Research question

If every dividend, safety net, or shorter workweek in this site's research has to be paid for by something, what is that something actually supposed to be — and has anyone actually started collecting it?

Abstract

Every mechanism examined so far on this site — a universal dividend, a public wealth fund, a shorter working week at full pay — assumes money arriving from somewhere. This paper asks the unglamorous question everyone skips past: who pays, and how. Bill Gates first floated a literal "robot tax" back in 2017. Nine years later, in April 2026, OpenAI's own policy blueprint proposed a menu of options — automation taxes, a shift from payroll toward capital gains, a public wealth fund partly funded by AI firms themselves.

Economists have since split the idea in two, in a way that matters more than it sounds: taxing what a robot sells to a customer is a normal, defensible consumption tax; taxing a company merely for owning the robot is a capital tax that risks discouraging the exact investment that grows the economy. Almost none of this has actually happened anywhere. The one real exception, South Korea's adjustment to its automation tax credit, is smaller and stranger than the "robot tax" headlines about it suggest.

Key findings
  • OpenAI's April 2026 policy blueprint proposed five mechanisms at once: a public wealth fund partly funded by AI firms, taxes on automation that displaces workers, a shift of the tax base from payroll toward capital gains and corporate income, government-backed 32-hour workweek pilots at full pay, and automatic safety-net triggers keyed to displacement data.
  • Economists Anton Korinek and Ben Lockwood, in prominent 2026 research, argue "robot tax" proposals conflate two different things: taxing robot-provided services sold to consumers is an ordinary, economically sound consumption tax, while taxing a firm simply for owning automation equipment is a capital tax that discourages the investment that drives productivity growth.
  • A separate proposal, from Anton Leicht and Dean Ball, argues the goal shouldn't be raising new revenue at all — it should be removing an existing bias, since payroll taxes already make human workers more expensive to employ than AI capital, which pays no equivalent tax.
  • Economists Falk and Tsoukalas have proposed a different justification entirely: when AI-driven layoffs happen, the company doing the laying off doesn't bear the cost of the reduced consumer spending that follows — a "demand externality" that, in their view, justifies a Pigouvian tax on automation, the same logic used to tax pollution.
  • South Korea is the one country that has actually implemented something in this space — not a literal per-robot tax, but a reduction to an existing automation-investment tax credit. Early empirical review suggests it may be a cost-effective way to support employment, though it is a modest fiscal adjustment, not the sweeping "tax on robots" the term usually implies.
  • The urgency behind all of this isn't hypothetical: Anthropic's Dario Amodei warned in 2025 that AI could eliminate a large share of entry-level white-collar jobs and push unemployment into double digits within a few years, around the same period Amazon, Meta, and UPS announced layoffs alongside record AI spending.
Three different questions hiding inside "robot tax"

Is it a sales tax or an investment tax? Korinek and Lockwood's distinction is the one that most public discussion skips. Taxing a robot-delivered service the way you'd tax any other purchase is uncontroversial. Taxing a company for the act of owning a machine is closer to taxing a factory for owning steel beams during the industrial revolution — which is exactly the comparison critics use to argue it would slow the investment the economy actually needs.

Is the goal revenue, or fairness between workers and machines? Leicht and Ball's argument isn't really about funding anything new. It's that payroll taxes already tilt the field against hiring a person over deploying AI, since only the person's wages get taxed. Their proposed fix is closer to a rebalancing than a new revenue source — deliberately designed so employers using AI to support workers, rather than replace them, would barely notice a change.

Is this about the job loss, or what happens after it? Falk and Tsoukalas's Pigouvian framing is the most novel of the three. It doesn't argue displacement itself is the problem to price — it argues the ripple effect is: a company that lays off workers doesn't pay for the fact that those workers can no longer buy anything from anyone else in the economy. That's a genuinely different justification than "machines should pay tax like workers do," and it borrows directly from how economists already justify carbon taxes.

The honest gap in the evidence

Every mechanism above except one is a proposal, a working paper, or a corporate policy document — not a law. OpenAI's own blueprint was explicitly described by Sam Altman as "a starting point, not a prescription," and specific Congressional response to it has so far been limited. That matters more here than in some of this site's other papers, because the entire premise of a universal dividend or a funded shorter workweek depends on one of these funding mechanisms actually existing at scale, and none currently does.

South Korea's automation tax credit adjustment is the sole real-world data point, and it's a narrower policy than the "robot tax" framing implies — an incremental change to investment incentives, not a new tax on owning or deploying a robot. Extrapolating from one modest fiscal tweak in one country to "this is how AI-era redistribution gets funded globally" would be a significant overreach; this paper doesn't make that leap, and any model that assumes a mature, revenue-generating automation tax already exists somewhere is describing a proposal, not a fact.

Why this matters for After Labour's research

This paper is the missing link behind Papers 003 and 006. The Human Dividend asked what form a citizen's claim on AI wealth should take; The 92 Percent Experiment showed what happens when productivity gains become time instead of income. Neither paper asked where the money comes from in the first place — this one does, and the honest answer is that the funding mechanism is currently the least developed part of the entire policy conversation, not the most.

Methodology

This paper synthesizes 2025–2026 academic working papers, policy documents, and journalistic reporting on automation and AI taxation proposals, rather than presenting original data collection or modelling. Where a proposal has not been implemented anywhere, it is described as a proposal, not as policy.

Assumptions
  • Corporate policy documents, including OpenAI's, are treated as evidence of a live, evolving conversation among affected firms — not as neutral or disinterested policy analysis, given the companies proposing these mechanisms would also be taxed under them.
  • South Korea's automation tax credit adjustment is treated as the most directly comparable real-world precedent available, despite being narrower in scope than most public "robot tax" discussion implies.
Limitations
  • No jurisdiction has implemented a broad-based tax specifically on AI capital, automation-driven displacement, or robot-provided services at meaningful scale; this paper describes proposals and one narrow precedent, not measured fiscal outcomes.
  • Academic and think-tank positions on this question are contested and often mutually exclusive — the consumption-tax, capital-tax, tax-neutrality, and Pigouvian-externality framings are not simply complementary views but genuinely different diagnoses of what problem a tax should solve.
  • This is a fast-moving policy area; corporate positions and legislative proposals cited here reflect the state of public reporting as of August 2026 and may have shifted materially by the time of reading.
Data sources
Citation

After Labour (2026). Nobody's Taxing the Robots Yet: Mapping the real proposals to fund an AI-era safety net.