Conversable Economist

At the start of World War II, the US economy relied almost exclusively on imported rubber as the key material for making, among other things, tires for cars and airplanes. The dependency was well-known, but in April 1942, when Japan cut off the foreign supply, the US was unprepared. Synthetic rubber ended up being part of the answer, but the rest of the answer involved a lot of muddling through. Alexander J. Field tells the story in “The World War II US Rubber Famine” (Business History Review, Autumn 2025, 99: 365-390). Field writes:

There were simply no satisfactory substitutes for rubber in a variety of critical uses, particularly tire carcasses and treads, the ultimate end use of 70% of rubber inputs. The severe shortage of natural rubber that resulted adversely affected the ability of the US military to project force and contributed to the disappointing record of wartime manufacturing productivity. … During World War II, the United States never escaped the threat of running out of rubber, which stood as a sword of Damocles over the entire economic and military effort. In 1944 the country almost ran out of natural rubber and would have, had the war continued into 1946.

One fun fact: Apparently the original World War II plan had been to launch a counterattack across the English Channel in 1943, rather than the D-Day invasion of June 1944. But in 1943, the lack of rubber meant that there weren’t enough landing vehicles for a counterattack to work.

A second fun fact: During World War II, “The rubber famine led directly to the imposition of a 35-mph speed limit and nationwide gas rationing in a country that, in the aggregate, was awash in petroleum. The intent was not to save fuel but to reduce tread wear on the tires installed on the nation’s 27 million automobiles and 5 million trucks, almost all of which were otherwise forecast to be off the roads within two years. These restrictions made it more difficult for people to get to and from work, contributing to absenteeism, and impacted the distribution of products by truck.”

The rubber famine is of interest for its own sake. But in addition, it echoes modern concerns about the risks of excessive dependence on key imported products–and thus may offer some food for thought on those issues as well. There were three possibilities for dealing with the rubber famine: grow rubber domestically, build up a stockpile of imported rubber, or produce synthetic rubber. Here’s what happened with each one:

Efforts to grow rubber plants from the far East in the western hemisphere had not worked well in the 1920s, but there was a plant called guayule that was a promising alternative natural source of latex–substitutable for natural rubber in many applications, and better than rubber for some of them. In 1930, a then-obscure Army major named Dwight D. Eisenhower toured the guayule plantations of Mexico, and wrote in his report: “Should our sea communications with [Southeast Asia] be cut in an emergency, shortage of rubber in the United States would rapidly become acute.” Eisenhower proposed in his 1930 report a US program to subsidize guayule. but at the time, nothing happened. On March 5, 1942, the US passed the Emergency Rubber Act, sometimes known as the Guayule Act, to subsidize farming of guayule in the the United States. But since guayule plants take four years of growth before they mature and produce latex, the law made no contribution to the wartime effort.

The US made efforts to build up its rubber stockpile starting in 1939, as war loomed, but by 1942 the total stockpile was equal to about one year of use. One problem was that the powers-that-be in Washington, DC, insisted that imported rubber travel by ship from the far East to New York, by way of the Panama Canal, rather than offloading in San Francisco and having the rubber cross the US by train. Having ships go all the way to New York was slightly cheaper than trains, but as a result, the ships carrying rubber were unable to go back-and-forth across the Pacific as often–and growth of the rubber stockpile was limited as a result..

The powers-that-be in Washington viewed research efforts in synthetic rubber as an overly costly sop to petrochemical companies up through February 1941, and the created a program that ended up costing as much as the Manhattan Project to develop the atomic bomb. As it turned out, total production of synthetic rubber during World War II was roughly two years of wartime quantity demanded–or bout twice as much as the stock of natural rubber entering the war.

As Field writes, the synthetic rubber progrm has been “uncritically lionized” in most writing about World War II, as if it just burst into existence and solved the problems. But aas Field writes: “The eventual availability of synthetic rubber in quantity did not end the rubber famine. Synthetic had to be blended with natural in the manufacture of almost all products, and in some cases synthetic could not be used at all.”

Moreover, the synthetic rubber program depended on a gas called butadiene, could be produced from petroleum or from alcohol. Standard Oil and other big petroleum companies favored a petroleum feedstock, although alcohol-based plants were also started. Field writes:

The first butadiene plant was not completed until April 1943. All three of the big alcohol plants opened that year, and they produced far more than their rated annual capacity. Of the five petroleum-based butadiene plants, the three largest did not begin production until 1944, and consistently produced below their rated capacity. … Without alcohol-based butadiene, it is hard to see how D-Day could have gone forward in June 1944. In retrospect, it is not clear that petroleum-based butadiene was needed at all to win the war. Butadiene from isobutylene was cheaper in the long run because, even though plants using this input were more expensive to construct, required more complex engineering, and relied on untested processes, the feedstock (petroleum) was ultimately cheaper. Due to huge agricultural surpluses accumulated during the Depression, however, the opportunity cost of ethanol was far lower during the war years, an advantage augmented by the much lower capital requirements of the process using it to produce butadiene.

The US government ended up establishing the foundations for a commercially successful synthetic rubber industry in the postwar period, one using petroleum as the principal feedstock, as Standard [Oil] intended. Given that synthetic rubber would be needed during the war, it would have been cheaper and faster to have focused from the outset on ethanol as the feedstock. Standard Oil bears responsibility for the emphasis on petroleum. Whether petroleum or ethanol was to be the feedstock, however, construction on the butadiene plants should have started earlier …

If Eisenhower’s recommendation had been followed back in 1930, or if the US rubber stockpile had been built up more expeditiously, then synthetic rubber would not have been needed to win World War II. If production of synthetic rubber had been alcohol-based from the start, it could have ramped up more quickly. If more rubber had been available, then the war might have ended sooner if the D-Day invasion could have happened in 1943.

The broader lesson is that a serious discussion of critical imported materials should never be a last-minute, against-the-deadline affair. Moreover, a serious discussion will need to resist pressure from domestic companies with a financial interest in the choices made. The US and Allied forces muddled through the rubber famine during World War II, but it was a nearer thing than most people realize. Field notes: “In 1945 the War Production Board forecast that, in the event of an invasion of Japan, the US would simply run out of natural rubber in 1946. Among other consequences, the US would then have been unable to manufacture airplane tires.” That fact surely played a role in the decision to use the atomic bomb.

There are certainly examples where new technology has replaced jobs. The US had 350,000 telephone switchboard operators in 1950, and the job just went away. The tractor played a big role in reducing the number of US farmers in the first half of the 20th century. Almost every village of any size had a blacksmith though much of the 19th century, but with the rise of industrial production, there weren’t enough of them left to be counted as a separate employment category in the 1900 census. And now here we are with the new artificial intelligence tools, and warnings that all manner of jobs that use computers–across a wide array of industries–could be at risk.

It seems obvious to me that many jobs will change as new technologies appear. Many of the tasks involved in my own job, running an academic economics journal since the late 1980s, changed substantially with the arrival of the internet, for example. But the job itself didn’t go away; indeed, the internet probably made me better at my job. For example, it’s a lot easier for me to look up cited articles from my desk than it was to walk over into the library stacks to find the article–and so I check many more articles as a result.

Thus, a key issue here is the extent to which the new AI tools replace workers outright, like telephone switchboard operators, or whether they allow workers to be more productive and effect–or even create the possibility for new and previously unimagined jobs. Daron Acemoglu, David Autor, and Simon Johnson work through these distinctions in “Building pro-worker artificial intelligence” (Hamilton Project at the Brookings Institution, February 2026). They write: “We define pro-worker technologies—including AI—as technologies that make human skills and expertise more valuable by expanding worker capabilities.”

They emphasize a key point about AI tools: such tools may actually be more useful when collaborating with humans. They write:

A modern AI system can ingest drone imagery and soil sensor data from a farm’s every acre, the complete sensor logs from a building’s HVAC system, or the detailed vital signs of a single patient observed over many months to support workers making high-stakes decisions. Drawing on this pretraining, AI tools can think alongside workers, identify relevant context, generate well-informed responses to questions, and present lucid, well-structured data to support decisionmaking. This is collaboration.

You may object: “If AI can behave like an expert, can’t it simply replace experts, thus automating their expertise into irrelevance?” In some cases, the answer is yes. But in many more cases, we think the answer is no: AI will prove more effective at collaboration than
at automation. Precisely because AI is not rule-bound, it is less trustworthy as an autonomous actor than a conventional computer system, and more valuable as a collaborator (Narayanan and Kapoor 2025).

To be useful, an automation tool must deliver near-flawless performance almost all the time. You would not tolerate a spreadsheet that hallucinated values, a robotic surgeon that glitched-out during bypass surgery, an agentic investing tool that squandered your money while you were not paying attention, or an AI-powered vending machine that gave away PlayStations and stocked live fish at the behest of persuasive customers (Stern 2025). For most of these tasks, the stakes are too consequential and the decisions too nuanced to be fully delegated to an automatic system that acts on its own discretion. The AI needs human expertise.

A collaboration tool does not need to be anywhere close to infallible to be useful. A doctor with a stethoscope can better diagnose a patient than the same doctor without one, and a contractor can pitch a squarer house frame with a laser level than they could by eyeballing it. These tools do not need to work flawlessly, because they do not promise to replace the expertise of their user. They make experts better at what they do—and extend their expertise to places it could not go unassisted. Rather than making expertise unnecessary, they render expertise more valuable by extending its efficacy and scope. It is this complementarity between machine capacity and human expertise that we believe imbues AI with vast pro-worker potential.

The authors provide concrete examples of pro-worker uses of AI for teachers, electricians’ assistants, patent examiners and others. They point out that the use of AI-assisted hearing aids might enable some workers to expand their on-the-job capabilities. However, they also worry that economic incentives and business habits may tend to emphasize AI applications that substitute for current jobs, rather than complement them. Thus, they argue for the public sector to nudge the incentives for pro-worker AI tools where possible. One of their examples stuck with me:

Indeed, the public sector already heavily shapes the path of technology in health care and education. For example, the federal Health Information Technology for Economic and Clinical Health (HITECH) Act of 2009 dramatically accelerated electronic health record adoption in U.S. hospitals through financial incentives and penalties. Within less than a decade, the
United States went from approximately 10 percent of hospitals with electronic health records to near-universal adoption (Office of the National Coordinator for Health Information Technology 2017). In a similar vein, the federal schools and libraries universal support (E-Rate) program, established by the Telecommunications Act of 1996, provides ongoing subsidies to schools and libraries for Internet connectivity. As of 2021, 95 percent of U.S. public school classrooms had WiFi (Munson 2023).

They also point out that the US corporate tax code treats investment in machinery more favorably than investment in, say, worker training and skills. That policy difference could be at least equalized. In this and other ways, the future uses of AI are not purely determined by technological advance, but instead by the incentives and beliefs of economic players–firms that develop AI technologies, firms that use them, managers thinking about how work should evolve, the the willingness of workers to build new skills.

When thinking about what makes an economy flourish, many of us tend to focus on the success stories of innovation and growth. After all, success stories involve an element of risk, which means a chance of failure. When it’s more expensive to fail, then avoiding the risk of failure–by avoiding innovative but risky business choices–starts to make sense. Yann Coatanlem and Oliver Coste put some meat on the bones of this idea in “Cost of Failure, Disruptive Innovation and Targeted Flexicurity: More evidence supporting targeted reforms” (Institute for European Policymaking at Bocconi University Working Paper, November 2025).

The authors focus on the situation of a large company that took a business risk that did not work out, and wants to restructure significantly–which in turn involves the costs of laying off a large number of employees. They have data on 250 such restucturings across a number of European countries and the United States. For each major restructuring, they go through financial and government reports to estimate the costs of laying a worker off, expressed in terms of “months of average employee compensation.” Thus, the illustrative figures show that in Germany, each layoff cost 31 months of average employee compensation; in France, 38 months; in Italy, 52 months; and in Spain, 63 months.

In comparison, when a similar exercise is carried out for US firms, the average cost was 7 months of employee compensation.

Here’s a concrete US example:

In the United States, the responsiveness of tech companies to a technological shock such as the unexpected success of ChatGPT is quite staggering. OpenAI released ChatGPT to the public on 30 November 2022. This breakthrough AI solution reached 1 million users within five days and 100 million users within two months — taking even the most seasoned stakeholders in Silicon Valley by surprise. The industry’s reaction unfolded within weeks: On 18 January 2023, Microsoft announced 10,000 layoffs, representing 5% of its workforce. The restructuring plan was completed by end of March 2023. On 20 January 2023, Google followed with 12,000 layoffs, or 6% of its workforce. The U.S. based employees were notified individually (by email) on the same day. On 14 March 2023, Meta announced 10,000 additional layoffs — a second round following the 11,000 job cuts announced on 9 November 2022 — bringing the total reduction to approximately 25% of its workforce. Most of the restructuring plan was completed by May 2023. This kind of technological shock is by no means unique. It is a recurring pattern in the tech industry—seen with the advent of cloud computing, smartphones, social networks, e-commerce, mobile phones and the internet.

Agility in laying off also means agility in hiring. U.S. tech companies did not cut tens of thousands of engineering jobs to scale back investment—quite the opposite. These workforce reductions were aimed at reallocating resources to accelerate innovation where most promising. They immediately began hiring thousands of AI engineers and invested heavily in AI computing capacity. Meta provides a striking example: after reducing its workforce by 25% in just six months, as seen before, the company hired approximately 10,000 engineers and ramped up its investment in AI supercomputers from around $1 billion in 2022 to $20 billion in 2023, $37 billion in 2024, and a projected $65 billion in 2025. Microsoft and Google are committing even larger sums to AI infrastructure. … In practice, European Employment Protection Law (EPL) often makes such strategic shifts nearly impossible.

A broader discussion of this subject by Pieter Garicano was titled “Why Europe doesn’t have a Tesla” (Works in Progress, February 7, 2026). If you are fortunate enough to be hired as a worker for Volkswagen in Germany, you are essentially guaranteed a job for life–because the costs of laying off such workers are so very high. But the tradeoff is that no company is like to start a new auto manufacturing facility in Germany–not because the potential rewards are too low, but because the potential costs of failure and restructuring are so high. More broadly, if the costs of restructuring a large technology company will be very high, then spending the R&D money seek out new innovations and products will be less attractive as well.

So are we forced into a harsh choice between protecting workers from layoffs and economic dynamism? Perhaps the tradeoff need not be as severe as one might fear. In the Coatanlem and Coste study, they note that the costs of restructuring are at US levels, or even lower, in Denmark, Sweden, and Switzerland. The reason seems tied to a set of policies they call “flexicurity.” (Michael Svarer and Claus Thustrup Kreiner provide an overview of “Danish Flexicurity: Rights and Duties” in the Fall 2022 issue of the Journal of Economic Perspectives, where I work as Managing Editor).

The basic “flexicurity” idea is that the government provides a package of higher unemployment benefits combined with support for job search and job training–and if a displaced worker find an alternative job relatively soon, the higher unemployment benefits start to decline. To put it another way, a company that wishes to restructure bears a relatively modest share of the total costs of the adjustment, and the government steps in with policies and incentives to hasten the adjustment. These steps are also called “active labor market policies,” which differentiates them from “passive labor market policies” like just paying unemployment benefits for a time, and the US government has lagged far behind other high-income countries in such efforts.

24 Min

Europe’s cutting edge firms are falling far behind the American frontier because of restrictive labor laws.

being hired in a Volkswagen factory in Germany is a lifetime job

what if Volkswagen wants to shift to producing EV’s, and existing workers don’t have the skills?

When recessions hit, the US Federal Reserve lowers its target interest rate–the “federal funds interest rate.” This interest rate applies to extremely safe borrowing: essentially, to overnight borrowing by large and safe financial institutions. The idea is that by altering this ultra-safe interest rate, other riskier interest rates will also be under pressure to adjust, so the lower federal funds rate will be passed through by a country’s financial sector to lower rates for mortgage lending, car lending,and the like (as well as to lower interest rates paid to those with deposits in banks or money market funds).

But in the Great Recession and its aftermath, the Fed ran into a problem: it took the federal funds interest rate to just barely above zero percent. Here, it ran into what economists sometimes call the “zero lower bound.” The Fed felt that it couldn’t make interest rates negative (instead of the bank paying interest to depositors, the depositors would pay interest to the bank?). Thus, the Fed felt as if it needed to try other policies for stimulating the economy with names like “quantitative easing,” “forward guidance,” and the like.

But the European Central Banks, along with central banks in Switzerland, Sweden, Denmark, and Japan, decided to break through the “zero lower bound” (often abbreviated as the ZLB) and try out a negative policy interest rate. I wrote about this when it was happening: for example, “Negative Interest Rates: Practical, but Limited” (March 15, 2021), ”Negative Interest Rates: Evidence and Practicalities” (August 8, 2017); ”What Else Can Central Banks Do?\” (October 24, 2016); and ”Fed Policy: Negative Interest Rates, Neo-Fisherian, or No Change” (August 26, 2016).

But now we have the advantage of a few years to look back at the experience. Michael McLeay, Silvana Tenreyro, and Lukas von dem Berge look back at the evidence in “Negative Rates and the Effective Lower Bound: Theory and Evidence” (Journal of the European Economic Association, 24: 1, February 2026, pp. 1–57). The paper is based on the FBBVA Lecture delivered at the 2025 meetings of the European Economic Association. The second half of the paper is technical stuff (that is, it’s a new model of a heterogeneous, oligopolistic banking sector, with high- and low-deposit banks, embedded in an “open-economy macroeconomic model, featuring exchange-rate and capital market transmission channels). But the basic fact patterns at the start of the paper, which the model seeks to capture, are an easy read.

The short answer is that when central banks moved the key policy interest rate (slightly) below zero, it did seem to provide additional economic stimulus without weird or disastraous consequences. Here are some top-line takeaways.

• “Empirical Observation 1. Pass-through of policy rates to household deposit rates is bounded by the ZLB. Meanwhile, corporate deposit and wholesale funding rates can fall below zero.”
• “Empirical Observation 2. At low or negative policy rates, aggregate pass-through to bank lending rates and volumes still occurs, though it is typically reduced and potentially delayed.” However, the authors cite studies that the pass-through rate of negative interest rates to lower interest rates on loans in the economy is not 1:1. The pass-through of negative interest rates in the euro area seems to have been more than 50%, but in some other countries, it was less than that.
• “Empirical Observation 3. Aggregate banking sector profitability is not necessarily adversely affected by negative policy rates, and may even improve owing to general equilibrium effects.” A concern with negative interest rates was that if banks were pressured to lend at extremely low interest rates, then the banks might go broke. But it turns out that the macroeconomic benefits banks received from lower interest rates stimulating the economy tended to offset the lower bank profits from lending at lower rates.

Of course, none of this means that it would be necessarily sensible policy for a central bank to make its policy interest rate deeply negative. But it turns out that businesses in particular can function with a negative interest rate at their bank–that is, they pay interest to the bank on their deposita at the bank. Also, as we have seen in the US, people can function with banks paying essentially zero interest on their deposits as well. I suspect the reason is that it’s worth something to business and to people to make and receive payments through the banking system (say, a business making payroll, or a person having paychecks automatically deposited). As a result, when a bank is paying zero or even slightly negative interest on deposits, the bank will not lose all of its deposits.

There are three main ways to tax carbon emissions: through a fuel tax (like a gasoline tax), through an explicit tax on carbon emissions, and through an emissions trading system that requires carbon emitters to have marketable permit for doing so. The OECD offers an overview of that is happening across these areas in 79 countries in “Effective Carbon Rates 2025: Recent Trends in Taxes on Energy Use and Carbon Pricing” (November 13, 2025).

Here’s a summary figure: the light-blue bars are the tax imposed on carbon emissions by a fuel tax; the darker blue bars are an explicit carbon tax; and the green bars are the costs imposed by an emissions trading system. The calculations take into account both the level of these taxes, and also the share of carbon emissions actually covered.

One obvious lesson is that fuel taxes remain the primary way in which carbon is taxed. Another lesson is that the countries with the highest carbon taxes tend to be in Europe. If you run your eye down the list looking for the United States, you will find it about three-quarters of the way down, between South Africa and Uruguay. But you will also note that China, by far the world’s largest emitter of carbon, falls even lower on the list. And India, with rapidly rising carbon emissions, is lower than China.

The recent Supreme Court decision that President Trump did not have arbitrary and unbounded powers he has claimed to impose or retrace tariffs came as no surprise, for a variety reasons laid out in earlier posts here, here, and here. But here I want to focus on a different claim: that the reason for tariffs was to reduce the US trade deficit.

Broadly speaking, there are two possible explanations for trade deficits. One possible explanation is that other countries are competing unfairly in some way, that this unfairness penalizes US exports and subsidizes US imports, and thus leads to a trade deficit. The other possible explanation argues that trade deficits are determined by macroeconomic factors like total levels of production and consumption, along with domestic savings and investment. From this view, government policies to subsidize or penalize certain industries will certainly affect those industries, and lead to shifts between industries. However, this second view argues, such policies won’t determine the trade deficit.

Well, President Trump started imposing tariffs last April. The effective tariff rate (that is, the average over all goods) rose substantially. The federal revenue collected by tariffs rose from near-zero to about 1% of GDP in the most recent federal budget projections–call it roughly $300 billion per year.

We have here a test of the two hypotheses about what causes trade deficits. Did the US government actions to disfavor imports in fact reduce the trade deficit? The US Department of Commerce just released the trade data for the full calendar year 2025, and here’s what the overall picture of imports and exports looks like:

It is simply a fact that the US trade deficit in 2025 barely budged, even with Trump’s tariffs. As the Census Bureau reports: “For 2025, the goods and services deficit was $901.5 billion, down $2.1 billion from $903.5 billion in 2024”.

You can break this down separately into goods and services. In goods, where Trump’s tariffs were focused, the US trade deficit got larger. In services, the US trade surplus got bigger. Census writes: “The 2025 decrease in the goods and services deficit reflected an increase in the goods deficit of $25.5 billion, or 2.1 percent, to $1,240.9 billion and an increase in the
services surplus of $27.6 billion, or 8.9 percent, to $339.5 billion.”

Yes, tariffs have effects. For example, they can reallocate where imports come from between countries. For example, imports of goods from China were $438 billion in 2024, and dropped to $308 billion in 2025. But reallocating imports and exports between countries is not the same as changing the overall balance of trade.

Indeed, the erratic and untargeted nature of Trump’s tariffs means that they have not even focused on the kinds of trade that matters most for the US economy. As one of the Census tables, the US economy had a trade deficit in advanced technology products of $297 billion in 2024 (that is, $763 billion in imports and $466 billion in exports). In 2025, with a blast of US tariffs, the US trade deficit in advanced technology products was even larger at $414 billion (that is, $965 billion in imports and $550 billion in exports).

It is a basic lesson in intro economics that trade deficits are determined by macroeconomic forces, not by levels of governmental fairness or unfairness about trade. I have tried to explain the macroeconomic argument on this blog a number of times, and I won’t bother trying again here. My point here is just to focus on the fact that if you believed that Trump’s tariffs would put a big dent in US trade deficits, your belief is being disproved by facts. Tariffs can readjust what goods and services are shipped, and to and from what places. But they are not in fact a way of addressing the overall US trade deficits.

On question about how artificial intelligence will affect jobs and productivity, the honest answer is that the actual empirical evidence is pretty thin. In part, this is because the new AI tools have arrived so recently, and are evolving so quickly, that don’t yet have much experience with their effects–or much data to separate the effects of AI from other changes in the economy. Eric Fruits and Kristian Stout offer an overview of what the emerging evidence has to say in “AI, Productivity, and Labor Markets: A Review of the Empirical Evidence” (International Center for Law & Economics Issue Brief, February 5, 2026). For those just getting up to speed on the subject, one especially useful part of the article is that the reference list is annotated–that is, there’s a brief discussion of what angle is taken by each of the references, with weblinks as possible.

The authors emphasize that most of the predictions about effects of AI are essentially driven by underlying assumptions: what share of tasks in in what share of jobs might be affected by AI; to what extent will AI complement workers and shift the tasks that they do in a given day, or replace them; how quickly will AI diffuse across varying sectors of the economy; how quickly will complementary investments be made to support the effects of AI (in everything from electric power generation to education to robotics); if AI requires large-scale organizational redesign, how long does that take; will AI expand overall output or make the distribution of income more unequal; and so on and so on. Choose your assumptions, and you can choose your outcome. As the authors point out, for example, its also possible to have effects that start in one direction and then swerve to another direction: for example, it’s straightforward to build a “J-curve” theory in which the adjustment to a new technology first has a negative effect during a transition period as firms face the cost of making adjustments, but then later has a very positive effect.

However, some themes are begining to emerge from actual research on the use of AI in different settings, like call centers, accountants, law students doing complex legal tasks, software developers, professional translators, randomized experiments with groups given certain tasks to perform, and others. Here are some themes that emerge:

First, such studies consistently show substantial gains in specific areas of both quantity and quality of output.

Second, the gains can be highly variable, even on tasks that may appear similar. For example, a given AI tool might be enormously helpful and reliable for certain tasks or up to a certain level of complexity–but then become highly unreliable beyond that point. Those using the tool may see that it works well in lots of cases, and then start trying to use it everywhere–even where it doesn’t work well. Thus, several of the studies emphasize the need for “verification protocols” to double-check quality of output and to provide feedback to users.

Third, another theme across these studies is “skill compression,” as the authors note: “Across settings, micro-level evidence points to skill compression. AI tools disproportionately boost output among lower-performing workers, narrowing performance gaps within job categories. This pattern recurs across productivity studies and carries distributional implications: AI can reduce inequality within occupations, even as it reshapes demand for certain entry-level roles.” In other words, access to AI tools may have the biggest benefit for the productivity of workers with less experience or skill, and smaller benefits for workers who already have greater skills. (This pattern may also help to explain some evidence that the number of jobs in sectors exposed to AI for young workers may have fallen: when AI boosts productivity of these younger workers, firms end up hiring fewer of them.)

Fourth, the studies of AI and jobs do not involve a random selection of jobs. Instead, the studies tend to focus on jobs where it seems intuitively plausible that the gains might be substantial. Thus, it would be clearly incorrect to extrapolate the patterns from individual studies across the economy as a whole.

Fifth, the new AI tools seem to be leading to a surge of new companies. Moreover, these companies are often smaller in terms of employment than average start-ups in the past. authors write: “AI substitutes for managerial, operational, and technical tasks that previously required additional hires or cofounders. By lowering the minimum viable team size, AI reduces entry costs, increases the number of entrants, and strengthens downstream competition.” The emerging pattern here may be large and concentrated “upstream” companies doing the foundational research on AI and owning the actual computing power, w while the new startups may be smaller competitive “downstream” users of this technology. Of course, the share of these start-ups that eventually succeed and grow it not yet clear.

Sixth, the growth of the new AI companies depend on other factors in the economy: availability of finance, a non-hostile and non-fragmented regulatory environment, a supply of science and engineering workers, capacity for training new workers, capacity for computing power and supporting infrastructure (like electricity and telecommunications).

I think we don’t yet have a clear sense of what these AI tools will actually be in the medium term of, say, 5-10 years. (See above discussion of how assumptions will determine predictions.) But I strongly suspect that no high-income country, if it wants to remain a high-income country, is going to be able to a to sidestep and avoid the new AI technologies. It’s worth remembering that in the long-run, the average standard of living in a society rises when the average productivity per worker rises. In turn, greater productivity pretty much always requires change, sometimes disruptive change.

I don’t actually expect that writing about US budget deficits and the accumulating federal debt, as I do from time to time, will have any near-term effect. But given the current trajectory, a time will come when decisions need to be made, and having a grip on the underlying dynamics is always a useful starting point. In that spirit of humble and diminished expectations, here are some figures from the just-released Congressional Budget Office report, “The Budget and Economic Outlook: 2026 to 2036” (February 2026).

This figure shows federal deficits as a percent of GDP, going back a half-century to 1976 and projected forward through 2036, based on current law. You can see that the current level fo deficits, and those projected for the next decade, are higher than any years except for the spikes during the Great Recession of 2007-09 and the Covid recession. If you are able to remember the concerns about the size of budget deficits during President Reagan’s term of office in the 1980s, we are comfortably above that level, and projected to stay above it.

The figure also divides the deficit into “primary” and “net interest”: that is, the “primary” deficit is what the deficit would look like if the federal government didn’t need to pay interest on past borrowing. The problem for the future, as you can see, is that the primary deficit is falling, but interest outlays are rising. This pattern creates a bad dynamic, where rising future deficits are being driven by rising interest payments.

I occasionally see some social media boasting about how revenue from the tariffs are going to fix federal deficit problems. Here’s a breakdown of federal revenues. Customs duties are tshow by a dark line near the bottom. You can see that they have definitely risen. This increase is taken into account, and projected into the future, by the deficit estimates given above.

On the spending side, it’s conventional to divide it into what is classified as mandatory, discretionary, and interest payments. Mandatory includes Social Security, federal health care programs, income support programs, retirement for federal employees, and some other items. Discretionary spending is at present about half made up of defense spending, and the other half made up of many smaller programs, including federal spending on education, transportation, veteran’s benefits, and other items. As a share of GDP, discreitional spending is falling over time, while mandatory spending is rising. In 2026, it’s already true that federal interest payments on past borrowign exceed national defense spending. Byt 2036, interest is projected to be more-or-less equal to all discretionary spending.

The costs of excessive federal debt are gradually becoming apparent in various ways. The inflation spike a few years ago was, in part, driven by the spike in federal deficits. The arguments between the Trump administration and the Federal Reserve over interest rate policy are driven, in part, by the recognition that high federal interest payments are crowding out other possibilities for greater spending or tax cuts. Concerns over large US trade deficits are created, in part, because the very large US budget deficits are creating an ongoing surge of US demand for imports. Concerns over slow rates of US economic growth are driven, in part, by concerns that high federal borrowing is crowding out some opportunities for private-sector investment. The current trajectory of federal deficits and debt seems sustainable for awhile longer, in the sense that it doesn’t seem likely to cause a near-term crisis, but that doesn’t make it a desirable path for the US economy to follow.

In the modern world, many of the topics about which people have opinions are some distance away from any actual experience or expertise that these same people have. In this situation, people are unlikely to form their opinions with a deep dive into relevant history, data, and academic research. Instead, they are more likely to turn to a mixture of personal experience and mental models. In a 2018 paper in Behavioral and Brain Sciences, Pascal Boyer and Michael Bang Petersen referred to this as “folk economics” and discuss “Folk-economic beliefs: An evolutionary cognitive model.”

More explicitly, the authors write in the abstract for their paper:

The domain of “folk-economics” consists in explicit beliefs about the economy held by laypeople, untrained in economics, about such topics as, for example, the causes of the wealth of nations, the benefits or drawbacks of markets and international trade, the effects of regulation, the origins of inequality, the connection between work and wages, the economic consequences of immigration, or the possible causes of unemployment. These beliefs are crucial in forming people’s political beliefs and in shaping their reception of different policies. Yet, they often conflict with elementary principles of economic theory and are often described as the consequences of ignorance, irrationality, or specific biases. … Here we propose that the cultural success of particular beliefs about the economy is predictable if we consider the influence of specialized, largely automatic inference systems that evolved as adaptations to ancestral human smallscale sociality. These systems, for which there is independent evidence, include free-rider detection, fairness-based partner choice, ownership intuitions, coalitional psychology, and more. Information about modern mass-market conditions activates these specific inference systems, resulting in particular intuitions, for example, that impersonal transactions are dangerous or that international trade is a zero-sum game. These intuitions in turn make specific policy proposals more likely than others to become intuitively
compelling, and as a consequence exert a crucial influence on political choices.

To give a sense of this line of thinking, here is a (nonexhaustive) list of eight folk-economic beliefs from the authors:

• International trade is zero-sum, has negative effects.
• Immigrants “steal” jobs.
• Immigrants abuse the welfare system.
• Necessary social welfare programs are abused by scroungers.
• Markets have a negative social impact.
• The profit motive is detrimental to general welfare.
• Labor is the source of value.
• Price-regulation has the intended effects.

This list is intended to be descriptive, rather than critical. Many of these beliefs have some mixture of truth and falsehood, depending on the setting. Folk economic beliefs need not be consistent with each other: for eample, there is some conflict between the beliefs that immigrants are simultaneously taking a large number of jobs while also not working and receiving welfare benefits. The beliefs need not govern the personal behavior of individuals, either: for example, someone who believes that labor is the source of value might also rent out an extra room to a college student. The central point is that these themes are part of how a given culture has evolved, and are widely believed in a way that influences political and policy beliefs. The authors write:

As an illustration, we can consider that optimization problems such as hunting, foraging, choosing the best mate, selecting nutritious foods, and garnering social support were present, and relevant to fitness, throughout human evolution. By contrast, urban life, mass communication, rapid long-distance travel, and mass-market economies only occurred for a small duration and only in some places at first. So it is more plausible that human minds were selected for systems geared to the first kind of adaptive problems, than to the second. … The human mind, in other words, contains a rudimentary exchange psychology, evolved by natural selection to help facilitate transactions. Although it evolved within ancestral small scale hunter–gatherer groups, the cues inherent in modern markets economies (transactions, bargaining, prices, etc.) also bring it online. However, market economies are a novelty at the scale of biological evolution, so we should not expect specific adaptations to their features, as the differences between ancestral exchange and the market are vast (Rubin 2003).

As an example of how this analysis works in practice, consider the first belief that that international trade is zero-sum or negative in its effects. The authors argue that this folk-economic belief is built on beliefs about coalitional dynamics that have been relevant for much of the history of homo sapiens. They write:

[T]he statement that international trade has negative consequences … contains several pieces of information likely to activate specific inference systems. Let us consider a news headline like “China sells more to the U.S. than to Russia.” Selling involves receiving resources and, importantly, resources in this case transfer from one nation to another. In psychological terms, nations are “imagined communities” (Anderson 1983) or, with the vocabulary presented above, nations are coalitions to the mind and, hence, mention of nations activates the coalitional psychological machinery (Gat 2006; Hechter 1987a). Nations are exclusive groups, citizens of a nation are assumed to have common interests, and nations are equipped with armies to fight each other. The activation of this machinery has the downstream consequence, we argue, that Americans will evaluate the transfer of resources to China – and, hence, the headline – negatively. As argued above, one key assumption of the coalitional system, once activated about two categories or groups, is that there is a zero-sum interaction between the mutually exclusive groups. As a consequence, there is a strong prior belief that any advantage to another group is detrimental to one’s own (Hiscox 2006). Any information to the effect that other groups are prosperous, or getting better, is equivalent to a threat-cue, indicating that our group stands to lose out. … To maintain stable and efficient coalitions, humans in many different contexts must have assumed that other groups’ advantage was a potential loss.

Viewing the “international trade is bad” belief as supported by coalitional psychology does not just explain the belief but also suggests novel testable predictions. In particular, we should expect the view that trade is bad to be particularly attractive when the trading crosses perceived coalitional boundaries. It is predicted to invariably occur in the context of, precisely, debates about trade between countries. American consumers may find it intuitive that the United States might suffer from Chinese prosperity, but, on this theory, they would find it less compelling that development in Vermont damages the economy of Texas. Similarly, the survival value of the belief might depend on the relationship between the countries. Trading between long-term allies (e.g., trading between Great Britain and the United States) should be viewed as less problematic than trading between rivals (e.g., trading between China and the United States), even if all else were equal.

The notion that human brains are still adjusting to a world of impersonal exchange makes sense too me. From this perspective, one of the challenges for economics in the world of public policy is to explain economic insights in a way that sidesteps and does not trigger folk economic beliefs.

In editing two papers on Asian American immigration for the Winter 2026 issue of the Journal of Economic Perspectives (where I work as Managing Editor), I found myself musing over two alternative histories: one about more such immigration, one about less.

Hannah Postel describes “Asian Immigration to the United States in Historical Perspective” (Journal of Economic Perspectives 40:1, 191–214) from the later decades of rhe 19th century up to the present. As she points out, Asians were the first group to be singled out for discriminatory treatment by US immigration laws, and the first to be excluded from immigrating. Postel writes:

The Chinese became a target of American racial animus and the nation’s first federal immigration laws during the late nineteenth century. By the 1870s, about 15,000 to 20,000 permanent Chinese immigrants were arriving each year … For perspective, the total US population in 1870 was about 38 million, or one-ninth its current level. The 1875 Page Act excluded Chinese women, thereby precluding Chinese family formation for almost a century. The 1882 Restriction Act limited future Chinese immigration to the United States and was extended in 1888 to “cut off circular Chinese migration, [pour] ten times the funding into enforcement, [eliminate] many avenues for fraud, and [develop] new systems of management and deportation” (Lew-Williams 2014, p. 27). New Chinese immigration then dropped to near-zero for decades. The Chinese immigrant population in the United States peaked at 106,688 in 1890, falling to a low of 45,757 in 1920.

US employers then turned to Japanese immigrants in the early 20th century, but that flow of immigration was shut down by a 1907 agreement, and then expanded to a continent-wide “barred zone” in 1917. The Phillipines had been ceded from Spain to the United States at the end of the 1898 Spanish-American war. By the early 1930s, more than 100,000 Filipinos had emigrated to Hawaii (which doesn’t become a US state until 1959) and another 30,000 had emigrated to California. This Asian immigration was then shut down by the 1934 Tydings-McDuffie Act, which as Postel writes “deemed Filipinos `foreign aliens’ subject to a quota of 50 new immigrants per year.”

Thus, my first alterate history is how the United States would have been shaped for several decades after the 1870s if the Asian exclusion laws had not been passed. For example, Asian immigration during this period would surely have been much larger. The population of west coast states like California (becomes a US state in 1850), Oregon (1859) and Washington (1889) might have grown much more rapidly. The share of seats for these states in the House of Representatives might have been substantially large, and the economies of the western states might have grown much more rapidly. The inherited cultural history of the western United States might have diverged substantially from US states that fought for the North during the US Civil War, as well as states that fought for the South.

The other paper in the JEP symposium by Gaurav Khanna, “From Asia, with Skills” (Journal of Economic Perspectives 40:1, 215–40) focuses on more recent experience of Asian American immigration. Overall, as Postel notes, “The Asian-identifying population is now nearly 8 percent of the US population, having more than doubled since 2000.” Khanna focuses on the insight that this wave of Asian American immigration has been predominantly high-skilled. He write:

Migration from countries in Asia in the last three decades or so—and especially China and India, as well as South Korea, the Philippines, and Japan—has occurred with unprecedented scale and impact. This wave of migration to the United States has been predominantly high-skill. In 2019, among working adults in the US labor force, 78 percent of Indian-born and 63 percent of Chinese-born workers held at least a four-year college degree, compared to 39 percent of US-born workers (ACS 2019). Between 1990 and 2019, the share of college-educated workers in the United States from these five Asian countries doubled to 7.3 percent. Over this period, migration from these countries contributed to over 38 percent of the growth in US employment of software developers, 25 percent of the increase in scientists and engineers, and 21 percent of the growth in physicians (ACS 2019).

Khanna goes into some detail about the conditions of demand for skilled workers in the US economy, the supply of skilled workers in these five Asian countries, greater economic opportunities in the US economy, changes in immigration laws since 1965, and other factors and events driving this change. My second alternative history involved musing about the shape of the US economy in recent decades if this supply of Asian immigration had remained highly constrained.

The overall picture of US colleges and universities would look rather different if international students from China and India had not increased by about 400,000 from 2007-2023. In technology fields, “In 2010, 27 percent of all foreign-born workers and 35 percent of foreign-born information technology workers in the United States initially arrived on a student visa. Asian-born students are heavily concentrated in science and engineering fields, such as computer science, engineering, and mathematics. … By 2010, one-third of US information technology workers were foreign-born, predominantly from India. By 2019, Indian-born individuals were 29 percent of all workers in science, technology, engineering, and mathematics jobs in the United States.”

In the area of entrepreneurship and corporate management, Khanna notes (citations omitted):

Asian-born immigrants have also played prominent roles in entrepreneurship, particularly Chinese and Indian immigrants in the tech sector. In the late 1990s, Chinese- and Indian born engineers were running one out of every four US tech startups, and a similar fraction had at least one immigrant founder. More generally, immigrants are twice as likely to start new businesses as natives. In 2023, 44.8 percent of US Fortune 500 companies were founded by immigrants (from anywhere in the world) or their children (AIC 2023); in 2022, 55 percent of America’s `unicorn’ startups (privately held startups with total stock market capitalization over $1 billion) had an immigrant founder. As of 2024, chief executive officers of giants like Google, Microsoft, Adobe, Cognizant, and IBM were born and educated in India. Indian-born CEOs of Fortune 500 companies are also found in other sectors of the economy, including technology (Micron Technology, Microchip Technology), consumer goods and retail (Starbucks, Chewy, Albertsons), technical services (FedEx, Cognizant, Honeywell, Jacobs Solutions), and pharmaceuticals (Vertex). … Looking ahead at the race for artificial intelligence startups, a recent analysis found that nearly two-thirds of the Forbes top 50 artificial intelligence startups had an immigrant founder, with Indian entrepreneurs leading the pack with 10 startups …

In the provision of health care, as Khanna notes, “In 2023, according to the Association of American Medical Colleges (2024), one in four active US physicians were international medical graduates. In 2016, migrants from Asia made up 17.7 percent of all US physicians and surgeons (Patel et al. 2018). Physicians are more likely to be from India, with recent growth in the number of physicians from China. The Philippines, on the other hand, is an important source of the nursing workforce in the United States. Just before the Covid-19 pandemic, there were 136,000 Filipino nurses and 78,000 Indian-born physicians in the United States. Migrants from Asia—and in particular, China, India, and the Philippines—represent 20 percent of the growth in the US physician population and 11 percent of the increase in the US nurse population between 1990 and 2019.”

When noting the actual contributions of immigrants to the US economy, one common response is along the lines that, if only the immigrants had not been here, those jobs would have been filled by native-born Americans. In some situations, for some jobs, I’m sure that’s true. But the opportunities and flexibility of the US economy is also a magnet for those with high skills from all over the world–including especially the two largest-population countries in the world of India and China. It does not seem to me that US medical schools, for example, have been standing by, ready and waiting to train massively more native-born doctors and nurses. I am dubious that a legion of native-born entrepreneurs or AI-talented managers have highly profitably business plans in hand–but are somehow being shut out by immigrants with skills in these areas.

Frankly, it doesn’t matter to me if potential native-born CEOs are displaced by immigrant talent. (I don’t spend much time worrying about the income and welfare of CEOs.) But there is strong evidence that new and successful companies increase the wage and job opportunities available to a wide swath of middle-class and lower-income workers, which matters to me a lot more. Asian American immigration seems clearly to be one reason why the sustained and solid growth of the US economy has in recent decades exceeded that of other high-income countries, like those in the European Union. I would be delighted to see the US education system, from K-12 and up through higher education, produce a sharp rise in the US share of high-skilled workers. A high-income economy is driven forward by high-skilled workers, and the more of them the better.