Imagine a company that is considering two possible ways to improve efficiency and productivity. One is to pay for many of its employees to go through a training program to learn new sets of useful skills. The other is to pay for new equipment that will replace many of the employees. Daron Acemoglu, Andrea Manera, and Pascual Restrepo argue that the US tax code tends to favor the second option. The technical version of their argument, "Does the U.S. Tax Code Favor Automation?" is published in most recent Brookings Papers on Economic Activity (Spring 2020, a short readable overview of the paper is also available at the link).They write (citations and footnotes omitted):

The most common perspective among economists is that even if automation is contributing to declining labor share and stagnant wages, the adoption of these new technologies is likely to be beneficial, and any adverse consequences thereof should be dealt with appropriate redistributive policies (and education and training investments). But could it be that the extent of automation is excessive, meaning that US businesses are adopting automation technologies beyond the socially optimal level? If this were the case, the policy responses to these major labor market trends would need to be rethought.

There are several reasons why the level of automation may be excessive. Perhaps most saliently, the US tax system is known to tax capital lightly and provide various subsidies to the use of capital in businesses. In this paper, we systematically document the asymmetric taxation of capital and labor in the US economy in the US tax system labor is much more heavily taxed than capital. ...

Mapping the complex range of taxes in the US to effective capital and labor taxes is not trivial. Nevertheless, under plausible scenarios (for example, depending on how much of healthcare and pension expenditures are valued by workers and the effects of means-tested benefits), we find that labor taxes in the US are in the range of 25.5-33.5%. Effective capital taxes on software and equipment, on the other hand, are much lower, about 10% in the 2010s and even lower, about 5%, after the 2017 tax reforms. We also show that effective taxes on software and equipment have experienced a sizable decline from a peak value of 20% in the year 2000.3 A major reason explaining this trend in capital taxation is the increased generosity [of] depreciation allowances ...

 I should emphasize that thispaper is part of an ongoing research effort by these authors to think about interactions between automation and jobs. I have blogged about a previous entry in this line of research in "Is Something Different This Time About the Effect of Technology on the Labor Market?" (May 6, 2019). I discussed there a paper by Daron Acemoglu and Pascual Restrepo titled"Automation and New Tasks: How Technology Displaces and Reinstates Labor."

In that paper, they suggest a framework in which automation can have three possible effects on the tasks that are involved in doing a job: a displacement effect, when automation replaces a task previously done by a worker; a productivity effect, in which the higher productivity from automation taking over certain tasks leads to more buying power in the economy, creating jobs in other sectors; and a reinstatement effect, when new technology reshuffles the production process in a way that leads to new tasks that will be done by labor. In this model, the effect of automation on labor is not predestined to be good, bad, or neutral. It depends on how these three factors interact.

In that context, the authors of the current paper suggest the theoretical possibility of an "automation tax," defined as "a higher tax on the use of capital in tasks where labor has a comparative advantage." They would combine this with a lower tax on other forms of capital, as well as on labor. In my own words, they are proposing that the tax code encourage the kind of automation that complements what workers do in a way that leads to sharp increases in productivity and output, but that the tax code not encourage the kind of automation that mostly just replaces workers with a real but only modest cost savings for the employer.

Of course, it's reasonable to note that a theoretical economic model can just create variables for these two kinds of automation, while a real world policy might face some difficult challenges in distinguishing between them. Still, the authors are trying to break out of a binary choice where automation is viewed as always good or always bad, and automation is instead being viewed as a range of choices that include automation that is more likely to be job-destroying or more likely to be job-creating. It feels to me like a potential distinction worth investigating.