The Automation Tax

In the AI Story

The policy logic rests on the displacement-reinstatement framework. If the social cost of displacement exceeds its private cost to the automating firm — through unemployment, retraining, reduced consumer spending, political instability — then the tax code that makes automation artificially cheap is producing too much of it. A Pigouvian correction, setting the price equal to the social marginal cost, restores the conditions under which firms' automation choices produce socially optimal outcomes.

The proposal has been criticized from both directions. Libertarian economists argue any tax on automation slows productivity growth. Acemoglu's response is that the current subsidy to automation is itself a distortion, and removing a subsidy is not a tax. Labor advocates argue the proposal is insufficient without complementary investment in reinstatement technologies and worker bargaining power. Acemoglu's response is that the tax is one instrument in a portfolio, not a complete program.

Applied to AI specifically, the proposal implies differential treatment between AI systems that augment human capability — machine usefulness — and those designed to replace human workers. The distinction is administratively hard but not impossible: tax authorities already make finer distinctions in R&D credit eligibility. The proposal would create commercial incentives for augmentation-oriented AI development, partially correcting the current bias toward replacement.

The deeper point is that the current tax treatment encodes assumptions about the social value of displacement that were never democratically ratified. The rapid-depreciation provisions were designed for an era when capital investment was scarce and labor plentiful. The inverse condition now prevails: capital is abundant, meaningful employment increasingly scarce. The tax code has not adapted, and the governance gap includes this failure of institutional updating.

Origin

The proposal appears in multiple Acemoglu papers from 2019 onward and was developed at length in the 2020 piece 'Does the US Tax Code Favor Automation?' with Andrea Manera and Pascual Restrepo in Brookings Papers on Economic Activity. It has been cited in US Senate testimony and in EU Commission documents on AI regulation.

Key Ideas

The current tax code subsidizes displacement. Depreciation rules and payroll tax structures make capital cheaper than labor by more than productivity differences alone would justify.

Pigouvian logic applies. If the social cost of automation exceeds the private cost, tax correction restores efficient decision-making — it does not penalize technology.

The instrument is calibrated, not prohibitive. The proposal removes artificial advantages for automation without creating artificial disadvantages, leaving technology choices to respond to genuine productivity differences.

Administrative tractability is sufficient, not perfect. Distinguishing augmenting from replacing AI is hard but already within the capacity of tax authorities that make comparable distinctions routinely.

Debates & Critiques

Conservative tax economists argue any differential treatment of capital types creates lobbying incentives that capture the policy apparatus. Acemoglu accepts the risk but argues the alternative — a uniform subsidy to capital — is itself captured by the firms that benefit from it, and the relevant question is which capture pattern produces better outcomes.