploration of jobs. What is more, any anomaly in one corner of the Health Care and Social Assistance sector (NAICS 62) is unlikely to explain weaknesses in other sectors. It's highly unlikely that a care worker would be misclassified into manufacturing, retail, or tech, for example. Across the board, the state's relative estimates appear in line with others; they simply add up to a story that is unique to the Golden State.

California has the highest foreign-born share of the population of any state. Perhaps the spike in care jobs somehow extends from President Trump's immigration crackdown, as once-informal jobs are formalized or native-born workers begin to occupy the positions vacated by departing migrants?

Here too, the evidence doesn't match up. The largest increase in care economy jobs appears in the first quarter of 2024, well before the election. The rate of increase in elderly services held steady around 3 percent per quarter from Q3 2024 through Q3 2025. And while it's true that����are����in the home health care industry (NAICS 621) and undocumented immigrants occupy����of healthcare support roles, we would expect to find evidence to support the formalization hypothesis in construction and other industries, too, where the immigrant share of the workforce is even higher. The fact still stands that few sectors are expanding in the Golden State beyond healthcare.

Prognosis

What does the present composition of job growth mean for the future of the California economy? It's hard to tell, but the rate at which the Golden State is shedding jobs in its highest-value sectors like tech and professional services while adding them in decidedly duller corners with more dismal wages is cause for concern. With a gubernatorial election this year, a billionaire tax possibly on the ballot, and soon-to-be-former Governor Newsom preparing for a potential presidential run, closer examination of the California model is also of national interest.

California's population is not especially old or feeble. Only 16.5 percent of its population is 65 or older, making it the sixth-youngest��state on this measure. Californians also����fewer chronic health conditions than almost anywhere else. So why is its healthcare sector so mighty?

One plausible partial explanation is that state policies are driving the outcome — in other words, that government subsidies are allowing more people to access more healthcare and social support, stoking demand.

And indeed, many of those new therapy jobs are probably supported by the state's Children and Youth Behavioral Health Initiative, a $4.6 billion dollar����launched in 2021 to connect youth with mental health services. Enrollment in Medi-Cal, the state's medicaid program, has increased significantly over the past decade as the state has continuously expanded coverage, including in recent years to undocumented populations. As a result, only 5.9 percent of Californians went without health insurance in 2024, a��. The extensive margin (new enrollment) isn't the only one at work; the intensive margin (use of healthcare services) is too. The state legislature����Medi-Cal spending per enrollee is increasing even faster than the number of enrolled. Both greater utilization and greater coverage are driving demand and creating jobs in the process.

So how should we think about the care economy in the broader context of the California labor market? Is it an economic engine, a useful jobs sponge? Or does it represent an unhealthy dependence? Is the care economy picking up the slack left by weakness in other sectors or crowding out growth?

And given how intimately involved the public sector is in healthcare finance, we have to ask where the public support is coming from and what tradeoffs are involved with raising resources from one corner of the economy to spend them in another.

And then, finally, we have to ask how the political economy changes when healthcare spending effectively becomes a jobs program.

We might start to get some answers to these questions over the coming months, as the Medicaid- and healthcare-related����of the One Big Beautiful Bill Act work their way through state coffers. Medi-Cal's expansion was not only fueled by the state's swelling tax receipts after the pandemic. It was also made possible by federal largesse and a tax on private plans that will be subject to heightened federal����going forward.

States that expanded Medicaid under the Affordable Care Act now have a nearly 3 percentage point greater share of their workforce in healthcare jobs than those that didn't on average. A dozen����have triggers that could discontinue or dial back their support if federal funding dries up — which it is poised to do. We may soon get a better sense of the extent to which the healthcare jobs juggernaut has been fueled by Washington, and whether it has legs strong enough to stand independently.

Where does that leave us? Healthcare has driven the plot in the national jobs story for several years running. The subplot of weak growth in other good jobs sectors may carry the narrative from here.

They say the future happens first in California. What we don't know is how healthy it will look.

Originally ��on Agglomerations, the Substack newsletter from the Economic Innovation Group.

By Sarah Eckhardt and Nathan Goldschlag

A recent����found that 42 percent of bachelor's degree students have reconsidered their degree choice because of Artificial Intelligence. Another 16 percent say they have��already��changed their field of study due to AI.

Students consider current labor market conditions when choosing a degree, so it isn't too surprising that AI could shape students' decisions today.^[1]

But the study does not ask students exactly how AI has shaped their choices, only whether it has. Are students shying away from fields that have more exposure to AI, perhaps worried that AI will shrink the number of jobs available to them? Or are students shifting towards those fields, preparing for a future in which they will have to be comfortable using AI?

To find out, we can check enrollment for groups of degrees based on the AI exposure of the jobs that students with those degrees are likely to take, as shown in Figure 1.

As is clear, undergraduates are flocking towards��the most-AI-exposed degrees, with enrollment in those degrees up 8 percent last year compared to 2017. This trend holds despite a����in Computer Science degrees, one of the most-AI-exposed degrees, but whose decline is more than offset by increases in other exposed degrees like Engineering.

(A quick primer on our methodology before we move on: To estimate AI exposure, the general approach taken by researchers is to look at the tasks that workers perform in a given occupation and count how many of those tasks AI is likely to be good at. The more tasks associated with a given occupation that AI can perform, the higher the AI exposure of a worker in that occupation. AI exposure measures have important��, most notably that exposure can mean either augmenting tasks or automating tasks. We estimate the AI exposure of degrees chosen by college students by using the typical occupations that graduates of each degree end up choosing.^[2] See the Appendix for more details.)

Why are students still pursuing AI-exposed degrees? Part of the answer is shown in Figure 2: Wages are highest for those with the most-AI-exposed degrees.

In addition to our earlier work��showing that AI-exposed occupations have higher incomes, this finding is also consistent with����from Morgan R. Frank, Alireza Javadian Sabet, Lisa Simon, Sarah H. Bana, and Renzhe Yu. They used data from LinkedIn profiles, combined with degree-level AI-exposure inferred from 3 million higher education course syllabi, and found that after the launch of ChatGPT, highly-exposed students enjoyed higher salaries and found jobs more quickly.

What About Recent Graduates?

While students today have the chance to switch degrees, those who graduated before the rise of generative AI are stuck with the degrees they have. It is theoretically plausible that young graduates with AI-exposed degrees face lower labor demand in their field and are forced to find work elsewhere. Think software developer turned retail manager. But is it in the data?

We can use American Community Survey (ACS) data to see the degrees and occupations of young graduates. The table below shows, for example, the top ten occupations of young graduates with a Computer and Information Sciences degree and how they have changed — or not changed — over time. The overwhelming majority of Computer Science graduates work in Computer and Mathematics occupations, the share having fallen less than a percentage point from the pre-COVID average to the years 2023 and 2024.

It doesn't seem like young graduates with Computer Science degrees are upending their careers to find work. What about those with other AI-exposed degrees?

We first looked at the top three most-common occupations for each degree during the years 2015–2019. For Computer Science degree holders, this would be Computer & Mathematical, Office & Administrative Support, and Management occupations from the table above. We can then determine how many young graduates with each degree in subsequent years have also ended up in those same three occupations. The results are in Figure 3.

Across all three groups — Computer Science graduates; all graduates with the most-AI-exposed degrees including Computer Science; and all graduates with any degree — the share of graduates going to the most-common occupations has changed little over time. This result is consistent with a more-complex but generalized measure of how occupation choices change over time, as shown in the Appendix.

Taking Stock

There are several possible explanations for why we don't see students avoiding AI-exposed degrees or young graduates working in occupations outside their field.

First, how AI is impacting the job market today is not at all clear. As Jed Kolko��, this research is still in its early innings. If students are looking at the labor market to decide what to study, it's not clear how AI would reorder their choice.

Second, the content of degrees or occupations themselves may change. If the curricula of AI-exposed degrees change to focus on skills complementary or resilient to AI, students may not feel the need to avoid those degrees. Similarly, the mix of tasks workers do in highly-AI-exposed occupations may be changing in ways that leave plenty of work for newly minted college graduates.

Third, as we noted above, the AI-exposure measures themselves have a lot of limitations. Knowing that a job contains tasks that an AI��might��be good at doesn't necessarily tell you what will happen to employment in that job. This could help explain why Computer Science had decreased enrollment while engineering, which is also classified as highly AI-exposed, had increased enrollment.

Finally, there might be some behavioral frictions that slow the reallocation of students and graduates to occupations outside their field. Rather than looking far and wide for work, young highly-AI-exposed graduates might hold out for work in their field, waiting in unemployment longer.

Perhaps next year's enrollment data will show movement away from AI-exposed degrees, but we doubt it. Not all "exposure" is the same. Occupation and sector-specific bottlenecks, diffusion dynamics, task reallocation, and output demand effects will jointly determine labor demand — and, in turn, the degrees students choose.

See our github with replication code��.

APPENDIX

Classifying degrees by AI exposure

We classify groupings of degrees by AI exposure using two pieces of information: (1) occupational-level AI-exposure measures and (2) the cross-sectional mapping of degrees to occupations. We utilize the Eloundou et al. (2024) GPT-4 beta scores, the most commonly used measure in the literature, as our occupational measure of AI exposure. This approach essentially walks occupational exposure back onto degrees based on how often individuals with each degree are observed in each occupation��before��the introduction of generative AI.

We start by using a sample of college graduate workers between the ages of 22 and 27, inclusive, in the 1-year ACS for years 2009 to 2019. Using these years avoids contamination with AI-induced changes in degree-to-occupation flows and changes during COVID. In each year, each individual in the sample is observed to be in one of 39 groupings of degrees in the ACS's classification scheme and the Census 2010 4-digit occupation code. We compute the weighted mean of occupational AI-exposure scores for each degree-occupation pair. We then use this weighted mean to classify the 39 degrees into 5 quintiles based on AI exposure.

The AI-exposure scores are crosswalked to 2018 census occupation codes following our methodology in��"AI and Jobs", and to Census 2010 codes based on a crosswalk constructed using ACS samples with codes available under both vintages (occ2010 and occ for post-2018 samples). In some analyses, like Table 1 and Figure 2, we collapse occupation detail to 24 major SOC occupation groups. The degree-level exposure measures are shown below.

The ACS data is not well-suited to measuring how responsive college enrollment counts are to the introduction of generative AI. The ACS only reports college majors for those who have received their diploma. Students who graduated in 2023 and before were unable to adapt their degree choice in response to developments in AI that occurred in that year or after. Students graduating in 2024 would have needed to select a major by their sophomore or junior years (2023 or 2022), and so may have already been anticipating AI disruption when making this selection, but many would have been locked in to a major before ChatGPT's release in November 2022. The 2025 cohort of college graduates with a bachelor's degree are the most likely to have been capable of taking AI into account in their degree selection, but the 1-year ACS runs only through 2024.

To address these issues, we rely on the����reports the number of enrollees — graduate and undergraduate — by declared major. This provides a more timely estimate of student majors. The clearinghouse received enrollment data submissions from 97 percent of all Title IV, degree-granting institutions in the U.S. in the fall of 2024. Coverage rates vary by year, but are handled with a weighing system described on their methodology pages��. Degree majors follow a different classification as the ACS, though most majors match up one-to-one.

Additional Figures

The figure below shows the change in degree-to-occupation shares for young college graduates with a Computer and Information Sciences degree. The figure shows the L1 distance, which is computed as the sum of absolute differences between shares in a given year and the base year. It answers the question: "How different is the occupation mix of students with a given degree now than it was in the past?"

If students were forced to look outside their field for a job, this should show up as an increase, even if temporary, at the end of each of these lines — the occupation mix would be different from prior years. Across different baselines, changes in degree-to-occupation shares are relatively stable over the last few years.

The chart below shows the percentage change in enrollment between 2024 and 2025 by degree and AI exposure. Bubble size is weighted by 2025 enrollment. We can see the 8 percent decline in enrollment in Computer and Information Sciences, which is among the most-exposed degrees. Engineering and business degrees, which are also highly exposed, saw rising enrollment. These two different trends net out to the slight increase in enrollment for all highly exposed degrees shown above. The linear trend is statistically insignificant at the 5% level (-17.17 [8.97]).

��

References

Blume-Kohout, M. E., & Clack, J. W. (2013). Are graduate students rational? Evidence from the market for biomedical scientists. PLoS One, 8(12), e82759.

Eloundou, T., Manning, S., Mishkin, P., & Rock, D. (2024). GPTs are GPTs: Labor market impact potential of LLMs. Science, 384(6702), 1306-1308.

Frank, M. R., Sabet, A. J., Simon, L., Bana, S. H., & Yu, R. (2026). AI-exposed jobs deteriorated before ChatGPT. arXiv preprint arXiv:2601.02554.

Long, M. C., Goldhaber, D., & Huntington-Klein, N. (2015). Do completed college majors respond to changes in wages?. Economics of Education Review, 49, 1-14.

Ryoo, J., & Rosen, S. (2004). The engineering labor market. Journal of political economy, 112(S1), S110-S140.