How MIT’s “Work of the Future” Project Gets It Wrong: Raising Taxes on Machinery and Software Would Kill Jobs and Hamper Wage Growth

Robert D. Atkinson October 5, 2020
October 5, 2020

(Ed. Note: The “Innovation Fact of the Week” appears as a regular feature in each edition of ITIF’s weekly email newsletter. Sign up today.)

The most important determinant of a nation’s economic wellbeing is the growth in per-capita GDP. The principle way that growth occurs is through productivity growth. And a principle way productivity grows is through automation technology.

Because a suite of new and improving technologies has emerged, which ITIF has termed “CAS” (connected, automated, and smart), many believe that the long U.S. productivity growth drought could soon end. And if we’re lucky, productivity growth could increase to the rates enjoyed in the last half of the 1990s and first half of the 2000s. But too many pundits, activists, and scholars have panicked, claiming automation will be all-powerful, and robots will mean the end of work. This in turn has led to an array of ill-informed and harmful policy proposals.

The latest case in this panic and neo-Luddite reaction: a new report from MIT’s “Work of the Future” project. The report, titled “Taxes, Automation, and the Future of Work,” is wrong in its analysis of the relationship between automation and employment; in its analysis of the relative taxes on labor and machinery; and most importantly, in its recommendation to increase the effective tax rate on machinery and equipment investments. Doing that would significantly lower wage growth and reduced U.S. international competitiveness.

This blog reviews major statements from the report (numbered in italics) and explains why each is wrong.

Faulty Reasoning on the Impact of Automation on the Labor Market

1. Automation can have very different effects on employment and wages than technologies that increase human productivity.

The authors provide no evidence for this claim, and it is patently illogical. Imagine a technology that helps bowling alley pinsetters better pick up bowling pins and put them in the right place, but still requires human pinsetters to be involved, versus automated pin setting machines, which do not. Clearly, the latter is more productive. The same is true for virtually every other automated technology: machines that shake trees to harvest fruits and nuts; automatic paint spray booths in car factories, etc. More and better automation boosts productivity.

2. Automation displaces workers from these tasks, creating a force toward lower labor demand. 

The authors’ argument is that by eliminating a job, automation leads to fewer jobs overall and lower labor demand. This may be true for some firms or industries, depending on the price elasticity of demand for their products. Farming produces as much food as a share of the economy as it did 100 years ago, but because of automation it employs a vastly reduced share of the workforce. But agricultural automation didn’t raise overall unemployment, because laid-off farm workers got jobs doing something else. As ITIF has shown, historical data and academic studies show that high productivity, including from automation, do not lead to reduced labor demand.

3. When capital can perform some tasks more cheaply than labor, automation reduces the firm’s costs and in effect makes it more productive. As the firm expands to take advantage of these productivity gains, it also demands more labor. We thus see the two countervailing effects of automation on labor: task displacement, which reduces labor demand, and productivity improvements, which increase labor demand. If the former force dominates, automation will reduce not only the labor share but also employment and earnings.

The authors succumb to what economists call the “lump of labor” fallacy, which holds that once a worker loses their job there is no additional demand for that worker. To use the agricultural example, this makes no sense. Agricultural automation lowers food prices, which leaves consumers more to spend on other items, and that in turns creates demand for labor.

When it comes to the notion that automation will reduce earnings, the authors make two errors. The first is the assertion that overall labor demand goes down, which reduces labor bargaining. There has been no evidence of that. In fact, historical evidence suggests the exact opposite. Periods of high productivity have been associated with lower unemployment rates and higher wage growth. The second is that they confuse wages and overall income. If a company automates, the prices for its products or services will increase less than the rate of inflation, or even decrease, meaning Americans become richer as consumers.

4. When firms decide whether to use machines or workers, they may ignore the possibility that workers themselves are much better off when employed, rather than when they are “on the dole” or out of the labor force.

Again, the authors repeat the lump of labor fallacy. If automation really put workers on the dole, then the societal cost-benefit calculus would be much different. But automation is the proverbial free lunch. Workers may suffer a short period of unemployment (hopefully compensated for by unemployment insurance and other supports), but the economy grows, producing not only what the worker used to produce, but also what the worker in their new job produces.

5. One implication of this substitution of machines for humans is that automation has an unambiguous negative impact on the labor share — how much of a firm’s, an industry’s, or the economy’s revenue goes to labor. By reducing the set of tasks performed by workers and increasing those performed by capital, automation always increases the capital share and reduces the labor share… In 1980, labor’s share of national income stood at 63.6%, meaning that nearly two-thirds of every dollar earned in the economy went into the pocket of workers. It fell to 56.6% by 2017.

It is surprising that the authors don’t consult the national income statistics from the U.S. Commerce Department’s Bureau of Economic Affairs. If they had, then they would have found that virtually all of the decline in labor share was offset by an increase in self-employment income (another form of labor income), and even more by an increase in rental income (the amount of rent renters pay and mortgages homeowners pay). As former Obama administration officials Jason Furman and Peter Orszag wrote: “the decline in the labor’s share of income is not due to an increase in the share of income going to productive capital—which has largely been stable—but instead is due to the increased share of income going to housing capital.

6. The problem is that not all automation creates this coveted productivity effect. In some cases, which Acemoglu and Restrepo call “so-so technologies,” the new automation isn’t much better in terms of productivity than the human labor it is replacing.

Companies may sometimes adopt automation that does not pay off. Companies make mistakes, especially with new and untested technologies. But the idea that they would do so systemically defies reason. Companies that did would lose market share or even go out of business. Besides, the idea that adopting technology that replaces a human task is less productive than a technology that complements a worker is illogical. Take the cases of automatic elevators, robots that pick and place, automated bottle fillers, etc. This is not to say that complementary automation cannot be productive, it is to say that claiming that there is good and bad production technology is wrong. All automation is good automation.

7. An example of so-so technology is self-checkout machines increasingly found at retail stores. Those machines shift the task of checking out groceries to consumers, eliminating the need for cashiers. And, since an average shopper is probably less adept at tallying groceries than a skilled cashier who doesn’t have to struggle to figure out if the tomatoes are organic, or if a can of soup is really on sale, it likely slows the checkout line rather than accelerates it.

The way to measure the productivity of service automation technology is not to compare workers’ productivity to the consumers’. Yes, consumers are less efficient. But that misses the point. The right question is, are consumers more efficient than cashiers, taking into account the time consumers wait for cashiers to check them out? When both workers’ and consumers’ time is included, self-service technology enables society to produce more with less labor.

Faulty Analysis of Relative Tax Rates on Labor and Machinery

After mistakenly arguing that automation increases unemployment and reduces labor demand, the MIT authors then mistakenly claim that the tax code favors investments in machinery over hiring workers.

8. These [tax disincentives] consist of payroll taxes and federal and local income taxes, both of which create a gap between what employers pay for labor and what workers receive.

Virtually all economists agree that employer-paid payroll taxes come at the expense of wages. Absent payroll taxes, wages would be higher. So payroll taxes do not raise the cost of labor. And to say that taxes workers pay on their incomes raise the cost of labor makes no sense. Employers do not pay the cost of income taxes; workers do. If society abolished the income tax and raised the same amount of revenue from property and sales taxes, then workers would still pay the same taxes on average.

9. Depreciation allowances enable corporations to deduct capital expenditures from their tax obligations and have been a major factor in reducing taxes on capital. But they have become much more generous over the last two decades, allowing firms to write off their expenditures immediately and fully.

Companies can immediately write off for tax purposes wages and expenditures on worker training. Moreover, depreciation allowances have generally not become more generous, at least compared to the period before 1986 when companies received a more generous tax credit for investing in new machinery.

10. Figuring a single effective tax for capital is more challenging…we do this by distinguishing between different types of corporations…

The authors argue that because the corporate tax rate has declined, it provides an unfair incentive to invest in machinery. But the corporate tax rate is no more a tax on capital (defined as machinery) as it is on labor. If a company hires a worker and it increases its profits by $10,000, then the company still pays corporate income tax on this income, just as it would if it buys a machine that increases profits by $10,000. The authors confuse taxes on capital income from taxes on machines per se.

11. [Ideally] there is no incentive to push one over the other without the corresponding cost reduction or productivity gain.

In fact, there should ideally be an incentive to push machinery, because economic research shows that there are large positive externalities to company investments in machinery and equipment that accrue to other companies and society overall. Providing incentives for machinery investment, including automation, maximizes overall economic welfare.

12. It is clear that structures capital (buildings) as well as capital types that work closely with labor, such as machine tools and many computer applications, should not be the target of automation taxes.

In fact, this is completely the opposite of what economic theory and research would suggest. There are no externalities from investing in buildings, but large ones from investing in capital equipment, suggesting that the taxes on the former should be higher than the latter.

13. Tax policy that leads to excessive automation not only costs jobs, but also is likely to reduce paychecks for the workers who remain.

This is clearly illogical, since companies that automate are now more productive and can afford to pay higher wages. In a review of the economic impact of industrial robots across 17 countries, Graetz and Michaels found that robots increased wages.

14. If we are correct that the tax code can spur excessive automation, why has the tax code evolved to favor capital over labor — and increasingly so? We don’t know for sure, but two factors seem to be important. Capital owners often have greater sway over policymakers than workers, so it is reasonable for them to receive more favorable treatment.

The authors confuse taxes on capital income and the income of the wealthy with tax incentives for investing on machinery. “Capital owners,” including rich people who own stock, may have more political sway than workers, but they have certainly not pushed for more tax incentives to invest in equipment. Most of them just want lower taxes, not lower taxes tied to requirements to invest in machines. Also, if they have such power, then why did Congress eliminate the investment tax credit in 1986?

15. But when it is excessive — for example, driven by tax incentives rather than for efficiency gains — [automation] harms labor and fails to improve productivity.

Given today’s weak tax incentives for investing in automation, and anemic productivity and capital investment rates, it is impossible for automation to be excessive.

16. Automation taxes should not be applied to cases where automation boosts productivity and adds jobs. Of course, it is a challenge for the tax code to identify these “marginal” cases.

One can see the tax fraud and definitional fights now. IRS: “But this machine replaced a worker.” Company: “No, it didn’t.” More troubling is that the authors want government to throw sand in the gears of progress by taxing automation technology, including artificial intelligence software.

17. The same argument applies to wealth taxes… it is not the type of tax that we are calling for here either. This is because, like capital taxes, wealth taxes would apply to all types of capital; and, further complicating matters in this case, they would only do so when these capital goods are bought or owned by the wealthiest investors.

If the authors really want to reduce inequality through the tax code, then they should not focus on taxing machinery investment, which increases growth and wages, but rather on taxing high incomes and wealth.

18. Changes to the tax code such as these not only would help bolster income for low-wage workers and create more jobs, but also could help lessen inequality.

This is wrong on all three counts. 1) Taxing automation would increase the number of low-wage jobs and slow productivity growth, both of which would hurt low-wage workers; 2) there is no negative correlation between automation and job growth; and 3) automation helps lower prices, which helps reduce inequality, since the higher-income Americans consume a much lower share of their incomes than low- and moderate-income Americans do.