Supporting case
A token-tax paper develops an audit mechanism for taxing AI use where it occurs
What happened
What happened and stage. Lucas Irwin, Tung-Yu Wu and Fazl Barez’s position paper was first submitted March 4, revised June 11, and included in the July 6–11, 2026 ICML proceedings. June version of the paper ↗.
The proposal is a surcharge on model inference at the point of sale. A “token” is a unit of model input/output, not a currency token. The paper proposes compute-provider records and a staged audit system: estimate/check usage externally, use model-category norms if necessary, and inspect internal generation records as a backstop. It favors capture where services are used rather than only where a model is hosted.
The authors acknowledge that external token audits and norm methods are unsolved, that languages/models tokenize differently, and that self-hosted models complicate enforcement. They call for modeling who bears the tax, effects on activity and comparison with computation-based taxes. This is a research roadmap, not a demonstrated collection system, revenue estimate or adopted tax.
This is eligible for Lane 9’s fiscal-idea track because it specifies a taxable transaction and verification architecture. It extends the current Korinek–Lockwood discussion with implementation questions rather than merely repeating a general call to “tax AI.”
My inference is that delegating verification to cloud firms could build fiscal capacity while also strengthening private intermediaries. Public audit access and treatment of independent deployment would determine that balance.
Dates and implementation stage
Actual underlying date 2026-03-04 (August study has month precision)
