°ż°ůľ±˛µľ±˛Ô˛ą±ô±ô˛âĚýĚýon Agglomerations, the Substack newsletter from the Economic Innovation Group.

By Adam Ozimek, Benjamin Glasner, and Jiaxin He

From national security, to productivity growth and innovation, to qualitative ideas of what a "good job" is, manufacturing holds a special place in the minds of policymakers.Ěý

Because the sector motivates so many policy objectives, it would seem we should know some basic facts, like how many manufacturing workers there really are. But while that sounds like a simple question, determining who counts as a manufacturing worker is actually a fraught empirical exercise with plenty of noise and gray space.

There are generally two ways to measure industry employment: surveying businesses and surveying workers. If you ask American businesses, there were around 12.5 million manufacturing workers in 2023. But if you ask workers, there were around 15 million—a difference of a whopping 2.5 million jobs.

Understanding the causes of this discrepancy is critical to approaching the topic of manufacturing's place in the American economy objectively. We discuss two methodological and reporting differences that mainly account for these "missing" 2.5 million manufacturing jobs and try to put them into the proper perspective.

1. Businesses report their industry to government statistical agencies at the "establishment" (location) level. Determining what constitutes a business "establishment" is ambiguous.
   
2. Workers often report their industry to agencies at the firm level or job level. Business categorizations have changed over time even while workers' jobs have not.
   

Grasping these factors and the discrepancies they generate helps to clarify the manufacturing jobs picture. While economists generally advise deferring to surveys of businesses, in this analysis we dig into why the disagreement occurs and argue that the higher, worker-reported numbers are meaningful as well. We think the "right answer" to how many manufacturing workers there are depends on the context and lay out here why sometimes the workers themselves are worth listening to.

How we count workers

To over-simplify just a bit, government statistical agencies ask businesses, "what kind of business are you and how many workers do you have?" and they ask workers, "what kind of business do you work at?"Ěý

Business surveys, usually called "establishment surveys," are filled out by the company. Businesses are asked to indicate the primary work done at a particular location or "establishment." Sometimes these data are derived from a literal survey filed to a government statistical agency, and other times the agencies start with information from tax returns. In some cases, businesses are legally required to reply and so we end up capturing information on nearly the entire universe of non-farm employer businesses.Ěý

Government agencies then use the information to categorize businesses into standardized codes using the North American Industry Classification System (NAICS). A handful of these codes are grouped into the wider category of "manufacturing."Ěý

Agencies then track employment over time at these businesses and aggregate that to the industry and/or geography level. This is how we measure, for example, how many manufacturing jobs there were in the Tuscaloosa, Alabama, metro area in September 1997 (14,814^[1]) or how many worked in food manufacturing nationwide in September 2025 (1.771 million^[2]). This broad approach is used for a variety of establishment surveys, including the Quarterly Census of Employment and Wages (QCEW), the Current Employment Statistics (CES), County Business Patterns (CBP), and the survey we will focus on most, the Business Dynamics Statistics (BDS). This approach is how the headline payroll growth is estimated with the CES every month on "Jobs Day."

The other main approach to measuring employment is to survey individuals instead of businesses. One example is the Current Population Survey (CPS) which is used to measure the unemployment rate. The Decennial Census and American Community Survey (ACS) are also done this way.

The ACS, which we will focus on most, asks the name of an individual's employer and also asks this about it:Ěý

"What kind of business or industry was this? Include the main activity, product, or service provided at the location where employed. (For example: elementary school, residential construction)."^[3]

The ACS then asks specifically if the employer is "mainly" in manufacturing, wholesale trade, retail trade, or other.

Figure 1: Actual ACS Questionnaire

Government statistical agencies categorize all of this information. Individual responses can then be aggregated to measure total industry employment.Ěý

When the data disagree

It turns out workers and businesses don't always agree about the industry they are in. To see this, we can compare the number of workers in each sector in two datasets: the ACS (worker reported) and BDS (business reported).^[4]

In 2023, the ACS recorded 15.1 million manufacturing jobs, while the BDS recorded 12.3 million, or 2.8 million fewer.^[5]

As the chart below shows, manufacturing employment is 22.7 percent larger in the ACS than in the BDS in 2023, while in most other industries the discrepancy is flipped.

This divergence is not just about picking the two surveys that are farthest apart. Other business surveys are fairly close to BDS at around 12 to 13 million jobs. The CPS, another individual-level survey, is very close to ACS at about 15.1 million. What we see is a general discrepancy between survey types. If you ask workers, there are at least 2.2 million more people employed in manufacturing than if you ask businesses, regardless which survey you choose.Ěý

If manufacturing employment were actually 22.7 percent bigger than we thought, that would have serious implications, perhaps weakening the widespread argument that manufacturing employment has been "hollowed out." To contextualize the discrepancy, the figure of 2.8 million manufacturing jobs is larger than the 2.5 million lost during the Great Recession, according to the Bureau of Labor Statistics' business survey.^[6] It's also larger than the 2 to 2.5 million estimated job losses from the China Shock.^[7] It would have made up for two-thirds of the 4.2 million total decline in manufacturing employment in business surveys from 2000 through 2023.

We can get even more detail on what missing manufacturing workers do from Emily Isenberg, Liana Landivar, and Esther Mezey (2013), who match ACS data to the Census Bureau's Longitudinal Employer Household Dynamics (LEHD) program, which covers the universe of firms and workers.^[9] They find that 23 percent of those identified in the LEHD as wholesale trade are counted as manufacturing in the ACS. The same is true for 28 percent of workers in management of companies, and 8 percent of workers in scientific, professional, and technical services.^[10]

What feels like manufacturing but isn't?Ěý

Do these discrepancies matter? Business surveys are generally seen as the more reliable estimate, since business owners or staff specifically designated with the task are expected to be more accurate than a random employee in identifying an establishment's industry. The government statistical agencies also do a lot of work to make sure these numbers are correct.

As John Haltiwanger, Henry Hyatt, and James Spletzer write, "The LEHD industry measures are of high quality from the establishment-level programs at BLS and Census. These agencies have a strong incentive to track industry carefully as their detailed industry statistics are critical for the NIPAs and productivity statistics."^[11]

There is also plenty of evidence that individuals make mistakes in reporting their industry even at the sectoral level. For example, economists Matthew Dey, Susan Houseman, and Anne Polivka document that temp workers commonly self-report the wrong employer and end up in the wrong industry as a result.^[12]

There is even more evidence that there is disagreement between types of surveys. Isenberg, Landivar, and Mezey found the same workers in both the ACS and LEHD and showed that the industry matched only 75 percent of the time.^[13]

But there are a few reasons to believe there is valuable information in the ACS, and that these workers in important ways really are manufacturing workers.Ěý

One reason is that many economists also think that the manufacturing sector is bigger than it seems.Ěý

Consider what Andrew Bernard and Teresa Fort "factoryless goods producing firms" or FGPFs.^[14] These are businesses that are involved in the production of goods, but may not be doing the raw assembly of the goods themselves. This can include pre-production activities like research and development (R&D), market research, product design, and product engineering. It can also include post-production activities like marketing, sales, logistics, and customer service. As Bernard and Fort write, "FGPFs are manufacturing-like as they perform many of the tasks and activities found in manufacturing firms."Ěý

In the U.S., Apple is the archetypal example of an FGPF. In the U.K., there is Dyson. Other examples include so-called "fabless" semiconductor companies like NVIDIA who design chips but don't make them.Ěý

Not only are many factoryless businesses "manufacturing-like," many of them were once involved in production as well. Apple, for example, used to assemble computers at its Elk Grove, California, campus. Today there is no assembly at that location, but thousands of Apple employees work there on logistics, distribution, repair, and customer support.^[15]

The site of Harley-Davidson's first factory in Milwaukee also does no production today, but is home to their corporate headquarters. This includes Harley-Davidson University, where employee training is done.Ěý

When a business stops producing goods at a specific establishment, that location stops "identifying" as manufacturing on business surveys. using one of those business surveys (of the type that includes the universe of businesses) showed that 40 percent of the lost manufacturing jobs from the China Shock were actually this type of phenomenon. ^[16] For these businesses, competition with Chinese imports didn't mean closing up shop. It just meant a shift away from production towards R&D, marketing, and other activities. In the Harley-Davidson example, all Harley-Davidson workers, in some sense, have a job in manufacturing, but according to establishment surveys only those at manufacturing sites do.

These sorts of employees—who work adjacent to manufacturing, but not in categorized establishments—make up a big chunk of the 2.2 to 2.8 million missing manufacturing workers.

It shouldn't be a huge surprise that some of the workers at these businesses think of themselves as working in manufacturing. After all, they are still contributing to the process of manufacturing goods, even if indirectly, and for many they might be doing the exact same non-production job they once did in a manufacturing establishment.Ěý

When the business is the same but definitions change

Changes in the operations of a business can clearly lead to missing manufacturing workers. Yet another issue is that sometimes the business hasn't actually changed at all, but definitions have. An important illustration of this occurred in 1997 when U.S. statistical agencies changed their industry codes, switching from the Standard Industrial Classification System, which had been in place but with evolutionary changes since the 1930s, to the more modernized NAICS codes.Ěý

This change was consequential for manufacturing in some datasets (importantly, not the BDS, which has consistent definitions over time). Previously when a manufacturing firm had a specific location that only engaged in R&D, the workers there would be counted as manufacturing. Under the NAICS system, those workers are instead classified based on the primary activity of their specific business location, not what the wider firm does. Workers at an R&D location for a manufacturing firm are not counted in manufacturing employment. The same is true of a location that is solely the headquarters of a manufacturing company.Ěý

The change in method makes it clear that what is counted as a manufacturing job is a statistical choice, and one that government statisticians have made differently in the past.Ěý

Altogether, from economists Teresa Fort and Shawn Klimek suggests that this change in industry definitions led to 1.4 million jobs being reclassified from manufacturing to services in the Census Bureau's Longitudinal Business Database (LBD).^[17] This likely represents another chunk of the missing manufacturing workers between at least some of the datasets.Ěý

What counts as a manufacturing establishment?

In some cases, the distinction between a manufacturing worker and a non-manufacturing worker seems almost arbitrary despite the huge policy implications.

Consider Boeing's 1,000-acre facility in Everett, Washington.

Boeing builds jets in the main assembly building, which is the largest building in the world. But there are 200 separate buildings at the facility that contain activities ranging from R&D to safety certifications. An important measurement question is whether the 1,000 acres constitute a single manufacturing establishment, or whether the various functions performed in the other buildings are treated as separate establishments, with each assigned its own industry code.

It's not clear which should be the case, nor is it clear which is the case for the various datasets we are considering. Yet the key issue of "how many manufacturing jobs we have" hinges on this question and many more instances like it.Ěý

Conclusion

The kind of work we care about varies by context. Whether a policy targets national security, productivity growth and innovation, or "good jobs" makes all the difference in how we should measure manufacturing employment.Ěý

Certainly when it comes to "good jobs," whether a worker thinks they are in manufacturing is more consequential than what the business itself thinks. If a local policymaker promises to bring back manufacturing to their town, and a firm makes 1,000 new hires in R&D, marketing, human resources, and sales for a product manufactured in Vietnam, would that be regarded as a policy success? It seems likely it would.Ěý

On the other hand, if a politician promised manufacturing jobs and instead a business was created that simply served as janitorial services for manufacturers, that would be unlikely to be counted as a policy success.Ěý

In the context of productivity and innovation, the R&D work that is often excluded from official definitions of manufacturing would seem to be just as relevant as—and arguably more so than—assembly. However, to the extent innovation is bolstered by co-location with production, having production on-site may matter as well.Ěý

For national security, one can make an argument in both directions. If the U.S. is cut off from assembly, then our ability to design, sell, and ship goods may be of little help. On the other hand, manufacturing R&D and logistics could be just as essential to military deployment.Ěý

Accepting that there is information in what workers tell us does not necessitate abandoning the importance of assembly itself or the general reliability of business-based surveys. However, we should make sure that when talking about manufacturing, we are clear that there is more than one measurement and definition that may be relevant. When it comes to what workers think they are doing, we have a lot more manufacturing than is commonly reported.Ěý

Ěý

Ěý

Ěý

By Sarah Eckhardt and Nathan Goldschlag

Recent tremors in the labor market are being pinned on Artificial Intelligence. A cooling job market for technology workers, for example, is of AI-induced job loss. So is the rising unemployment rate among .

Is your job at risk? Well, if you are reading this, chances are you do a lot of your work on a laptop — and , we are told, will be the first to go.

We can’t predict the future, but the good news is that we have all the data we need to assess whether AI is causing significant job losses right now. In this study we do two things with the data that we have not seen done in other investigations of it.

First, we conduct multiple analyses of several possible labor market effects beyond simply the impact of AI on employment or unemployment.

Second, we conducted our analyses using five different measures of AI exposure drawn from four different research papers. Most research and analyses of AI’s effects to this point have used only one such measure. (For simplicity and brevity, the main text of this study presents the findings from just one measure as well, but in the Appendix we explain how our findings are similar across the other four measures.)

We believe the consistency of our findings across all of the different labor market analyses and across the five different measures of exposure gives them an added weight.

—â¶Ä”â¶Ä”â¶Ä”â¶Ä”â¶Ä”â¶Ä”â¶Ä”â¶Ä”â¶Ä”â¶Ä”â¶Ä”â¶Ä”â¶Ä”â¶Ä”â¶Ä”â¶Ä”â¶Ä”

Who is Exposed?

Before we start looking at employment outcomes, we first need to know which workers are likely to be affected by AI and which workers are not. To assess exposure, a number of researchers have used occupation descriptions and information about the tasks that workers perform and those that AI is likely to be good at. The more tasks associated with a given occupation that AI can perform, the more exposed is a worker in that occupation.

In Table 1 below, we show the most and least exposed occupations according to the AI Occupational Exposure (AIOE) developed by researchers Edward Felten, Manav Raj, and Robert Seamans. (See the appendix for a detailed description and comparison of the various measures of AI exposure used in other studies. We also explain why we chose the AIOE for our analyses, though it’s worth noting that our findings were similar across all the available measures.)

At the extreme ends, both the most and least exposed occupations are intuitive. The most exposed occupations include genetic counselors, who assess risk for inherited conditions using gene panels and family histories, and financial examiners, who authenticate records to ensure compliance with laws governing financial institutions. These tasks, which are heavily reliant on the standardized processing of textual information, are things AI is likely to be good at.

On the other end of the spectrum are dancers and construction helpers, who hold materials or tools and clean work areas on a construction site. These tasks, which rely on in-person physical dexterity, are naturally less exposed to AI. At least for now, chatbots cannot pirouette.

For most of our analyses we will start by splitting occupations into five equally sized groups based on their AIOE score, with quintile 1 being the least AI-exposed and quintile 5 being the most exposed. Table 2 gives a sense of what AI-exposed workers look like.

One pattern is clear in the data: highly exposed workers are doing better in the labor market than less exposed workers. Workers more exposed to AI are better paid, more likely to have Bachelor’s or graduate degrees, and less likely to be unemployed than less exposed workers.

The gender split for the quintile of workers most exposed to AI is nearly even, while 70 percent of workers who are least exposed are men.

These characteristics are important to keep in mind when thinking about the subsequent analyses, as differential patterns in labor market outcomes between the most and least exposed workers will be heavily influenced by their differences in educational attainment.

I. Is the Job Apocalypse Here?

In Figure 1 below, we calculate the unemployment rate for our five groups of workers based on AI exposure.

What it shows is that although the unemployment rate for the most AI-exposed workers is indeed rising, it is actually rising even faster for the least exposed workers.

To be more precise, between 2022 and the beginning of 2025 the unemployment rate for the quintile of the most exposed workers rose by 0.30 percentage points. For the quintile of the least exposed workers, it climbed by 0.94 percentage points. The unemployment rate also climbed during this period for the other three quintiles.

By the most obvious measure, then, the effect of AI on jobs is invisible. But what about less obvious measures?

II. Bailing on Work?

Even though we see little evidence of AI’s impact on unemployment, maybe it’s because we are looking underneath the wrong lamppost. Rather than falling into unemployment, workers displaced by AI might be exiting the labor force entirely. Older workers, for example, might see the AI writing on the wall and decide to pack up and retire.

In Figure 2 we widen our lens, showing the percent of workers exiting the labor force by AI-exposure groups. Specifically, for each worker in a given AI-exposure group, we look at the probability that the worker is not in the labor force a year later.

Once again the disruptive effect of AI fails to appear. Highly exposed workers aren’t running for the exits and, in fact, the share exiting the labor force is actually the lowest for the most AI-exposed workers, and has been roughly flat since 2022.

III. Lateral Movement

Another response to AI-driven job displacement could be occupational switching. That is, maybe we don’t see rising unemployment or falling participation among highly exposed workers because they’re changing careers into less exposed occupations. Maybe computer programmers are becoming ballerinas.

To find out, we can again use the year-ahead framing. For a worker in a highly AI-exposed occupation in a given month, what is the probability that a year later they 1) work in a different occupation, or 2) work in a lower AI-exposed occupation?

Answering the first question gives us a sense of overall occupational switching by AI exposure. If AI diffusion is not uniform among our most-exposed workers, then highly exposed workers may switch occupations but remain in the same exposure group. Answering the second question offers a sense of whether highly exposed workers are fleeing to occupations with less of a risk that AI will affect their work.

As shown in Figure 3 below, highly AI-exposed workers tend to change occupations more often, but that probability is roughly flat since mid 2022 and remains lower than it was pre-COVID. It doesn’t look like workers more exposed to AI are changing occupations more than they used to.

As for the second question, the likelihood that the most AI-exposed workers switch to less exposed occupations began falling in 2019, dipped even lower during the pandemic, and now remains well below its 2018 level.

It simply does not look like AI-exposed workers are retreating to occupations with less exposure.

Another swing, another miss.

IV. Broadening the Scope: Firm Behavior

Yet another possibility is that we can’t see the impact of AI on the labor market because, to this point, we have focused on data about workers. Perhaps the trend we are looking for is happening at the firm level.

Firms might be reallocating tasks across workers in such a way that higher unemployment doesn’t impact the most AI exposed workers.

What might this look like? Suppose a firm has one high-skilled worker and two low-skilled workers. Imagine that AI can perform a third of the tasks done by the high-skilled worker. The firm may respond to this by reallocating tasks from one or both of the low skill workers to the high skill worker, reducing the firm’s total employment but not putting the AI-exposed worker out of a job. In this scenario, AI does have a dis-employment effect, but it does not fall on the AI-exposed worker.

This also highlights why it matters that our AIOE measure does not distinguish between AI exposure that displaces workers and exposure that augments them. The high-skilled worker, in our example, might use AI to conduct the tasks previously done by the low-skilled workers. Aided by AI, high-skilled workers might have broader roles and work in .

How could we see this in the data? A reasonable place to start is industry-level employment. Industries with more exposed workers should see declining employment as firms get smaller, even if the AI-exposed workers stick around.

Figure 5 groups industries based on their employment of AI-exposed workers, similar to how we grouped occupations. It shows total employment for industries in each group. Once again the effect of AI doesn’t leap from the page. Industries with the most AI-exposed workers account for a larger share of total employment and have enjoyed a steady increase post-COVID.

Something that an industry analysis like this one will struggle to capture is AI’s effect on new hires. Much has been made of the rising unemployment rates among . Since new labor market entrants, like recent college graduates, are a small portion of the labor force, they may not show up in this industry-level analysis.^[1] To understand the effects on young workers and recent college graduates, we need to look at them directly.

Is it the case that young workers or recent college graduates going into occupations more exposed to AI have worse labor market outcomes?

It appears not. The two solid lines in Figure 6 show the unemployment rate for the most AI-exposed young workers (blue) and recent college graduates (yellow). The associated dashed lines show the unemployment rate for less exposed young workers and recent college graduates.^[2]

Unemployment rates have been creeping up for young workers and recent graduates alike, whether they are AI-exposed or not.

V. Diffusion: The Last Hope

No big AI effects so far, but this could be because the actual diffusion of AI in the economy is slower than we expected. The Census Bureau asks businesses whether or not they use AI in the production of goods or services. A year and a half ago, only 5 percent said yes. The share is rising fast and is now at about 9 percent, but that remains a low figure. (See the appendix for a visualization of this trend. And note also that using AI in the production of goods and services is different from using it for incidental tasks like summarizing memos.)

That economy-wide 9 percent doesn’t tell the whole story though. There are some sectors where AI use is much higher. About 27 percent of businesses in the information sector say they use AI.

Perhaps the effects of AI will finally appear in the employment trends of the information sector. We can look for them in Figure 7, which shows employment in each subsector of the overall information sector.

The subsector with the highest use rate, publishing industries, where 36 percent of businesses report using AI, had a post-COVID surge in employment that then retreated and leveled off above its pre-COVID level, but is still below the 2021–23 trend.

Data processing and computing, which has the second highest AI use rate (35 percent), was rising until 2023 and has been flat since.

Two cases where employment had been growing and then flattened in the past few years — so have we finally found the effects of AI? If so, the effect is extremely modest, more of a brake, perhaps even just a temporary break, on employment growth in a few subsectors of the economy rather than anything that suggests a meaningful threat to jobs.

More importantly, though, there is no way to know whether AI specifically is the cause of the flatlining in those subsectors. In both cases, for instance, employment growth was quite healthy in the year right after COVID. Maybe companies simply realized they had expanded too quickly and wanted to slow down. Or it could be something else. The lack of causal evidence is a problem, in fact, not just for all of our analyses presented here but also for any investigation that uses the available measures of AI exposure. We expand on the point in the next section.

Still Haven’t Found What We’re Looking For

Why, no matter how we cut the data, do we not see any meaningful AI impacts in the labor market? We consider two possibilities.

Garbage-In Garbage-Out?

It could be that our occupation and industry classification doesn’t do a great job of differentiating where AI is actually having an impact. Lucky for us, there are a number of academic studies that attempt to identify AI-exposure in the labor market, some of which use significantly different methods of identifying what workers are likely to be affected by AI. (See appendix.) Some use human-curated lists of exposure and some even use AI to classify which tasks are exposed to AI.

We’ve run our analyses with each of these different measures and the results are broadly consistent. The trendlines in the unemployment rates for the most AI-exposed workers, for example, look fairly similar across all of the measures. The same goes for exiting the labor force, changing occupations, and switching to lower exposure occupations.

Having said that, we can detect a slight statistical difference in recent unemployment rates between AI-exposed and not using two of the measures. Figure 8 below shows the difference in unemployment rates between the most exposed and everyone else between the 2022-2023 and 2024-2025 periods.

The measure from scholars Tyna Eloundou, Sam Manning, Pamela Mishkin, and Daniel Rock — labeled “Eloundou et al. (2024) (gpt-4)” in Figure 8— shows about a 0.2 percentage point increase in unemployment (relative to the less exposed workers).

The measure from Andrea Eisfeldt, Gregor Schuber, and Miao Ben Zhang — “Eisfeldt et al. (2023)” — shows about a 0.3 percentage point increase (relative to the less exposed workers).

Ironically, these are two measures that use large language models to assess AI exposure. So perhaps, if we squint, AI is telling us that AI is increasing unemployment. But the differences found by these two measures are really quite small.

Finally, it’s important to again emphasize the point about causality. In truth, by merely looking for any visual sign of AI effects in our analyses, we set the bar quite low — and yet even that low bar was not cleared.

Even if it had been cleared, the most we would have been able to say is that highly AI-exposed workers were experiencing the labor market differently than other workers. But there might have been other reasons besides AI for that to be the case.

Remember that AI-exposed workers are very different from other workers. They tend to have higher education, lower unemployment rates, and higher wages. It could be that those characteristics — or any other — drive differences in employment trends rather than AI.

Occam’s Razor?

We don’t see the effects because they have not materialized — at least not yet. Official statistics on AI diffusion, as noted already, tend to be lower than popular discussion might lead you to believe.

We also know, from , that the vast majority of firms (roughly 95 percent) report that AI had no net impact on their employment, and about the same number said it led to an increase in employment as said it reduced employment.

When firms were asked how they expect their employment to change 6 months into the future based on AI use, roughly 6.5 percent expected an increase and 6.1 percent expected a decrease.

Ultimately, in the short and medium term, AI may end up having a bigger impact on tasks than on employment. About 27 percent of AI-using firms reported that they used AI to replace worker tasks.

Even if AI isn’t totally reordering the labor market today, it may yet have a different kind of big effect: reorganizing how workers spend their time.

AppendixĚý

A1. Defining AI ExposureĚý

The academic literature has produced several measures of worker exposure to AI. Typically, these measures start by analyzing worker tasks, identifying tasks that are “exposed” to AI (which can mean multiple things) and then rolling those tasks up to occupations.

Once exposure is defined at the occupation-level, you can take that measure into worker-level data and, through industry-occupation matrices, to industries. Below we describe several of the AI exposure measures we’ve collected for this piece.

I. The first AI measure we consider, and the primary one used in this piece, was produced by Felten et al. (2021). We chose this measure for several reasons. First, it provides ability-level exposure measures, which makes the crosswalking between occupation code schemes described below more accurate. Second, the authors provide industry-level exposure measures, which proved useful in the analyses above.

Their exposure measure is built on a subset of common applications of AI based on categories provided by the Electronic Frontier Foundation (EFF) AI Progress Measurement project.

EFF maintains statistics about the progress of AI across different applications. Felten et al. (2021) combine EFF data with data on workplace abilities from the Occupational Information Network (O*NET) database from the Department of Labor. The authors link EFF’s AI progress measures to O*NET data by surveying approximately 2,000 “gig” workers via Amazon’s Mechanical Turk (mTurk) web service. Each respondent was asked to consider how an AI application is related to each O*NET ability. This provides a measure of application-ability relatedness, which is then summed to create an ability-level exposure score.

Felten et al. (2021) then compute occupation-level exposure by summing ability-level AI exposure by ability weighed prevalence and importance as measured by O*NET, then scale it by the weighted sum of prevalence and importance of all abilities to obtain a relative exposure measure.

Once at the occupation level, the authors use industry-occupation matrices to measure industry-level exposure, which they combine with county-industry data to measure geographic AI exposure.

Importantly, this measure is agnostic as to whether AI is a complement or substitute to workers, meaning that higher exposure could indicate increased productivity or a higher risk of job displacement.

II. The second AI measure we consider was developed by Eloundou et al. (2024). The authors use a rubric to determine whether a specific task is exposed to AI. That rubric classifies tasks based upon whether generative AI would reduce the time required for a human to perform a specific task.

Tasks are placed in three groups: (E0) tasks with no or minimal time reduction, (E1) tasks that could be completed 50 percent faster with access to an LLM, and (E2) tasks where an LLM alone would not reduce the time to complete by 50 percent but additional software on top of an LLM could reduce the time to complete by at least half, or (E3) tasks could not be completed 50 percent faster but would be given access to an image generation system.

Eloundou et al. (2024) classify tasks into E0-3 in two ways. First, using human annotations, with the authors and experienced annotators that have experience using OpenAI’s LLMs. Second, using GPT-4 to classify tasks. The authors construct several measures using different combinations of these classifications. We focus on the author’s preferred “beta” measure, which is the sum of tasks classified as E1 and 0.5 times the sum of tasks classified as E2 or E3, divided by the total number of tasks for an occupation. The 0.5 weight on (E2/E3) places less weight on tasks where complementary tools and investments are required. The authors note several weaknesses of their classification approach, but also find strong agreement between the human and GPT-4-based classifications.

The human-based classifications of Eloundou et al. (2024) are conceptually related to the Felten et al. (2021) measure, which also used surveys of humans to determine exposure. The GPT-4 measure, which uses AI to classify exposure to AI, is methodologically distinct from the Felten et al. (2021) measure.

III. Eisfeldt et al. (2023) study how firm-level exposure to generative AI impacts firm values. The authors estimate firm-level exposure to AI in two steps. First, using a process similar to Eloundou et al. (2024), the authors use Open AI’s GPT 3.5 Turbo to classify whether a task can be done by the current generative AI or by future instantiations of generative AI conditional on complementary investments (E0-E2 as in Eloundou et al. (2024)). The authors exclude E3 and compute exposure as the count of tasks classified as E1 plus 0.5 times the count of tasks classified as E2 divided by the total number of tasks.

There are three primary differences between the AI exposure measures produced by Eisfeldt et al. (2023) and Eloundou et al. (2024). First, Eloundou et al. (2024) include image capability dependent automation (combining E2 and E3). Second, Eloundou et al. (2024) uses ChatGPT-4 to perform the classifications while Eisfeldt et al. (2023) use GPT 3.5 Turbo. Finally, Eisfeldt et al. (2023) distinguish between core and supplementary tasks. The core task-based measure is meant to more accurately capture labor substitution, which would more likely generate unemployment effects. For this reason, we focus on the core-based measure in our analyses.

IV. The final AI exposure measure we consider was developed by Webb (2022). Webb (2022) uses text data from both patents and job descriptions to measure how much patenting in a particular technology has been directed at the tasks of a given occupation. Using verb-noun pairs, Webb (2022) quantifies the overlap between patents and tasks, aggregating the frequency of overlap at the task-level to the occupation-level. After validating the methodology with case studies focusing on robots and software, Webb (2022) creates an AI-exposure measure by focusing on AI patents and identifying the tasks, and in turn occupations, that have the most overlap with those patents. The Webb (2022) measure is methodologically distinct from the other measures because of its focus on patent documents. Advances in AI technology that overlap with worker tasks that are not mentioned in the patent corpus would not be captured by the Webb (2022) measure.

Unfortunately, with the exception of Webb (2022), these measures do not use an occupation coding scheme available in the CPS. To address this, we apply a series of crosswalks to translate each AI exposure measure’s native occupation scheme to occupation codes available in the CPS, which we describe in Appendix A2.

A2. CrosswalkingĚýĚý

The Current Population Survey identifies occupations using the Census’s , while the AI exposure measures are based on either Standard Occupational Classification (SOC) codes, or David Dorn’s occupation code system.Ěý

David Dorn’s codes are designed to be mapped onto the Census’s 1990 occupation code system, for which Dorn provides a crosswalk.^[3] Using IPUMS’s OCC90 variable, which provides a harmonized 1990 occupation code for all respondents regardless of survey year, Webb (2022) exposure measures can be mapped directly onto our CPS sample from IPUMS.

In addition to occupation, industry, and geographic exposure, Felten et al. (2021) provide an ability-level exposure measure. This ability-level exposure eases the translation between occupation coding schemes.Ěý We directly translate Felten et al. (2021)’s measure onto the most current 2018-based Census occupation codes by crosswalking O*NET 2018 SOC codes to Census 2018 codes using Census’s many-to-many official crosswalk, then weighing ability-level exposure using O*NET’s importance weights at the occupation level.

SOC-to-Census Crosswalks: Approach 1

For the Eisfeldt et al. (2023) and Eloundou et al. (2024) measures, we explore two approaches to translating from their native SOC schemes to Census 2018 occupation codes. Our first approach uses the SOC 2018 to Census 2018 official crosswalk^[4] and computes the average exposure across all SOC 2018 codes that map to a given Census 2018 code. A similar process is used for translating SOC 2010 occupation measures to Census 2018 occupation measures.

SOC-to-Census Crosswalks: Approach 2

We explored an alternative approach to crosswalking SOC codes to Census 2018 codes. In this alternative approach, we disaggregate the SOC codes into O*NET tasks, carrying along the AI-exposure weights, and then re-aggregate to Census 2018 occupation codes using O*NET ability importance weights. The initial disaggregation creates a many-to-many relationship between AI-exposure measures and tasks because many occupations share tasks.

Since Felten et al. (2021) provide ability-level exposure scores, we can use that information to evaluate the two crosswalking approaches. We treat the ability-level Felten (2021) translation of SOC 2010 to Census 2018 occupation codes as “truth” and compare it to the Census 2018 occupation-level exposure measures resulting from both Approach 1 and Approach 2. We find that the measures reliant on occupation-to-occupation crosswalks, Approach 1, produces AI exposure measures most similar to the ability-weighted “truth” measure.

In all analyses using the Eisfeldt et al. (2023) and Eloundou et al. (2024) measures we rely on crosswalk Approach 1 to move from the native SOC occupation codes to Census 2018 occupation codes.

A3. Labor Market Outcomes and Alternative Measures of AI Exposure

Figure 10 shows the most AI-exposed (quintile 5) for each AI-exposure measure along with the unemployment rate for less exposed workers (dashed line).

Figure 11 shows the probability of exiting the labor force for the most AI-exposed (quintile 5) for each AI-exposure measure along with the exit rate for less exposed workers (dashed line).

Figure 12 shows the probability of changing occupations for the most AI-exposed (quintile 5) for each AI-exposure measure along with the less exposed workers (dashed line).

Figure 13 shows the probability of changing to a lower exposure occupation for the most AI-exposed (quintile 5) for each AI-exposure measure along with the less exposed workers (dashed line).

The industry-level analysis cannot be done directly with any of the exposure measures except Felten et al. (2021). One contribution of Felten et al. (2021) is the translation of occupation-based AI-exposure to NAICS industries.

A4. Additional FiguresĚý

Figure 14 shows the percent of businesses that report using AI in the prior two weeks.

Figure 15 shows the percent of businesses that report using AI in the prior two weeks, and the percent that expect to use AI in the next six months, by major sector for the most recent period, June 18th 2025.

References

Eisfeldt, A. L., Schubert, G., & Zhang, M. B. (2023). Generative AI and firm values (No. w31222). National Bureau of Economic Research.

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

Felten, E., Raj, M., & Seamans, R. (2021). Occupational, industry, and geographic exposure to artificial intelligence: A novel dataset and its potential uses. Strategic Management Journal, 42(12), 2195-2217.

Webb, M. (2019). The impact of artificial intelligence on the labor market. Available at SSRN 3482150.

By Adam Ozimek

Protectionists love talking about the auto industry. Believing it offers a potent example of the harms of globalization, their arguments have long been politically attractive to politicians on both left and right. Most recently they have justified the Trump administration’s 25 percent tariffs on auto imports by emphasizing the long-term decline of the industry.

It is time to set the record straight.Ěý

The protectionist argument for insulating the American auto industry from foreign competition not only draws the wrong lessons from history, it gets the history itself wrong. It rests on four myths, all of which I debunk in this analysis:

1. The U.S. auto industry has collapsed.
2. Globalization caused the death of Detroit.
3. Japanese imports nearly destroyed the auto industry in the early 1980s…
4. … until auto protectionism saved it.Ěý

Once these myths are set aside in favor of a clear, accurate understanding of the auto sector and its history, there is no reason to be optimistic that the Trump administration’s protectionist approach to the sector will work as intended. Indeed the case for it falls apart entirely.

I. The Auto Industry: Doing Just Fine?

The backdrop for President Trump’s tariffs is the idea that the auto industry has been decimated by globalization. “No one anymore, on the left or the right, denies that globalization has fractured the U.S., both economically and socially,” Joe Nocera at the Free Press. “It has hollowed out once-prosperous regions like the furniture-making areas of North Carolina and the auto manufacturing towns of the Midwest.”

The situation is so dire, Steven Miller , that “if we stayed on this current path, within a few years there would have been no US automobile industry.”

Such claims of vast deindustrialization ring true for certain manufacturing industries. Some goods really did stop, or mostly stopped, being made in the United States because of globalization.

Apparel, for example, is a quintessential globalized good, its factories shifting across the globe in search of the lowest labor costs. The United States once made a lot of clothes. Today it more than 90 percent fewer workers in apparel than it used to, and 90 percent less of the output. Apparel was a classic “China Shock” industry, where imports caused substantial and long-lasting economic disruptions in the parts of the country where it used to be concentrated. ^[1]

But the domestic auto industry is different. It remains alive and well, with 10.5 million vehicles assembled in American factories last year. This number is down from the peak of the post-NAFTA boom period, but it is well above the depressed years of the 2000s and nearly equals the average of 10.3 million annual vehicles made in the pre-NAFTA period dating back to 1969.

And these production numbers actually fail to capture the true strength of the industry. The economic value of the cars being made has climbed substantially through the years. As a result, real value added and industrial production — two different ways of measuring actual output — are now at all-time highs.Ěý

What about jobs? The auto industry today employs 1 million workers. Between 1950 and the signing of NAFTA in 1993, it averaged 1.1 million workers, just slightly higher. ^[2]

So much for Myth 1, the persistent notion that the American auto industry has collapsed. With output at all-time highs and employment hardly lower than in the time before NAFTA (the biggest globalizing event for the auto industry in recent decades), the evidence goes hard in the other direction.Ěý

II. The Detroit Whodunnit

But we are left with a puzzle. The perception that the auto industry has been decimated — and decimated specifically by globalization — is widespread. Where does it come from?

The likely answer is that in Detroit, the decline of the auto industry is certainly not a myth. But its very real decline was caused by competition not with the rest of the world, but with the rest of the United States.

The deindustrialization of Detroit is typically understood as a phenomenon of the 1970s and 1980s, and it is therefore blamed on the growth of trade during this period. But the fact is that auto investment and employment had started moving out of Detroit decades earlier.Ěý

I pieced together data from a variety of sources, which shows that auto manufacturing employment in the City of Detroit had already peaked in 1950, at just over 220,000 workers. ^[3]

By 1970 the biggest declines had already occurred, with employment falling by more than half, to fewer than 100,000 jobs.Ěý

An important nuance is that many of these lost jobs migrated to other parts of Michigan, at least for a while. So while auto employment was collapsing in Detroit, the rest of Michigan managed to hold auto employment stable for another five decades until the 2000s, when it started falling everywhere in the state.

The steady employment outside of Michigan did not, however, do anything to offset the big declines in Detroit. For all of Michigan, both Detroit and ex-Detroit combined, employment also peaked in 1950 and, except for a couple of blips, has been mostly in decline since. Because although the Big 3 (Ford, General Motors, and Chrysler) were moving some jobs from Detroit to other parts of Michigan, they were moving even more jobs out of the state entirely.

The exodus from the city proper is also evident in the investments of the Big 3. From 1946 to 1956, they built in the broader Detroit metro area, but not a single one within the City of Detroit itself. In 1940, 58 percent of Ford’s assembly plants were located in the City of Detroit. By 1956, none were. ^[4]

The historical record paints the picture. Henry Ford II announced in 1950 that his company’s investments would no longer be concentrated in their established industrial centers. By the mid-1960s, Ford had made major investments not just in the southern states of Alabama, Tennessee, and Georgia, but also in New York and New Jersey.Ěý

For its part, GM made investments in Indiana, Ohio, Illinois, New Jersey, Mississippi, and California, while Chrysler invested in New York, Delaware, Indiana, and Ohio — all by the late 1950s.

These investments and the jobs that followed them did not occur as a result of foreign car companies establishing their factories in the south and other parts of the country. That came much later. Long before the Big 3 were seriously challenged by foreign automakers, they themselves had begun decentralizing auto investment out the industrial hub of Michigan and spreading jobs and investment across the rest of the country. ^[5]

So globalization couldn’t kill Detroit because Detroit was already dead. Or at least dying.

And it really was a story of decline in Detroit almost exclusively. From 1950 to 1980, auto employment even in most of the Rust Belt states expanded. The gains in Ohio, Indiana, and Illinois exceeded those of any southern state.

If not globalization, what drove this push out of Detroit?Ěý

One big factor was the militancy of local unions. Consider what happened at Ford’s . Located in the Detroit suburb of Dearborn, it was once the largest integrated factory in the country with 93 structures, 90 miles of railroad track, and 53,000 machine tools. Its workers belonged to the United Auto Workers (UAW) Local 600 union. Throughout the 1940s, the Local 600 organized hundreds of strikes that would routinely shut down production at River Rouge.Ěý

Citing this union activity as the cause, Ford Motor company relocated production to other parts of the country — building a stamping plant in Buffalo, NY, an engine factory in Brook Park, OH, and other factories in Tennessee and California.Ěý

Between 1941 and 1960, employment in the River Rouge Complex fell from 90,000 workers to just 30,000. ^[6]

The Big 3 were not avoiding all unions. In a 1950 speech about Ford’s new investments to the Buffalo Chamber of Commerce, Henry Ford II praised Buffalo labor leaders as forward-looking, and Buffalo itself as “the place where an organization can get work done — where good production cooperation is possible.”^[7] It was specifically the local unions of Detroit that compelled the Big 3 to find new places to set up factories.Ěý

The decentralization trend accelerated in the 1970s when foreign automakers began establishing their own factories in the United States, and then continued for decades beyond the 1980s.Ěý

The desire of the foreign automakers to avoid Michigan and its unions was even stronger than it was for the Big 3. A letter from Mitsubishi to U.S. Representative Mary Rose Oaker, written in 1985 to explain why Cleveland, Ohio had been rejected as the site of their first plant, makes their rationale explicit: “The rule of thumb we have been using in our site selection process is to avoid going right into the heart of any existing heavily automobile industrial region.” ^[8]Ěý

And while Nissan chose Smyrna, TN in 1980 for a variety of reasons, Tennessee’s status as a right-to-work state was a big one, as it was “far enough from the industrial North to be beyond the reach of the UAW.” ^[9] As the first manager of that plant said, “You won’t get the cooperation necessary to build a quality product with the union.” ^[10]

Michigan now has about 280,000 fewer auto jobs than it did in the 1950s, a decline of roughly 60 percent. ^[11]ĚýFor the United States as a whole, auto employment is only down 4.7 percent — further showing that the struggles of Detroit and Michigan are less about the decline of the American auto industry and more about its relocation elsewhere.Ěý

Another way of understanding the trend: If Michigan had simply maintained the same share of American auto jobs as it had in the 1950s, meaning it did not lose any production to other states, then it would only have lost 21,000 auto jobs since then, not the 280,000 it actually did lose.Ěý

Should the federal government have done more to help Detroit weather its deindustrialization? Maybe. A strong argument can certainly be made for it. But no argument can be made that protectionism would have helped, as tariffs and other trade barriers have absolutely nothing to do with internal competition between states.

III. Japanese Imports and the Crash of 1979–82

While the struggles of Detroit started back in the 1950s, a genuine crisis for the wider American auto industry did not arrive until nearly the end of the 1970s. But when it finally hit, it hit hard.

More than 400,000 auto workers lost their jobs between 1978 and 1982 — a staggering 30 percent decline — as the industry confronted its biggest challenge since the Great Depression. Domestic auto production had collapsed. American companies made fewer than 7 million vehicles in 1982, a 46 percent decline from their output just four years earlier.Ěý

Some of the loudest voices from across the political spectrum placed the blame on Japanese imports. Demands for protectionist policy came not just from anti-trade crusaders but also from important figures inside the presidential administrations of both Jimmy Carter and Ronald Reagan. (More on the effects of their policies later.)Ěý

Was globalization, and specifically the flood of auto imports from Japan, responsible for the crisis?Ěý

The first point to make is that foreign competition did not arrive overnight. It had grown steadily over a long period. The first auto import was the Volkswagen Beetle, from West Germany, in 1949. Cars from the European automakers Austin Healey, MG, Jaguar, and Volvo arrived next. ^[12] In 1957 came the first Toyota imports, followed by the Datsun Bluebird from Nissan the next year.

These early imports were no threat to domestic automakers, who were dismissive of the new competition — especially the Japanese imports, which Detroit perceived as “shoddy, tinny” cars that “rattled and fell apart.” ^[13]

That initial perception was largely right. Toyota’s first imports were designed for low-speed driving on Japan’s famously bumpy urban roads. The fast speeds of American highways caused them to overheat and, yes, sometimes rattle and fall apart. Toyota stopped importing them within a few years and went back to the drawing board. ^[14]

But gradually the pressure from foreign competition increased. In the 1950s, the Big 3 carmakers were a comfortable oligopoly and faced little pressure to innovate. That would soon change.

As of 1957, only 4 percent of cars sold in the United States were foreign imports. ^[15]) Exactly a decade later, the share had tripled to 12 percent, and foreign imports had climbed to more than a million cars per year for the first time. The Volkswagen Beetle continued leading the way as the nation’s best-selling import. ^[16]

Japanese companies had meanwhile pioneered new and shockingly effective production methods, pushing their cars to the forefront of quality and cost. The cars were also, finally, well matched to American roads and drivers. In 1974, imports from Japan by themselves exceeded a million cars, boosting total foreign imports to more than 2.5 million — a fifth of all cars sold in the United States. ^{[17] [18]}

And yet despite the globalization of the industry, the steep rise in imports had failed to devastate American producers. Instead, domestic production and employment from the late 1960s to the late 1970s had risen together with imports.

In 1978, when imports hit a record high of more than 3 million vehicles, American auto producers had their best year ever, with 12.9 million cars assembled.Ěý

If globalization had been destined to eviscerate the auto industry, it would have happened much sooner.Ěý

As for what did bring about the crisis of 1979–82, the likely culprit was the cratering of consumer demand throughout the American economy, which itself had various causes.Ěý

An oil shock in 1979 had pushed inflation above 10 percent. The Federal Reserve responded by raising interest rates to nearly 18 percent in March of 1980 and then keeping them elevated for several more years. The unemployment rate spiked and nearly reached 11 percent at the end of 1982, by far its highest level since the Great Depression. Quite simply, fewer people were making money, and the cost to borrow it was prohibitive.Ěý

That auto sales fell as they did should therefore have been no surprise.Ěý

Domestic auto companies took the biggest hit. Roughly 40 percent fewer vehicles made in the United States were sold in 1982 than four years earlier. Disappearing revenues forced American automakers to cut $10 billion in costs, leading to factory closures and layoffs. ^[19]

Chrysler was pushed to the brink of bankruptcy and needed a federal bailout from the Carter administration in the form of a $1.5 billion dollar loan guarantee. Ford stock fell by more than half, and its credit was downgraded by ratings agencies. ^[20]) (Ironically, Ford might also have faced bankruptcy, or at least been forced into a merger, were it not for its profitable overseas operations. In a sense it was saved by globalization. But protectionists tend to leave out that bit of the story.) ^[21]

During this time, sales of imports were more resilient, rising from 3.1 million vehicles to 3.3 million between 1978 and 1982 (see Table 1 above). Because the sales of American-made vehicles had fallen so much, importers dramatically increased their share of the market — setting the stage for what came next.ĚýĚý

Import competition, especially from Japan, became the scapegoat for the domestic industry’s problems. Demands for protectionism then followed. Ford and the United Auto Workers filed a complaint with the U.S. International Trade Commission (ITC), which included a warning from the UAW president: “The auto industry here is in danger of losing up to one-third of the U.S. market permanently if action is not taken rapidly.” ^[22])

But the ITC rejected the claim, denying that imports were a “substantial” cause of the industry’s problems. ^[23] Simple math, with the help of a thought experiment, strongly supports the ITC’s argument.

Counterfactual 1: Imagine a counterfactual world in which demand had not crashed. Annual sales of vehicles remained steady at 1978 levels throughout the crisis years. But in this counterfactual, foreign importers still increased their market share by the exact same amount as they did in reality, rising from 20 to 31 percent of all sales in 1982.

In such a world, American producers would have sold 10.6 million vehicles in 1982. This is fewer than the number sold in 1978, but only by 14 percent — which may not be great, but is still just a third of the decline that happened in real life (a 41 percent decline).Ěý

Put another way, American producers in this counterfactual would have sold roughly the same number of vehicles as in 1976, just three years before the crisis started.ĚýĚý

Counterfactual 2: Now imagine an entirely different counterfactual, one in which demand does crash (as indeed it did in real life), but foreign importers fail to gain any market share, remaining at 20 percent of vehicles sold in the United States.Ěý

In this world, American automakers would have sold 8.4 million vehicles, a 32 percent decline — much closer to the actual catastrophe that happened in the real world.Ěý

Conclusion: Forced to choose between the collapse of demand (a 32 percent decline) or the rise of imports (14 percent decline), it is clear which counterfactual path the industry would have chosen.Ěý

—â¶Ä”â¶Ä”â¶Ä”â¶Ä”â¶Ä”â¶Ä”â¶Ä”â¶Ä”â¶Ä”â¶Ä”â¶Ä”â¶Ä”

A close look at truck sales offers another reason to doubt imports were the problem. ^[24]

Imports of light trucks rose 5 percent between 1978 and 1982, which was weaker growth than the 11 percent rise for imports of all vehicles. Yet despite the slower growth of truck imports, sales of domestic trucks actually fell by more than sales of domestic vehicles overall (see Table 2).ĚýĚý

Macroeconomic conditions — the weak economy, rising gas prices, high unemployment — were going to be a problem for the industry regardless of imports.Ěý

While rising imports left the domestic auto industry more fragile than it otherwise would have been, they were just one among many problems. A lack of competition in the post-war period had lulled the Big 3 into a period of sclerosis, low innovation, and labor costs that had grown much faster than productivity. ^[25] The problems were most visible at Chrysler, which, in 1978, a year of record-high auto demand, lost $205 million and found itself $1 billion in debt. ^[26]

The domestic automakers also struggled because they simply weren’t making enough of the small cars that customers wanted when gas prices skyrocketed. They believed that small cars were a “profitless hole.” Their unwillingness to make a quality, affordable, small car in the 1960s was so obvious that UAW president Walter Reuther asked Lyndon Johnson to suspend antitrust enforcement and allow the Big 3 to collaborate on one. ^[27]

The growing share of imports no doubt placed competitive pressure on domestic producers. But in a healthy macroeconomic environment in which total demand was strong, as it had been for the decade prior to the crisis, there would have been no crisis regardless of rising imports.Ěý

The dominant effect of consumer demand on the American auto industry can also be shown using statistics and a simple chart. In Figure 8 below, for each year between 1968 to 2024, the number of vehicles sold in the United States is plotted against the number of vehicles assembled inside the country (a measure of production).Ěý

The precise relationship is that every 1 percent increase in total vehicles sold is associated with a statistically significant 1.2 percent increase in domestic production.Ěý

In contrast, neither the import share of sales nor the percent change in imports has a statistically significant effect after controlling for total sales (Table 3). In fact, if total sales are not controlled for, higher imports are correlated with higher domestic production.Ěý

A good market swamps everything, benefitting producers both domestic and foreign.

IV. A Protectionist Success Story?

Nonetheless, the scapegoating of Japanese imports and the mass layoffs at the Big 3 led to mounting pressure on policymakers to respond.Ěý

The Reagan administration found itself caught between its free trade principles and its campaign promise to help American auto workers. As a compromise, the administration pressured the Japanese government into imposing voluntary export restraints (VER), which limited exports of Japanese autos into the United States. For three years, starting in 1981, Japanese carmakers combined were allowed to sell no more than 1.68 million cars per year to American buyers.Ěý

The goal was to give the American auto industry “breathing room” — a phrase that pops up again and again in and the . The industry was in a crisis, and this was the plan to help it adjust.

Did the policy succeed?Ěý

There are two reasons to doubt it. First, it had no significant effect on the primary goal, which was to help the Big 3 get through the crisis. Second, contrary to protectionist claims, the policy is not why Japanese automakers now produce so many cars from inside the United States.Ěý

The conclusion of the crisis

The most important reason the American auto industry eventually did get through the macroeconomic crisis was that the macroeconomic crisis ended. December of 1982 was the last month of the recession. By then, interest rates had fallen all the way to their 1979 levels, while real gas prices (gas prices adjusted for inflation) had fallen halfway back to their pre-crisis levels. ^[28]) ^[29])

The economic stage was set for the auto industry to start recovering in 1983. That is exactly what it did, with total vehicle sales that year returning to 96 percent of the pre-crisis average. By 1984 they were 12 percent above it. ^[30] And after having lost a combined $8 billion during the crisis, profits for the Big 3 hit a record $10 billion in 1984. ^[31]

Profits were so high, in fact, that they enabled the UAW to negotiate a deal that included, for the first time, the infamous “jobs bank” requiring the automakers to find new employment for laid-off workers and continue paying them until then.Ěý

Another factor helping domestic producers was their rollout — better late than never — of fuel-efficient small cars with front wheel drive, which started in 1979 and 1980. These cars were introduced, notably, before the Voluntary Export Restraints on Japan went into effect. Included among them were the Chrysler K and GM X series. ^[32]

While the improved economy and industry adjustments put an end to the crisis for the Big 3, it’s far less likely that VER played much of a role. The policy had little impact until the crisis was already over. The collapse in demand was so severe that even if the export restraints had been introduced earlier, during the worst years of the crisis, the VER’s limit on Japanese imports would not have been binding.Ěý

In other words, automakers importing from Japan would have sold close to the maximum allowable 1.68 million cars between 1979 and 1982 anyways. How do we know? Econometric evidence shows that VER only started raising the cost of Japanese imports — demonstrating the effect of having curbed their supply in the American market — in 1984 and 1985. And the effect did not become statistically significant until 1986. ^[33]

A report from the International Trade Commission is consistent with only minor effects in the first few years, increasing domestic production by an estimated 75,000 units in 1981 and 128,000 in 1982. ^[34] These figures amount to just 1 percent of total industry sales those years.Ěý

The ITC concluded that by 1984 the impact had become more substantial, with domestic auto sales 620,000 vehicles higher than they would have been without VER. By this time, of course, the crisis was long over.Ěý

This story is also reflected in perceptions at the time. At a hearing of the U.S. Congress Joint Economic Committee in 1985, Acting U.S. Trade Representative Michael Smith said “one could argue that for the first year or so the price only gradually increased. By the time of the fourth year, the additionality was very clear.” ^[35]

Despite the accumulated evidence that VER had no meaningful effect until after the crisis had passed, it was extended beyond its initial term of three years. It remained in place until 1994, more than a decade beyond the conclusion of the crisis it was meant to address.

Japanese factories, inevitable

That protectionism did not achieve its intended purpose might seem like a closed case for declaring the policy a failure. But protectionists have another reason they love the policy of Voluntary Export Restraints: they believe it caused Japanese automakers to build plants inside the United States, thereby employing American workers rather than competing against them.Ěý

As with the notion that VER led the way out of the crisis, this post hoc rationalization is unsupported by the actual evidence.Ěý

The protectionist argument relies on two assumptions. The first is that protectionism pushes firms to invest in the United States. The second is that this investment then leads to permanent long-term changes for the domestic industry even though the protection itself is temporary. The temporary protectionism in this case is the VER, and the permanent long-term change is that Japanese automakers relocated their production inside American borders instead of just importing more and more.Ěý

One problem for this theory is that Japanese investment was definitely not the Reagan administration’s goal when it negotiated this policy. Nor was it the goal of the Big 3 when lobbying for it. They simply wanted “breathing room” for American automakers to get through the crisis.Ěý

The UAW had been appealing to Japanese automakers to invest in the United States throughout the 1970s, but for them too the goal was not to attract new investment. Its aim instead was to increase unionized employment. And just like the Reagan administration and the Big 3, the union didn’t get what it wanted either. By the late 1980s, the UAW was suggesting that even the American factories of Japanese companies should be constrained by quotas. ^[36])

As such, if VER worked to increase Japanese investment and employment in the American south, it did so while failing to achieve what those who actually fought for it wanted. Hardly a rousing endorsement of the idea that protectionism “works.” ^[37]

But more importantly, the theory is simply wrong. The United States was not on a path to entirely import-based production, so it follows that the VER did not prevent this outcome by bringing the Japanese automaker to America. There are many good reasons, on the other hand, to believe that powerful magnets would have pulled the Japanese automakers to eventually produce vehicles in the United States even if VER had never existed.Ěý

The first is that producing in the country where the vehicles will be consumed not only reduces the risk of future protectionism, it also reduces exchange rate risk. Exchange rate fluctuations nearly bankrupted Jaguar in the United Kingdom and Saab in Sweden in the late 1980s, resulting in their absorption by Ford and GM respectively. ^[38]

Second, the economics of the auto industry compel production to move closer to the buyerĚý once demand reaches a certain level. Automobiles are what economists Thomas Klier and James Rubinstein call a “bulk-gaining industry.” They write:

“An assembled motor vehicle occupies a much greater volume and is more expensive to ship than the sum of its individual parts. Consequently, carmakers have selected assembly plant sites that minimize their costs of shipping finished vehicles to dealerships.” ^[39]

Estimates of economies of scale in the industry suggest that an auto assembly plant should produce around 200,000 to 300,000 units to be cost efficient; engine plants should produce around 400,000 units; and transmission plants more than 500,000 units. ^[40]

It therefore makes sense for producers to wait until a local market demonstrates a persistent and sustainable demand for their products before making large investments needed to produce at efficient quantities. When vehicles are demanded in small quantities, it makes more sense to import.ĚýĚý

Evidence for this dynamic can be seen using data from a recent Quarterly Journal of Economics paper on the auto industry. ^[41] In 2018, the most recent year that data is available, there were 74 models of car sold in the United States by non-Japanese foreign auto companies. Only 11 percent of these models were assembled domestically, and another 5 percent in Mexico and Canada combined. But among the models with more than 100,000 vehicles sold, 60 percent were assembled in America. Looking at vans, trucks, and SUVs, roughly half the models with sales above 100,000 vehicles were assembled in the United States versus just 15 percent of all models.Ěý

Looking across all 314 models in the data for 2018 (Figure 9) shows that as sales go up, vehicles become far more likely to be assembled inside the United States.Ěý

Japanese automakers were coming because they were producing hit cars, and it is simply good business to make hit cars near the customers. At best, VER might have slightly accelerated a process that was occurring anyway — though as Jordan McGillis at City Journal points out, to whatever extent that is true, it happened at the expense of consumers looking for affordable, high-mileage cars. ^[42]

Indeed, the economics of the industry was already attracting investment from foreign automakers long before VER.Ěý

• Volvo had announced in 1973 that it would build a $100 million factory in Virginia with the capacity to assemble 200,000 cars a year and would employ 3,500 workers.
  ^[43]
  The factory was built, but the plan was ultimately cancelled as a result of falling demand, and the factory switched to making buses instead. ^[44]
• Volkswagen opened a $250 million factory in Westmoreland, PA in 1978 to assemble Rabbits. ^[45]
• Honda was manufacturing motorcycles in the United States by 1979. When the motorcycle factory was confirmed in 1977, Honda executives said that if production went as expected, “it is our present intention to start manufacturing automobiles by expanding the plant site.” ^[46]
• The decision of whether to build an American plant was in front of the Nissan board for 10 years before the oil shock of 1979. The hesitance was about the concern that Detroit would fail to get its act back together and leave them with factories that were not competitive. ^[47]

Undoubtedly, the 1980s and 1990s brought a much more rapid pace of factory building in America by foreign automakers. But if avoiding VER had truly been the reason for it, the trend would have been limited to Japanese automakers. It wasn’t.Ěý

Instead, both Japanese and non-Japanese automakers from abroad were shifting production to the United States or the other NAFTA countries whenever production of a given vehicle reached a high enough level. Automakers first tested the market with imports at smaller volumes, and then accommodated higher volume models with regional production.

Japanese automakers were coming to the United States in such big numbers because, in addition to having become the most productive auto companies in the world, they were making cars that were hugely popular with American drivers. The same is true for Korean and many European automakers now — even though the VER never applied to them and it expired more than three decades ago.

Finally, I have focused so much of this analysis on VER and the events of the early 1980s because they loom so large in the story told by protectionists, but I have not yet mentioned perhaps the most trenchant example of auto protectionism that continues to this day. Originally intended as a temporary policy, a retaliation against Europe for their tariffs on American chicken imports, the United States has upheld a 25 percent tariff on imported light trucks since 1964. ^[48] The result has been less competition, fewer choices, and higher prices for American buyers of these vehicles. ^[49]

VER lasted a decade longer than it was meant to, but at least it finally did end. The light truck tariff reveals the danger of implementing a policy that initially is not meant to be permanent but turns out that way nonetheless. A new protectionist policy can be sticky, leading to political apathy and status quo bias as policymakers accustom themselves to it, fearing the backlash from the protected industry if they try to undo it. In the case of the truck tariff, a policy response in an unrelated trade war ended up remaining in place for more than six decades — and counting.Ěý

—â¶Ä”â¶Ä”â¶Ä”â¶Ä”â¶Ä”â¶Ä”â¶Ä”â¶Ä”â¶Ä”â¶Ä”â¶Ä”â¶Ä”

To summarize, the policy of Voluntary Export Restraints failed in its intended goal of giving American automakers breathing room during the crisis of 1979–82. It then did give them a windfall — by forcing American drivers to pay higher prices — after the crisis had passed. Data and the economics of the auto industry support the likelihood that the Japanese companies were destined to eventually open American factories in greater numbers regardless of VER, just as other major auto producers did after VER had expired.Ěý

As industrial policies go, this one was hardly a riveting success to be emulated.Ěý

Conclusion

This history matters. By enacting 25 percent tariffs on auto imports, the Trump administration has once again decided that globalization is the problem and protectionism the solution.Ěý

Supporters will point to the decimation of the American auto industry as demonstrating the need for trade barriers. This claim is false.ĚýĚý

Protectionists will doubtless continue pointing to the entrenched economic struggles of Detroit and Michigan as proof that globalization was indeed the villain. This too is false.Ěý

And they will point to VER as an example of how temporary, targeted protectionism works in the auto industry to generate substantial new industrial growth. False yet again.Ěý

The arguments in favor of tariffs are not just based on mistaken interpretations of the history of the auto industry. They are based on a deep and persistent misreading of the history itself.Ěý

A clear and accurate understanding of the past reveals the flaws in the policies of the present.