The Case for an AI Token Tax

The episode opens with the host framing the token tax debate as an increasingly important policy conversation, triggered by Elizabeth Warren's Time magazine op-ed and Senate candidate Mallory McMorrow's comprehensive AI worker protection plan. McMorrow proposes a modest per-token fee on commercial AI usage, arguing that as AI scales to billions of queries per day, even a fraction-of-a-cent charge becomes a meaningful funding stream without raising taxes on workers. Warren similarly calls for taxing AI data centers via an energy excise tax, framing it as a way to ensure AI's economic benefits are broadly shared rather than concentrated among the wealthy.

The host notes surprising support from tech figures: Mark Cuban proposed a federal token tax under 50 cents per million tokens, arguing it would incentivize optimization, reduce energy use, and generate tens of billions annually. DuckDuckGo's Gabriel Weinberg advocated collecting the tax now and holding funds in a lockbox for displaced workers, and Anthropic's Dario Amodei floated a 3% revenue share model, acknowledging it wasn't in his financial interest but calling it reasonable.

The philosophical case for a token tax rests on the idea that as AI agents increasingly perform tasks previously done by taxable human workers, the existing tax base—built around labor income—erodes. Tokens are proposed as an administratable proxy for AI labor, analogous to hours worked for humans. The host argues a token tax could theoretically create fiscal neutrality between hiring humans and deploying AI agents, removing what is effectively a tax-incentivized preference for automation over human employment.

However, the episode gives substantial weight to counterarguments. Critic David Friedman identifies tokens as a deeply flawed economic proxy—one million tokens can represent anything from spam to high-value legal analysis, varying in economic worth by orders of magnitude. He also raises the 'tokenizer endogeneity problem,' where different providers tokenize the same content at wildly different rates, creating arbitrary discrimination. The secular 200x annual decline in per-token prices means a fixed tax would either become confiscatory or collapse as a revenue source. Geographically, a U.S.-only token tax would function as a subsidy for foreign AI providers, pushing American customers toward overseas inference platforms.

A Brookings-sponsored working paper is cited as arguing that if a token tax is appropriate at all, it should apply only at the point of final consumption—integrated into VAT/sales tax infrastructure—not as a blanket intermediate production tax, which would distort investment and adoption. The host adds his own concern that a token tax would create a known-ROI bias, pushing firms toward efficiency AI over exploratory, potentially higher-value use cases, while also entrenching large incumbents who can self-host or negotiate discounts. The episode concludes with the host expressing openness to the broader philosophical shift in tax structure that AI may necessitate, while suggesting a token tax as currently conceived is too blunt an instrument, and that the real value lies in having this uncomfortable policy conversation at all.