Artificial intelligence (AI) holds enormous promise. Properly deployed, with suitable guardrails against its risks, AI is likely to raise living standards by enhancing productivity growth and to lead to new scientific discoveries, the development of new cures for all kinds of disease, and the creation of new products and services. Yet even as China accelerates efforts to develop and diffuse AI throughout its economy, public attitudes in the United States are markedly more wary. In 2025, nearly equal shares of Republicans (50%) and Democrats (51%) said that the growing use of AI in daily life made them more concerned than excited.1 This apprehension is reflected in broad opposition to local AI data center construction, which 71% of Americans opposed in a 2026 Gallup poll, including majorities of Republicans, independents, and Democrats, and in concern that AI can enable misinformation, fraud, and other misuse and may disrupt jobs and careers.2 Since that poll, concerns about AI’s potential existential risks have intensified, following public warnings from current and former employees of leading AI companies that current AI systems have insufficient oversight.3
I focus here, however, on just the job-displacement fears, which are likely to persist because AI use nonetheless will continue to grow even if data center construction and AI model development in the U.S. are slowed or halted. Even recent warnings that AI could take over the internet seem unlikely, by themselves, to halt AI adoption. AI is also being deployed to strengthen cyber defenses, and surveys document its growing use by businesses and individuals.4 So the job threat will remain. In July 2026, over 1,000 economists signed an online petition organized by Stanford’s Erik Brynjolfsson urging action to deal with the threat of future labor market disruption posed by AI.5 Even if those worries eventually prove unfounded as they so far appear to be—and even if over the longer run our economy and workers will adapt to AI just as they have adapted to previous technological advances—it makes sense for the government to have policies in place to ease the pain of transitioning to a post-AI economy.
In the following sections I lay out three such transition proposals—wage insurance, retraining assistance, and entry-level tax credits—as well as their annual costs under different scenarios, and suggest ways to pay for them. Under a reasonable set of assumptions, labor market disruption equivalent to a sustained 6% unemployment rate would require about $60 billion a year to finance these policies. Each policy addresses a different dimension, and so I view these as complementary proposals that, together, could serve as insurance that protects our economy and its workers from the potential downside of AI-driven job and income losses.
At minimum, to the extent the AI revolution has been causing a temporary jobs boom in the creation and usage of AI, including the industries constructing the infrastructure, those new jobs are not suitable for many, perhaps most, workers whose jobs AI may replace.6 Having the government take on an insurance role to ease the AI job transition for these displaced workers is no different from other government insurance programs in place for other low-probability, high-consequence events: for farmers against risks to their crops, for individuals and businesses against the risk of floods, let alone the many forms of insurance provided by the private insurance companies against risks to property and health.
In the case of AI-induced job and income risk, however, the private insurance market is likely to fall short for two reasons. First, there is no actuarial basis for private insurers to set premiums because future AI-induced job and income displacement could be so different from past technological change. Second, private insurance only works if the events insured against are independent from each other; private insurers will not take on highly correlated risks. That is the situation with AI-induced job and income losses: if they materialize, they are likely to be too highly correlated across people and areas of the country for private insurers to offer insurance.
To some, a universal basic income (UBI) is the answer. However, UBI would be far too expensive at the scale it might be needed. Unlike trade or automation, whose job and income displacement effects tend to be concentrated on middle-income manufacturing workers, AI threatens not only those workers (through AI-assisted robots), but potentially a broad swath of white-collar workers with substantially higher incomes. Accordingly, even a UBI program promising a poverty-level floor income, which has been estimated to have a net annual cost of approximately $784 billion after accounting for taxes clawed back (as of 2025),7 would do nothing to quell the anxieties of many, perhaps most, higher-income AI-displaced workers.
Fortunately, there are other more effective and politically viable approaches. One such program is wage or income insurance, which would reimburse displaced workers for some portion of their income loss for some period of time after taking a new, but lower-paying job. Another set of ideas would assist workers to retrain for new jobs and careers, either through government lifetime retraining accounts for individuals themselves and/or tax credits for employees and/or employers who provide such training. A third idea is to subsidize the hiring of entry-level workers, who would not be eligible for either wage insurance or retraining assistance.
So far, however, these programs have been advocated only in general terms, with minimal details and no cost estimates. The purpose of this report is to fill in those blanks, as well as to provide more detailed thoughts about how the three programs would be paid for. In doing so, I will draw on my own prior published work, spanning over three decades, with multiple authors, in advocating for both wage insurance and, separately, from my own earlier proposal for lifetime retraining accounts.
Wage loss insurance
The most recent comprehensive survey of wage insurance and its effects by Benjamin G. Hyman, Brian K. Kovak and Adam Leive (“HKL”)8 identifies the origins of the concept of wage insurance for displaced workers in a chapter of a Brookings publication, “Economic Choices,” in 1984, co-authored by Robert Lawrence.9 Two years later I teamed up with Lawrence to offer a few details of such a program targeted only for workers displaced by trade in a book titled “Saving Free Trade.”10 We proposed that trade-displaced workers receive payments up to 50% of their income loss once they obtained a new job, subject to an unspecified ceiling, and for an unspecified limited time. We suggested that the reinsurance percentage might be phased down over the eligibility time limit. We offered the proposal not as a replacement for unemployment insurance, but as a supplement, the idea being to encourage displaced workers to find jobs more quickly in order to receive the insurance payments. To finance the program, we suggested converting existing import quotas on textiles, sugar, and steel into tariffs.
Fifteen years later I teamed with a different co-author, Lori Kletzer, not only to flesh out more details about a wage insurance program, but to estimate its annual cost.11 In particular, we proposed a universal program, not tied to any specific reason for job loss, for all displaced workers (those involuntarily unemployed for reasons other than cause) with tenure of at least two years, with insurance making up 50% of the income loss upon the worker’s obtaining a new full-time job, up to a $10,000 annual maximum. Insurance payments would be triggered on the date of reemployment, providing incentives for quick reemployment, even as the newly employed workers could continue to look for better jobs. The proposal was largely designed for displaced manufacturing workers, victims more of technological advances than trade. Available evidence back then indicated that displaced service-sector workers tended to suffer little or no wage losses, or minimal wage losses, when they found their new jobs. On reasonable assumptions, we estimated then that such a universal program, serving approximately 800,000 eligible displaced workers who suffered an average 22% cut in wages upon taking new jobs would cost approximately $3 billion, at a national unemployment rate of 4.9%. Five years later, I teamed with Lael Brainard and Nicholas Warren to update the annual cost estimate for an eligible population of approximately 500,000, which we pegged at $3.5 billion given a 4.7% national unemployment rate.12 The costs were higher than those estimated in my study with Kletzer, even with a smaller eligible population, because the estimated annual earnings loss in absolute dollars was roughly double the amount in that study (approximately $17,000 versus the earlier $8,800 wage cut).13 We also suggested that the program could easily be financed by a small increase in the federal unemployment tax (FUTA).
Remarkably, wage insurance, albeit limited to trade-displaced workers and not universally extended, quickly found its way into public policy in the U.S., following Canada’s experimentation with it in 1995. In 2002, the administration of George W. Bush worked with Congress to introduce a version of wage insurance as an alternative Trade Adjustment Assistance for workers over the age of 50 determined to be trade-displaced, who found new work in less than 26 weeks from the date of unemployment, and were compensated for 50% of their wage loss up to a lifetime (not annual) maximum of $10,000. President Obama, with congressional approval, later extended that program, before proposing in his last State of the Union address in 2016 that the program be extended to all displaced workers regardless of cause.14
For illustrative purposes, I now model the gross costs of a generous wage insurance program available to all long-tenured displaced workers earning below $250,000 in their last job. The Displaced Worker Survey in recent years only reports on workers with at least three years’ tenure at their prior jobs, and that is the eligibility criterion used here. The insurance covers such workers who obtain either a new full-time or part-time job. The compensation for both is capped at a level that produces an average income loss for estimation purposes here of $30,000 per year. The $30,000 average loss roughly scales for inflation the $17,000 wage loss figure that Brainard, Litan, and Warren estimated for the year 2001. A more precise average wage loss figure cannot be determined from the current Displaced Worker Survey, which only breaks down the number of workers who suffer income loss of less than or more than $20,000. A more granular analysis of the average wage losses would be required to set actual compensation caps. The estimates here apply the same wage losses to displaced workers who may prefer part-time work initially, while they are looking for new full-time employment, in order to encourage such searches. Based on the 2023–25 Displaced Worker Survey, 66% of long-term displaced workers found new full-time work, 51% of whom took a cut in wages on reemployment.15
The proposed program also has some additional features. Workers must be involuntarily displaced for reasons other than cause, the definition used for unemployment insurance (UI) programs. Benefits are provided for three years (not two, as in the earlier proposed program), from the date of reemployment, with 80% of earnings loss compensated in the first year, 60% in the second year, and 40% in the third year, with no annual or lifetime cap.
I have estimated the gross and net costs of wage insurance under the foregoing assumptions, under three different national unemployment rates: 4% (approximately the rate at this writing), 6% (mild recession and/or relatively mild AI impact), and 8% (severe recession and/or severe AI impact). The costs for the higher unemployment scenarios are scaled linearly from the 4% base case, so the 6% scenario implies 50% higher cost than the 4% unemployment scenario and the 8% scenario implies a doubling of the cost under the 4% scenario. The net costs offset the gross amount for (1) an illustrative two months’ quicker reemployment rate with wage insurance in place, and (2) additional tax revenues collected on such reemployment, at an illustrative tax rate of 15%, reflecting a conservative estimate of the average tax rate likely paid by displaced workers. The costs are “steady-state” costs, or those that take full account of all three years of compensation.
Table 1 provides net cost estimates using the assumptions described above, under three different scenarios. Table 1 shows that in the “base case” estimates the net costs range from $25 billion under 4% unemployment to $50 billion under 8% unemployment. The annual cost estimates are conservative, since they assume that the accompanying level of unemployment is sustained indefinitely, which would not be true as workers and labor markets adjust over time. The annual cost estimates are expressed in 2025 dollars; if the wage loss compensation amounts are scaled with inflation, as they should be, the 2025 cost estimates should be viewed as real costs going forward. In the midpoint 6% unemployment scenario, the annual net costs of wage insurance are $38 billion, or a little more than 0.1% of GDP at year-end 2025.
These steady-state cost estimates may be understated, however, to the extent that policymakers extend coverage to workers with less than three years’ tenure at their prior job. On the other hand, it is possible that the presence of wage insurance could raise reemployment incomes relative to what occurred in the past. The HKL study of the limited wage insurance program found that it “increased employment probabilities by 8 to 17 percentage points during the two years following displacement before fading to zero after four years.” The study also found that earnings replacement rates increased by 10 percentage points under the program and cumulative earnings increased by over $18,000 during the four years following displacement, a 26.6% increase relative to the average four-year cumulative earnings of ineligible workers.16 Most striking is their finding that the program more than paid for itself: the net cost to the government was negative because of reduced UI payments and tax receipts on the higher earnings. Although the authors caution that this favorable result could be the consequence of the limited nature of the program,17 the prior experiment with more limited wage insurance suggests that there could be an additional tax receipts offset due to enhanced earnings that is not reflected in the net cost estimates for the universal program suggested here.
Retraining assistance
Not all displaced workers will look immediately for a new job. Some may opt to retrain for a different career path by going back to school or entering a retraining program.
In principle, the federal government can support retraining either through much expanded tax credits for all or a portion of an individual’s expenditures on retraining, or through a version of “lifetime retraining accounts.” The reason for expanding the existing tax credits for retraining is that they are very limited. Federal law now offers a Lifetime Learning Credit (LLC) of up to $2,000 annually for qualifying postsecondary job-skills coursework, but its nonrefundability (it doesn’t help workers who pay no federal income tax), limited eligible expenses, and income phaseout materially limit its usefulness as a displacement-retraining benefit.18 The American Opportunity Tax Credit (AOTC) is a federal, partially refundable income-tax credit for the early years of postsecondary education, but because it is designed chiefly to help students, and often their parents, pay the costs of beginning an undergraduate degree or recognized credential, it does not function as a general, lifetime retraining benefit.19 Given these limitations, Davidson and Slaughter propose an unspecified, presumably additional, tax credit to either or both displaced employees and employers for retraining.20 Anthropic proposes unspecified “incentives for firms who retain and redeploy their workers.”21
Although tax credits are relatively easy to administer, since they build on the existing tax code, they have drawbacks. When extended to employers, there is no easy way to ensure that the credit is not rewarding training that employers would otherwise undertake. When extended to employees, even if they are refundable in the full amount of any retraining, tax credits can’t help those who don’t have the cash or borrowing capacity to pay for the costs upfront, though the One Big Beautiful Bill Act of 2025 (OBBBA) extended Pell Grants to short-term training programs that alleviate some of this concern.
Given the liquidity issues associated with paying for lengthier careers, I and others have advocated instead a lifetime training account (LTA) available to every working person that could be drawn down for retraining, at multiple stages of a worker’s life. About 10 years ago, Ed Alden and I suggested a flat $100,000 loan account only, with federal regulation aimed at ensuring that the training programs meet minimum standards.22 Shortly thereafter, the Aspen Institute proposed a retraining savings account that would be matched with federal dollars.23 Earlier, Robert LaLonde (University of Chicago) and Daniel Sullivan (Federal Reserve Bank of Chicago) proposed increasing Pell Grants for training-ready displaced workers rather than creating accounts per se.24
I propose here instead for illustrative purposes a much more targeted version of my original proposal with Alden, focusing on the 1.1 million displaced workers each year with at least three years’ tenure on their prior job. This feature aligns with the eligibility criteria for the wage insurance proposal. The annual flow is scaled linearly to each unemployment scenario off a 4.0% baseline, roughly the 2024 average. As with the wage insurance estimates, the cost figures are scaled linearly with the unemployment rate. I model worker take-up rates of 20% (low), 30% (medium), and 40% (high), respectively. The LTA program I envision would provide grants to cover training. One way to estimate that cost is to use the average in-district cost of $4,150 for attending community college, and assume a two-year retraining period, to arrive at annual training costs of $8,300.25 Displaced workers may or may not require a living cost subsidy, depending on whether they have a working spouse or partner. Living costs under the proposal would be paid for through an income-contingent loan—one that is paid back at a certain percentage of future income. I assume here a gross annual borrowing of $30,000, and a subsidy rate of 50%, meaning that on average only half the loans would be repaid. This scores the loan portion of the LTA, for those who take it, at $15,000 per year.
Table 2 provides total LTA costs under the foregoing assumptions, under the three different unemployment and take-up rate scenarios.
The midpoint of the table shows annual LTA costs of $11.5 billion, under a medium take up rate with a 6% unemployment rate equivalent. Over all the scenarios, the costs range from a low of $5.1 billion to a high of $20.5 billion.
Entry-level job tax credit
Early evidence is mixed as to the impact of AI on entry-level jobs so far. Brynjolfsson, Chandar, and Chen find a relative post-ChatGPT employment decline among U.S. workers ages 22–25 in highly AI-exposed occupations.26 In contrast, Hernæs and Kostøl find no statistically precise corresponding decline across the complete distribution of AI-exposed occupations among young Norwegian private-sector workers,27 and Danish matched employer–worker evidence finds no difference in entry-level hiring between AI-adopting and nonadopting firms.28 A subsequent study by Fairlie and Wu, using data for 2026, comes to a similar conclusion.29
Nonetheless, there are widespread fears that in the future AI will negatively affect entry-level jobs. The federal government can address this problem in a cost-effective way without limiting any hiring incentives to AI-specific sectors, which can be difficult to define. I do so here by modeling a variation of the tax deduction for hiring young workers advocated by Leicht and Ball.30
More specifically, I propose a first-year 50% tax credit for hiring recent graduates of two- or four-year colleges within 12 months of their earning their degree who have been hired into a job paying more than $40,000 annually, but not more than $100,000. The wage floor is designed to address real entry-level jobs. The wage ceiling is designed to exclude jobs that are not truly entry-level. Both the floor and ceiling obviously can be adjusted in the legislative process, and both numbers should be indexed to inflation.
Table 3 depicts the annual cost of the proposed entry-level job tax credit, across a reasonable range for the annual take-up rate. The cost estimates are built off an annual cohort of graduates from two- and four-year colleges of 3 million,31 an employed rate of 72%,32 and my estimate that of those employed, 65% of those have annual earnings between $40,000 and $100,000. Under these assumptions, approximately 45% of each young cohort, or about 1 million young people each year, would be eligible for an employer tax credit. The table shows the annual costs ranging from approximately $6 billion to $17 billion, with a midpoint of $11 billion.
The take-up rates are likely to be inversely proportional to the level of the unemployment rate, with more take-up expected when unemployment rates are low and employers more willing to hire young people for entry-level jobs, and less take-up when unemployment is high. This pattern is opposite to that of the cost estimates for wage insurance and LTA, which rise as the national unemployment rate rises (or the displacement effect of AI is greater).
Paying for transition insurance
So how will the government pay for the different policies suggested here? It would be nice if the HKL result, that at least wage insurance can actually pay for itself through reduced UI payments and higher tax collections, could be generalized to a program serving all displaced workers. Such an outcome is possible, or at the very least, the Congressional Budget Office, the official federal scorekeeper, might offset some of the gross costs of the wage insurance component with these credits. Still, some portion of the wage insurance, plus the estimated costs of the LTAs and entry-level job tax credit will need to be financed.
In theory, the amount of financing could vary with projected outlays for a given year, which in turn depends on the degree of AI job displacement and/or macroeconomic projections (here proxied by the unemployment rate). That way, again at least in principle, there would be no net fiscal impact on the budget each year. But as a political matter, the notion that Congress must appropriate sufficient funds every year to pay for wage insurance/LTAs/entry-level job tax credit is a nonstarter, especially in years of recession, when Congress would be loath to pass anything that looks like a tax increase. For this reason, I suggest that when launching the package of “AI transition insurance,” Congress raise sufficient funds through any of the mechanisms suggested below to pay at least for the 6% unemployment scenario, the version of wage insurance outlined here, and the medium take up rates for LTA withdrawals and the entry-level job tax credit—assumptions that would produce a total annual net cost, and hence need for annual financing, in the $60 billion range. This is a bit smaller than annual earned income tax credit expenditures,33 and totals about 0.8% of total federal outlays.34 Net annual surpluses (receipts minus the outlays for the program) in any individual year would go toward reducing the deficit, which at the current 6% of GDP range clearly needs reducing. In worse years, for example if AI displacement is higher, the program would not be fully paid for, which is the right outcome because then the absence of full financing acts, on net, as a countercyclical stabilizer.
Now, to the “how” question. It is tempting to say that only AI companies should pay for the transition insurance program. But who counts as an AI company and why limit it to them? Just the companies that make large language models (LLMs)? Companies that make robots, increasingly powered by AI? Companies or individuals who use AI to become more productive? There are no clear answers to these questions, but the cloudiness suggests we look to broader-based sources of revenue. For related reasons, I reject calls for taxes on AI “tokens” or on “robots,” not just because they directly impair innovation, but because they would be highly complex to design and administer.
Here are some possible alternative ideas. Davidson and Slaughter suggest a new 25% payroll tax on all equity compensation, to get at the fact that most executive compensation is paid in the form of stock or stock options, rather than in cash.35 Although such a tax reaches beyond those working for “AI companies,” however defined, it would cover those who stand to benefit from the AI boom. The authors estimate such a tax would generate at least $100 billion in annual revenue.
Alternatively, a second idea suggested by Greg Ip of the Wall Street Journal,36 Nobel Prize winners Paul Krugman37 and Daron Acemoglu (with co-authors Andrea Manera and Pascual Restrepo) is to rectify the current imbalance of taxation on labor and capital, by increasing the taxes on physical investment.38 As Ip points out, since the 2017 tax bill, while the marginal tax on labor has dropped a bit, from 29% to 27%, the marginal tax rate on new equipment is now just 4%, down from 14% before that legislation.39 This glaring disparity calls out for rectification especially because of the very limited effect that tax rates are having on physical investment and in turn GDP.40 Moreover, all three authors point to the secular decline in labor’s share of national income and the corresponding rise of the capital share (represented by profits), trends which AI is likely to accentuate. Hence at least a modest tax increase on capital seems in order. It is not clear what form such an increase should take—a reversal of the 100% expensing of capital investment “made permanent” in the OBBBA and/or an increase in the corporate tax rate generally—but it ought to total roughly $60 billion annually to fund the full transition insurance program.
The $60 billion increase is less substantial than it may appear. Given that corporate income tax receipts for 2025 were approximately $450 billion,41 the amount required to fully fund the AI transition program would represent a 13% increase in average corporate taxes, or an increase of about 2.5 percentage points of corporate income, from the current 21% to 23.5%. The higher average rate would still be well below the top marginal rate of 35%, which was replaced by the 21% flat rate in the Tax Cuts and Jobs Act of 2017 during the first Trump term.42 If assessed on corporate income, generally, the tax increase would fall on a combination of shareholders, customers, and workers. That is a reason to favor the Davidson and Slaughter tax on equity compensation, whose incidence would be split among those executives and the corporations paying them, resulting in less pass-on to corporate customers and labor compared to a tax on corporations directly.
Closing caveat: Status loss of potentially many displaced workers
I close with one important caveat and commentary on it. The transition insurance program outlined here would not eliminate the loss of status many displaced workers suffer when forced to pursue new careers and jobs. A career is not only an income stream; it’s a source of identity, mastery, and social standing, and asking a 50-year-old professional to move into hands-on work they were raised to see as a step down is asking them to absorb a real status loss.
The only honest answer to this legitimate concern is that over time AI will change the culture, which will assign more value to hands-on work and careers it entails, such as electricians, plumbers, and nurses. Some portion of those white-collar workers who will learn these skills also will be entrepreneurial and will build businesses that employ others in the skilled trades that AI or artificial general intelligence (AGI) will not hollow out. Indeed, insufficient attention is given to the several million entrepreneurs, many if not most without fancy college degrees, who have done precisely that over the last few decades.43 In the AI age, those opportunities for building business empires in “hands-on” activities, especially in the service sector, will continue to exist. Others may find comfort in the fact that the work AI leaves for humans is disproportionately the work that requires care, judgment, and trust, which is to say the work that is most recognizably human. But the hard reality is that the AI genie is out of the bottle, and there is no sugarcoating that fact. The transition insurance program here at least might help many people adjust to the new reality.
Admittedly, some may balk at the $60 billion a year price tag, which should be viewed as the equivalent of an insurance premium paid to avert a self-defeating populist backlash against AI, or even the capitalist system that has produced it. The stakes are that high, and the fears about future AI job and income displacement are sufficiently legitimate, in my view, to justify the payment of that premium.
The Brookings Institution is committed to quality, independence, and impact.
We are supported by a diverse array of funders. In line with our values and policies, each Brookings publication represents the sole views of its author(s).