Taxing tokens, robots or rents: disagreement about the right base

Gates (26 August 2026) proposes taxes on AI tokens and robots: make automated substitution slightly less attractive and fund worker support; target so medicine and education are not inadvertently burdened. He does not specify a rate, threshold, unit of billable tokens, or how to distinguish substitutes from complements. The claim that payroll taxes and capital expensing create unequal incentives is a reason to examine the existing code, not proof that all employer investment is untaxed: corporate income, asset returns and depreciation matter too. Link to policy synthesis.

Brollo et al., IMF staff discussion note (June 2024) argues a special generative-AI tax to slow adoption is difficult to administer and risks deterring labor-augmenting uses; urges revisiting asymmetric capital allowances, strengthening capital-income and economic-rent taxation, plus social protection. Its conclusion is a policy analysis of scenarios, not an estimated effect of Gates’s hypothetical 2026 levy. Thümmel, Optimal Taxation of Robots (JEEA, 2023) calibrates industrial robots to historical US wage effects: depending on distribution and robot price, an optimal subsidy can turn into a small tax; reforming income taxes dominates incremental welfare gains from separately taxing robots when available. Guerreiro, Rebelo and Teles, Should Robots Be Taxed? (2017/2022) instead finds a temporary robot tax while routine workers cannot change occupation; Thümmel explains the sign changes when the model includes a third occupational group. Neither transfers a calibrated rate to tokens/LLMs.

Faivre and Cen, Taxing Artificial Intelligence (July 2026) helpfully separate correction, redistribution and regulatory-capacity funding: a token/API-usage excise is measurable for a provider but taxes use in care and schools too, can be passed downstream and is not a proxy for verified worker harm; a broad profit/rent tax better targets gains but need not discourage substitution at the margin. The paper is a design survey and argument, not an evaluated token-tax field experiment; many claims and examples are proposals. Defined upstream remitter differs from ultimate bearer; offshoring, self-hosted inference, mixed services, reclassification and charitable exemptions complicate tracking. Distinct harms (local water, dangerous use, displaced worker) imply distinct measurable bases.

What would falsify a chosen design? Compare incidence and substitution across affected API clients, small organizations, educational/health deployments and foreign/self-hosted models; link revenues and audited recipients to firm-level hiring, output and worker earnings. If first evidence shows few displaced workers receive transfers while small useful uses absorb the burden, neither of Gates’s two goals is achieved. If broad profit taxes raise revenue but do not slow clearly harmful substitution, distinguish redistribution from correction rather than declaring all tax options equivalent. No direct estimate of 2026 token-levy incidence found here; more primary empirical fiscal evidence is needed before prescribing rates.